Aydin's Workbench Blog https://blog.aydins-workbench.com/ Notes on statistical modelling, research software, public data questions, and project-adjacent thinking. quarto-1.9.38 Tue, 16 Jun 2026 00:00:00 GMT How Panel Pilot Chooses What To Zoom Into Aydin https://blog.aydins-workbench.com/posts/panel-pilot-panel-detection/ Panel-by-panel reading sounds like it should be simple. Find the boxes, zoom into the boxes, move to the next box.

It is not simple.

Panel Pilot started from a practical annoyance: manga apps can pull from huge libraries through source extensions, but on a phone the full page is often too small to read comfortably. The actual job was not just “detect panels”. It was:

  1. choose readable regions,
  2. move through them in the right order,
  3. keep enough surrounding context that the page still feels like a comic,
  4. avoid getting stuck when detection is wrong,
  5. handle manga, webtoons and western comics without making the reader feel like three unrelated apps.

This post is a build note on the methods we tried and what the current prototype landed on.

The Reader Model

The reader does not crop the image into separate files. It keeps the page image intact and treats each panel as a normalized rectangle:

{
  x: 0.12,
  y: 0.34,
  w: 0.42,
  h: 0.18
}

When the user moves panel-by-panel, the app expands that rectangle by the current padding setting, computes the scale needed to fit it inside the reader viewport, and translates the full image so the target region is centered.

That matters because it preserves context. A pure crop can feel clean, but manga and comics often rely on gutters, speech tails, nearby reactions, and the rhythm of the page. Panel Pilot therefore lets the user set context padding. Zero percent means “show me the selected region only”. A larger value means “give me a little of the page around it”.

The whole detector stack feeds that one camera model.

First Attempt: One Detector For Everything

The first obvious approach was to take a page, find the white gutters, and treat the content blocks between those gutters as panels.

That works surprisingly often for conventional manga pages:

  • manga pages usually have clear black ink on pale paper,
  • panel gutters are often bright and low-variance,
  • rectangular panels are common,
  • right-to-left reading order can be approximated by row and column grouping.

But it fails quickly once the layout becomes more interesting. Title pages may have huge blank areas. Action pages may have overlapping art, black backgrounds, angled panel borders, or speech bubbles that look like little panels. Western comics often have caption boxes, diagonal cuts, inset panels and painted backgrounds where “find the gutters” is not a strong enough idea.

So the current system is not one detector. It is a small committee of detectors, plus rules for choosing between them.

Manga Mode

Manga mode is the most mature path.

The page is scaled down for analysis, then converted into luminance-based masks:

  • a dark mask for anything that looks like ink or art,
  • a black mask for strong panel borders and heavy linework,
  • a light mask for pale panel interiors and page background,
  • row and column statistics for brightness, variance and dark-pixel runs.

From those masks, Panel Pilot tries several panel candidates.

Recursive Region Splitting

The main manga detector recursively splits the page into regions. It looks for likely horizontal or vertical separations, then checks the content bounds inside each proposed region.

This works well when panels are rectangular and separated by gutters. It also avoids treating every speech bubble or texture as a separate panel because candidate regions have to pass size and density checks.

Connected Light Regions

The second candidate path looks for connected light regions. In a typical black-and-white manga page, panel interiors often form large pale regions surrounded by dark ink or borders.

This can catch panel shapes the gutter detector misses, but it can also over-detect. Speech bubbles, bright clothing, sky, paper texture and sound effects can look like “light regions”. We keep it as a useful alternative, not as the default truth.

Grid/Gutter Splitting

The third candidate path builds row and column cuts from quiet gutters. It then takes the Cartesian product of those cuts, checks each cell for content, and keeps plausible cells.

This is useful when the recursive detector merges panels or when connected components produce too many fragments. It is also where a recent bug came from: an empty grid cell should be skipped, not abort the whole detector. That little distinction matters more than it has any right to.

Slanted Manga Separators

The newest manga addition handles pages where a major separator is not horizontal. Some manga pages use angled borders: visually obvious to a human, but invisible to a detector that only searches for straight row or column gutters.

The slanted pass scans for long near-horizontal diagonal black dividers. When it finds credible separators, it splits the page into diagonal bands and then runs the usual region splitting inside those bands.

This is still an approximation. The reader regions remain rectangles, not polygons. But it catches the important practical case: “there are two or three real panel bands here, and the standard detector merged them because the divider was sloped.”

Choosing The Manga Candidate Set

The detector does not simply concatenate all candidates. It chooses.

The current priority is roughly:

  1. prefer slanted panels if they create a plausible set and reduce oversized merged regions;
  2. use recursive panels when they produce a small plausible set;
  3. fall back to split panels when connected components over-detect;
  4. use connected panels when they are plausible and ordered;
  5. otherwise return a full-page fallback.

After that, panels are merged if they overlap heavily, filtered for plausible size, sorted, and repaired for reading order.

For right-to-left manga, the order check is as important as detection. It should be rare to move rightward within a row or upward within a column. The sorter therefore groups panels into bands, orders columns RTL or LTR, and applies a small repair pass for local reading-order violations.

Webtoon Mode

Webtoon mode needed a different model.

At first, a naive “move one screen-height down” approach seemed tempting. It is simple, but it cuts through speech bubbles, faces and scene transitions. The point of panel mode is readability, not scrolling with extra ceremony.

The current webtoon method still detects regions, but with different assumptions:

  • the content is one continuous vertical strip,
  • the reader should keep the full strip width,
  • regions may overlap,
  • page boundaries from the source should not be treated as meaningful story boundaries.

That last point is important. Many sources split a webtoon chapter into multiple image files, but visually the chapter is one long strip. A panel or speech bubble can cross the boundary between source images. So Panel Pilot builds a continuous synthetic page.

The first two source images are loaded immediately, so a panel that crosses the first boundary can be seen. The rest of the chapter loads in the background and is appended to the strip without changing the visual width. As new pages arrive, the app preserves the current vertical position by finding the nearest panel center in the expanded strip.

For detection, webtoon mode analyzes rows rather than boxes. It classifies rows as quiet or active based on brightness, variance, ink and edge changes. Long quiet runs become candidate cuts. Active regions become vertical panels with full width. Very tall regions are split into readable chunks with overlap, preferably near quiet rows.

That landed better than fixed scrolling because it centers on content while still behaving like a vertical reader.

Comic Beta

Western comics are the least solved mode, and the app labels this honestly as beta.

We tried reusing manga mode with left-to-right ordering. It works on clean rectangular pages, but it breaks down on dynamic layouts: painted backgrounds, caption strips, inset panels, overlapping art, diagonal borders and pages where the story focus is not the same thing as a bordered rectangle.

The current comic beta keeps the same luminance and gutter candidates as manga, but adds text-like component detection and several flow fallbacks:

  • full-page mode for covers and backmatter,
  • caption-strip windows for pages that are mostly text bands,
  • readable text windows when text clusters are the clearest navigation targets,
  • story-flow regions when panel detection produces montage slices or tiny fragments,
  • focus regions when connected components are more useful than panel boxes.

This is more a “readable region selector” than a true comic panel detector. That is intentional for now. On a phone, the practical question is not always “what is the canonical panel?” It is often “what region lets me read the words and understand the scene?”

For western comics, I expect future versions to need stronger text and speech-bubble understanding. A page can be visually complex but still have a clear reading path through balloon clusters. The current beta is a step toward that, not the final answer.

Defensive Fallbacks

One lesson from building this interactively: detector failures must not break the reader.

The app now sanitizes every panel list before storing it:

  • invalid coordinates are removed,
  • rectangles are clamped to the page,
  • empty output falls back to full page,
  • sparse tiny detections on mostly blank pages fall back to full page,
  • movement repairs a bad current panel list before reading it.

That sounds boring, but it is the difference between “this page was not detected well” and “the app stopped responding”. A panel reader can tolerate occasional full-page fallback. It cannot tolerate navigation state becoming invalid.

Testing What We Could See

The testing loop became visual very quickly.

For manga, we used Frieren and Chainsaw Man pages because they include a useful spread of normal layouts, title pages, speech-heavy pages, action pages and occasional angled divisions. For webtoon mode, we used long-strip manhwa chapters because they stress page-boundary stitching and tall-region splitting. For western comics, Darth Vader pages were useful precisely because they were unpleasant: dynamic layouts, caption-heavy pages, and art that makes rectangular panel assumptions wobble.

The most useful test artifact was an overlay/contact sheet: render the page, draw detected regions, label the reading order, then inspect whether the flow makes sense. A count alone is not enough. Six detected panels can still be wrong if the order jumps rightward in an RTL row, splits a speech bubble, or zooms into a blank caption edge.

We also added a “report bad panels” button in the app. That saves the current page, selected region and detector metadata so future failures can be reviewed directly instead of reconstructed from memory.

Visual Examples From The Test Set

These are small annotated excerpts from the detector test runs, not full-resolution reading assets. The colored boxes are the regions Panel Pilot would zoom into, and the small black labels show the reading order.

Manga: Chainsaw Man

Chainsaw Man was useful for manga mode because it has many conventional rectangular pages, but it also makes reading order easy to inspect. In the example below, the top row is read right-to-left: region 1 on the upper right, region 2 below it, then region 3 and 4 on the left. After that, the reader moves down the page.

A low-resolution Chainsaw Man page excerpt with colored panel boxes and numbered reading-order labels.

Chainsaw Man manga-mode detector overlay showing right-to-left reading order across rectangular panels.

This is the case where manga mode should feel almost invisible. The detector is not trying to understand the story; it is finding plausible panel regions, grouping them into rows, and then applying RTL ordering inside each row. The important behavior is that the reader rarely moves rightward inside a row.

Webtoon: Solo Leveling

Solo Leveling pushed the webtoon path in a different direction. The source gives the chapter as multiple image files, but the reader has to treat those files as one continuous vertical strip. A fixed screen-height scroll would land halfway through faces, speech bubbles and scene transitions.

Three low-resolution Solo Leveling strip excerpts side by side, each showing vertical webtoon slices separated by red detector boundaries.

Solo Leveling webtoon-mode detector slices from one continuous vertical strip.

The webtoon method keeps the strip width fixed, then chooses vertical slices from quiet and active row analysis. The regions can overlap. That is deliberate: for a long vertical comic, overlapping slices are usually better than cutting a speech bubble or character beat in half.

Comic Beta: Darth Vader (2020)

Darth Vader (2020) chapter 33 is a good stress test for western comics because the page design is doing a lot: diagonal motion, inset panels, large sound effects, captions and overlapping focal regions. A manga-style “find neat rectangles” pass is not enough.

A low-resolution Darth Vader 2020 page excerpt with colored overlapping focus regions and numbered labels.

Darth Vader 2020 comic-beta detector overlay showing overlapping focus regions on a dynamic western comic page.

This is why the comic beta became a readable-region selector rather than a pure panel detector. Some boxes overlap, and some are aimed at text or visual focus rather than strict bordered panels. It is still the least polished mode, but the objective is different: choose zoom targets where the reader can actually read the words and follow the page flow on a phone.

What We Landed On

The current prototype uses three modes:

Mode Core idea Current status
Manga Multiple rectangular detectors, candidate selection, RTL/LTR ordering, slanted separator pass Most usable
Webtoon Continuous vertical strip, row-based active/quiet segmentation, overlapping readable slices Usable, still tuning slice height
Comic beta Hybrid region selector using panels, text-like components and flow fallbacks Useful but not yet elegant

The most important design choice was to stop treating panel detection as a single computer-vision answer. It is closer to a reading interface problem. The app needs enough detection to choose a good camera target, enough ordering to preserve story flow, and enough fallback behavior to keep reading when the detector is uncertain.

The next big step is probably better text awareness: not full OCR for transcription, but enough detection to know whether a speech bubble or caption is fully readable inside at least one region. That would give the detector a more reader-centered objective than “did we find the panel border?”

For now, the working rule is simple: a panel selection method is good if the next tap feels obvious, the words are readable, and the reader does not have to think about the detector.

]]> software computer-vision reading tools https://blog.aydins-workbench.com/posts/panel-pilot-panel-detection/ Tue, 16 Jun 2026 00:00:00 GMT A First Pass At Australian Dwelling Prices Aydin https://blog.aydins-workbench.com/posts/housing-prices-first-pass/ Housing debates have a way of turning into vibes very quickly. Prices feel high, supply feels tight, first-home buyer policy gets loud, and then everyone starts arguing from their favourite chart.

So this is a small grounding exercise: use the latest ABS dwelling-value data, look at the national dwelling stock and mean prices by state, then write down the policy questions that seem worth chasing next.

Data

The main source is ABS Total Value of Dwellings, March quarter 2026. The release estimates the total value, number and mean price of Australia’s residential dwellings. I also keep Lending Indicators, Building Approvals, and Treasury’s home ownership support page in view because prices alone do not tell us whether buyers are being helped, stretched, or both.

Code 
download_if_missing <- function(url, path) {
  if (!file.exists(path)) {
    download.file(url, path, mode = "wb", quiet = TRUE)
  }
  path
}

as_number <- function(x) {
  parse_number(as.character(x), na = c("", "NA", "n/a", "-"))
}

quarter_date_from_excel <- function(x) {
  as.Date(as.numeric(x), origin = "1899-12-30")
}

state_from_item <- function(item) {
  str_split_fixed(item, ";", 3)[, 2] |>
    str_squish()
}

measure_from_item <- function(item) {
  case_when(
    str_detect(item, "Value of dwelling stock; Owned by All Sectors") ~ "Total dwelling stock value",
    str_detect(item, "Mean price of residential dwellings") ~ "Mean dwelling price",
    str_detect(item, "Number of residential dwellings") ~ "Number of dwellings",
    TRUE ~ NA_character_
  )
}

workbook_path <- download_if_missing(
  source_urls$dwellings_workbook,
  file.path(raw_dir, "abs-total-value-dwellings-mar-quarter-2026.xlsx")
)

raw <- read_excel(workbook_path, sheet = "Data1", col_names = FALSE)

series_meta <- tibble(
  column = seq_along(raw),
  item = as.character(unlist(raw[1, ])),
  unit = as.character(unlist(raw[2, ])),
  measure = measure_from_item(item),
  state = state_from_item(item)
) |>
  filter(!is.na(measure), !is.na(state), state != "")

dwelling_series <- raw |>
  slice(-(1:10)) |>
  transmute(row_id = row_number(), quarter_date = quarter_date_from_excel(...1)) |>
  bind_cols(raw |> slice(-(1:10)) |> select(all_of(series_meta$column))) |>
  pivot_longer(-c(row_id, quarter_date), names_to = "column_name", values_to = "value") |>
  mutate(
    column = as.integer(str_remove(column_name, "\\.\\.\\.")),
    value = as_number(value)
  ) |>
  left_join(series_meta, by = "column") |>
  filter(!is.na(value), !is.na(quarter_date), !is.na(measure)) |>
  mutate(
    value = case_when(
      measure == "Total dwelling stock value" ~ value / 1000,
      measure == "Mean dwelling price" ~ value * 1000,
      measure == "Number of dwellings" ~ value * 1000,
      TRUE ~ value
    )
  )

latest_quarter <- max(dwelling_series$quarter_date, na.rm = TRUE)

latest_state_prices <- dwelling_series |>
  filter(measure == "Mean dwelling price", quarter_date %in% sort(unique(quarter_date), decreasing = TRUE)[1:2], state != "Australia") |>
  select(quarter_date, state, mean_price = value) |>
  pivot_wider(names_from = quarter_date, values_from = mean_price) |>
  rename(previous_price = 2, latest_price = 3) |>
  mutate(
    quarterly_change = latest_price / previous_price - 1,
    dollar_change = latest_price - previous_price
  )

What Has Happened Nationally?

Code 
dwelling_series |>
  filter(measure == "Total dwelling stock value", state == "Australia", quarter_date >= as.Date("2021-03-31")) |>
  ggplot(aes(quarter_date, value)) +
  geom_line(linewidth = 0.9, colour = "#126782") +
  geom_point(size = 2, colour = "#126782") +
  scale_y_continuous(labels = dollar_format(suffix = "t", scale = 1 / 1000)) +
  labs(x = NULL, y = "Total dwelling stock value")

ABS total value of Australian residential dwelling stock.
Code 
dwelling_series |>
  filter(measure == "Mean dwelling price", state == "Australia", quarter_date >= as.Date("2021-03-31")) |>
  ggplot(aes(quarter_date, value)) +
  geom_line(linewidth = 0.9, colour = "#7a4b9d") +
  geom_point(size = 2, colour = "#7a4b9d") +
  scale_y_continuous(labels = dollar) +
  labs(x = NULL, y = "Mean dwelling price")

ABS mean dwelling price for Australia.

The March quarter 2026 release puts the national mean dwelling price at over $1.1 million. That is not a typical transaction price and it is not an affordability measure, but it is a useful whole-market signal: the stock keeps getting more expensive, even before we ask who can actually buy into it.

Which States Moved Most Recently?

Code 
latest_state_prices |>
  ggplot(aes(fct_reorder(state, latest_price), latest_price, fill = quarterly_change)) +
  geom_col() +
  coord_flip() +
  scale_y_continuous(labels = dollar) +
  scale_fill_gradient2(labels = percent, low = "#8c2d04", mid = "grey80", high = "#006d2c", midpoint = 0) +
  labs(x = NULL, y = "Mean dwelling price", fill = "Quarterly change")

Latest ABS mean dwelling prices by state and territory.
Code 
latest_state_prices |>
  arrange(desc(quarterly_change)) |>
  transmute(
    state,
    latest_mean_price = dollar(latest_price, accuracy = 100),
    quarterly_change = percent(quarterly_change, accuracy = 0.1),
    dollar_change = dollar(dollar_change, accuracy = 100)
  ) |>
  kable()
state latest_mean_price quarterly_change dollar_change
New South Wales $11,495,200 0.5% $54,200

Policy Questions This Sets Up

Treasury says the 5% Deposit Scheme was expanded from 1 October 2025, with uncapped places, removed income caps and higher property price caps. That creates a nice empirical question, but not a simple one.

The questions I would ask next:

  • Did first-home buyer loan commitments rise after the deposit-scheme expansion, and was that concentrated in states where prices were already rising fastest?
  • Did the policy mostly pull forward demand, or did it change the composition of buyers entering the market?
  • Are building approvals moving in the same places as dwelling prices, or is supply still lagging the demand signal?
  • Are investor loans and first-home buyer loans moving together or pushing against each other?
  • If mean prices rise after access support expands, how much of the buyer benefit is eaten by price pressure?

Those need lending and building-approval data joined to this price series. This post is just the first brick.

Caveats

  • Mean dwelling price is a stock estimate, not a median sale price or a first-home buyer price.
  • State averages hide massive variation inside cities, regions and property types.
  • The latest ABS dwelling estimates are preliminary and can be revised.
  • Prices do not measure affordability on their own. We need income, rents, interest rates, deposit constraints, lending standards and supply.
  • Policy timing matters. A demand-side policy can show up in lending before it shows up in prices, and supply responses can be much slower.
]]> housing cost-of-living australia policy abs https://blog.aydins-workbench.com/posts/housing-prices-first-pass/ Thu, 11 Jun 2026 00:00:00 GMT Are NDIS Participant Outcomes Changing? Aydin https://blog.aydins-workbench.com/posts/ndis-outcomes-spend-service-access/ The simplest outcomes question is not about spend. It is this:

Can we see any change in outcomes for NDIS participants?

This post starts there. It uses the public participant outcomes files to ask what outcomes are measured, what the current national levels look like, and how reported outcomes change after plan reassessment. The analysis is descriptive. It shows changes in aggregate survey outcomes; it does not prove that the NDIS caused those changes for any individual participant.

The companion budget and utilisation model is here.

Code 
clean_names <- function(x) {
  x |>
    str_to_lower() |>
    str_replace_all("&", "and") |>
    str_replace_all("[^a-z0-9]+", "_") |>
    str_replace_all("^_|_$", "")
}

clean_df <- function(x) {
  names(x) <- clean_names(names(x))
  x
}

as_number <- function(x) {
  if (is.numeric(x)) return(x)
  parse_number(
    as.character(x),
    na = c("", "NA", "na", "n/a", "Not applicable", "No data available", "\\")
  )
}

as_rate <- function(x) {
  out <- as_number(x)
  ifelse(out > 1.5, out / 100, out)
}

safe_median <- function(x) {
  x <- x[is.finite(x)]
  if (length(x) == 0) return(NA_real_)
  median(x)
}

format_optional_percent <- function(x, accuracy = 1) {
  ifelse(is.na(x), NA_character_, percent(x, accuracy = accuracy))
}

read_page <- function(url) {
  paste(readLines(url, warn = FALSE, encoding = "UTF-8"), collapse = "\n")
}

extract_links <- function(url) {
  html <- read_page(url)
  base_url <- str_replace(url, "^(https?://[^/]+).*$", "\\1")
  matches <- gregexpr("<a[^>]+href=[\"']([^\"']+)[\"'][^>]*>(.*?)</a>", html, perl = TRUE)
  pieces <- regmatches(html, matches)[[1]]
  if (length(pieces) == 0) return(tibble(label = character(), url = character()))

  tibble(raw = pieces) |>
    mutate(
      href = str_match(raw, "href=[\"']([^\"']+)[\"']")[, 2],
      label = raw |>
        str_replace_all("<[^>]+>", " ") |>
        str_squish(),
      url = case_when(
        str_detect(href, "^https?://") ~ href,
        str_detect(href, "^/") ~ paste0(base_url, href),
        TRUE ~ paste0(str_remove(url, "/[^/]*$"), "/", href)
      )
    ) |>
    select(label, url)
}

download_one <- function(url, filename) {
  path <- file.path(raw_dir, filename)
  if (!file.exists(path)) {
    download.file(url, path, mode = "wb", quiet = TRUE)
  }
  path
}

download_named <- function(page_url, label_regex, filename) {
  path <- file.path(raw_dir, filename)
  if (file.exists(path)) return(path)

  link <- extract_links(page_url) |>
    filter(str_detect(label, regex(label_regex, ignore_case = TRUE))) |>
    slice(1)

  if (nrow(link) == 0) return(NA_character_)
  download_one(link$url[[1]], filename)
}

cohort_levels <- c(
  "Early childhood",
  "School age",
  "Young people",
  "Adults",
  "Family/carer of children",
  "Family/carer of adults"
)

standardise_cohort <- function(questionnaire) {
  case_when(
    questionnaire == "Participant 0 to before school" ~ "Early childhood",
    questionnaire == "Participant starting school to 14" ~ "School age",
    questionnaire == "Participant 15 to 24" ~ "Young people",
    questionnaire == "Participant 25 and over" ~ "Adults",
    questionnaire == "Family/carer of participant 0 to 14" ~ "Family/carer of children",
    questionnaire == "Family/carer of participant 15 and over" ~ "Family/carer of adults",
    TRUE ~ questionnaire
  )
}

classify_outcome_theme <- function(indicator) {
  text <- str_to_lower(indicator)
  case_when(
    str_detect(text, "concerns|development|functional|self-care|independent|daily living") ~ "Development and daily living",
    str_detect(text, "communicat|tell them what|what they want") ~ "Communication",
    str_detect(text, "choice|control|decisions|advocacy|advocate|rights|choose who supports|choose what they do") ~ "Choice and control",
    str_detect(text, "friend|relationship|community|social|family life|meet more people|be more involved|welcomed|included") ~ "Relationships and community",
    str_detect(text, "home|safe") ~ "Home and safety",
    str_detect(text, "health|wellbeing|well-being") ~ "Health and wellbeing",
    str_detect(text, "school|education|training|learn|course|job|employment|paid job|volunteer|mainstream settings") ~ "Learning and work",
    str_detect(text, "services|specialist|programs and activities") ~ "Service access",
    TRUE ~ "Other outcomes"
  )
}

classify_direction <- function(indicator) {
  text <- str_to_lower(indicator)
  case_when(
    str_detect(text, "concerns in 6 or more|with no friends|no friends other than|unable to do .*wanted|difficulties accessing") &
      !str_detect(text, "did not have any difficulties") ~ "Lower is better",
    str_detect(text, "want more choice and control") ~ "Lower may be better",
    str_detect(text, "^has the ndis|involvement with the ndis|ndis helped") ~ "Reported improvement",
    str_detect(text, "^%|^of those") ~ "Higher is better",
    TRUE ~ "Interpret with care"
  )
}

direction_adjust_change <- function(change, direction) {
  case_when(
    direction == "Lower is better" ~ -change,
    direction == "Higher is better" ~ change,
    TRUE ~ NA_real_
  )
}

Data

The two source files are the national NDIS participant outcomes extracts:

  1. Baseline outcomes: current reported levels by state, respondent group and indicator.
  2. Longitudinal outcomes: baseline and reassessment values for people with different numbers of plan reassessments.

The main charts use the national rows only, where StateCd == "ALL". Participant cohorts are shown first. Family and carer outcomes are kept as a short companion note at the end.

Code 
baseline_outcomes_path <- download_named(
  page_urls$participant,
  "^Baseline Outcomes data",
  "baseline_outcomes.csv"
)

longitudinal_outcomes_path <- download_named(
  page_urls$participant,
  "^Longitudinal Outcomes data",
  "longitudinal_outcomes.csv"
)

baseline_raw <- read_csv(
  baseline_outcomes_path,
  na = c("", "NA", "na", "n/a"),
  show_col_types = FALSE
) |>
  clean_df() |>
  mutate(
    report_date = dmy(rprtdt),
    percentage = as_rate(percentage),
    cohort_label = factor(standardise_cohort(questionnaire), levels = cohort_levels),
    outcome_theme = classify_outcome_theme(indicator_description),
    direction = classify_direction(indicator_description),
    respondent_type = if_else(str_detect(questionnaire, "^Participant"), "Participant", "Family/carer")
  )

longitudinal_raw <- read_csv(
  longitudinal_outcomes_path,
  na = c("", "NA", "na", "n/a"),
  show_col_types = FALSE
) |>
  clean_df() |>
  mutate(
    report_date = dmy(rprtdt),
    across(starts_with("percentage_"), as_rate),
    cohort_label = factor(standardise_cohort(questionnaire), levels = cohort_levels),
    outcome_theme = classify_outcome_theme(indicator_description),
    direction = classify_direction(indicator_description),
    respondent_type = if_else(str_detect(questionnaire, "^Participant"), "Participant", "Family/carer")
  )

participant_baseline <- baseline_raw |>
  filter(statecd == "ALL", respondent_type == "Participant")

participant_longitudinal <- longitudinal_raw |>
  filter(statecd == "ALL", respondent_type == "Participant")

family_longitudinal <- longitudinal_raw |>
  filter(statecd == "ALL", respondent_type == "Family/carer")

source_summary <- tibble(
  file = c("Baseline outcomes", "Longitudinal outcomes"),
  report_date = c(max(baseline_raw$report_date, na.rm = TRUE), max(longitudinal_raw$report_date, na.rm = TRUE)),
  national_participant_rows = c(nrow(participant_baseline), nrow(participant_longitudinal)),
  participant_cohorts = c(n_distinct(participant_baseline$cohort_label), n_distinct(participant_longitudinal$cohort_label))
)
Code 
source_summary |>
  mutate(
    report_date = format(report_date, "%d %B %Y"),
    national_participant_rows = comma(national_participant_rows)
  ) |>
  kable()
file report_date national_participant_rows participant_cohorts
Baseline outcomes 31 March 2026 42 4
Longitudinal outcomes 31 March 2026 670 4

What Outcomes Are Measured?

The participant files cover four broad age cohorts. The outcomes are not identical across ages: early-childhood items focus on development and communication, school-age items add education and independence, and adult items add choice, home, health, education, work and volunteering.

Code 
participant_outcome_map <- bind_rows(
  participant_baseline |>
    transmute(cohort_label, outcome_theme, indicator_description, direction),
  participant_longitudinal |>
    transmute(cohort_label, outcome_theme, indicator_description, direction)
) |>
  distinct()

outcome_theme_table <- participant_outcome_map |>
  group_by(cohort_label, outcome_theme) |>
  summarise(
    outcomes = n_distinct(indicator_description),
    example_outcomes = str_c(head(unique(indicator_description), 2), collapse = "; "),
    .groups = "drop"
  ) |>
  arrange(cohort_label, outcome_theme)

outcome_theme_table |>
  mutate(cohort_label = as.character(cohort_label)) |>
  kable(col.names = c("Cohort", "Outcome theme", "Outcomes", "Example outcomes"))
Cohort Outcome theme Outcomes Example outcomes
Early childhood Communication 2 % who say their child is able to tell them what he/she wants; Has the NDIS helped increase your child’s ability to communicate what they want?
Early childhood Development and daily living 2 % with concerns in 6 or more of the areas: gross motor skills, fine motor skills, self-care, eating, social interaction, communication, cognitive development, sensory processing; Has the NDIS improved your child’s development?
Early childhood Relationships and community 5 % of children who can make friends with people outside the family; % of children who participate in age appropriate community, cultural or religious activities
Early childhood Service access 1 Has the NDIS improved your child’s access to specialist services?
School age Choice and control 1 % of children who have a genuine say in decisions about themselves
School age Development and daily living 3 % developing functional, learning and coping skills appropriate to their ability and circumstances; % who say their child is becoming more independent
School age Learning and work 2 % of children attending school in a mainstream class; Has the NDIS improved your child’s access to education?
School age Relationships and community 7 % of children who can make friends with people outside the family; % of children who spend time after school and on weekends with friends and/or in mainstream programs
Young people Choice and control 6 % who are happy with the level of independence/control they have now; % who choose who supports them
Young people Development and daily living 1 Has the NDIS helped you with daily living activities?
Young people Health and wellbeing 3 % who rate their health as good, very good or excellent; % who did not have any difficulties accessing health services
Young people Home and safety 3 % who are happy with their home; % who feel safe or very safe in their home
Young people Learning and work 5 % who currently attend or previously attended school in a mainstream class; % who have a paid job
Young people Relationships and community 4 % with no friends other than family or paid staff; % who have been actively involved in a community, cultural or religious group in the last 12 months
Adults Choice and control 5 % who choose who supports them; % who choose what they do each day
Adults Development and daily living 2 % who participate in education, training or skill development; Has the NDIS helped you with daily living activities?
Adults Health and wellbeing 3 % who rate their health as good, very good or excellent; % who did not have any difficulties accessing health services
Adults Home and safety 3 % who are happy with their home; % who feel safe or very safe in their home
Adults Learning and work 6 Of those who participate, % who do so in mainstream settings; % unable to do a course or training they wanted to do in the last 12 months
Adults Relationships and community 4 % with no friends other than family or paid staff; % who have been actively involved in a community, cultural or religious group in the last 12 months

The direction labels are deliberately conservative. Most percentage items are easier to read as higher is better, but some indicators are clearly negative: concerns in many areas, having no friends other than family or paid staff, or being unable to do a course or training the participant wanted to do.

Code 
participant_outcome_map |>
  count(direction, sort = TRUE, name = "outcomes") |>
  kable(col.names = c("Direction label", "Outcome rows"))
Direction label Outcome rows
Higher is better 35
Reported improvement 25
Lower is better 4
Interpret with care 2
Lower may be better 2

Current Levels

The baseline file gives the latest national level for each participant indicator. This is the best starting point before asking whether anything changed.

Code 
baseline_theme_summary <- participant_baseline |>
  group_by(cohort_label, outcome_theme) |>
  summarise(
    outcomes = n_distinct(indicator_description),
    median_level = safe_median(percentage),
    p25 = quantile(percentage, 0.25, na.rm = TRUE),
    p75 = quantile(percentage, 0.75, na.rm = TRUE),
    .groups = "drop"
  ) |>
  filter(is.finite(median_level))

baseline_theme_summary |>
  ggplot(aes(fct_reorder(outcome_theme, median_level), median_level)) +
  geom_col(fill = "#2f6f73") +
  geom_errorbar(aes(ymin = p25, ymax = p75), width = 0.2, colour = "grey35") +
  coord_flip() +
  facet_wrap(~ cohort_label, scales = "free_y") +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Baseline percentage")

Median national baseline outcome level by participant cohort and outcome theme.
Code 
participant_baseline |>
  arrange(cohort_label, outcome_theme, indicator_number) |>
  transmute(
    Cohort = as.character(cohort_label),
    Theme = outcome_theme,
    Outcome = indicator_description,
    Direction = direction,
    Level = percent(percentage, accuracy = 1)
  ) |>
  kable()
Cohort Theme Outcome Direction Level
Early childhood Communication % who say their child is able to tell them what he/she wants Higher is better 68%
Early childhood Development and daily living % with concerns in 6 or more of the areas: gross motor skills, fine motor skills, self-care, eating, social interaction, communication, cognitive development, sensory processing Lower is better 67%
Early childhood Relationships and community % of children who can make friends with people outside the family Higher is better 56%
Early childhood Relationships and community % of children who participate in age appropriate community, cultural or religious activities Higher is better 47%
Early childhood Relationships and community Of these, % who are welcomed or actively included Interpret with care 63%
School age Choice and control % of children who have a genuine say in decisions about themselves Higher is better 79%
School age Development and daily living % developing functional, learning and coping skills appropriate to their ability and circumstances Higher is better 24%
School age Development and daily living % who say their child is becoming more independent Higher is better 39%
School age Learning and work % of children attending school in a mainstream class Higher is better 78%
School age Relationships and community % of children who can make friends with people outside the family Higher is better 66%
School age Relationships and community % of children who spend time after school and on weekends with friends and/or in mainstream programs Higher is better 42%
School age Relationships and community Of these, % who are welcomed or actively included Interpret with care 74%
School age Relationships and community % of children who spend time with friends without an adult present Higher is better 13%
Young people Choice and control % who are happy with the level of independence/control they have now Higher is better 30%
Young people Choice and control % who choose who supports them Higher is better 41%
Young people Choice and control % who choose what they do each day Higher is better 51%
Young people Choice and control % who had been given the opportunity to participate in a self-advocacy group meeting Higher is better 19%
Young people Choice and control % who want more choice and control in their life Lower may be better 82%
Young people Health and wellbeing % who rate their health as good, very good or excellent Higher is better 64%
Young people Health and wellbeing % who did not have any difficulties accessing health services Higher is better 65%
Young people Home and safety % who are happy with their home Higher is better 78%
Young people Home and safety % who feel safe or very safe in their home Higher is better 84%
Young people Learning and work % who currently attend or previously attended school in a mainstream class Higher is better 46%
Young people Learning and work % who have a paid job Higher is better 18%
Young people Learning and work % who volunteer Higher is better 9%
Young people Relationships and community % with no friends other than family or paid staff Lower is better 33%
Young people Relationships and community % who have been actively involved in a community, cultural or religious group in the last 12 months Higher is better 28%
Adults Choice and control % who choose who supports them Higher is better 66%
Adults Choice and control % who choose what they do each day Higher is better 73%
Adults Choice and control % who had been given the opportunity to participate in a self-advocacy group meeting Higher is better 24%
Adults Choice and control % who want more choice and control in their life Lower may be better 78%
Adults Development and daily living % who participate in education, training or skill development Higher is better 9%
Adults Health and wellbeing % who rate their health as good, very good or excellent Higher is better 39%
Adults Health and wellbeing % who did not have any difficulties accessing health services Higher is better 59%
Adults Home and safety % who are happy with their home Higher is better 67%
Adults Home and safety % who feel safe or very safe in their home Higher is better 67%
Adults Learning and work Of those who participate, % who do so in mainstream settings Higher is better 69%
Adults Learning and work % unable to do a course or training they wanted to do in the last 12 months Lower is better 36%
Adults Learning and work % who have a paid job Higher is better 23%
Adults Learning and work % who volunteer Higher is better 10%
Adults Relationships and community % with no friends other than family or paid staff Lower is better 34%
Adults Relationships and community % who have been actively involved in a community, cultural or religious group in the last 12 months Higher is better 30%

At baseline, the adult indicators show some obvious gaps as well as strengths. For example, many adults report choice over daily decisions and supports, but paid work and volunteering are much lower-level outcomes. For children, the outcome set is different: it is more about development, communication, friendships, community inclusion and school.

Change After Reassessment

The cleanest first longitudinal comparison is baseline to reassessment 1 among national participant rows with one plan reassessment. This keeps the comparison simple and avoids mixing people with very different reassessment histories.

Code 
participant_reassessment1 <- participant_longitudinal |>
  filter(number_of_plan_reassessments == 1) |>
  transmute(
    cohort_label,
    outcome_theme,
    indicator_number,
    indicator_description,
    direction,
    baseline = percentage_baseline,
    reassessment_1 = percentage_reassessment_1,
    raw_change = reassessment_1 - baseline,
    direction_adjusted_change = direction_adjust_change(raw_change, direction)
  )

comparable_reassessment1 <- participant_reassessment1 |>
  filter(is.finite(baseline), is.finite(reassessment_1), is.finite(direction_adjusted_change))
Code 
reassessment1_summary <- comparable_reassessment1 |>
  group_by(cohort_label, outcome_theme) |>
  summarise(
    outcomes = n_distinct(indicator_description),
    median_change = safe_median(direction_adjusted_change),
    p25 = quantile(direction_adjusted_change, 0.25, na.rm = TRUE),
    p75 = quantile(direction_adjusted_change, 0.75, na.rm = TRUE),
    .groups = "drop"
  ) |>
  filter(is.finite(median_change))

reassessment1_summary |>
  ggplot(aes(fct_reorder(outcome_theme, median_change), median_change)) +
  geom_hline(yintercept = 0, colour = "grey60") +
  geom_col(fill = "#7a5c2e") +
  geom_errorbar(aes(ymin = p25, ymax = p75), width = 0.2, colour = "grey35") +
  coord_flip() +
  facet_wrap(~ cohort_label, scales = "free_y") +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Direction-adjusted percentage-point change")

Median direction-adjusted change from baseline to reassessment 1.

Most of the comparable indicators move in a favourable direction after the first reassessment, but the size varies. The table below keeps the raw baseline and reassessment percentages visible, then adds a direction-adjusted change only where the direction is clear.

Code 
comparable_reassessment1 |>
  mutate(abs_change = abs(direction_adjusted_change)) |>
  arrange(desc(abs_change)) |>
  slice_head(n = 25) |>
  transmute(
    Cohort = as.character(cohort_label),
    Theme = outcome_theme,
    Outcome = indicator_description,
    Direction = direction,
    Baseline = percent(baseline, accuracy = 1),
    `Reassessment 1` = percent(reassessment_1, accuracy = 1),
    `Raw change` = percent(raw_change, accuracy = 0.1),
    `Adjusted change` = percent(direction_adjusted_change, accuracy = 0.1)
  ) |>
  kable()
Cohort Theme Outcome Direction Baseline Reassessment 1 Raw change Adjusted change
Early childhood Communication % who say their child is able to tell them what he/she wants Higher is better 62% 77% 15.0% 15.0%
Early childhood Relationships and community % of children who can make friends with people outside the family Higher is better 49% 59% 10.0% 10.0%
Early childhood Development and daily living % with concerns in 6 or more of the areas: gross motor skills, fine motor skills, self-care, eating, social interaction, communication, cognitive development, sensory processing Lower is better 69% 76% 7.0% -7.0%
School age Development and daily living % who say their child is becoming more independent Higher is better 43% 49% 6.0% 6.0%
Young people Learning and work % who have a paid job Higher is better 15% 19% 4.0% 4.0%
School age Learning and work % of children attending school in a mainstream class Higher is better 81% 78% -3.0% -3.0%
Young people Choice and control % who are happy with the level of independence/control they have now Higher is better 29% 32% 3.0% 3.0%
Young people Choice and control % who choose who supports them Higher is better 36% 39% 3.0% 3.0%
Adults Relationships and community % who have been actively involved in a community, cultural or religious group in the last 12 months Higher is better 29% 32% 3.0% 3.0%
Young people Choice and control % who choose what they do each day Higher is better 46% 49% 3.0% 3.0%
Young people Relationships and community % who have been actively involved in a community, cultural or religious group in the last 12 months Higher is better 28% 31% 3.0% 3.0%
School age Relationships and community % of children who can make friends with people outside the family Higher is better 67% 70% 3.0% 3.0%
Young people Health and wellbeing % who did not have any difficulties accessing health services Higher is better 70% 72% 2.0% 2.0%
Adults Home and safety % who are happy with their home Higher is better 66% 68% 2.0% 2.0%
Adults Health and wellbeing % who did not have any difficulties accessing health services Higher is better 59% 61% 2.0% 2.0%
Adults Learning and work % unable to do a course or training they wanted to do in the last 12 months Lower is better 37% 35% -2.0% 2.0%
Early childhood Relationships and community % of children who participate in age appropriate community, cultural or religious activities Higher is better 45% 47% 2.0% 2.0%
School age Development and daily living % developing functional, learning and coping skills appropriate to their ability and circumstances Higher is better 28% 30% 2.0% 2.0%
School age Choice and control % of children who have a genuine say in decisions about themselves Higher is better 79% 80% 1.0% 1.0%
School age Relationships and community % of children who spend time after school and on weekends with friends and/or in mainstream programs Higher is better 38% 39% 1.0% 1.0%
Young people Relationships and community % with no friends other than family or paid staff Lower is better 34% 33% -1.0% 1.0%
Young people Health and wellbeing % who rate their health as good, very good or excellent Higher is better 67% 66% -1.0% -1.0%
Young people Learning and work % who currently attend or previously attended school in a mainstream class Higher is better 46% 45% -1.0% -1.0%
Adults Choice and control % who choose what they do each day Higher is better 74% 75% 1.0% 1.0%
Adults Home and safety % who feel safe or very safe in their home Higher is better 64% 65% 1.0% 1.0%

Some longitudinal items ask directly whether the NDIS helped or improved something. These do not have a meaningful baseline percentage, so they are better read as reassessment-level responses rather than baseline-to-reassessment changes.

Code 
direct_improvement_items <- participant_reassessment1 |>
  filter(is.na(baseline), is.finite(reassessment_1), direction == "Reported improvement") |>
  arrange(cohort_label, outcome_theme, indicator_number)

direct_improvement_items |>
  transmute(
    Cohort = as.character(cohort_label),
    Theme = outcome_theme,
    Outcome = indicator_description,
    `Reported improvement at reassessment 1` = percent(reassessment_1, accuracy = 1)
  ) |>
  kable()
Cohort Theme Outcome Reported improvement at reassessment 1
Early childhood Communication Has the NDIS helped increase your child’s ability to communicate what they want? 85%
Early childhood Development and daily living Has the NDIS improved your child’s development? 92%
Early childhood Relationships and community Has the NDIS improved how your child fits into family life? 84%
Early childhood Relationships and community Has the NDIS improved how your child fits into community life? 71%
Early childhood Service access Has the NDIS improved your child’s access to specialist services? 93%
School age Development and daily living Has the NDIS helped your child to become more independent? 72%
School age Learning and work Has the NDIS improved your child’s access to education? 53%
School age Relationships and community Has the NDIS improved your child’s relationships with family and friends? 63%
School age Relationships and community Has the NDIS improved your child’s social and recreational life? 57%
Young people Choice and control Has the NDIS helped you have more choices and more control over your life? 69%
Young people Development and daily living Has the NDIS helped you with daily living activities? 66%
Young people Health and wellbeing Has your involvement with the NDIS improved your health and wellbeing? 52%
Young people Home and safety Has your involvement with the NDIS helped you to choose a home that’s right for you? 27%
Young people Learning and work Has your involvement with the NDIS helped you to learn things you want to learn or to take courses you want to take? 39%
Young people Learning and work Has your involvement with the NDIS helped you find a job that’s right for you? 23%
Young people Relationships and community Has the NDIS helped you to meet more people? 50%
Young people Relationships and community Has the NDIS helped you be more involved? 58%
Adults Choice and control Has the NDIS helped you have more choices and more control over your life? 78%
Adults Development and daily living Has the NDIS helped you with daily living activities? 80%
Adults Health and wellbeing Has your involvement with the NDIS improved your health and wellbeing? 62%
Adults Home and safety Has your involvement with the NDIS helped you to choose a home that’s right for you? 36%
Adults Learning and work Has your involvement with the NDIS helped you to learn things you want to learn or to take courses you want to take? 33%
Adults Learning and work Has your involvement with the NDIS helped you find a job that’s right for you? 21%
Adults Relationships and community Has the NDIS helped you to meet more people? 57%
Adults Relationships and community Has the NDIS helped you be more involved? 65%

Later Reassessments

The longitudinal file also includes reassessment 2 and later columns. Those are useful as a rough trend, but they should be read carefully: later reassessment rows are not a clean participant-level panel in this public extract, and the mix of participants changes as reassessment history gets longer.

For the later windows below, “reassessment 5” means the group with five plan reassessments, comparing its baseline value with its reassessment 5 value. It is a useful descriptive window, but it should not be read as the exact same participants shown in the reassessment 1 chart.

Code 
reassessment_cols <- names(participant_longitudinal) |>
  keep(\(nm) str_detect(nm, "^percentage_reassessment_"))

participant_trend <- participant_longitudinal |>
  filter(
    is.finite(percentage_baseline),
    direction %in% c("Higher is better", "Lower is better")
  ) |>
  pivot_longer(
    all_of(reassessment_cols),
    names_to = "reassessment",
    values_to = "reassessment_value"
  ) |>
  mutate(
    reassessment_number = as.integer(str_extract(reassessment, "[0-9]+")),
    raw_change = reassessment_value - percentage_baseline,
    direction_adjusted_change = direction_adjust_change(raw_change, direction)
  ) |>
  filter(
    reassessment_number == number_of_plan_reassessments,
    is.finite(direction_adjusted_change)
  )

trend_summary <- participant_trend |>
  group_by(reassessment_number) |>
  summarise(
    rows = n(),
    outcome_cohort_pairs = n_distinct(str_c(cohort_label, indicator_number, sep = "::")),
    median_change = safe_median(direction_adjusted_change),
    p25 = quantile(direction_adjusted_change, 0.25, na.rm = TRUE),
    p75 = quantile(direction_adjusted_change, 0.75, na.rm = TRUE),
    .groups = "drop"
  ) |>
  filter(rows >= 20)

if (nrow(trend_summary) > 0) {
  trend_summary |>
    ggplot(aes(reassessment_number, median_change)) +
    geom_hline(yintercept = 0, colour = "grey60") +
    geom_ribbon(aes(ymin = p25, ymax = p75), fill = "#b7c7a3", alpha = 0.45) +
    geom_line(colour = "#395f3b", linewidth = 0.8) +
    geom_point(colour = "#395f3b", size = 2) +
    scale_x_continuous(breaks = trend_summary$reassessment_number) +
    scale_y_continuous(labels = percent) +
    labs(x = "Reassessment number", y = "Direction-adjusted change from baseline")
} else {
  tibble(note = "There are not enough later reassessment rows to draw a stable trend.") |>
    kable()
}

Median direction-adjusted change by reassessment number.
Code 
trend_summary |>
  transmute(
    Reassessment = reassessment_number,
    Rows = comma(rows),
    `Outcome/cohort pairs` = outcome_cohort_pairs,
    `Median adjusted change` = percent(median_change, accuracy = 0.1),
    `25th percentile` = percent(p25, accuracy = 0.1),
    `75th percentile` = percent(p75, accuracy = 0.1)
  ) |>
  kable()
Reassessment Rows Outcome/cohort pairs Median adjusted change 25th percentile 75th percentile
1 39 39 1.0% 0.0% 2.5%
2 39 39 1.0% 0.0% 3.0%
3 39 39 1.0% -0.5% 5.5%
4 39 39 2.0% -1.0% 7.5%
5 39 39 2.0% -1.0% 8.5%
6 35 35 2.0% -1.0% 7.0%
7 35 35 2.0% -1.0% 8.5%
8 34 34 0.5% -1.7% 9.0%
9 27 27 3.0% -1.0% 12.5%

To make the later reassessment pattern easier to read, the next chart repeats the reassessment 1 idea for two later windows: one around the middle of the available reassessment history, and the latest window with enough comparable outcome rows.

Code 
later_windows <- trend_summary |>
  filter(reassessment_number > 1) |>
  arrange(reassessment_number)

selected_later_reassessments <- if (nrow(later_windows) > 0) {
  unique(c(
    later_windows$reassessment_number[[ceiling(nrow(later_windows) / 2)]],
    max(later_windows$reassessment_number)
  ))
} else {
  integer()
}

selected_later_summary <- participant_trend |>
  filter(reassessment_number %in% selected_later_reassessments) |>
  mutate(window = str_c("Reassessment ", reassessment_number)) |>
  group_by(window, reassessment_number, cohort_label, outcome_theme) |>
  summarise(
    rows = n(),
    median_change = safe_median(direction_adjusted_change),
    p25 = quantile(direction_adjusted_change, 0.25, na.rm = TRUE),
    p75 = quantile(direction_adjusted_change, 0.75, na.rm = TRUE),
    .groups = "drop"
  ) |>
  filter(is.finite(median_change)) |>
  mutate(cohort_theme = str_c(cohort_label, ": ", outcome_theme))

if (nrow(selected_later_summary) > 0) {
  selected_later_summary |>
    ggplot(aes(fct_reorder(cohort_theme, median_change), median_change)) +
    geom_hline(yintercept = 0, colour = "grey60") +
    geom_col(fill = "#7a5c2e") +
    geom_errorbar(aes(ymin = p25, ymax = p75), width = 0.2, colour = "grey35") +
    coord_flip() +
    facet_wrap(~ window, scales = "free_y", ncol = 1) +
    scale_y_continuous(labels = percent) +
    labs(x = NULL, y = "Direction-adjusted percentage-point change")
} else {
  tibble(note = "There are not enough later reassessment windows to draw selected comparisons.") |>
    kable()
}

Direction-adjusted change for middle and latest reassessment windows.

The detailed table keeps the raw baseline and later reassessment values visible. Positive adjusted changes mean the result moved in the favourable direction after accounting for whether a higher or lower raw percentage is better.

Code 
participant_trend |>
  filter(reassessment_number %in% selected_later_reassessments) |>
  mutate(abs_change = abs(direction_adjusted_change)) |>
  group_by(reassessment_number) |>
  slice_max(abs_change, n = 12, with_ties = FALSE) |>
  ungroup() |>
  arrange(reassessment_number, desc(abs_change)) |>
  transmute(
    Reassessment = reassessment_number,
    Cohort = as.character(cohort_label),
    Theme = outcome_theme,
    Outcome = indicator_description,
    Direction = direction,
    Baseline = percent(percentage_baseline, accuracy = 1),
    `Selected reassessment` = percent(reassessment_value, accuracy = 1),
    `Raw change` = percent(raw_change, accuracy = 0.1),
    `Adjusted change` = percent(direction_adjusted_change, accuracy = 0.1)
  ) |>
  kable()
Reassessment Cohort Theme Outcome Direction Baseline Selected reassessment Raw change Adjusted change
5 Early childhood Communication % who say their child is able to tell them what he/she wants Higher is better 17% 72% 55.0% 55.0%
5 Early childhood Relationships and community % of children who can make friends with people outside the family Higher is better 35% 67% 32.0% 32.0%
5 Early childhood Development and daily living % with concerns in 6 or more of the areas: gross motor skills, fine motor skills, self-care, eating, social interaction, communication, cognitive development, sensory processing Lower is better 49% 79% 30.0% -30.0%
5 Young people Learning and work % who have a paid job Higher is better 7% 27% 20.0% 20.0%
5 Early childhood Relationships and community % of children who participate in age appropriate community, cultural or religious activities Higher is better 35% 50% 15.0% 15.0%
5 Young people Choice and control % who choose what they do each day Higher is better 35% 50% 15.0% 15.0%
5 Young people Choice and control % who choose who supports them Higher is better 27% 41% 14.0% 14.0%
5 School age Choice and control % of children who have a genuine say in decisions about themselves Higher is better 62% 74% 12.0% 12.0%
5 School age Learning and work % of children attending school in a mainstream class Higher is better 69% 58% -11.0% -11.0%
5 Young people Home and safety % who are happy with their home Higher is better 85% 75% -10.0% -10.0%
5 Adults Learning and work % unable to do a course or training they wanted to do in the last 12 months Lower is better 38% 28% -10.0% 10.0%
5 School age Development and daily living % who say their child is becoming more independent Higher is better 42% 51% 9.0% 9.0%
9 Young people Learning and work % who have a paid job Higher is better 3% 30% 27.0% 27.0%
9 Young people Choice and control % who choose what they do each day Higher is better 18% 42% 24.0% 24.0%
9 Young people Choice and control % who choose who supports them Higher is better 12% 33% 21.0% 21.0%
9 Young people Choice and control % who had been given the opportunity to participate in a self-advocacy group meeting Higher is better 10% 29% 19.0% 19.0%
9 Young people Relationships and community % with no friends other than family or paid staff Lower is better 44% 29% -15.0% 15.0%
9 Adults Learning and work % unable to do a course or training they wanted to do in the last 12 months Lower is better 34% 19% -15.0% 15.0%
9 Young people Home and safety % who are happy with their home Higher is better 94% 79% -15.0% -15.0%
9 Adults Relationships and community % who have been actively involved in a community, cultural or religious group in the last 12 months Higher is better 36% 49% 13.0% 13.0%
9 Adults Development and daily living % who participate in education, training or skill development Higher is better 22% 9% -13.0% -13.0%
9 Adults Learning and work Of those who participate, % who do so in mainstream settings Higher is better 49% 61% 12.0% 12.0%
9 Young people Health and wellbeing % who did not have any difficulties accessing health services Higher is better 70% 79% 9.0% 9.0%
9 Adults Health and wellbeing % who rate their health as good, very good or excellent Higher is better 59% 51% -8.0% -8.0%

The later-reassessment pattern is best read as directional context, not as a precise individual trajectory. Still, it helps separate the first question from the bigger spend question: the public data does show reported outcome movement after entry, even before we try to link those changes to budgets, service markets or regional access.

Family And Carer Note

Family and carer outcomes are part of the public outcomes files, but they answer a different question. They are about the circumstances of the family or carer around the participant, not the participant’s own outcome level. For this reason, they sit beside the participant story rather than inside it.

Code 
family_reassessment1 <- family_longitudinal |>
  filter(number_of_plan_reassessments == 1) |>
  transmute(
    cohort_label,
    outcome_theme,
    indicator_description,
    direction,
    baseline = percentage_baseline,
    reassessment_1 = percentage_reassessment_1,
    raw_change = reassessment_1 - baseline,
    direction_adjusted_change = direction_adjust_change(raw_change, direction)
  )

family_summary <- family_reassessment1 |>
  group_by(cohort_label) |>
  summarise(
    comparable_outcomes = sum(is.finite(direction_adjusted_change)),
    median_adjusted_change = safe_median(direction_adjusted_change),
    direct_improvement_items = sum(is.na(baseline) & is.finite(reassessment_1)),
    median_direct_improvement = safe_median(reassessment_1[is.na(baseline) & is.finite(reassessment_1)]),
    .groups = "drop"
  )

family_summary |>
  transmute(
    Cohort = as.character(cohort_label),
    `Comparable outcomes` = comparable_outcomes,
    `Median adjusted change` = format_optional_percent(median_adjusted_change, accuracy = 0.1),
    `Direct improvement items` = direct_improvement_items,
    `Median direct improvement response` = format_optional_percent(median_direct_improvement, accuracy = 1)
  ) |>
  kable()
Cohort Comparable outcomes Median adjusted change Direct improvement items Median direct improvement response
Family/carer of children 12 2.0% 5 78%
Family/carer of adults 12 1.0% 4 61%

The family and carer rows are useful, especially for questions about informal care, work, confidence, health and service access. They should not be blended with participant rows when the main question is whether participant outcomes themselves changed.

Caveats

  • These are aggregate public survey rows, not participant-level records.
  • The reassessment comparisons are descriptive. They do not prove causation.
  • Later reassessment rows may reflect a different participant mix from earlier reassessment rows.
  • Direction-adjusting outcomes helps readability, but it does not make all indicators equally important.
  • Some items ask whether the NDIS helped or improved something. Those are reassessment responses, not true baseline-to-change measures.
  • Spend, provider access and regional service-market questions are deliberately left out of this simplified version.
]]> ndis outcomes disability australia https://blog.aydins-workbench.com/posts/ndis-outcomes-spend-service-access/ Wed, 10 Jun 2026 00:00:00 GMT Did the 1 November 2025 Bulk Billing Changes Shift Access? Aydin https://blog.aydins-workbench.com/posts/bulk-billing-changes-impact/ The 1 November 2025 bulk billing changes were designed to make it easier for general practices to bulk bill more patients. The narrowest measurable question is whether the share of GP attendances that were bulk billed moved after the change. The broader access question is harder: a higher bulk billing rate is not the same thing as more appointments, shorter waits or less delayed care.

This post treats access as two linked signals:

  1. the GP bulk billing rate, which is the main outcome;
  2. service-use and payment checks, so we do not confuse a billing-mix change with an access change.

The analysis is deliberately framed as an early signal. The policy began on 1 November 2025, the AIHW dashboard is updated quarterly, and AIHW treats the last 3 months of data as preliminary.

Code 
ensure_aihw_tableau(refresh = FALSE)

manifest <- read_csv(file.path(raw_dir, "aihw_tableau_manifest.csv"), show_col_types = FALSE)

national <- read_tableau("national") |>
  standardise_rate_sheet() |>
  mutate(month_index = as.integer(interval(min(service_month), service_month) %/% months(1)))

age_rates <- read_tableau("age") |>
  standardise_rate_sheet("age_group_series_alias", "Age group") |>
  mutate(group = str_replace(group, " - ", "-"))

seifa_rates <- read_tableau("seifa") |>
  standardise_rate_sheet("seifa_series_alias", "Socioeconomic quintile")

benefits_raw <- read_tableau("benefits")
benefits <- benefits_raw |>
  transmute(
    service_month = as.Date(.data[[first_col(benefits_raw, "^month_service_month_value$")]]),
    series = .data[[first_col(benefits_raw, "^chart_4_series_alias$")]],
    value = as.numeric(.data[[first_col(benefits_raw, "^sum_pivot_field_values_value$")]])
  ) |>
  filter(!is.na(service_month), is.finite(value), !is.na(series)) |>
  distinct()

lga_latest <- read_tableau("latest_lga_map") |>
  transmute(
    state = state1_alias,
    lga_code = as.integer(lga_code21_alias),
    lga_name = attr_lga_name_alias,
    latest_bulk_billing_rate = as_number(sum_pivot_field_values_copy_alias),
    remoteness_area = attr_main_ra_alias,
    seifa_quintile = as.integer(attr_main_seifa_alias)
  ) |>
  filter(is.finite(latest_bulk_billing_rate), !is.na(lga_code))

phn_path <- download_if_missing(source_urls$phn_quarterly, file.path(raw_dir, "phn_bulk_billing.xlsx"))
sa3_path <- download_if_missing(source_urls$sa3_quarterly, file.path(raw_dir, "sa3_quarterly.xlsx"))
mmm_path <- download_if_missing(source_urls$mmm_quarterly, file.path(raw_dir, "mmm_quarterly.xlsx"))
pcst_path <- download_if_missing(source_urls$primary_care_service_type, file.path(raw_dir, "primary_care_service_type.xlsx"))
phn_lga_path <- download_if_missing(source_urls$phn_lga, file.path(raw_dir, "phn_lga_concordance.xlsx"))

phn_quarterly <- read_excel(phn_path, sheet = "By BTOS and PHN", skip = 4) |>
  clean_df() |>
  transmute(
    state,
    phn = primary_health_network,
    quarter,
    service_type = broad_type_of_service,
    phn_bulk_billing_rate = as.numeric(mbs_bulk_billing_rate)
  ) |>
  filter(service_type == "Total GP Non-Referred Attendances", !is.na(phn_bulk_billing_rate))

phn_age <- read_excel(phn_path, sheet = "GP NRA by PHN and Age Group", skip = 4) |>
  clean_df() |>
  select(state, phn = primary_health_network, quarter, contains("bulk_billing_rate")) |>
  pivot_longer(contains("bulk_billing_rate"), names_to = "age_group", values_to = "bulk_billing_rate") |>
  mutate(
    age_group = case_when(
      str_detect(age_group, "under_16") ~ "0-15",
      str_detect(age_group, "16_to_64") ~ "16-64",
      str_detect(age_group, "65") ~ "65+",
      str_detect(age_group, "all_ages") ~ "All ages",
      TRUE ~ age_group
    ),
    bulk_billing_rate = as.numeric(bulk_billing_rate)
  ) |>
  filter(!is.na(bulk_billing_rate))

sa3_visits_raw <- read_excel(sa3_path, sheet = "GP NRA by SA3", skip = 4) |>
  clean_df()

sa3_visits <- sa3_visits_raw |>
  transmute(
    state,
    sa3 = sa3,
    quarter,
    quarter_date = quarter_to_date(quarter),
    service_type = broad_type_of_service,
    services = as.numeric(services),
    benefits = as.numeric(benefits),
    bulk_billed_services = as.numeric(bulk_billed_services),
    patient_billed_services = as.numeric(patient_billed_services),
    bulk_billing_rate = as.numeric(mbs_bulk_billing_rate),
    avg_patient_contribution_patient_billed = as.numeric(avg_patient_contribution_per_service_out_of_hospital_patient_billed),
    avg_patient_contribution_total = as.numeric(avg_patient_contribution_per_service_out_of_hospital_total)
  ) |>
  filter(
    service_type == "Total GP Non-Referred Attendances",
    !is.na(quarter_date),
    is.finite(services),
    services > 0,
    !sa3 %in% c("Australia", state),
    !str_detect(state, regex("unknown|other territories", ignore_case = TRUE)),
    !str_detect(sa3, regex("^unknown|unknown state", ignore_case = TRUE))
  ) |>
  mutate(
    sa3_key = paste(state, sa3, sep = " - "),
    quarter_status = case_when(
      quarter_date < transition_quarter_date ~ "Pre-policy",
      quarter_date == transition_quarter_date ~ "Transition quarter",
      quarter_date >= clean_post_quarter_date ~ "Clean post-policy",
      TRUE ~ NA_character_
    )
  )

sa3_exposure <- sa3_visits |>
  filter(
    quarter_date >= max(quarter_date[quarter_date < transition_quarter_date], na.rm = TRUE) %m-% months(9),
    quarter_date < transition_quarter_date
  ) |>
  group_by(sa3_key) |>
  summarise(
    baseline_quarters = n_distinct(quarter_date),
    baseline_services = mean(services, na.rm = TRUE),
    pre_policy_bulk_billing_rate = weighted_rate(bulk_billed_services, services),
    pre_policy_patient_contribution = weighted.mean(
      avg_patient_contribution_patient_billed,
      w = patient_billed_services,
      na.rm = TRUE
    ),
    .groups = "drop"
  ) |>
  mutate(
    pre_policy_non_bulk_billed_share = 1 - pre_policy_bulk_billing_rate,
    pre_policy_gap_10pp = pre_policy_non_bulk_billed_share / 0.1
  )

sa3_panel <- sa3_visits |>
  left_join(sa3_exposure, by = "sa3_key") |>
  mutate(
    log_services = log1p(services),
    clean_post_policy = quarter_date >= clean_post_quarter_date,
    exposure_band = cut(
      pre_policy_non_bulk_billed_share,
      breaks = quantile(pre_policy_non_bulk_billed_share, probs = c(0, 1 / 3, 2 / 3, 1), na.rm = TRUE),
      include.lowest = TRUE,
      labels = c("Lower room to move", "Middle", "Higher room to move")
    ),
    rate_check = bulk_billed_services / services,
    rate_difference = bulk_billing_rate - rate_check
  ) |>
  filter(
    is.finite(pre_policy_gap_10pp),
    is.finite(baseline_services),
    baseline_quarters == 4
  )

mmm_visits <- read_excel(mmm_path, sheet = "All Services", skip = 4) |>
  clean_df() |>
  transmute(
    modified_monash = modified_monash_group,
    quarter,
    quarter_date = quarter_to_date(quarter),
    service_type = broad_type_of_service,
    services = as.numeric(services),
    bulk_billed_services = as.numeric(bulk_billed_services),
    patient_billed_services = as.numeric(patient_billed_services),
    bulk_billing_rate = as.numeric(mbs_bulk_billing_rate)
  ) |>
  filter(
    service_type == "Total GP Non-Referred Attendances",
    !is.na(quarter_date),
    is.finite(services),
    services > 0,
    !modified_monash %in% c("Australia")
  ) |>
  mutate(
    quarter_status = case_when(
      quarter_date < transition_quarter_date ~ "Pre-policy",
      quarter_date == transition_quarter_date ~ "Transition quarter",
      quarter_date >= clean_post_quarter_date ~ "Clean post-policy",
      TRUE ~ NA_character_
    )
  )

pcst_quarters <- read_excel(pcst_path, sheet = "State", skip = 4) |>
  clean_df() |>
  transmute(
    quarter,
    quarter_date = quarter_to_date(quarter),
    service_type = broad_type_of_service,
    primary_care_service_type,
    services = as.numeric(services)
  ) |>
  filter(!is.na(quarter_date), service_type == "Total GP Non-Referred Attendances")

phn_lga <- read_excel(phn_lga_path) |>
  clean_df() |>
  transmute(
    lga_code = as.integer(lga_code_2021),
    lga_name = lga_name_2021,
    phn_code = phn_code_2023,
    phn = phn_name_2023,
    ratio = as.numeric(ratio_from_to)
  )

lga_phn_latest <- lga_latest |>
  left_join(phn_lga, by = "lga_code", suffix = c("", "_phn")) |>
  group_by(phn) |>
  summarise(
    lgas = n_distinct(lga_code),
    latest_lga_mean_rate = weighted.mean(latest_bulk_billing_rate, w = coalesce(ratio, 1), na.rm = TRUE),
    .groups = "drop"
  ) |>
  filter(!is.na(phn))

Source Audit

The AIHW monthly dashboard is the main source because it is dated by month of service and has the policy date built into the time series. The Department’s PHN workbook is added as a quarterly cross-check because it has current PHN and age information, but it is not identical to the AIHW dashboard.

Code 
audit <- tibble(
  source = c(
    "AIHW Tableau monthly dashboard",
    "Department quarterly PHN workbook",
    "Department quarterly SA3 workbook",
    "Department quarterly MMM workbook",
    "Department primary care service type workbook",
    "PHN 2023 to LGA 2021 concordance",
    "Modified Monash Model 2019 feature service"
  ),
  used_for = c(
    "Monthly national, age, SEIFA, benefits and latest LGA map",
    "Quarterly PHN and PHN by age cross-check",
    "Quarterly SA3 GP visit-volume panel",
    "Quarterly Modified Monash visit-volume cross-check",
    "Primary care service type source audit",
    "Approximate latest LGA-to-PHN grouping",
    "Documented as the correct MMM source; not spatially joined in this render"
  ),
  latest_period = c(
    format(max(national$service_month), "%B %Y"),
    phn_quarterly$quarter[[1]],
    sa3_panel |>
      filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
      pull(quarter) |>
      first(),
    mmm_visits |>
      filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
      pull(quarter) |>
      first(),
    pcst_quarters |>
      filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
      pull(quarter) |>
      first(),
    "2021 LGA to 2023 PHN",
    "2019 MMM polygons"
  ),
  caveat = c(
    "Last 3 months are preliminary and may revise",
    "Quarterly and processed-date style statistics may differ from AIHW",
    "December quarter 2025 is a transition quarter, not a clean post-policy period",
    "Coarser than SA3 but directly reports Modified Monash categories",
    "Used for source triangulation, not the main model",
    "Latest LGA means are not attendance-weighted",
    "A faithful MMM join needs spatial or mesh-block concordance work"
  )
)

audit |> kable()
source used_for latest_period caveat
AIHW Tableau monthly dashboard Monthly national, age, SEIFA, benefits and latest LGA map March 2026 Last 3 months are preliminary and may revise
Department quarterly PHN workbook Quarterly PHN and PHN by age cross-check 2025-26 Q3 (Mar Qtr) Quarterly and processed-date style statistics may differ from AIHW
Department quarterly SA3 workbook Quarterly SA3 GP visit-volume panel 2025-26 Q3 (Mar Qtr) December quarter 2025 is a transition quarter, not a clean post-policy period
Department quarterly MMM workbook Quarterly Modified Monash visit-volume cross-check 2025-26 Q3 (Mar Qtr) Coarser than SA3 but directly reports Modified Monash categories
Department primary care service type workbook Primary care service type source audit 2025-26 Q3 (Mar Qtr) Used for source triangulation, not the main model
PHN 2023 to LGA 2021 concordance Approximate latest LGA-to-PHN grouping 2021 LGA to 2023 PHN Latest LGA means are not attendance-weighted
Modified Monash Model 2019 feature service Documented as the correct MMM source; not spatially joined in this render 2019 MMM polygons A faithful MMM join needs spatial or mesh-block concordance work
Code 
manifest |> kable()
label sheet rows columns path source_workbook
national GP attendances 1 506 22 aihw_tableau_national.csv HWE-97-MBS-GP-bulk-billing-data_20260526
age GP attendances 2 1518 21 aihw_tableau_age.csv HWE-97-MBS-GP-bulk-billing-data_20260526
seifa GP attendances 3 2530 22 aihw_tableau_seifa.csv HWE-97-MBS-GP-bulk-billing-data_20260526
benefits GP attendances 4 1518 24 aihw_tableau_benefits.csv HWE-97-MBS-GP-bulk-billing-data_20260526
latest_lga_map Map 515 8 aihw_tableau_latest_lga_map.csv HWE-97-MBS-GP-bulk-billing-data_20260526

The extract found 506 monthly observations from February 1984 to March 2026. There are 6 months in the immediate pre-window and 5 months after the policy start.

National Signal

The national monthly rate is the cleanest first test. A one-off step after November is suggestive, but the stronger signal is a step that is larger than the pre-existing trend and is visible in the 3-month average.

Code 
national_summary <- window_compare(national) |>
  mutate(
    across(starts_with("rate"), \(x) percent(x, accuracy = 0.1)),
    change = percent(change, accuracy = 0.1)
  )

national_summary |> kable()
dimension group rate_May-Oct 2025 rate_Nov 2025 onward months_May-Oct 2025 months_Nov 2025 onward change
All All patients 78.1% 81.7% 6 5 3.6%
Code 
national_plot <- national |>
  arrange(service_month) |>
  mutate(
    rolling_3_month = (bulk_billing_rate + lag(bulk_billing_rate, 1) + lag(bulk_billing_rate, 2)) / 3
  ) |>
  filter(service_month >= as.Date("2023-01-01")) |>
  pivot_longer(c(bulk_billing_rate, rolling_3_month), names_to = "measure", values_to = "rate")

ggplot(national_plot, aes(service_month, rate, colour = measure)) +
  geom_vline(xintercept = policy_date, linetype = "dashed", colour = "grey40") +
  geom_line(linewidth = 0.8, na.rm = TRUE) +
  scale_y_continuous(labels = percent) +
  scale_colour_manual(
    values = c(bulk_billing_rate = "#1b6ca8", rolling_3_month = "#c44e52"),
    labels = c(bulk_billing_rate = "Monthly", rolling_3_month = "Rolling 3 months")
  ) +
  labs(x = NULL, y = "Bulk billing rate", colour = NULL)

National GP bulk billing rate, with a 3-month rolling average.
Code 
model_data <- national |>
  filter(service_month >= as.Date("2023-01-01")) |>
  mutate(
    post = service_month >= policy_date,
    months_since_policy = pmax(0, as.integer(interval(policy_date, service_month) %/% months(1))),
    month_index = as.integer(interval(min(service_month), service_month) %/% months(1))
  )

its_model <- lm(bulk_billing_rate ~ month_index + post + months_since_policy, data = model_data)

tidy(its_model, conf.int = TRUE) |>
  mutate(
    estimate = percent(estimate, accuracy = 0.01),
    conf.low = percent(conf.low, accuracy = 0.01),
    conf.high = percent(conf.high, accuracy = 0.01)
  ) |>
  kable()
term estimate std.error statistic p.value conf.low conf.high
(Intercept) 77.56% 0.0031667 244.9187761 0.0000000 76.92% 78.20%
month_index 0.01% 0.0001650 0.5360189 0.5953359 -0.02% 0.04%
postTRUE 3.24% 0.0080242 4.0401600 0.0002779 1.61% 4.87%
months_since_policy 0.30% 0.0029888 1.0059201 0.3213582 -0.31% 0.91%

The postTRUE coefficient is the immediate step after November 2025, after allowing for a linear pre/post time trend. This is useful as a screening estimate, not as a final causal estimate.

Who Moved?

The policy should matter most for patients who were not already covered by the old child/concession-heavy incentive structure. In the public AIHW age bands, that makes 16-64 the group to watch.

Code 
age_compare <- window_compare(age_rates) |>
  mutate(
    across(starts_with("rate"), \(x) percent(x, accuracy = 0.1)),
    change = percent(change, accuracy = 0.1)
  )

age_compare |> kable()
dimension group rate_May-Oct 2025 rate_Nov 2025 onward months_May-Oct 2025 months_Nov 2025 onward change
Age group 16-64 69.6% 75.9% 6 5 6.2%
Age group 65+ 87.4% 88.4% 6 5 1.0%
Age group 0-15 90.6% 90.8% 6 5 0.3%
Code 
age_rates |>
  filter(service_month >= as.Date("2023-01-01")) |>
  ggplot(aes(service_month, bulk_billing_rate, colour = group)) +
  geom_vline(xintercept = policy_date, linetype = "dashed", colour = "grey40") +
  geom_line(linewidth = 0.8) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Bulk billing rate", colour = "Age group")

Bulk billing rates by age group.
Code 
seifa_compare <- window_compare(seifa_rates) |>
  mutate(
    across(starts_with("rate"), \(x) percent(x, accuracy = 0.1)),
    change = percent(change, accuracy = 0.1)
  )

seifa_compare |> kable()
dimension group rate_May-Oct 2025 rate_Nov 2025 onward months_May-Oct 2025 months_Nov 2025 onward change
Socioeconomic quintile 2 81.3% 85.6% 6 5 4.3%
Socioeconomic quintile 3 80.6% 84.6% 6 5 4.0%
Socioeconomic quintile 4 74.4% 78.1% 6 5 3.7%
Socioeconomic quintile 1 (lowest socioeconomic area) 88.4% 92.1% 6 5 3.7%
Socioeconomic quintile 5 (highest socioeconomic area) 67.2% 69.7% 6 5 2.5%
Code 
seifa_rates |>
  filter(service_month >= as.Date("2023-01-01")) |>
  mutate(group = fct_relevel(group, "1 (lowest socioeconomic area)", "2", "3", "4", "5 (highest socioeconomic area)")) |>
  ggplot(aes(service_month, bulk_billing_rate, colour = group)) +
  geom_vline(xintercept = policy_date, linetype = "dashed", colour = "grey40") +
  geom_line(linewidth = 0.8) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Bulk billing rate", colour = "SEIFA quintile")

Bulk billing rates by socioeconomic quintile.

Payment Check

AIHW’s benefits measure includes Medicare benefits for bulk billed GP attendances and the matched bulk billing incentive items. It does not include the separate Bulk Billing Practice Incentive Program payments.

Code 
benefits |>
  filter(service_month >= as.Date("2023-01-01")) |>
  ggplot(aes(service_month, value, colour = series)) +
  geom_vline(xintercept = policy_date, linetype = "dashed", colour = "grey40") +
  geom_line(linewidth = 0.8) +
  scale_y_continuous(labels = dollar) +
  labs(x = NULL, y = "Monthly benefit amount", colour = NULL)

Bulk billed GP attendance benefits and related bulk billing incentives.
Code 
benefit_wide <- benefits |>
  filter(service_month >= as.Date("2025-05-01")) |>
  pivot_wider(names_from = series, values_from = value)

if (all(c("Bulk billing incentives", "Total benefit") %in% names(benefit_wide))) {
  benefit_wide |>
    transmute(
      service_month,
      incentive_share = `Bulk billing incentives` / `Total benefit`
    ) |>
    arrange(service_month) |>
    mutate(incentive_share = percent(incentive_share, accuracy = 0.1)) |>
    kable()
} else {
  tibble(note = "The expected incentive and total benefit series were not both present in the Tableau export.") |>
    kable()
}
service_month incentive_share
2025-05-01 18.7%
2025-06-01 18.5%
2025-07-01 19.2%
2025-08-01 19.1%
2025-09-01 19.2%
2025-10-01 18.9%
2025-11-01 27.1%
2025-12-01 27.7%
2026-01-01 26.9%
2026-02-01 26.6%
2026-03-01 27.0%

Geography

The AIHW dashboard gives a latest-month LGA map. That is not enough for an LGA-level before/after model, but it is enough to see where the current access surface is high or low. The PHN grouping below is an approximate LGA-to-PHN summary of latest LGA rates, not an attendance-weighted PHN statistic.

Code 
bind_rows(
  lga_latest |>
    arrange(latest_bulk_billing_rate) |>
    slice_head(n = 10) |>
    mutate(rank_type = "Lowest latest LGA rates"),
  lga_latest |>
    arrange(desc(latest_bulk_billing_rate)) |>
    slice_head(n = 10) |>
    mutate(rank_type = "Highest latest LGA rates")
) |>
  transmute(
    rank_type,
    state,
    lga_name,
    remoteness_area,
    seifa_quintile,
    latest_bulk_billing_rate = percent(latest_bulk_billing_rate, accuracy = 0.1)
  ) |>
  kable()
rank_type state lga_name remoteness_area seifa_quintile latest_bulk_billing_rate
Lowest latest LGA rates WA Peppermint Grove 0 (Major cities) NA 47.7%
Lowest latest LGA rates WA Cottesloe 0 (Major cities) NA 47.9%
Lowest latest LGA rates WA Mosman Park 0 (Major cities) NA 51.6%
Lowest latest LGA rates WA Subiaco 0 (Major cities) NA 51.9%
Lowest latest LGA rates NSW Mosman 0 (Major cities) NA 53.4%
Lowest latest LGA rates WA Claremont 0 (Major cities) NA 54.3%
Lowest latest LGA rates WA Nedlands 0 (Major cities) NA 54.6%
Lowest latest LGA rates ACT Unincorporated ACT 0 (Major cities) NA 55.0%
Lowest latest LGA rates WA Cambridge 0 (Major cities) NA 55.4%
Lowest latest LGA rates WA East Fremantle 0 (Major cities) NA 55.4%
Highest latest LGA rates Queensland Torres Strait Island 4 (Very remote) NA 98.9%
Highest latest LGA rates Queensland Torres 4 (Very remote) NA 98.9%
Highest latest LGA rates NSW Fairfield 0 (Major cities) NA 98.9%
Highest latest LGA rates SA Whyalla 2 (Outer regional) NA 98.8%
Highest latest LGA rates NT West Daly 3 (Remote) NA 98.7%
Highest latest LGA rates Tasmania Flinders (Tas.) 4 (Very remote) 2 98.4%
Highest latest LGA rates NT MacDonnell 4 (Very remote) NA 98.3%
Highest latest LGA rates NT Central Desert 4 (Very remote) NA 98.3%
Highest latest LGA rates NSW Cumberland 0 (Major cities) NA 98.3%
Highest latest LGA rates NSW Campbelltown (NSW) 0 (Major cities) 2 98.2%
Code 
lga_latest |>
  filter(!is.na(remoteness_area)) |>
  ggplot(aes(fct_reorder(remoteness_area, latest_bulk_billing_rate, .fun = median, na.rm = TRUE), latest_bulk_billing_rate)) +
  geom_boxplot(fill = "#d9e8f5", colour = "grey35", outlier.alpha = 0.35) +
  coord_flip() +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Latest LGA bulk billing rate")

Latest LGA bulk billing rates by AIHW remoteness field.
Code 
lga_phn_latest |>
  arrange(latest_lga_mean_rate) |>
  slice_head(n = 15) |>
  mutate(latest_lga_mean_rate = percent(latest_lga_mean_rate, accuracy = 0.1)) |>
  kable()
phn lgas latest_lga_mean_rate
Australian Capital Territory 1 55.0%
Perth North 18 61.9%
Northern Sydney 10 71.6%
Perth South 15 75.5%
Adelaide 18 75.8%
South Eastern Melbourne 14 77.9%
Central and Eastern Sydney 13 79.4%
Brisbane South 4 79.9%
Eastern Melbourne 13 80.2%
Central Queensland, Wide Bay, Sunshine Coast 11 82.4%
North Western Melbourne 13 82.8%
South Eastern NSW 12 83.8%
Country WA 98 83.9%
Hunter New England and Central Coast 23 83.9%
Brisbane North 5 84.1%

The Modified Monash Model source is available as a polygon feature service, but this render does not force a low-quality LGA-only MMM approximation. A proper MMM join should use a spatial intersection or a finer ABS geography concordance, then aggregate to the patient geography used by the Medicare data.

PHN Cross-Check

The Department’s PHN workbook is quarterly. It is useful as a second view of whether the post-policy shift is visible across PHNs, and whether the working-age pattern appears there too.

Code 
phn_quarterly |>
  filter(phn %in% c("Australia", "NSW", "VIC", "QLD", "SA", "WA", "TAS", "ACT", "NT")) |>
  arrange(quarter, state, phn) |>
  tail(20) |>
  mutate(phn_bulk_billing_rate = percent(phn_bulk_billing_rate, accuracy = 0.1)) |>
  kable()
state phn quarter service_type phn_bulk_billing_rate
SA SA 2024-25 Q4 (Jun Qtr) Total GP Non-Referred Attendances 76.5%
WA WA 2024-25 Q4 (Jun Qtr) Total GP Non-Referred Attendances 71.9%
ACT ACT 2025-26 Q1 (Sep Qtr) Total GP Non-Referred Attendances 51.4%
Australia Australia 2025-26 Q1 (Sep Qtr) Total GP Non-Referred Attendances 77.6%
NSW NSW 2025-26 Q1 (Sep Qtr) Total GP Non-Referred Attendances 82.2%
NT NT 2025-26 Q1 (Sep Qtr) Total GP Non-Referred Attendances 76.6%
SA SA 2025-26 Q1 (Sep Qtr) Total GP Non-Referred Attendances 74.5%
WA WA 2025-26 Q1 (Sep Qtr) Total GP Non-Referred Attendances 69.9%
ACT ACT 2025-26 Q2 (Dec Qtr) Total GP Non-Referred Attendances 53.0%
Australia Australia 2025-26 Q2 (Dec Qtr) Total GP Non-Referred Attendances 80.1%
NSW NSW 2025-26 Q2 (Dec Qtr) Total GP Non-Referred Attendances 84.2%
NT NT 2025-26 Q2 (Dec Qtr) Total GP Non-Referred Attendances 85.6%
SA SA 2025-26 Q2 (Dec Qtr) Total GP Non-Referred Attendances 78.2%
WA WA 2025-26 Q2 (Dec Qtr) Total GP Non-Referred Attendances 72.2%
ACT ACT 2025-26 Q3 (Mar Qtr) Total GP Non-Referred Attendances 54.1%
Australia Australia 2025-26 Q3 (Mar Qtr) Total GP Non-Referred Attendances 81.9%
NSW NSW 2025-26 Q3 (Mar Qtr) Total GP Non-Referred Attendances 85.6%
NT NT 2025-26 Q3 (Mar Qtr) Total GP Non-Referred Attendances 89.8%
SA SA 2025-26 Q3 (Mar Qtr) Total GP Non-Referred Attendances 80.4%
WA WA 2025-26 Q3 (Mar Qtr) Total GP Non-Referred Attendances 74.0%
Code 
latest_phn_quarter <- phn_quarterly$quarter[[1]]

phn_quarterly |>
  filter(quarter == latest_phn_quarter, !phn %in% c("Australia", state)) |>
  arrange(phn_bulk_billing_rate) |>
  slice_head(n = 15) |>
  transmute(
    quarter,
    state,
    phn,
    phn_bulk_billing_rate = percent(phn_bulk_billing_rate, accuracy = 0.1)
  ) |>
  kable()
quarter state phn phn_bulk_billing_rate
2025-26 Q3 (Mar Qtr) ACT Australian Capital Territory 54.1%
2025-26 Q3 (Mar Qtr) WA Perth North 68.0%
2025-26 Q3 (Mar Qtr) Qld Brisbane North 70.9%
2025-26 Q3 (Mar Qtr) NSW Northern Sydney 73.5%
2025-26 Q3 (Mar Qtr) WA Perth South 76.4%
2025-26 Q3 (Mar Qtr) NSW Hunter New England and Central Coast 77.8%
2025-26 Q3 (Mar Qtr) Tas Tasmania 78.0%
2025-26 Q3 (Mar Qtr) SA Adelaide 79.1%
2025-26 Q3 (Mar Qtr) NSW Central and Eastern Sydney 79.4%
2025-26 Q3 (Mar Qtr) Qld Brisbane South 79.6%
2025-26 Q3 (Mar Qtr) Qld Central Queensland, Wide Bay, Sunshine Coast 79.6%
2025-26 Q3 (Mar Qtr) Vic Eastern Melbourne 80.7%
2025-26 Q3 (Mar Qtr) WA Country WA 81.6%
2025-26 Q3 (Mar Qtr) Vic South Eastern Melbourne 81.7%
2025-26 Q3 (Mar Qtr) Qld Northern Queensland 82.0%
Code 
phn_age |>
  filter(quarter == latest_phn_quarter, phn == "Australia", age_group != "All ages") |>
  transmute(
    quarter,
    age_group,
    bulk_billing_rate = percent(bulk_billing_rate, accuracy = 0.1)
  ) |>
  kable()
quarter age_group bulk_billing_rate
2025-26 Q3 (Mar Qtr) 0-15 90.8%
2025-26 Q3 (Mar Qtr) 16-64 76.2%
2025-26 Q3 (Mar Qtr) 65+ 88.5%

Visit Access

The next test is stricter: did areas with more room for the incentive to shift billing also see more GP attendances? For this, the lowest-level public panel in the current official releases is SA3 by quarter. The December quarter 2025 is mixed, because October was pre-policy and November-December were post-policy, so it is labelled as a transition quarter and excluded from the fixed-effects model.

Code 
exposure_baseline_quarters <- sa3_panel |>
  filter(quarter_date < transition_quarter_date) |>
  distinct(quarter, quarter_date) |>
  arrange(desc(quarter_date)) |>
  slice_head(n = 4) |>
  arrange(quarter_date)

visit_access_audit <- tibble(
  metric = c(
    "Latest SA3 quarter",
    "Usable pre-policy quarters",
    "Exposure baseline window",
    "Transition quarter",
    "Visit model window starts",
    "Clean post-policy quarters",
    "SA3s in model panel",
    "Rows with missing model fields",
    "Largest rate validation difference"
  ),
  value = c(
    sa3_panel |>
      filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
      pull(quarter) |>
      first(),
    n_distinct(sa3_panel$quarter_date[sa3_panel$quarter_date < transition_quarter_date]),
    paste(exposure_baseline_quarters$quarter, collapse = " to "),
    "2025-26 Q2 (Dec Qtr), excluded from the model",
    sa3_panel |>
      filter(quarter_date == visit_model_start_quarter_date) |>
      pull(quarter) |>
      first(),
    n_distinct(sa3_panel$quarter_date[sa3_panel$quarter_date >= clean_post_quarter_date]),
    n_distinct(sa3_panel$sa3_key),
    sum(!complete.cases(sa3_panel[, c("services", "bulk_billed_services", "bulk_billing_rate", "pre_policy_gap_10pp")])),
    percent(max(abs(sa3_panel$rate_difference), na.rm = TRUE), accuracy = 0.001)
  )
)

visit_access_audit |> kable()
metric value
Latest SA3 quarter 2025-26 Q3 (Mar Qtr)
Usable pre-policy quarters 21
Exposure baseline window 2024-25 Q2 (Dec Qtr) to 2024-25 Q3 (Mar Qtr) to 2024-25 Q4 (Jun Qtr) to 2025-26 Q1 (Sep Qtr)
Transition quarter 2025-26 Q2 (Dec Qtr), excluded from the model
Visit model window starts 2023-24 Q1 (Sep Qtr)
Clean post-policy quarters 1
SA3s in model panel 332
Rows with missing model fields 0
Largest rate validation difference 0.000%
Code 
sa3_national_visits <- sa3_panel |>
  group_by(quarter_date, quarter_status) |>
  summarise(services = sum(services, na.rm = TRUE), .groups = "drop")

ggplot(sa3_national_visits, aes(quarter_date, services)) +
  geom_vline(xintercept = transition_quarter_date, linetype = "dotted", colour = "grey45") +
  geom_vline(xintercept = clean_post_quarter_date, linetype = "dashed", colour = "grey25") +
  geom_line(colour = "#1b6ca8", linewidth = 0.8) +
  geom_point(aes(shape = quarter_status), colour = "#1b6ca8", size = 2) +
  scale_y_continuous(labels = comma) +
  labs(x = NULL, y = "GP NRA services", shape = NULL)

Quarterly GP non-referred attendance services summed across SA3s.
Code 
sa3_growth <- sa3_panel |>
  mutate(window = case_when(
    quarter_date %in% exposure_baseline_quarters$quarter_date ~ "baseline",
    quarter_date >= clean_post_quarter_date ~ "post",
    TRUE ~ NA_character_
  )) |>
  filter(!is.na(window)) |>
  group_by(sa3_key, state, sa3, exposure_band, pre_policy_non_bulk_billed_share, baseline_services, window) |>
  summarise(
    services = mean(services, na.rm = TRUE),
    bb_rate = weighted_rate(bulk_billed_services, services),
    .groups = "drop"
  ) |>
  pivot_wider(names_from = window, values_from = c(services, bb_rate)) |>
  mutate(
    visit_growth = services_post / services_baseline - 1,
    bulk_billing_rate_change = bb_rate_post - bb_rate_baseline,
    exposure_label = as.character(exposure_band)
  ) |>
  filter(is.finite(visit_growth), is.finite(pre_policy_non_bulk_billed_share))

ggplot(sa3_growth, aes(visit_growth)) +
  geom_histogram(bins = 30, fill = "#d9e8f5", colour = "white") +
  geom_vline(xintercept = 0, linetype = "dashed", colour = "grey35") +
  scale_x_continuous(labels = percent) +
  labs(x = "Visit growth", y = "SA3s")

SA3 visit growth from the pre-policy exposure window to clean post-policy quarters.
Code 
band_trends <- sa3_panel |>
  group_by(exposure_band, quarter_date, quarter_status) |>
  summarise(services = sum(services, na.rm = TRUE), .groups = "drop") |>
  group_by(exposure_band) |>
  mutate(
    baseline_services = mean(services[quarter_date %in% exposure_baseline_quarters$quarter_date], na.rm = TRUE),
    visit_index = services / baseline_services
  ) |>
  ungroup()

ggplot(band_trends, aes(quarter_date, visit_index, colour = exposure_band)) +
  geom_vline(xintercept = transition_quarter_date, linetype = "dotted", colour = "grey45") +
  geom_vline(xintercept = clean_post_quarter_date, linetype = "dashed", colour = "grey25") +
  geom_line(linewidth = 0.8, na.rm = TRUE) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Visit volume index", colour = "Pre-policy non-bulk-billed share")

Longer-run quarterly GP visit volume indexed to the four-quarter pre-policy exposure window.

The 2021-22 peak is real in the official service counts, not an indexing artefact. It lines up with the pandemic-era disruption and rebound in GP claiming, so it is a poor baseline for the 2025 policy question. The chart below keeps the same index but starts after that shock; the visit-volume models also use this flatter window.

Code 
band_trends |>
  filter(quarter_date >= visit_model_start_quarter_date) |>
  ggplot(aes(quarter_date, visit_index, colour = exposure_band)) +
  geom_vline(xintercept = transition_quarter_date, linetype = "dotted", colour = "grey45") +
  geom_vline(xintercept = clean_post_quarter_date, linetype = "dashed", colour = "grey25") +
  geom_line(linewidth = 0.8, na.rm = TRUE) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Visit volume index", colour = "Pre-policy non-bulk-billed share")

Post-spike quarterly GP visit volume indexed to the four-quarter pre-policy exposure window.
Code 
ggplot(sa3_growth, aes(pre_policy_non_bulk_billed_share, visit_growth)) +
  geom_hline(yintercept = 0, linetype = "dashed", colour = "grey50") +
  geom_point(aes(size = baseline_services), alpha = 0.35, colour = "#1b6ca8") +
  geom_smooth(aes(weight = baseline_services), method = "lm", se = TRUE, colour = "#c44e52") +
  scale_x_continuous(labels = percent) +
  scale_y_continuous(labels = percent) +
  scale_size_continuous(labels = comma, range = c(1, 7)) +
  labs(
    x = "Pre-policy non-bulk-billed share",
    y = "Visit growth",
    size = "Baseline services"
  )

Areas with lower pre-policy bulk billing had more room for the incentives to shift billing behaviour.
Code 
ggplot(sa3_growth, aes(bulk_billing_rate_change, visit_growth)) +
  geom_hline(yintercept = 0, linetype = "dashed", colour = "grey50") +
  geom_vline(xintercept = 0, linetype = "dashed", colour = "grey50") +
  geom_point(aes(size = baseline_services), alpha = 0.35, colour = "#1b6ca8") +
  geom_smooth(aes(weight = baseline_services), method = "lm", se = TRUE, colour = "#c44e52") +
  scale_x_continuous(labels = percent) +
  scale_y_continuous(labels = percent) +
  scale_size_continuous(labels = comma, range = c(1, 7)) +
  labs(
    x = "Observed bulk billing rate change",
    y = "Visit growth",
    size = "Baseline services"
  )

Observed SA3 bulk billing rate change versus visit growth.
Code 
access_model_panel <- sa3_panel |>
  filter(
    quarter_date >= visit_model_start_quarter_date,
    quarter_date != transition_quarter_date
  ) |>
  mutate(post_x_gap_10pp = as.integer(clean_post_policy) * pre_policy_gap_10pp)

fit_access_model <- function(outcome, label, outcome_type) {
  model_data <- access_model_panel |>
    filter(
      is.finite(.data[[outcome]]),
      is.finite(post_x_gap_10pp),
      is.finite(baseline_services),
      baseline_services > 0
    )

  model <- lm(
    reformulate(
      c("factor(sa3_key)", "factor(quarter_date)", "post_x_gap_10pp"),
      response = outcome
    ),
    data = model_data,
    weights = baseline_services
  )

  tidy(model, conf.int = TRUE) |>
    filter(term == "post_x_gap_10pp") |>
    transmute(
      outcome = label,
      interpretation = "Clean post-policy change per 10 percentage point higher pre-policy non-bulk-billed share",
      estimate = format_model_estimate(estimate, outcome_type),
      conf_low = format_model_estimate(conf.low, outcome_type),
      conf_high = format_model_estimate(conf.high, outcome_type),
      rows = nrow(model_data),
      sa3s = n_distinct(model_data$sa3_key)
    )
}

access_model_summary <- bind_rows(
  fit_access_model("log_services", "GP visit volume", "log"),
  fit_access_model("bulk_billing_rate", "Bulk billing rate", "rate"),
  fit_access_model("avg_patient_contribution_patient_billed", "Patient contribution among patient-billed services", "dollar")
)

access_model_summary |> kable()
outcome interpretation estimate conf_low conf_high rows sa3s
GP visit volume Clean post-policy change per 10 percentage point higher pre-policy non-bulk-billed share -0.4% -0.6% -0.1% 3320 332
Bulk billing rate Clean post-policy change per 10 percentage point higher pre-policy non-bulk-billed share -0.43% -0.59% -0.27% 3320 332
Patient contribution among patient-billed services Clean post-policy change per 10 percentage point higher pre-policy non-bulk-billed share -$1.11 -$1.30 -$0.93 3320 332

The first row is the key access test. A positive estimate would mean higher-exposure areas had faster visit-volume growth after the clean post-policy quarter. The second and third rows are mechanism checks: the policy story is stronger if the same high-exposure areas also show larger bulk billing rate gains or lower patient contributions.

Code 
reference_quarter <- max(access_model_panel$quarter_date[access_model_panel$quarter_date < transition_quarter_date], na.rm = TRUE)

event_study_data <- access_model_panel |>
  select(sa3_key, quarter_date, log_services, pre_policy_gap_10pp, baseline_services) |>
  left_join(
    access_model_panel |>
      filter(quarter_date == reference_quarter) |>
      select(sa3_key, reference_log_services = log_services),
    by = "sa3_key"
  ) |>
  mutate(delta_log_services = log_services - reference_log_services) |>
  filter(is.finite(delta_log_services), is.finite(pre_policy_gap_10pp), is.finite(baseline_services))

event_study <- event_study_data |>
  group_by(quarter_date) |>
  group_modify(\(.x, .y) {
    tidy(
      lm(delta_log_services ~ pre_policy_gap_10pp, data = .x, weights = baseline_services),
      conf.int = TRUE
    ) |>
      filter(term == "pre_policy_gap_10pp")
  }) |>
  ungroup() |>
  mutate(
    estimate_pct = exp(estimate) - 1,
    conf_low_pct = exp(conf.low) - 1,
    conf_high_pct = exp(conf.high) - 1,
    quarter_status = case_when(
      quarter_date < transition_quarter_date ~ "Pre-policy",
      quarter_date >= clean_post_quarter_date ~ "Clean post-policy",
      TRUE ~ "Transition quarter"
    )
  )

ggplot(event_study, aes(quarter_date, estimate_pct)) +
  geom_hline(yintercept = 0, colour = "grey55") +
  geom_vline(xintercept = transition_quarter_date, linetype = "dotted", colour = "grey45") +
  geom_vline(xintercept = clean_post_quarter_date, linetype = "dashed", colour = "grey25") +
  geom_ribbon(aes(ymin = conf_low_pct, ymax = conf_high_pct), fill = "#d9e8f5", alpha = 0.7) +
  geom_line(colour = "#1b6ca8", linewidth = 0.8) +
  geom_point(aes(shape = quarter_status), colour = "#1b6ca8", size = 2) +
  scale_y_continuous(labels = percent) +
  labs(
    x = NULL,
    y = "Differential visit change per 10pp exposure",
    shape = NULL
  )

Differential SA3 visit-volume change by pre-policy non-bulk-billed share, relative to September quarter 2025.
Code 
mmm_trends <- mmm_visits |>
  filter(modified_monash != "Unknown") |>
  group_by(modified_monash, quarter_date, quarter_status) |>
  summarise(services = sum(services, na.rm = TRUE), .groups = "drop") |>
  group_by(modified_monash) |>
  mutate(
    baseline_services = mean(services[quarter_date %in% exposure_baseline_quarters$quarter_date], na.rm = TRUE),
    visit_index = services / baseline_services
  ) |>
  ungroup()

ggplot(mmm_trends, aes(quarter_date, visit_index, colour = modified_monash)) +
  geom_vline(xintercept = transition_quarter_date, linetype = "dotted", colour = "grey45") +
  geom_vline(xintercept = clean_post_quarter_date, linetype = "dashed", colour = "grey25") +
  geom_line(linewidth = 0.8, na.rm = TRUE) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Visit volume index", colour = "Modified Monash")

Modified Monash GP visit volume indexed to the four-quarter pre-policy exposure window.
Code 
unknown_mmm <- mmm_visits |>
  group_by(quarter, quarter_date) |>
  summarise(
    total_services = sum(services, na.rm = TRUE),
    unknown_services = sum(services[modified_monash == "Unknown"], na.rm = TRUE),
    unknown_share = unknown_services / total_services,
    .groups = "drop"
  ) |>
  filter(quarter_date >= as.Date("2023-09-01")) |>
  arrange(quarter_date)

unknown_mmm |>
  mutate(
    total_services = comma(total_services),
    unknown_services = comma(unknown_services),
    unknown_share = percent(unknown_share, accuracy = 0.001)
  ) |>
  kable()
quarter quarter_date total_services unknown_services unknown_share
2023-24 Q1 (Sep Qtr) 2023-09-01 41,611,756 419 0.001%
2023-24 Q2 (Dec Qtr) 2023-12-01 39,169,904 402 0.001%
2023-24 Q3 (Mar Qtr) 2024-03-01 39,909,885 861 0.002%
2023-24 Q4 (Jun Qtr) 2024-06-01 43,838,981 9,111 0.021%
2024-25 Q1 (Sep Qtr) 2024-09-01 43,278,054 8,829 0.020%
2024-25 Q2 (Dec Qtr) 2024-12-01 40,206,213 8,403 0.021%
2024-25 Q3 (Mar Qtr) 2025-03-01 40,481,593 8,076 0.020%
2024-25 Q4 (Jun Qtr) 2025-06-01 43,733,040 7,893 0.018%
2025-26 Q1 (Sep Qtr) 2025-09-01 42,347,110 7,466 0.018%
2025-26 Q2 (Dec Qtr) 2025-12-01 39,768,686 7,030 0.018%
2025-26 Q3 (Mar Qtr) 2026-03-01 39,968,536 6,986 0.017%

The Unknown Modified Monash category is excluded from the indexed MMM chart because it is tiny in absolute terms and appears to have a classification break around the June quarter 2024. It rises from hundreds of services per quarter to roughly 7,000-9,000, but still remains around 0.021% of GP non-referred attendances. That makes it visually noisy as an index and not a meaningful access geography.

Code 
sa3_growth |>
  group_by(exposure_band) |>
  summarise(
    sa3s = n(),
    median_pre_policy_non_bulk_billed_share = median(pre_policy_non_bulk_billed_share, na.rm = TRUE),
    median_visit_growth = median(visit_growth, na.rm = TRUE),
    weighted_visit_growth = weighted.mean(visit_growth, w = baseline_services, na.rm = TRUE),
    .groups = "drop"
  ) |>
  mutate(
    median_pre_policy_non_bulk_billed_share = percent(median_pre_policy_non_bulk_billed_share, accuracy = 0.1),
    median_visit_growth = percent(median_visit_growth, accuracy = 0.1),
    weighted_visit_growth = percent(weighted_visit_growth, accuracy = 0.1)
  ) |>
  kable()
exposure_band sa3s median_pre_policy_non_bulk_billed_share median_visit_growth weighted_visit_growth
Lower room to move 111 14.7% -3.8% -3.8%
Middle 111 22.9% -4.5% -3.8%
Higher room to move 110 34.3% -5.3% -5.0%

Early Answer

Code 
national_delta <- window_compare(national)$change[[1]]
age_delta <- window_compare(age_rates) |>
  select(group, change) |>
  arrange(desc(change))
seifa_delta <- window_compare(seifa_rates) |>
  select(group, change) |>
  arrange(desc(change))
visit_access_estimate <- access_model_summary |>
  filter(outcome == "GP visit volume") |>
  pull(estimate) |>
  first()

tibble(
  question = c(
    "Did the national rate move after 1 November 2025?",
    "Was the largest age signal in working-age adults?",
    "Did lower and higher socioeconomic areas both move?",
    "Did higher-exposure SA3s also show more GP visits?",
    "Can we call this access rather than billing?"
  ),
  early_read = c(
    paste0("National before/after change: ", percent(national_delta, accuracy = 0.1), "."),
    paste0("Largest age-band change: ", age_delta$group[[1]], " at ", percent(age_delta$change[[1]], accuracy = 0.1), "."),
    paste0("Largest SEIFA change: ", seifa_delta$group[[1]], " at ", percent(seifa_delta$change[[1]], accuracy = 0.1), "."),
    paste0("SA3 fixed-effects visit estimate: ", visit_access_estimate, " per 10 percentage point higher pre-policy non-bulk-billed share."),
    "Not fully. The data show billing and service-use/payment signals, not unmet need or appointment availability."
  )
) |>
  kable()
question early_read
Did the national rate move after 1 November 2025? National before/after change: 3.6%.
Was the largest age signal in working-age adults? Largest age-band change: 16-64 at 6.2%.
Did lower and higher socioeconomic areas both move? Largest SEIFA change: 2 at 4.3%.
Did higher-exposure SA3s also show more GP visits? SA3 fixed-effects visit estimate: -0.4% per 10 percentage point higher pre-policy non-bulk-billed share.
Can we call this access rather than billing? Not fully. The data show billing and service-use/payment signals, not unmet need or appointment availability.

The early read is that the 1 November 2025 changes should be treated as having shifted measured bulk billing access if the post-November lift is visible nationally, strongest for 16-64, and not limited to places that already had high bulk billing. The broader access case is stronger if the SA3 visit-volume model shows higher-exposure areas also gaining visits after the clean post-policy quarter. The evidence is weaker if the rise is small relative to normal monthly volatility, if it does not persist in the rolling average, or if service-use checks deteriorate.

Caveats

  • AIHW’s monthly dashboard is based on date of service; Department Medicare statistics can differ because official Medicare statistics are generally processed-date based.
  • AIHW notes that the last 3 months of monthly dashboard data are preliminary.
  • Bulk billing rate is a claim-level access proxy, not a direct measure of unmet need, waiting time or clinic availability.
  • The AIHW benefits series excludes Bulk Billing Practice Incentive Program payments.
  • LGA statistics are based on patient location. People can travel outside their LGA for care.
  • The SA3 visit model uses service counts rather than per-capita rates. SA3 fixed effects absorb stable differences between areas, but population growth can still affect visit volumes.
  • The December quarter 2025 is excluded from the visit model because it mixes one pre-policy month with two post-policy months.
  • PHN and LGA summaries in this post mix monthly AIHW dashboard data and quarterly Department statistics; they are used as triangulation, not as one single modelled dataset.
  • Modified Monash should be added with a proper spatial or fine-geography concordance before making MMM-specific claims.
]]> medicare bulk-billing health-policy australia https://blog.aydins-workbench.com/posts/bulk-billing-changes-impact/ Thu, 28 May 2026 00:00:00 GMT Questions, Data Sources and Half-Decent Rabbit Holes Aydin https://blog.aydins-workbench.com/posts/australian-data-sources/ This site is basically a place to ask public policy questions and see whether public data can give us better context, maybe even the occasional answer.

The vibe is: pick a real policy question, find the best public dataset I can, do the analysis transparently, and be honest about what the data cannot tell us. A lot of this is for fun, interest, and keeping the data muscles from getting too creaky. So, as ever: grain of salt, ideally iodised.

Public datasets are useful, but they are rarely complete. They usually show service use, claims, budgets, or reported outcomes. They do not magically reveal need, quality, unmet demand, implementation messiness, or the lived experience behind the row counts. Still, they are often enough to make a policy conversation a bit less foggy.

How I Want Posts To Work

The rough template is:

  1. Ask one question.
  2. Find one credible dataset.
  3. Make a couple of clear charts.
  4. Explain the main pattern.
  5. Say what would change my mind.

That last part matters. The goal is to be interesting without pretending to be the final word.

Datasets Worth Keeping Handy

Health, Care And Disability

Population, Places And Context

Housing And Cost Of Living

Running Policy Question List

Health Access

  • Did the expanded bulk billing incentives from 1 November 2025 change GP bulk billing rates, appointment volumes, or just the billing mix?
  • Did the new Bulk Billing Practice Incentive Program show up differently in metropolitan, regional and remote areas?
  • Are areas with low GP bulk billing also showing more emergency department pressure?
  • Are MBS telehealth patterns settling into a stable access role, or fading back toward pre-pandemic geography?
  • Do areas with lower specialist service use have higher potentially preventable hospitalisations?
  • After PBS co-payment changes, are chronic-disease medicine claims changing more for concession-heavy areas?

NDIS, Disability And Outcomes

  • Is NDIS participation growing faster than the population, and where is that growth most concentrated?
  • Are utilisation gaps more about provider supply, participant mix, thin markets, administrative friction, or plan timing?
  • Which disability groups are driving payment growth, and are those changes broad-based or concentrated in particular supports?
  • As new NDIS framework planning begins from mid-2026, do plan budgets become more consistent for similar public-data groups?
  • Can participant outcome indicators be treated as investment signals, not just nice-to-have survey outcomes?
  • Are regions with stronger service access also showing better NDIS outcomes over time?

Aged Care

  • After the new Aged Care Act and Support at Home started on 1 November 2025, do hospital discharge and aged-care access indicators move differently by region?
  • Are home-care service volumes changing in ways that suggest substitution away from residential care or hospital stays?
  • Which regions have the biggest mismatch between ageing population growth and aged-care service availability?

Housing And Cost Of Living

  • After the 5% Deposit Scheme expansion from 1 October 2025, did first-home buyer loan commitments rise in the places where prices were already moving fastest?
  • Are first-home buyer supports being offset by higher mean dwelling prices, or are they mainly changing who can enter the market?
  • Are building approvals rising in the places with the fastest dwelling-price growth, or are supply signals lagging demand signals?
  • Are investor lending changes leading or following price changes across states?
  • Do rents, dwelling prices, income and population growth tell the same story by region, or are they pulling in different directions?

Climate, Housing And Health

  • Which urban areas combine heat exposure, poor housing quality, low income and higher chronic disease risk?
  • Do energy-bill relief or efficiency programs show up in hardship, health or consumption indicators?
  • Are climate-related health risks visible in emergency presentations, ambulance callouts or mortality data by region?

Modelling Notes

  • Treat MBS and PBS as claims and service-use datasets, not direct measures of health need.
  • Treat budgets, payments and outcomes as different things. They overlap, but they do not answer the same question.
  • For regional modelling, add ABS context: age structure, income, labour force status, housing, remoteness, disability prevalence, language and area disadvantage.
  • Use time-aware validation when possible: train on earlier periods, test on later ones, and keep policy-change dates explicit.
  • Do not turn prediction into allocation advice. For public writing, models are better framed as descriptive, diagnostic or forecasting tools.

Post Selection Checklist

Before starting an analysis:

  • Is the question specific enough for one post?
  • Is the dataset public and stable enough to rerender?
  • Are the definitions readable for someone who is not already living inside the spreadsheet?
  • Can the analysis be reproduced from code?
  • Is there a caveat section that actually says something useful?

That is the whole operating system, really. Ask the question, check the data, stay curious, do not overclaim.

]]> data australia health policy methods https://blog.aydins-workbench.com/posts/australian-data-sources/ Tue, 26 May 2026 00:00:00 GMT What Is Driving NDIS Spend Growth? Aydin https://blog.aydins-workbench.com/posts/ndis-budget-utilisation-prediction/ The National Disability Insurance Scheme is usually discussed through one big number: total spend. That is not wrong, but it hides several different mechanisms. Spend can grow because more people enter the scheme, because average plan budgets rise, because participants use more of their approved supports, or because the participant mix changes toward groups with higher support needs.

This post uses the short public quarterly window to tell a descriptive spend-growth story:

  1. How quickly is observed NDIS spend growing?
  2. Is scheme participation growing faster than the Australian population?
  3. Is spend pressure coming more from participant growth, average budgets, or utilisation?
  4. Which disability groups appear largest, most expensive, and fastest growing?
  5. What public explanatory factors help describe grouped quarterly spend?

The models are descriptive, not funding advice and not a participant-level forecast. I deliberately avoid lagged payment and lagged budget features in the overall spend model: those improve fit, but they mostly tell us that last quarter predicts this quarter. The aim here is to describe observable drivers.

Data Sources

The workflow downloads source files from the NDIS participant datasets, NDIS provider datasets, and NDIS payments datasets. It focuses on plan budgets, utilisation, providers, market concentration, payments, and LGA mapping files. It also uses ABS estimated resident population from National, state and territory population, then tries to add regional context from SEIFA 2021 and population age structure. If the ABS files change shape, the NDIS-only analysis still runs and reports the missing context.

Code 
clean_names <- function(x) {
  x |>
    str_to_lower() |>
    str_replace_all("&", "and") |>
    str_replace_all("[^a-z0-9]+", "_") |>
    str_replace_all("^_|_$", "")
}

clean_df <- function(x) {
  names(x) <- clean_names(names(x))
  x
}

as_number <- function(x) {
  if (is.numeric(x)) return(x)
  x_chr <- as.character(x)
  out <- parse_number(x_chr, na = c("", "NA", "n/a", "Not applicable", "No data available"))
  ifelse(str_detect(x_chr, "^\\s*<"), NA_real_, out)
}

as_rate <- function(x) {
  out <- as_number(x)
  ifelse(out > 1.5, out / 100, out)
}

safe_sum <- function(x) {
  x <- x[is.finite(x)]
  if (length(x) == 0) return(NA_real_)
  sum(x)
}

safe_mean <- function(x) {
  x <- x[is.finite(x)]
  if (length(x) == 0) return(NA_real_)
  mean(x)
}

safe_median <- function(x) {
  x <- x[is.finite(x)]
  if (length(x) == 0) return(NA_real_)
  median(x)
}

safe_mode_character <- function(x) {
  x <- as.character(x)
  x <- x[!is.na(x) & x != "" & x != "ALL"]
  if (length(x) == 0) return(NA_character_)
  names(sort(table(x), decreasing = TRUE))[[1]]
}

collapse_top_values <- function(x, n = 4) {
  x <- as.character(x)
  x <- unique(x[!is.na(x) & x != "" & x != "ALL"])
  if (length(x) == 0) return(NA_character_)
  paste(head(sort(x), n), collapse = ", ")
}

safe_weighted_mean <- function(x, w) {
  x <- as.numeric(x)
  w <- as.numeric(w)
  keep <- is.finite(x) & is.finite(w) & w > 0
  if (any(keep)) return(weighted.mean(x[keep], w[keep]))
  safe_mean(x)
}

has_enough_data <- function(data, cols, min_rows = 2) {
  if (nrow(data) < min_rows) return(FALSE)
  all(map_lgl(cols, \(col) col %in% names(data) && sum(is.finite(data[[col]])) >= min_rows))
}

format_optional_number <- function(x, accuracy = 0.1, suffix = "") {
  ifelse(is.na(x) | !is.finite(x), "Not joined", paste0(number(x, accuracy = accuracy), suffix))
}

format_optional_percent <- function(x, accuracy = 0.1) {
  ifelse(is.na(x) | !is.finite(x), "Not joined", percent(x, accuracy = accuracy))
}

index_to_first <- function(x) {
  bases <- x[is.finite(x) & x > 0]
  base <- if (length(bases) > 0) bases[[1]] else NA_real_
  if (length(base) == 0 || is.na(base)) return(rep(NA_real_, length(x)))
  x / base - 1
}

pretty_feature <- function(x) {
  recode(
    x,
    quarter_index = "time",
    log_participants = "participant count",
    log_providers = "active providers",
    log_participants_per_provider = "participants per active provider",
    payment_share_top10 = "top-10 provider payment share",
    irsd_score = "SEIFA disadvantage score",
    irsd_decile = "SEIFA disadvantage decile",
    ier_score = "economic resources score",
    ier_decile = "economic resources decile",
    ieo_score = "education/occupation score",
    ieo_decile = "education/occupation decile",
    median_age = "median local age",
    pct_0_14 = "local population aged 0-14",
    pct_15_64 = "local population aged 15-64",
    pct_65_plus = "local population aged 65+",
    log_lga_population = "local population",
    remoteness_rank = "remoteness",
    remoteness_major_cities_share = "major-cities share",
    remoteness_inner_regional_share = "inner-regional share",
    remoteness_outer_regional_share = "outer-regional share",
    remoteness_remote_share = "remote share",
    remoteness_very_remote_share = "very-remote share",
    log_budget = "average support budget",
    state = "state",
    service_district = "service district",
    age_group = "age group",
    disability_group = "disability group",
    support_class = "support class",
    remoteness = "remoteness",
    .default = str_replace_all(x, "_", " ")
  )
}

top_rf_predictors <- function(model, n = 5) {
  importance(model) |>
    as.data.frame() |>
    tibble::rownames_to_column("feature") |>
    arrange(desc(IncNodePurity)) |>
    slice_head(n = n) |>
    pull(feature) |>
    pretty_feature() |>
    paste(collapse = ", ")
}

first_col <- function(data, patterns, required = TRUE) {
  hits <- names(data)[map_lgl(names(data), \(nm) any(str_detect(nm, regex(paste(patterns, collapse = "|"), ignore_case = TRUE))))]
  if (length(hits) == 0) {
    if (required) stop("Could not find column matching: ", paste(patterns, collapse = ", "))
    return(NA_character_)
  }
  hits[[1]]
}

sample_rows <- function(data, n = 8000) {
  data |> slice_sample(n = min(nrow(data), n))
}

normalise_service_district <- function(x) {
  x |>
    as.character() |>
    str_replace("^[A-Z]{2,3}~", "") |>
    str_squish()
}

read_page <- function(url) {
  paste(readLines(url, warn = FALSE, encoding = "UTF-8"), collapse = "\n")
}

extract_links <- function(url) {
  html <- read_page(url)
  base_url <- str_replace(url, "^(https?://[^/]+).*$", "\\1")
  matches <- gregexpr("<a[^>]+href=[\"']([^\"']+)[\"'][^>]*>(.*?)</a>", html, perl = TRUE)
  pieces <- regmatches(html, matches)[[1]]
  if (length(pieces) == 0) return(tibble(label = character(), url = character()))
  tibble(raw = pieces) |>
    mutate(
      href = str_match(raw, "href=[\"']([^\"']+)[\"']")[, 2],
      label = raw |>
        str_replace_all("<[^>]+>", " ") |>
        str_squish(),
      url = case_when(
        str_detect(href, "^https?://") ~ href,
        str_detect(href, "^/") ~ paste0(base_url, href),
        TRUE ~ paste0(str_remove(url, "/[^/]*$"), "/", href)
      )
    ) |>
    select(label, url)
}

quarter_from_label <- function(label) {
  month <- str_match(label, "(March|June|September|December)")[, 2]
  year <- str_match(label, "(20[0-9]{2})")[, 2]
  tibble(quarter_label = paste(month, year)) |>
    left_join(analysis_quarters, by = "quarter_label")
}

download_one <- function(url, filename) {
  path <- file.path(raw_dir, filename)
  if (!file.exists(path) || (str_detect(filename, "\\.xlsx?$") && !is_xlsx_file(path))) {
    download.file(url, path, mode = "wb", quiet = TRUE)
  }
  path
}

download_quarterly <- function(page_url, label_regex, prefix) {
  links <- extract_links(page_url) |>
    filter(str_detect(label, regex(label_regex, ignore_case = TRUE))) |>
    mutate(row_id = row_number()) |>
    bind_cols(quarter_from_label(extract_links(page_url) |>
      filter(str_detect(label, regex(label_regex, ignore_case = TRUE))) |>
      pull(label))) |>
    select(-row_id) |>
    filter(!is.na(quarter_date), quarter_label %in% analysis_quarters$quarter_label) |>
    distinct(quarter_label, .keep_all = TRUE) |>
    arrange(quarter_date)

  missing <- setdiff(analysis_quarters$quarter_label, links$quarter_label)
  if (length(missing) > 0) {
    message("Missing visible links for ", prefix, ": ", paste(missing, collapse = ", "))
  }

  links |>
    mutate(
      path = map2_chr(url, quarter_label, \(u, q) {
        download_one(u, paste0(prefix, "_", str_replace_all(str_to_lower(q), " ", "_"), ".csv"))
      })
    )
}

is_xlsx_file <- function(path) {
  sig <- readBin(path, what = "raw", n = 4)
  identical(as.integer(sig[1:2]), as.integer(charToRaw("PK")))
}

read_public_table <- function(path) {
  out <- if (is_xlsx_file(path) || str_detect(path, regex("\\.xls$|\\.xlsx$", ignore_case = TRUE))) {
    read_excel(path) |> clean_df()
  } else {
    read_csv(path, show_col_types = FALSE) |> clean_df()
  }
  out |> mutate(across(everything(), as.character))
}

read_quarterly_csvs <- function(links) {
  links |>
    mutate(data = map(path, read_public_table)) |>
    select(quarter_label, quarter_date, quarter_index, data) |>
    unnest(data)
}

download_named <- function(page_url, label_regex, filename) {
  link <- extract_links(page_url) |>
    filter(str_detect(label, regex(label_regex, ignore_case = TRUE))) |>
    slice(1)
  if (nrow(link) == 0) return(NA_character_)
  download_one(link$url[[1]], filename)
}

extension_from_url <- function(url, default = ".csv") {
  ext <- str_match(str_to_lower(url), "\\.(csv|xlsx|xls|zip)(\\?|$)")[, 2]
  ifelse(is.na(ext), default, paste0(".", ext))
}

download_matching <- function(page_url, label_regex, prefix, limit = Inf) {
  links <- extract_links(page_url) |>
    filter(str_detect(label, regex(label_regex, ignore_case = TRUE))) |>
    distinct(label, url) |>
    slice_head(n = limit)

  if (nrow(links) == 0) {
    return(tibble(label = character(), url = character(), path = character()))
  }

  links |>
    mutate(
      path = pmap_chr(list(url, label, row_number()), \(u, lab, i) {
        safe_label <- lab |>
          str_to_lower() |>
          str_replace_all("[^a-z0-9]+", "_") |>
          str_replace_all("^_|_$", "") |>
          str_trunc(70, ellipsis = "")
        download_one(u, paste0(prefix, "_", str_pad(i, 2, pad = "0"), "_", safe_label, extension_from_url(u)))
      })
    )
}
Code 
budget_links <- download_quarterly(
  page_urls$participant,
  "^Participant numbers and plan budgets data",
  "participant_plan_budgets"
)

utilisation_links <- download_quarterly(
  page_urls$participant,
  "^Utilisation of Plan budgets data",
  "plan_utilisation"
)

provider_links <- download_quarterly(
  page_urls$provider,
  "^Active providers data",
  "active_providers"
)

budget_raw <- read_quarterly_csvs(budget_links)
utilisation_raw <- read_quarterly_csvs(utilisation_links)
provider_raw <- read_quarterly_csvs(provider_links)

lga_path <- download_named(page_urls$participant, "^Participants by LGA data", "participants_by_lga.csv")
service_lga_path <- download_named(page_urls$participant, "^Service District to LGA 2021 data", "service_district_to_lga_2021.csv")
market_path <- download_named(page_urls$provider, "^Market Concentration data", "market_concentration.csv")
payments_links <- download_quarterly(page_urls$payments, "^Payments data", "payments")
payments_raw <- read_quarterly_csvs(payments_links)

baseline_outcomes_path <- download_named(page_urls$participant, "^Baseline Outcomes data", "baseline_outcomes.csv")
longitudinal_outcomes_path <- download_named(page_urls$participant, "^Longitudinal Outcomes data", "longitudinal_outcomes.csv")
regional_outcome_links <- download_matching(page_urls$outcomes, "by Service District Accessible", "regional_outcomes_service_district")
Code 
standardise_budget <- function(data) {
  state <- first_col(data, c("^state", "state_territory"), required = FALSE)
  if (is.na(state)) state <- first_col(data, c("state_cd", "statecd"), required = FALSE)
  service <- first_col(data, c("service_district", "srvc_dstrct_nm", "srvcdstrctnm"), required = FALSE)
  age <- first_col(data, c("^age", "age_group", "age_band", "age_bnd", "agebnd"), required = FALSE)
  disability <- first_col(data, c("disability", "dsblty_grp_nm", "dsbltygrpnm"), required = FALSE)
  support <- first_col(data, c("support_class", "support_category", "supp_class", "suppclass", "support"), required = FALSE)
  participants <- first_col(data, c("active.*participant", "actv_prtcpnt", "actvprtcpnt", "number.*participant", "participant.*count", "participants"), required = FALSE)
  budget <- first_col(data, c("average.*support", "average.*budget", "average.*annual", "avg.*annual", "avg_anlsd_cmtd_supp_bdgt", "avganlsdcmtdsuppbdgt", "committed"), required = TRUE)

  data |>
    transmute(
      quarter_label, quarter_date, quarter_index,
      state = if (!is.na(state)) .data[[state]] else NA_character_,
      service_district = if (!is.na(service)) normalise_service_district(.data[[service]]) else NA_character_,
      age_group = if (!is.na(age)) .data[[age]] else NA_character_,
      disability_group = if (!is.na(disability)) .data[[disability]] else NA_character_,
      support_class = if (!is.na(support)) .data[[support]] else NA_character_,
      participant_count = if (!is.na(participants)) as_number(.data[[participants]]) else NA_real_,
      avg_support_budget = as_number(.data[[budget]])
    ) |>
    filter(!is.na(avg_support_budget), state != "ALL", service_district != "ALL", age_group != "ALL", disability_group != "ALL", support_class != "ALL")
}

standardise_national_participants <- function(data) {
  state <- first_col(data, c("^state", "state_territory"), required = FALSE)
  if (is.na(state)) state <- first_col(data, c("state_cd", "statecd"), required = FALSE)
  service <- first_col(data, c("service_district", "srvc_dstrct_nm", "srvcdstrctnm"), required = FALSE)
  age <- first_col(data, c("^age", "age_group", "age_band", "age_bnd", "agebnd"), required = FALSE)
  disability <- first_col(data, c("disability", "dsblty_grp_nm", "dsbltygrpnm"), required = FALSE)
  support <- first_col(data, c("support_class", "support_category", "supp_class", "suppclass", "support"), required = FALSE)
  participants <- first_col(data, c("active.*participant", "actv_prtcpnt", "actvprtcpnt", "number.*participant", "participant.*count", "participants"), required = FALSE)

  if (is.na(participants)) {
    return(tibble(
      quarter_label = character(), quarter_date = as.Date(character()),
      quarter_index = numeric(), ndis_participants = numeric(),
      participant_basis = character()
    ))
  }

  data |>
    transmute(
      quarter_label, quarter_date, quarter_index,
      state = if (!is.na(state)) .data[[state]] else NA_character_,
      service_district = if (!is.na(service)) normalise_service_district(.data[[service]]) else NA_character_,
      age_group = if (!is.na(age)) .data[[age]] else NA_character_,
      disability_group = if (!is.na(disability)) .data[[disability]] else NA_character_,
      support_class = if (!is.na(support)) .data[[support]] else NA_character_,
      participant_count = as_number(.data[[participants]])
    ) |>
    filter(is.finite(participant_count), participant_count > 0) |>
    mutate(
      detail_count = rowSums(across(
        c(state, service_district, age_group, disability_group, support_class),
        \(x) !is.na(x) & x != "ALL"
      ))
    ) |>
    group_by(quarter_label, quarter_date, quarter_index) |>
    filter(detail_count == min(detail_count, na.rm = TRUE)) |>
    summarise(
      ndis_participants = sum(participant_count, na.rm = TRUE),
      detail_count = min(detail_count, na.rm = TRUE),
      participant_basis = if_else(detail_count == 0, "National aggregate row", "Least-detailed rows summed"),
      .groups = "drop"
    )
}

standardise_utilisation <- function(data) {
  state <- first_col(data, c("^state", "state_territory"), required = FALSE)
  if (is.na(state)) state <- first_col(data, c("state_cd", "statecd"), required = FALSE)
  service <- first_col(data, c("service_district", "srvc_dstrct_nm", "srvcdstrctnm"), required = FALSE)
  age <- first_col(data, c("^age", "age_group", "age_band", "age_bnd", "agebnd"), required = FALSE)
  disability <- first_col(data, c("disability", "dsblty_grp_nm", "dsbltygrpnm"), required = FALSE)
  support <- first_col(data, c("support_class", "support_category", "supp_class", "suppclass", "support"), required = FALSE)
  sil_sda <- first_col(data, c("sil_or_sda", "silorsda", "sil.*sda"), required = FALSE)
  utilisation <- first_col(data, c("utilisation", "utilization", "utlstn"), required = TRUE)

  data |>
    transmute(
      quarter_label, quarter_date, quarter_index,
      state = if (!is.na(state)) .data[[state]] else NA_character_,
      service_district = if (!is.na(service)) normalise_service_district(.data[[service]]) else NA_character_,
      age_group = if (!is.na(age)) .data[[age]] else NA_character_,
      disability_group = if (!is.na(disability)) .data[[disability]] else NA_character_,
      support_class = if (!is.na(support)) .data[[support]] else NA_character_,
      sil_or_sda = if (!is.na(sil_sda)) .data[[sil_sda]] else NA_character_,
      utilisation_rate = as_rate(.data[[utilisation]])
    ) |>
    filter(!is.na(utilisation_rate), state != "ALL", service_district != "ALL", age_group != "ALL", disability_group != "ALL", support_class != "ALL")
}

standardise_provider <- function(data) {
  state <- first_col(data, c("^state", "state_territory"), required = FALSE)
  if (is.na(state)) state <- first_col(data, c("state_cd", "statecd"), required = FALSE)
  age <- first_col(data, c("^age", "age_group", "age_band", "age_bnd", "agebnd"), required = FALSE)
  disability <- first_col(data, c("disability", "dsblty_grp_nm", "dsbltygrpnm"), required = FALSE)
  support <- first_col(data, c("support_class", "support_category", "supp_class", "suppclass", "support"), required = FALSE)
  providers <- first_col(data, c("active.*provider", "prvdr_cnt", "prvdrcnt", "number.*provider", "provider.*count", "providers"), required = TRUE)

  data |>
    transmute(
      quarter_label, quarter_date, quarter_index,
      state = if (!is.na(state)) .data[[state]] else NA_character_,
      age_group = if (!is.na(age)) .data[[age]] else NA_character_,
      disability_group = if (!is.na(disability)) .data[[disability]] else NA_character_,
      support_class = if (!is.na(support)) .data[[support]] else NA_character_,
      active_providers = as_number(.data[[providers]])
    ) |>
    filter(state != "ALL", age_group != "ALL", disability_group != "ALL", support_class != "ALL") |>
    group_by(quarter_label, quarter_date, quarter_index, state, age_group, disability_group, support_class) |>
    summarise(active_providers = max(active_providers, na.rm = TRUE), .groups = "drop")
}

standardise_payments <- function(data) {
  support_class <- first_col(data, c("^supp_class", "^suppclass"), required = FALSE)
  support_category <- first_col(data, c("supp_cat_nm", "suppcatnm", "support_category"), required = FALSE)
  support_item_number <- first_col(data, c("supp_item_nmbr", "suppitemnmbr", "item_number"), required = FALSE)
  support_item_desc <- first_col(data, c("supp_item_desc", "suppitemdesc", "item_desc"), required = FALSE)
  state <- first_col(data, c("rsds_in_state_cd", "rsdsinstatecd", "^state"), required = FALSE)
  service <- first_col(data, c("rsds_in_srvc_dstrct_nm", "rsdsinsrvcdstrctnm", "service_district"), required = FALSE)
  disability <- first_col(data, c("ndis_dsblty_grp_nm", "ndisdsbltygrpnm", "disability"), required = FALSE)
  age <- first_col(data, c("ndia_age_bnd", "ndiaagebnd", "age"), required = FALSE)
  payments <- first_col(data, c("pmt_amt", "pmtamt", "payment"), required = TRUE)
  participants <- first_col(data, c("count_of_participants", "countofparticipants", "participant"), required = FALSE)

  data |>
    transmute(
      quarter_label, quarter_date, quarter_index,
      support_class = if (!is.na(support_class)) .data[[support_class]] else NA_character_,
      support_category = if (!is.na(support_category)) .data[[support_category]] else NA_character_,
      support_item_number = if (!is.na(support_item_number)) .data[[support_item_number]] else NA_character_,
      support_item_desc = if (!is.na(support_item_desc)) .data[[support_item_desc]] else NA_character_,
      state = if (!is.na(state)) .data[[state]] else NA_character_,
      service_district = if (!is.na(service)) normalise_service_district(.data[[service]]) else NA_character_,
      disability_group = if (!is.na(disability)) .data[[disability]] else NA_character_,
      age_group = if (!is.na(age)) .data[[age]] else NA_character_,
      payment_amount = as_number(.data[[payments]]),
      payment_participants = if (!is.na(participants)) as_number(.data[[participants]]) else NA_real_
    ) |>
    mutate(payment_per_participant = payment_amount / payment_participants)
}

standardise_market <- function(path) {
  if (is.na(path) || !file.exists(path)) {
    return(tibble(
      market_quarter_date = as.Date(character()),
      state = character(),
      service_district = character(),
      support_class = character(),
      payment_share_top10 = numeric(),
      payment_band = character()
    ))
  }

  data <- read_public_table(path)
  report <- first_col(data, c("rprt_dt", "rprtdt"), required = FALSE)
  state <- first_col(data, c("^state", "state_cd", "statecd"), required = FALSE)
  service <- first_col(data, c("service_district", "srvc_dstrct_nm", "srvcdstrctnm"), required = FALSE)
  support <- first_col(data, c("support_class", "supp_class", "suppclass"), required = FALSE)
  share <- first_col(data, c("pymnt_share_of_top10", "pymntshareoftop10", "share.*top10"), required = FALSE)
  band <- first_col(data, c("pymnt_bnd", "pymntbnd", "payment.*band"), required = FALSE)

  data |>
    transmute(
      market_report = if (!is.na(report)) .data[[report]] else NA_character_,
      state = if (!is.na(state)) .data[[state]] else NA_character_,
      service_district = if (!is.na(service)) normalise_service_district(.data[[service]]) else NA_character_,
      support_class = if (!is.na(support)) .data[[support]] else NA_character_,
      payment_share_top10 = if (!is.na(share)) as_rate(.data[[share]]) else NA_real_,
      payment_band = if (!is.na(band)) .data[[band]] else NA_character_
    ) |>
    mutate(
      market_quarter_date = dmy(str_replace_all(market_report, "([0-9]{2})([A-Z]{3})([0-9]{4})", "\\1-\\2-\\3"), quiet = TRUE)
    ) |>
    filter(state != "ALL", service_district != "ALL", support_class != "ALL")
}

read_optional_table <- function(path) {
  if (is.na(path) || !file.exists(path)) return(tibble())
  tryCatch(read_public_table(path), error = \(e) tibble())
}

standardise_outcomes <- function(path, source_label) {
  data <- read_optional_table(path)
  if (nrow(data) == 0) {
    return(tibble(
      source = character(), state = character(), respondent_group = character(),
      domain = character(), indicator = character(), outcome_value = numeric(), sample_n = numeric()
    ))
  }

  state <- first_col(data, c("^state", "state_territory", "state_cd", "statecd"), required = FALSE)
  respondent <- first_col(data, c("questionnaire", "respondent", "cohort", "participant", "family", "carer"), required = FALSE)
  domain <- first_col(data, c("domain", "area", "outcome_area"), required = FALSE)
  indicator <- first_col(data, c("indicator_description", "indicator.*description", "indicator", "measure", "question", "outcome"), required = FALSE)
  value <- first_col(data, c("percent", "percentage", "proportion", "value", "result", "rate"), required = FALSE)
  sample <- first_col(data, c("sample", "denominator", "respondents", "^n$"), required = FALSE)

  if (is.na(value)) {
    numeric_candidates <- names(data)[map_lgl(data, \(x) any(!is.na(as_number(x))))]
    value <- numeric_candidates[[min(length(numeric_candidates), 1)]]
  }

  data |>
    transmute(
      source = source_label,
      state = if (!is.na(state)) .data[[state]] else NA_character_,
      respondent_group = if (!is.na(respondent)) .data[[respondent]] else NA_character_,
      domain = if (!is.na(domain)) .data[[domain]] else NA_character_,
      indicator = if (!is.na(indicator)) .data[[indicator]] else NA_character_,
      outcome_value = if (!is.na(value)) as_rate(.data[[value]]) else NA_real_,
      sample_n = if (!is.na(sample)) as_number(.data[[sample]]) else NA_real_
    ) |>
    mutate(
      domain = case_when(
        !is.na(domain) ~ domain,
        str_detect(str_to_lower(indicator), "choice|control|decide|say|want") ~ "Choice and control",
        str_detect(str_to_lower(indicator), "health|wellbeing|well-being") ~ "Health and wellbeing",
        str_detect(str_to_lower(indicator), "work|job|employment|learning|school|education") ~ "Work and learning",
        str_detect(str_to_lower(indicator), "community|social|relationship|friend|family") ~ "Community and relationships",
        str_detect(str_to_lower(indicator), "daily|self-care|home|living|independent") ~ "Daily living",
        TRUE ~ "Other outcomes"
      )
    ) |>
    filter(!is.na(outcome_value), outcome_value >= 0, outcome_value <= 1)
}

read_excel_all_sheets <- function(path) {
  tryCatch({
    sheets <- excel_sheets(path)
    map_dfr(sheets, \(sheet) {
      read_excel(path, sheet = sheet) |>
        clean_df() |>
        mutate(across(everything(), as.character), sheet = sheet, .before = 1)
    })
  }, error = \(e) tibble())
}

standardise_regional_outcomes_file <- function(path, label) {
  data <- read_excel_all_sheets(path)
  if (nrow(data) == 0) {
    return(tibble(
      source_file = character(), state = character(), service_district = character(),
      respondent_group = character(), domain = character(), indicator = character(),
      outcome_value = numeric(), sample_n = numeric()
    ))
  }

  service <- first_col(data, c("service_district", "srvc_dstrct", "district"), required = FALSE)
  state <- first_col(data, c("^state", "state_territory", "state_cd"), required = FALSE)
  respondent <- first_col(data, c("questionnaire", "respondent", "participant", "family", "carer"), required = FALSE)
  domain <- first_col(data, c("domain", "area", "outcome_area"), required = FALSE)
  indicator <- first_col(data, c("indicator", "measure", "question", "outcome"), required = FALSE)
  sample <- first_col(data, c("sample", "denominator", "respondents", "^n$"), required = FALSE)

  value_cols <- names(data)[map_lgl(data, \(x) any(!is.na(as_number(x))))] |>
    setdiff(c(sample))

  if (is.na(service) || length(value_cols) == 0) {
    return(tibble(
      source_file = character(), state = character(), service_district = character(),
      respondent_group = character(), domain = character(), indicator = character(),
      outcome_value = numeric(), sample_n = numeric()
    ))
  }

  data |>
    transmute(
      source_file = label,
      state = if (!is.na(state)) .data[[state]] else str_match(label, "Dashboard ([A-Z]{2,3}) by")[, 2],
      service_district = normalise_service_district(.data[[service]]),
      respondent_group = if (!is.na(respondent)) .data[[respondent]] else sheet,
      domain = if (!is.na(domain)) .data[[domain]] else sheet,
      indicator = if (!is.na(indicator)) .data[[indicator]] else NA_character_,
      sample_n = if (!is.na(sample)) as_number(.data[[sample]]) else NA_real_,
      across(all_of(value_cols), as_rate)
    ) |>
    pivot_longer(all_of(value_cols), names_to = "measure", values_to = "outcome_value") |>
    mutate(indicator = coalesce(indicator, measure)) |>
    select(-measure) |>
    filter(!is.na(service_district), service_district != "ALL", !is.na(outcome_value), outcome_value >= 0, outcome_value <= 1)
}

budget <- standardise_budget(budget_raw)
national_participants <- standardise_national_participants(budget_raw)
utilisation <- standardise_utilisation(utilisation_raw)
providers <- standardise_provider(provider_raw)
payments <- standardise_payments(payments_raw)
market_concentration <- standardise_market(market_path)
baseline_outcomes <- standardise_outcomes(baseline_outcomes_path, "Baseline outcomes")
longitudinal_outcomes <- standardise_outcomes(longitudinal_outcomes_path, "Longitudinal outcomes")
national_outcomes <- bind_rows(baseline_outcomes, longitudinal_outcomes)
regional_outcomes <- pmap_dfr(
  list(regional_outcome_links$path, regional_outcome_links$label),
  standardise_regional_outcomes_file
)

latest_market <- market_concentration |>
  filter(market_quarter_date == max(market_quarter_date, na.rm = TRUE)) |>
  select(state, service_district, support_class, payment_share_top10, payment_band)

utilisation_for_model <- utilisation |>
  filter(is.na(sil_or_sda) | sil_or_sda == "ALL")

model_base <- budget |>
  left_join(
    utilisation_for_model,
    by = c("quarter_label", "quarter_date", "quarter_index", "state", "service_district", "age_group", "disability_group", "support_class")
  ) |>
  left_join(
    providers,
    by = c("quarter_label", "quarter_date", "quarter_index", "state", "age_group", "disability_group", "support_class")
  ) |>
  left_join(
    latest_market,
    by = c("state", "service_district", "support_class")
  )

payments_support_class <- payments |>
  filter(
    state == "ALL", service_district == "ALL", disability_group == "ALL", age_group == "ALL",
    support_class != "ALL", support_category == "ALL", support_item_number == "ALL"
  )

payments_category <- payments |>
  filter(
    state == "ALL", service_district == "ALL", disability_group == "ALL", age_group == "ALL",
    support_class != "ALL", support_category != "ALL", support_item_number == "ALL"
  )

payments_item <- payments |>
  filter(
    state == "ALL", service_district == "ALL", disability_group == "ALL", age_group == "ALL",
    support_class != "ALL", support_category != "ALL", support_item_number != "ALL"
  )

payments_grouped <- payments |>
  filter(
    state != "ALL", service_district != "ALL", disability_group != "ALL", age_group != "ALL",
    support_class != "ALL", support_category == "ALL", support_item_number == "ALL"
  ) |>
  group_by(quarter_label, quarter_date, quarter_index, state, service_district, disability_group, age_group, support_class) |>
  summarise(
    payment_amount = sum(payment_amount, na.rm = TRUE),
    payment_participants = sum(payment_participants, na.rm = TRUE),
    .groups = "drop"
  ) |>
  mutate(payment_per_participant = payment_amount / payment_participants)

model_base <- model_base |>
  left_join(
    payments_grouped,
    by = c("quarter_label", "quarter_date", "quarter_index", "state", "service_district", "age_group", "disability_group", "support_class")
  ) |>
  mutate(
    active_providers_per100 = active_providers / participant_count * 100,
    participants_per_provider = if_else(active_providers > 0, participant_count / active_providers, NA_real_)
  )

The current public pages may not expose every historic quarter forever. The workflow records what it found at render time:

Code 
tibble(
  dataset = c("Plan budgets", "Utilisation", "Active providers", "Payments"),
  quarters_found = c(
    paste(budget_links$quarter_label, collapse = ", "),
    paste(utilisation_links$quarter_label, collapse = ", "),
    paste(provider_links$quarter_label, collapse = ", "),
    paste(payments_links$quarter_label, collapse = ", ")
  )
) |>
  kable()
dataset quarters_found
Plan budgets June 2024, September 2024, December 2024, March 2025, June 2025, September 2025, December 2025
Utilisation June 2024, September 2024, December 2024, March 2025, June 2025, September 2025, December 2025
Active providers June 2024, September 2024, December 2024, March 2025, June 2025, September 2025, December 2025
Payments June 2024, September 2024, December 2024, March 2025, June 2025, September 2025, December 2025

ABS Regional Context

The NDIS plan budget data is organised around NDIA service districts. To bring in regional context, the workflow uses the NDIS service-district-to-LGA mapping and then adds ABS LGA indicators where available. For this first pass, the most useful lightweight context is socioeconomic position and age structure.

Code 
read_optional_csv <- function(path) {
  if (is.na(path) || !file.exists(path)) return(tibble())
  read_csv(path, show_col_types = FALSE) |> clean_df()
}

service_lga <- read_optional_csv(service_lga_path)
participants_lga <- read_optional_csv(lga_path)

service_context <- tibble(
  state = character(),
  service_district = character(),
  lga_code = character(),
  lga_name = character(),
  remoteness = character()
)

if (nrow(service_lga) > 0) {
  state_col <- first_col(service_lga, c("^state", "state_territory", "state_cd", "statecd", "rsds.*state"), required = FALSE)
  service_col <- first_col(service_lga, c("service_district", "srvc_dstrct", "srvcdstrct", "rsds.*srvc.*dstrct"), required = FALSE)
  lga_col <- first_col(service_lga, c("lga.*code", "lga_code", "lgacd"), required = FALSE)
  lga_name_col <- first_col(service_lga, c("lga.*name", "lganm", "local.*government.*area", "area.*name"), required = FALSE)
  remote_col <- first_col(service_lga, c("remoteness", "remote"), required = FALSE)

  if (!is.na(service_col) && !is.na(lga_col)) {
    service_context <- service_lga |>
      transmute(
        state = if (!is.na(state_col)) .data[[state_col]] else NA_character_,
        service_district = normalise_service_district(.data[[service_col]]),
        lga_code = str_pad(as.character(as_number(.data[[lga_col]])), width = 5, side = "left", pad = "0"),
        lga_name = if (!is.na(lga_name_col)) .data[[lga_name_col]] else NA_character_,
        remoteness = if (!is.na(remote_col)) .data[[remote_col]] else NA_character_
      ) |>
      filter(!is.na(service_district), !is.na(lga_code)) |>
      distinct()
  }
}

service_lga_participants <- tibble(
  state = character(),
  service_district = character(),
  participant_lgas = numeric(),
  participant_lga_names = character(),
  lga_participants = numeric(),
  latest_lga_report = as.Date(character()),
  remoteness = character(),
  lga_avg_support_budget = numeric()
)

if (nrow(participants_lga) > 0) {
  report_col <- first_col(participants_lga, c("report", "rprt"), required = FALSE)
  service_col <- first_col(participants_lga, c("service_district", "srvc_dstrct", "srvcdstrct", "rsds.*srvc.*dstrct"), required = FALSE)
  lga_name_col <- first_col(participants_lga, c("lga.*name", "lganm", "local.*government.*area", "area.*name"), required = FALSE)
  state_col <- first_col(participants_lga, c("^state", "state_territory", "state_cd", "statecd"), required = FALSE)
  remote_col <- first_col(participants_lga, c("remoteness", "remote"), required = FALSE)
  participants_col <- first_col(participants_lga, c("active.*participant", "actv_prtcpnt", "actvprtcpnt", "participant.*count", "participants", "prtcpnt"), required = FALSE)
  budget_col <- first_col(participants_lga, c("average.*support", "average.*budget", "avg.*annual", "avg_anlsd", "committed", "cmtd"), required = FALSE)

  if (!is.na(service_col)) {
    service_lga_participants <- participants_lga |>
      transmute(
        latest_lga_report = if (!is.na(report_col)) dmy(.data[[report_col]], quiet = TRUE) else as.Date(NA),
        state = if (!is.na(state_col)) .data[[state_col]] else NA_character_,
        service_district = normalise_service_district(.data[[service_col]]),
        lga_name = if (!is.na(lga_name_col)) .data[[lga_name_col]] else NA_character_,
        remoteness = if (!is.na(remote_col)) .data[[remote_col]] else NA_character_,
        lga_participants = if (!is.na(participants_col)) as_number(.data[[participants_col]]) else NA_real_,
        lga_avg_support_budget = if (!is.na(budget_col)) as_number(.data[[budget_col]]) else NA_real_
      ) |>
      filter(!is.na(service_district), service_district != "ALL", service_district != "Other") |>
      group_by(state, service_district) |>
      filter(is.na(latest_lga_report) | latest_lga_report == max(latest_lga_report, na.rm = TRUE)) |>
      summarise(
        participant_lgas = n_distinct(lga_name),
        participant_lga_names = collapse_top_values(lga_name),
        lga_participants = sum(lga_participants, na.rm = TRUE),
        latest_lga_report = max(latest_lga_report, na.rm = TRUE),
        remoteness = safe_mode_character(remoteness),
        lga_avg_support_budget = safe_weighted_mean(lga_avg_support_budget, lga_participants),
        .groups = "drop"
      ) |>
      mutate(
        latest_lga_report = if_else(is.infinite(latest_lga_report), as.Date(NA), latest_lga_report),
        lga_participants = if_else(is.infinite(lga_participants), NA_real_, lga_participants),
        lga_avg_support_budget = if_else(is.nan(lga_avg_support_budget), NA_real_, lga_avg_support_budget)
      )
  }
}

abs_note <- "ABS context was not joined in this render."

seifa_file <- tryCatch(
  download_named(page_urls$seifa, "Local Government Area.*Indexes.*SEIFA|LGA.*Indexes.*SEIFA", "abs_seifa_lga.xlsx"),
  error = \(e) NA_character_
)

population_file <- tryCatch(
  download_named(page_urls$population_age, "Median age, sex ratio and broad age groups, by LGA.*2021.*Final", "abs_lga_age_structure.xlsx"),
  error = \(e) NA_character_
)

abs_lga <- tibble(
  lga_code = character(),
  irsd_score = numeric(),
  irsd_decile = numeric(),
  ier_score = numeric(),
  ier_decile = numeric(),
  ieo_score = numeric(),
  ieo_decile = numeric(),
  median_age = numeric(),
  pct_0_14 = numeric(),
  pct_15_64 = numeric(),
  pct_65_plus = numeric(),
  lga_population = numeric()
)

lga_remoteness <- tibble(
  lga_code = character(),
  remoteness = character(),
  remoteness_rank = numeric(),
  remoteness_major_cities_share = numeric(),
  remoteness_inner_regional_share = numeric(),
  remoteness_outer_regional_share = numeric(),
  remoteness_remote_share = numeric(),
  remoteness_very_remote_share = numeric()
)

seifa_file <- tryCatch(
  download_one(page_urls$seifa_lga, "abs_seifa_lga.xlsx"),
  error = \(e) seifa_file
)

if (!is.na(seifa_file) && file.exists(seifa_file) && is_xlsx_file(seifa_file)) {
  seifa <- read_excel(seifa_file, sheet = "Table 1", skip = 6, col_names = FALSE) |>
    clean_df()

  abs_lga <- seifa |>
    transmute(
      lga_code = str_pad(as.character(as_number(.data[["1"]])), width = 5, side = "left", pad = "0"),
      irsd_score = as_number(.data[["3"]]),
      irsd_decile = as_number(.data[["4"]]),
      ier_score = as_number(.data[["5"]]),
      ier_decile = as_number(.data[["6"]]),
      ieo_score = as_number(.data[["7"]]),
      ieo_decile = as_number(.data[["8"]])
    ) |>
    filter(!is.na(lga_code), is.finite(irsd_score)) |>
    distinct(lga_code, .keep_all = TRUE)
  abs_note <- "ABS SEIFA LGA context joined."
}

population_file <- tryCatch(
  download_one(page_urls$population_age_lga, "abs_lga_age_structure.xlsx"),
  error = \(e) population_file
)

if (!is.na(population_file) && file.exists(population_file) && is_xlsx_file(population_file)) {
  pop_lga <- read_excel(population_file, sheet = "Table 1", skip = 7, col_names = FALSE) |>
    clean_df() |>
    transmute(
      lga_code = str_pad(as.character(as_number(.data[["3"]])), width = 5, side = "left", pad = "0"),
      lga_population = as_number(.data[["7"]]),
      median_age = as_number(.data[["11"]]),
      pct_0_14 = as_number(.data[["13"]]),
      pct_15_64 = as_number(.data[["14"]]),
      pct_65_plus = as_number(.data[["15"]])
    ) |>
    filter(!is.na(lga_code), is.finite(lga_population)) |>
    distinct(lga_code, .keep_all = TRUE)

  abs_lga <- full_join(abs_lga, pop_lga, by = "lga_code")
  abs_note <- paste(abs_note, "ABS age-structure context also joined.")
}

mb_file <- tryCatch(
  download_one(page_urls$asgs_mb_2021, "abs_mb_2021_allocation.xlsx"),
  error = \(e) NA_character_
)

lga_allocation_file <- tryCatch(
  download_one(page_urls$asgs_lga_2021, "abs_lga_2021_aust.xlsx"),
  error = \(e) NA_character_
)

ra_file <- tryCatch(
  download_one(page_urls$asgs_ra_2021, "abs_ra_2021_aust.xlsx"),
  error = \(e) NA_character_
)

if (
  !is.na(mb_file) && file.exists(mb_file) && is_xlsx_file(mb_file) &&
  !is.na(lga_allocation_file) && file.exists(lga_allocation_file) && is_xlsx_file(lga_allocation_file) &&
  !is.na(ra_file) && file.exists(ra_file) && is_xlsx_file(ra_file)
) {
  mb_sa1 <- read_excel(mb_file, sheet = 1) |>
    clean_df() |>
    transmute(
      mb_code = as.character(mb_code_2021),
      sa1_code = as.character(sa1_code_2021)
    )

  lga_mb <- read_excel(lga_allocation_file, sheet = 1) |>
    clean_df() |>
    transmute(
      mb_code = as.character(mb_code_2021),
      lga_code = str_pad(as.character(as_number(lga_code_2021)), width = 5, side = "left", pad = "0"),
      lga_area_sqkm = as_number(area_albers_sqkm)
    ) |>
    filter(!is.na(lga_code), is.finite(lga_area_sqkm), lga_area_sqkm > 0)

  sa1_ra <- read_excel(ra_file, sheet = 1) |>
    clean_df() |>
    transmute(
      sa1_code = as.character(sa1_code_2021),
      remoteness_code = as_number(ra_code_2021),
      remoteness = str_remove(as.character(ra_name_2021), " of Australia$")
    ) |>
    filter(!is.na(sa1_code), !is.na(remoteness))

  lga_ra_long <- lga_mb |>
    left_join(mb_sa1, by = "mb_code") |>
    left_join(sa1_ra, by = "sa1_code") |>
    filter(!is.na(remoteness)) |>
    group_by(lga_code, remoteness, remoteness_code) |>
    summarise(area_sqkm = sum(lga_area_sqkm, na.rm = TRUE), .groups = "drop") |>
    group_by(lga_code) |>
    mutate(area_share = area_sqkm / sum(area_sqkm, na.rm = TRUE)) |>
    ungroup()

  lga_remoteness <- lga_ra_long |>
    group_by(lga_code) |>
    summarise(
      remoteness = remoteness[which.max(area_share)],
      remoteness_rank = weighted.mean(remoteness_code, area_share, na.rm = TRUE),
      remoteness_major_cities_share = sum(area_share[remoteness == "Major Cities"], na.rm = TRUE),
      remoteness_inner_regional_share = sum(area_share[remoteness == "Inner Regional"], na.rm = TRUE),
      remoteness_outer_regional_share = sum(area_share[remoteness == "Outer Regional"], na.rm = TRUE),
      remoteness_remote_share = sum(area_share[remoteness == "Remote"], na.rm = TRUE),
      remoteness_very_remote_share = sum(area_share[remoteness == "Very Remote"], na.rm = TRUE),
      .groups = "drop"
    )

  abs_note <- paste(abs_note, "ABS remoteness context also joined.")
}

service_abs <- service_context |>
  left_join(lga_remoteness, by = "lga_code", suffix = c("", "_abs")) |>
  mutate(remoteness = coalesce(remoteness, remoteness_abs)) |>
  left_join(abs_lga, by = "lga_code") |>
  group_by(state, service_district) |>
  summarise(
    irsd_score = safe_weighted_mean(irsd_score, lga_population),
    irsd_decile = safe_weighted_mean(irsd_decile, lga_population),
    ier_score = safe_weighted_mean(ier_score, lga_population),
    ier_decile = safe_weighted_mean(ier_decile, lga_population),
    ieo_score = safe_weighted_mean(ieo_score, lga_population),
    ieo_decile = safe_weighted_mean(ieo_decile, lga_population),
    median_age = safe_weighted_mean(median_age, lga_population),
    pct_0_14 = safe_weighted_mean(pct_0_14, lga_population),
    pct_15_64 = safe_weighted_mean(pct_15_64, lga_population),
    pct_65_plus = safe_weighted_mean(pct_65_plus, lga_population),
    lga_population = sum(lga_population, na.rm = TRUE),
    remoteness_rank = safe_weighted_mean(remoteness_rank, lga_population),
    remoteness_major_cities_share = safe_weighted_mean(remoteness_major_cities_share, lga_population),
    remoteness_inner_regional_share = safe_weighted_mean(remoteness_inner_regional_share, lga_population),
    remoteness_outer_regional_share = safe_weighted_mean(remoteness_outer_regional_share, lga_population),
    remoteness_remote_share = safe_weighted_mean(remoteness_remote_share, lga_population),
    remoteness_very_remote_share = safe_weighted_mean(remoteness_very_remote_share, lga_population),
    mapped_lgas = n_distinct(lga_code),
    mapped_lga_names = collapse_top_values(lga_name),
    remoteness = safe_mode_character(remoteness),
    .groups = "drop"
  ) |>
  left_join(service_lga_participants, by = c("state", "service_district"), suffix = c("", "_participant_file")) |>
  mutate(
    max_remoteness_share = pmax(
      remoteness_major_cities_share,
      remoteness_inner_regional_share,
      remoteness_outer_regional_share,
      remoteness_remote_share,
      remoteness_very_remote_share,
      na.rm = TRUE
    ),
    dominant_remoteness = case_when(
      !is.finite(max_remoteness_share) ~ NA_character_,
      remoteness_major_cities_share == max_remoteness_share ~ "Major Cities",
      remoteness_inner_regional_share == max_remoteness_share ~ "Inner Regional",
      remoteness_outer_regional_share == max_remoteness_share ~ "Outer Regional",
      remoteness_remote_share == max_remoteness_share ~ "Remote",
      remoteness_very_remote_share == max_remoteness_share ~ "Very Remote",
      TRUE ~ NA_character_
    ),
    remoteness = coalesce(dominant_remoteness, remoteness, remoteness_participant_file),
    mapped_lgas = if_else(is.finite(participant_lgas), pmax(mapped_lgas, participant_lgas, na.rm = TRUE), mapped_lgas),
    mapped_lga_names = coalesce(mapped_lga_names, participant_lga_names),
    lga_population = if_else(lga_population == 0, NA_real_, lga_population),
    lga_participants = if_else(lga_participants == 0, NA_real_, lga_participants),
    across(c(irsd_score, irsd_decile, ier_score, ier_decile, ieo_score, ieo_decile, median_age, pct_0_14, pct_15_64, pct_65_plus, lga_avg_support_budget, remoteness_rank, remoteness_major_cities_share, remoteness_inner_regional_share, remoteness_outer_regional_share, remoteness_remote_share, remoteness_very_remote_share), \(x) ifelse(is.nan(x), NA_real_, x))
  )

model_data <- model_base |>
  left_join(service_abs, by = c("state", "service_district"))

model_service_keys <- model_data |>
  filter(!is.na(state), !is.na(service_district)) |>
  distinct(state, service_district)

service_context_covered <- service_context |>
  distinct(state, service_district) |>
  semi_join(model_service_keys, by = c("state", "service_district"))

service_abs_covered <- service_abs |>
  semi_join(model_service_keys, by = c("state", "service_district"))

join_diagnostics <- tibble(
  check = c(
    "Budget rows with utilisation",
    "Budget rows with active providers",
    "Budget rows with grouped payments",
    "Budget rows with market concentration",
    "Service districts with LGA mapping",
    "Service districts with ABS SEIFA/age context",
    "Service districts with remoteness context",
    "Service districts with LGA participant context"
  ),
  rows_or_groups = c(
    sum(!is.na(model_data$utilisation_rate)),
    sum(!is.na(model_data$active_providers)),
    sum(!is.na(model_data$payment_amount)),
    sum(!is.na(model_data$payment_share_top10)),
    nrow(service_context_covered),
    nrow(service_abs_covered |> filter(!is.na(irsd_score) | !is.na(median_age))),
    nrow(service_abs_covered |> filter(!is.na(remoteness))),
    nrow(service_abs_covered |> filter(!is.na(lga_participants)))
  ),
  denominator = c(
    nrow(model_data),
    nrow(model_data),
    nrow(model_data),
    nrow(model_data),
    nrow(model_service_keys),
    nrow(model_service_keys),
    nrow(model_service_keys),
    nrow(model_service_keys)
  )
) |>
  mutate(coverage = rows_or_groups / denominator)

outcome_spend_snapshot <- model_data |>
  filter(quarter_date == as.Date("2024-06-30")) |>
  group_by(state, service_district) |>
  summarise(
    participants = sum(participant_count, na.rm = TRUE),
    avg_budget = safe_weighted_mean(avg_support_budget, participant_count),
    utilisation = safe_weighted_mean(utilisation_rate, participant_count),
    payment_per_participant = safe_weighted_mean(payment_per_participant, participant_count),
    participants_per_provider = safe_weighted_mean(participants_per_provider, participant_count),
    payment_share_top10 = safe_mean(payment_share_top10),
    irsd_decile = safe_mean(irsd_decile),
    ier_decile = safe_mean(ier_decile),
    ieo_decile = safe_mean(ieo_decile),
    remoteness_rank = safe_mean(remoteness_rank),
    median_age = safe_mean(median_age),
    pct_0_14 = safe_mean(pct_0_14),
    pct_65_plus = safe_mean(pct_65_plus),
    lga_population = safe_mean(lga_population),
    .groups = "drop"
  ) |>
  mutate(across(c(avg_budget, utilisation, payment_per_participant, participants_per_provider, payment_share_top10, irsd_decile, ier_decile, ieo_decile, remoteness_rank, median_age, pct_0_14, pct_65_plus, lga_population), \(x) ifelse(is.nan(x), NA_real_, x)))

state_outcome_spend_snapshot <- model_data |>
  filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
  group_by(state) |>
  summarise(
    participants = sum(participant_count, na.rm = TRUE),
    avg_budget = safe_weighted_mean(avg_support_budget, participant_count),
    utilisation = safe_weighted_mean(utilisation_rate, participant_count),
    payment_per_participant = safe_weighted_mean(payment_per_participant, participant_count),
    participants_per_provider = safe_weighted_mean(participants_per_provider, participant_count),
    payment_share_top10 = safe_mean(payment_share_top10),
    .groups = "drop"
  ) |>
  filter(!is.na(state), state != "ALL")

regional_outcomes_spend <- regional_outcomes |>
  left_join(outcome_spend_snapshot, by = c("state", "service_district")) |>
  filter(!is.na(payment_per_participant), !is.na(outcome_value))

state_outcomes_spend <- national_outcomes |>
  filter(!is.na(state), state != "ALL") |>
  left_join(state_outcome_spend_snapshot, by = "state") |>
  filter(!is.na(payment_per_participant), !is.na(outcome_value))

state_outcome_domain <- state_outcomes_spend |>
  group_by(source, state, domain) |>
  summarise(
    indicators = n_distinct(indicator),
    outcome_value = safe_mean(outcome_value),
    participants = max(participants, na.rm = TRUE),
    avg_budget = safe_mean(avg_budget),
    utilisation = safe_mean(utilisation),
    payment_per_participant = safe_mean(payment_per_participant),
    participants_per_provider = safe_mean(participants_per_provider),
    payment_share_top10 = safe_mean(payment_share_top10),
    .groups = "drop"
  ) |>
  mutate(
    participants = if_else(is.infinite(participants), NA_real_, participants),
    outcome_per_1000_dollars = outcome_value / (payment_per_participant / 1000)
  )

outcome_spend_join <- tibble(
  check = c(
    "Service districts with dashboard outcomes",
    "Dashboard outcome rows joined to June 2024 district spend",
    "State outcome rows joined to latest state spend"
  ),
  rows_or_groups = c(
    n_distinct(regional_outcomes$service_district),
    nrow(regional_outcomes_spend),
    nrow(state_outcomes_spend)
  ),
  denominator = c(
    n_distinct(outcome_spend_snapshot$service_district),
    nrow(regional_outcomes),
    nrow(national_outcomes |> filter(!is.na(state), state != "ALL"))
  ),
  coverage = if_else(denominator > 0, rows_or_groups / denominator, NA_real_)
)

abs_note
[1] "ABS SEIFA LGA context joined. ABS age-structure context also joined. ABS remoteness context also joined."

Source Rows and Join Diagnostics

Code 
tibble(
  table = c("Budget rows", "Utilisation rows", "Provider rows", "Payments rows", "Market concentration rows", "Model rows"),
  rows = c(nrow(budget), nrow(utilisation), nrow(providers), nrow(payments), nrow(market_concentration), nrow(model_data))
) |>
  kable()
table rows
Budget rows 105041
Utilisation rows 223624
Provider rows 18349
Payments rows 816299
Market concentration rows 7228
Model rows 105041
Code 
join_diagnostics |>
  mutate(coverage = percent(coverage, accuracy = 0.1)) |>
  kable()
check rows_or_groups denominator coverage
Budget rows with utilisation 62709 105041 59.7%
Budget rows with active providers 61506 105041 58.6%
Budget rows with grouped payments 63823 105041 60.8%
Budget rows with market concentration 1112 105041 1.1%
Service districts with LGA mapping 80 84 95.2%
Service districts with ABS SEIFA/age context 80 84 95.2%
Service districts with remoteness context 80 84 95.2%
Service districts with LGA participant context 80 84 95.2%

Overall Spend Growth

Across the visible payments extracts, observed NDIS payments increased by 18.7%, equivalent to about 2.9% per quarter. This is a short window, so I treat it as the recent pace in the public files rather than a long-run trend.

Code 
if (nrow(overall_payment_trend) > 1) {
  ggplot(overall_payment_trend, aes(quarter_date, total_payments)) +
    geom_line(linewidth = 0.9, colour = "#0b6b57") +
    geom_point(size = 2, colour = "#0b6b57") +
    scale_y_continuous(labels = dollar) +
    labs(x = NULL, y = "Total quarterly payments")
} else {
  tibble(note = "Not enough payment quarters to draw the total spend chart.") |> kable()
}

Total NDIS payments in the visible quarterly extracts.

How Fast Is Participation Growing?

Before modelling budgets or utilisation, the first question is scale: is NDIS participation growing faster than the Australian population? The table below compares national NDIS participant counts from the public participant budget extract with ABS estimated resident population. The ABS quarterly population release currently overlaps the NDIS analysis window through September 2025, so the December 2025 NDIS quarter is not used in this comparison.

Code 
abs_population <- tibble::tribble(
  ~quarter_label, ~quarter_date, ~australian_population,
  "March 2024", "2024-03-31", 27113517,
  "June 2024", "2024-06-30", 27194286,
  "September 2024", "2024-09-30", 27301149,
  "December 2024", "2024-12-31", 27388133,
  "March 2025", "2025-03-31", 27531443,
  "June 2025", "2025-06-30", 27613654,
  "September 2025", "2025-09-30", 27724744
) |>
  mutate(
    quarter_date = as.Date(quarter_date),
    population_source = "ABS National, state and territory population, September 2025 release"
  )

population_participation <- national_participants |>
  inner_join(abs_population, by = c("quarter_label", "quarter_date")) |>
  arrange(quarter_date) |>
  mutate(
    ndis_share_population = ndis_participants / australian_population,
    ndis_participant_growth = ndis_participants / first(ndis_participants) - 1,
    population_growth = australian_population / first(australian_population) - 1,
    share_change_percentage_points = (ndis_share_population - first(ndis_share_population)) * 100
  )

participation_summary <- if (nrow(population_participation) > 0) {
  population_participation |>
    summarise(
      start_quarter = first(quarter_label),
      end_quarter = last(quarter_label),
      population_growth = last(population_growth),
      ndis_participant_growth = last(ndis_participant_growth),
      start_share = first(ndis_share_population),
      end_share = last(ndis_share_population)
    )
} else {
  tibble(
    start_quarter = NA_character_,
    end_quarter = NA_character_,
    population_growth = NA_real_,
    ndis_participant_growth = NA_real_,
    start_share = NA_real_,
    end_share = NA_real_
  )
}

if (nrow(population_participation) > 0) {
  population_participation |>
    transmute(
      quarter = quarter_label,
      australian_population = comma(australian_population),
      ndis_participants = comma(ndis_participants),
      population_growth = percent(population_growth, accuracy = 0.1),
      ndis_participant_growth = percent(ndis_participant_growth, accuracy = 0.1),
      ndis_share_population = percent(ndis_share_population, accuracy = 0.01),
      share_change_percentage_points = number(share_change_percentage_points, accuracy = 0.01)
    ) |>
    kable()
} else {
  tibble(note = "National NDIS participant counts could not be aligned to the ABS population denominator in this render.") |>
    kable()
}
quarter australian_population ndis_participants population_growth ndis_participant_growth ndis_share_population share_change_percentage_points
June 2024 27,194,286 661,267 0.0% 0.0% 2.43% 0.00
September 2024 27,301,149 680,123 0.4% 2.9% 2.49% 0.06
December 2024 27,388,133 692,822 0.7% 4.8% 2.53% 0.10
March 2025 27,531,443 716,998 1.2% 8.4% 2.60% 0.17
June 2025 27,613,654 739,414 1.5% 11.8% 2.68% 0.25
September 2025 27,724,744 751,446 2.0% 13.6% 2.71% 0.28

Over the overlapping ABS and NDIS window, the Australian population grew 2.0% while NDIS participant counts grew 13.6%. That lifted observed NDIS participation from 2.43% of the population in June 2024 to 2.71% by September 2025.

Code 
if (nrow(population_participation) > 1) {
  population_participation |>
    select(quarter_date, population_growth, ndis_participant_growth) |>
    pivot_longer(
      c(population_growth, ndis_participant_growth),
      names_to = "series",
      values_to = "growth"
    ) |>
    mutate(
      series = recode(
        series,
        population_growth = "Australian population",
        ndis_participant_growth = "NDIS participants"
      )
    ) |>
    ggplot(aes(quarter_date, growth, colour = series)) +
    geom_hline(yintercept = 0, colour = "grey80") +
    geom_line(linewidth = 0.9) +
    geom_point(size = 2) +
    scale_y_continuous(labels = percent) +
    labs(x = NULL, y = "Growth since March 2024", colour = NULL)
} else {
  tibble(note = "Not enough overlapping population and NDIS quarters to draw the indexed growth chart.") |>
    kable()
}

NDIS participant growth compared with Australian population growth, indexed to March 2024.
Code 
if (nrow(population_participation) > 1) {
  population_participation |>
    ggplot(aes(quarter_date, ndis_share_population)) +
    geom_line(linewidth = 0.9, colour = "#1b6ca8") +
    geom_point(size = 2, colour = "#1b6ca8") +
    scale_y_continuous(labels = percent) +
    labs(x = NULL, y = "NDIS participants as share of population")
} else {
  tibble(note = "Not enough overlapping population and NDIS quarters to draw the participation-share chart.") |>
    kable()
}

Share of the Australian population participating in the NDIS.
Code 
budget_trend_data <- budget |>
  group_by(quarter_date) |>
  summarise(
    participants = sum(participant_count, na.rm = TRUE),
    avg_budget = safe_weighted_mean(avg_support_budget, participant_count),
    .groups = "drop"
  ) |>
  filter(!is.na(avg_budget))

if (has_enough_data(budget_trend_data, c("avg_budget"))) {
  ggplot(budget_trend_data, aes(quarter_date, avg_budget)) +
    geom_line(linewidth = 0.8, colour = "#1b6ca8") +
    geom_point(size = 2, colour = "#1b6ca8") +
    scale_y_continuous(labels = dollar) +
    labs(x = NULL, y = "Average annualised budget", title = "Average support budgets over time")
} else {
  tibble(note = "Not enough non-missing average budget data to draw the trend chart.") |> kable()
}

Average support budget and participant counts across the visible quarterly extracts.
Code 
utilisation_trend_data <- utilisation |>
  group_by(quarter_date) |>
  summarise(avg_utilisation = safe_mean(utilisation_rate), .groups = "drop") |>
  arrange(quarter_date)

utilisation_summary <- utilisation_trend_data |>
  summarise(
    start_utilisation = first(avg_utilisation),
    end_utilisation = last(avg_utilisation),
    min_utilisation = min(avg_utilisation, na.rm = TRUE),
    max_utilisation = max(avg_utilisation, na.rm = TRUE)
  )

utilisation_trend_data |>
  ggplot(aes(quarter_date, avg_utilisation)) +
  geom_line(linewidth = 0.8, colour = "#7a4b9d") +
  geom_point(size = 2, colour = "#7a4b9d") +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Average utilisation", title = "Plan utilisation over time")

Plan utilisation across the visible quarterly extracts.

Payments are approximately participants multiplied by approved budgets multiplied by utilisation. In this short series, utilisation moves from 66.1% to 67.1%, with a range of 66.0% to 68.7%. That points the descriptive story toward participant growth and average budget growth rather than a large utilisation-rate shift.

Code 
budget_pressure <- budget |>
  group_by(quarter_date) |>
  summarise(
    participants = sum(participant_count, na.rm = TRUE),
    avg_budget = safe_weighted_mean(avg_support_budget, participant_count),
    implied_budget_pool = sum(participant_count * avg_support_budget, na.rm = TRUE),
    .groups = "drop"
  ) |>
  arrange(quarter_date) |>
  mutate(
    participant_growth = participants / first(participants) - 1,
    avg_budget_growth = avg_budget / first(avg_budget) - 1,
    implied_pool_growth = implied_budget_pool / first(implied_budget_pool) - 1
  )

budget_pressure |>
  select(quarter_date, participant_growth, avg_budget_growth, implied_pool_growth) |>
  pivot_longer(-quarter_date, names_to = "measure", values_to = "growth") |>
  mutate(
    measure = recode(
      measure,
      participant_growth = "Participants",
      avg_budget_growth = "Average support budget",
      implied_pool_growth = "Participant x budget pool"
    )
  ) |>
  ggplot(aes(quarter_date, growth, colour = measure)) +
  geom_hline(yintercept = 0, colour = "grey80") +
  geom_line(linewidth = 0.8) +
  geom_point(size = 2) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Growth since first available quarter", colour = NULL)

Participant count and average support budget growth, indexed to the first available quarter.
Code 
budget |>
  filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
  group_by(support_class) |>
  summarise(
    rows = n(),
    usable_participants = sum(participant_count, na.rm = TRUE),
    median_avg_budget = safe_median(avg_support_budget),
    mean_avg_budget = safe_mean(avg_support_budget),
    .groups = "drop"
  ) |>
  left_join(
    payments_support_class |>
      filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
      group_by(support_class) |>
      summarise(total_payments = sum(payment_amount, na.rm = TRUE), .groups = "drop"),
    by = "support_class"
  ) |>
  arrange(desc(total_payments)) |>
  mutate(
    usable_participants = comma(usable_participants),
    median_avg_budget = dollar(median_avg_budget, accuracy = 1),
    mean_avg_budget = dollar(mean_avg_budget, accuracy = 1),
    total_payments = dollar(total_payments, accuracy = 1)
  ) |>
  kable()
support_class rows usable_participants median_avg_budget mean_avg_budget total_payments
Core 5857 524,476 $93,000 $103,684 $38,509,796,000
Capital 3446 106,679 $19,500 $20,195 $1,419,899,000
CapacityBuilding 5949 633,591 $20,000 $20,354 NA

Suppressed participant counts such as <11 are treated as missing. Weighted summaries use non-suppressed participant counts where they are available and otherwise fall back to unweighted summaries.

Disability Mix and Payment Growth

The payments files are the most useful addition for spend analysis because they contain actual payment amounts by disability group as well as support class, support category, item, region and age band. The disability-group rows below use national aggregate rows where available, so nested ALL rows are not double-counted.

The support-level tables are still useful audit checks, but the main story below focuses on disability groups.

Code 
payments_support_class |>
  group_by(quarter_date, support_class) |>
  summarise(payment_amount = sum(payment_amount, na.rm = TRUE), .groups = "drop") |>
  group_by(support_class) |>
  arrange(quarter_date, .by_group = TRUE) |>
  mutate(index = payment_amount / first(payment_amount) - 1) |>
  ungroup() |>
  ggplot(aes(quarter_date, index, colour = support_class)) +
  geom_hline(yintercept = 0, colour = "grey80") +
  geom_line(linewidth = 0.8) +
  geom_point(size = 2) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Growth since first available quarter", colour = "Support class")

Payment growth by support class, indexed to the first available quarter.
Code 
first_payment_qtr <- min(payments_support_class$quarter_date, na.rm = TRUE)
last_payment_qtr <- max(payments_support_class$quarter_date, na.rm = TRUE)

payment_change <- function(data, group_vars) {
  if (nrow(data) == 0) {
    return(tibble(
      !!!set_names(rep(list(character()), length(group_vars)), group_vars),
      payment_amount_first = numeric(),
      payment_amount_last = numeric(),
      participants_first = numeric(),
      participants_last = numeric(),
      payment_change = numeric(),
      participant_change = numeric(),
      payment_per_participant_first = numeric(),
      payment_per_participant_last = numeric(),
      payment_per_participant_change = numeric()
    ))
  }

  wide <- data |>
    filter(quarter_date %in% c(first_payment_qtr, last_payment_qtr)) |>
    group_by(across(all_of(group_vars)), quarter_date) |>
    summarise(
      payment_amount = sum(payment_amount, na.rm = TRUE),
      participants = sum(payment_participants, na.rm = TRUE),
      .groups = "drop"
    ) |>
    mutate(period = if_else(quarter_date == first_payment_qtr, "first", "last")) |>
    select(-quarter_date) |>
    pivot_wider(
      names_from = period,
      values_from = c(payment_amount, participants),
      values_fill = 0
    )

  needed_cols <- c("payment_amount_first", "payment_amount_last", "participants_first", "participants_last")
  for (col in needed_cols) {
    if (!col %in% names(wide)) wide[[col]] <- 0
  }

  wide |>
    mutate(
      payment_amount_first = coalesce(payment_amount_first, 0),
      payment_amount_last = coalesce(payment_amount_last, 0),
      participants_first = coalesce(participants_first, 0),
      participants_last = coalesce(participants_last, 0),
      payment_change = payment_amount_last - payment_amount_first,
      participant_change = participants_last - participants_first,
      payment_per_participant_first = payment_amount_first / participants_first,
      payment_per_participant_last = payment_amount_last / participants_last,
      payment_per_participant_change = payment_per_participant_last - payment_per_participant_first
    ) |>
    arrange(desc(abs(payment_change)))
}

support_class_change <- payment_change(payments_support_class, "support_class")

support_class_change |>
  mutate(
    across(c(payment_amount_first, payment_amount_last, payment_change), \(x) dollar(x, accuracy = 1)),
    across(c(participants_first, participants_last, participant_change), comma),
    across(c(payment_per_participant_first, payment_per_participant_last, payment_per_participant_change), \(x) dollar(x, accuracy = 1))
  ) |>
  kable()
support_class payment_amount_first payment_amount_last participants_first participants_last payment_change participant_change payment_per_participant_first payment_per_participant_last payment_per_participant_change
Core $32,425,627,000 $38,509,796,000 512,210 552,928 $6,084,169,000 40,718 $63,305 $69,647 $6,342
Capacity Building $7,802,961,000 $9,028,315,000 637,317 737,282 $1,225,354,000 99,965 $12,243 $12,245 $2
Capital $1,018,264,000 $1,419,899,000 80,160 86,369 $401,635,000 6,209 $12,703 $16,440 $3,737
Missing $7,371,000 $3,716,000 151,095 290,973 -$3,655,000 139,878 $49 $13 -$36
Code 
category_change <- payment_change(payments_category, c("support_class", "support_category"))

category_change |>
  slice_max(abs(payment_change), n = 15) |>
  mutate(
    payment_change = dollar(payment_change, accuracy = 1),
    payment_amount_last = dollar(payment_amount_last, accuracy = 1),
    participant_change = comma(participant_change),
    payment_per_participant_change = dollar(payment_per_participant_change, accuracy = 1)
  ) |>
  select(support_class, support_category, payment_amount_last, payment_change, participant_change, payment_per_participant_change) |>
  kable()
support_class support_category payment_amount_last payment_change participant_change payment_per_participant_change
Core Daily Activities $25,446,659,000 $4,005,204,000 29,401 $5,735
Core Social Community and Civic Participation $11,563,196,000 $2,090,588,000 23,533 $3,977
Capacity Building CB Daily Activity $5,790,549,000 $614,241,000 79,004 -$117
Capacity Building CB Relationships $954,703,000 $323,821,000 18,386 $1,801
Capital Assistive Technology $789,391,000 $203,350,000 6,627 $1,945
Capital Home Modifications $630,508,000 $198,285,000 1,070 $7,770
Capacity Building Support Coordination $1,198,990,000 $126,501,000 38,708 -$44
Capacity Building CB Choice and Control $645,490,000 $86,859,000 97,518 -$74
Core Consumables $620,377,000 -$64,787,000 -29,903 $15
Core Transport $879,563,000 $53,163,000 4,449 $146
Capacity Building CB Employment $177,050,000 $45,336,000 -4,672 $3,146
Capacity Building CB Social Community and Civic Participation $233,283,000 $35,722,000 -2,067 $915
Capacity Building CB Health and Wellbeing $26,683,000 -$7,512,000 -5,506 $161
Missing Missing $3,716,000 -$3,655,000 139,878 -$36
Capacity Building CB Home Living $1,216,000 $212,000 -43 $431
Code 
item_change <- payment_change(payments_item, c("support_class", "support_category", "support_item_number", "support_item_desc"))

if (nrow(item_change) > 0) {
  item_change |>
    slice_max(abs(payment_change), n = 15) |>
    mutate(
      support_item_desc = str_trunc(support_item_desc, 70),
      payment_amount_last = dollar(payment_amount_last, accuracy = 1),
      payment_change = dollar(payment_change, accuracy = 1),
      participant_change = comma(participant_change),
      payment_per_participant_change = dollar(payment_per_participant_change, accuracy = 1)
    ) |>
    select(support_class, support_category, support_item_number, support_item_desc, payment_amount_last, payment_change, participant_change, payment_per_participant_change) |>
    kable()
} else {
  tibble(note = "The current public payments extract did not expose national item-level rows under the no-double-counting filters used here.") |>
    kable()
}
note
The current public payments extract did not expose national item-level rows under the no-double-counting filters used here.
Code 
payments_disability_candidates <- payments |>
  filter(
    state == "ALL", service_district == "ALL", disability_group != "ALL", age_group == "ALL",
    support_category == "ALL", support_item_number == "ALL"
  )

if (any(payments_disability_candidates$support_class == "ALL", na.rm = TRUE)) {
  payments_disability <- payments_disability_candidates |>
    filter(support_class == "ALL") |>
    group_by(quarter_label, quarter_date, quarter_index, disability_group) |>
    summarise(
      payment_amount = sum(payment_amount, na.rm = TRUE),
      payment_participants = max(payment_participants, na.rm = TRUE),
      .groups = "drop"
    )
  disability_payment_basis <- "National disability-group aggregate rows"
} else {
  payments_disability <- payments_disability_candidates |>
    filter(support_class != "ALL") |>
    group_by(quarter_label, quarter_date, quarter_index, disability_group) |>
    summarise(
      payment_amount = sum(payment_amount, na.rm = TRUE),
      payment_participants = sum(payment_participants, na.rm = TRUE),
      .groups = "drop"
    )
  disability_payment_basis <- "Support-class rows summed within disability group"
}

payments_disability <- payments_disability |>
  mutate(
    payment_participants = if_else(is.infinite(payment_participants), NA_real_, payment_participants),
    payment_per_participant = payment_amount / payment_participants
  )

tibble(
  basis = disability_payment_basis,
  disability_groups = n_distinct(payments_disability$disability_group),
  quarters = paste(sort(unique(payments_disability$quarter_label)), collapse = ", ")
) |>
  kable()
basis disability_groups quarters
Support-class rows summed within disability group 18 December 2024, December 2025, June 2024, June 2025, March 2025, September 2024, September 2025

The participant-count view shows scale: groups that are large enough to move total spend even if average payments are moderate.

Code 
largest_by_count |>
  transmute(
    disability_group,
    payment_participants = comma(payment_participants),
    total_payments = dollar(payment_amount, accuracy = 1),
    payment_per_participant = dollar(payment_per_participant, accuracy = 1)
  ) |>
  kable()
disability_group payment_participants total_payments payment_per_participant
Autism 630,030 $10,565,618,000 $16,770
Intellectual Disability 213,590 $10,992,967,000 $51,468
Developmental delay 172,567 $891,824,000 $5,168
Psychosocial disability 144,402 $6,054,874,000 $41,931
Other Neurological 75,576 $4,158,314,000 $55,022
Hearing Impairment 58,663 $283,182,000 $4,827
Other Physical 57,391 $1,534,717,000 $26,741
ABI 51,816 $3,441,433,000 $66,416
Cerebral Palsy 49,537 $3,175,466,000 $64,103
Global developmental delay 38,648 $315,257,000 $8,157

The total-payments view combines scale and intensity: these are the groups contributing the largest payment dollars in the latest quarter.

Code 
largest_by_spend |>
  transmute(
    disability_group,
    total_payments = dollar(payment_amount, accuracy = 1),
    payment_participants = comma(payment_participants),
    payment_per_participant = dollar(payment_per_participant, accuracy = 1)
  ) |>
  kable()
disability_group total_payments payment_participants payment_per_participant
Intellectual Disability $10,992,967,000 213,590 $51,468
Autism $10,565,618,000 630,030 $16,770
Psychosocial disability $6,054,874,000 144,402 $41,931
Other Neurological $4,158,314,000 75,576 $55,022
ABI $3,441,433,000 51,816 $66,416
Cerebral Palsy $3,175,466,000 49,537 $64,103
Stroke $1,746,058,000 30,248 $57,725
Down Syndrome $1,612,746,000 26,879 $60,000
Other Physical $1,534,717,000 57,391 $26,741
Other $1,255,897,000 36,546 $34,365

The average-payment view is filtered to groups above the first quartile of participant counts, so it does not get dominated by very small groups with noisy averages.

Code 
largest_by_average_spend |>
  transmute(
    disability_group,
    payment_per_participant = dollar(payment_per_participant, accuracy = 1),
    payment_participants = comma(payment_participants),
    total_payments = dollar(payment_amount, accuracy = 1)
  ) |>
  kable()
disability_group payment_per_participant payment_participants total_payments
ABI $66,416 51,816 $3,441,433,000
Cerebral Palsy $64,103 49,537 $3,175,466,000
Stroke $57,725 30,248 $1,746,058,000
Other Neurological $55,022 75,576 $4,158,314,000
Intellectual Disability $51,468 213,590 $10,992,967,000
Psychosocial disability $41,931 144,402 $6,054,874,000
Multiple Sclerosis $39,677 31,629 $1,254,956,000
Other $34,365 36,546 $1,255,897,000
Other Physical $26,741 57,391 $1,534,717,000
Autism $16,770 630,030 $10,565,618,000
Code 
disability_change <- payment_change(payments_disability, "disability_group")

total_disability_payment_change <- sum(disability_change$payment_change, na.rm = TRUE)

disability_driver <- disability_change |>
  mutate(
    participant_contribution = (participants_last - participants_first) *
      ((payment_per_participant_first + payment_per_participant_last) / 2),
    intensity_contribution = (payment_per_participant_last - payment_per_participant_first) *
      ((participants_first + participants_last) / 2),
    change_share = payment_change / total_disability_payment_change,
    stroke_flag = str_detect(str_to_lower(disability_group), "\\bstroke\\b")
  ) |>
  arrange(desc(abs(payment_change)))

if (nrow(disability_driver) > 0) {
  disability_driver |>
    slice_max(abs(payment_change), n = 15) |>
    transmute(
      disability_group,
      payment_amount_last = dollar(payment_amount_last, accuracy = 1),
      payment_change = dollar(payment_change, accuracy = 1),
      change_share = percent(change_share, accuracy = 0.1),
      participant_change = comma(participant_change),
      participant_contribution = dollar(participant_contribution, accuracy = 1),
      intensity_contribution = dollar(intensity_contribution, accuracy = 1),
      payment_per_participant_change = dollar(payment_per_participant_change, accuracy = 1)
    ) |>
    kable()
} else {
  tibble(note = "No disability-group payment rows were available under the no-double-counting filters used here.") |>
    kable()
}
disability_group payment_amount_last payment_change change_share participant_change participant_contribution intensity_contribution payment_per_participant_change
Autism $10,565,618,000 $2,293,245,000 29.7% 173,093 $3,018,221,951 -$724,976,951 -$1,334
Intellectual Disability $10,992,967,000 $1,362,626,000 17.7% 19,472 $984,098,993 $378,527,007 $1,857
Other Neurological $4,158,314,000 $844,032,000 10.9% 14,054 $765,191,910 $78,840,090 $1,150
Psychosocial disability $6,054,874,000 $812,576,000 10.5% 10,473 $424,539,006 $388,036,994 $2,788
ABI $3,441,433,000 $536,420,000 6.9% 6,727 $440,096,603 $96,323,397 $1,988
Stroke $1,746,058,000 $409,681,000 5.3% 5,248 $291,735,853 $117,945,147 $4,270
Cerebral Palsy $3,175,466,000 $406,537,000 5.3% 3,541 $220,077,145 $186,459,855 $3,904
Other $1,255,897,000 $397,298,000 5.1% 11,275 $385,269,522 $12,028,478 $389
Multiple Sclerosis $1,254,956,000 $177,850,000 2.3% 4,389 $173,845,483 $4,004,517 $136
Down Syndrome $1,612,746,000 $167,282,000 2.2% 1,166 $67,753,650 $99,528,350 $3,785
Spinal Cord Injury $1,132,552,000 $150,387,000 1.9% 2,184 $128,799,772 $21,587,228 $1,205
Other Physical $1,534,717,000 $140,440,000 1.8% 4,015 $106,122,913 $34,317,087 $620
Developmental delay $891,824,000 -$83,173,000 -1.1% 15,852 $90,272,802 -$173,445,802 -$1,053
Visual Impairment $504,113,000 $62,762,000 0.8% 2,015 $41,165,690 $21,596,310 $934
Hearing Impairment $283,182,000 $24,455,000 0.3% 5,694 $27,649,398 -$3,194,398 -$57
Code 
top_driver_groups <- disability_driver |>
  filter(is.finite(payment_change)) |>
  slice_max(abs(payment_change), n = 7) |>
  pull(disability_group)

stroke_groups <- disability_driver |>
  filter(stroke_flag) |>
  pull(disability_group)

driver_groups <- unique(c(stroke_groups, top_driver_groups)) |>
  head(8)

disability_trend <- payments_disability |>
  filter(disability_group %in% driver_groups) |>
  arrange(disability_group, quarter_date) |>
  group_by(disability_group) |>
  mutate(
    payment_growth = index_to_first(payment_amount),
    participant_growth = index_to_first(payment_participants),
    payment_per_participant_growth = index_to_first(payment_per_participant)
  ) |>
  ungroup()

if (nrow(disability_trend) > 0) {
  disability_trend |>
    select(quarter_date, disability_group, payment_growth, participant_growth, payment_per_participant_growth) |>
    pivot_longer(
      ends_with("_growth"),
      names_to = "measure",
      values_to = "growth"
    ) |>
    mutate(
      measure = recode(
        measure,
        payment_growth = "Payments",
        participant_growth = "Participants",
        payment_per_participant_growth = "Payments per participant"
      )
    ) |>
    filter(is.finite(growth)) |>
    ggplot(aes(quarter_date, growth, colour = disability_group)) +
    geom_hline(yintercept = 0, colour = "grey80") +
    geom_line(linewidth = 0.8) +
    geom_point(size = 1.8) +
    facet_wrap(~ measure) +
    scale_y_continuous(labels = percent) +
    labs(x = NULL, y = "Growth since first available quarter", colour = "Disability group")
} else {
  tibble(note = "No stroke or top-driver disability-group trend rows were available to plot.") |>
    kable()
}

Payment, participant and payment-per-participant trends for stroke and the largest disability-group growth contributors.
Code 
if (any(disability_driver$stroke_flag, na.rm = TRUE)) {
  disability_driver |>
    filter(stroke_flag) |>
    transmute(
      disability_group,
      payment_amount_first = dollar(payment_amount_first, accuracy = 1),
      payment_amount_last = dollar(payment_amount_last, accuracy = 1),
      payment_change = dollar(payment_change, accuracy = 1),
      change_share = percent(change_share, accuracy = 0.1),
      participants_first = comma(participants_first),
      participants_last = comma(participants_last),
      participant_change = comma(participant_change),
      payment_per_participant_change = dollar(payment_per_participant_change, accuracy = 1)
    ) |>
    kable()
} else {
  tibble(note = "Stroke was not exposed as a separate disability group in the current public payment extract.") |>
    kable()
}
disability_group payment_amount_first payment_amount_last payment_change change_share participants_first participants_last participant_change payment_per_participant_change
Stroke $1,336,377,000 $1,746,058,000 $409,681,000 5.3% 25,000 30,248 5,248 $4,270

The payment tables are useful, but the time window is short. These are growth-contribution signals, not stable long-run trends. The disability-group decomposition separates participant growth from payment intensity, but it still reflects aggregate public payment rows rather than participant-level need or plan history.

Budgets, Payments and Utilisation

Budget, utilisation and payments answer related but different questions. A high annualised support budget does not necessarily mean a high payment amount in the observed quarter, and low utilisation can reflect need, timing, provider availability or administrative barriers.

Code 
latest_model_data <- model_data |>
  filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
  filter(!is.na(avg_support_budget), !is.na(utilisation_rate), !is.na(payment_amount), !is.na(payment_per_participant)) |>
  mutate(
    utilisation_rate = pmin(utilisation_rate, 1),
    high_budget_cut = quantile(avg_support_budget, 0.75, na.rm = TRUE),
    low_utilisation_cut = quantile(utilisation_rate, 0.25, na.rm = TRUE),
    high_utilisation_cut = quantile(utilisation_rate, 0.75, na.rm = TRUE),
    high_payment_cut = quantile(payment_per_participant, 0.75, na.rm = TRUE),
    signal = case_when(
      avg_support_budget >= high_budget_cut & utilisation_rate <= low_utilisation_cut ~ "High budget, low utilisation",
      utilisation_rate >= high_utilisation_cut & payment_per_participant >= high_payment_cut ~ "High utilisation, high payment",
      payment_per_participant >= high_payment_cut & avg_support_budget < high_budget_cut ~ "High payment, moderate budget",
      TRUE ~ NA_character_
    )
  )
Code 
tibble(
  signal = c("High budget, low utilisation", "High utilisation, high payment", "High payment, moderate budget"),
  definition = c(
    "Average support budget is in the top quartile while utilisation is in the bottom quartile.",
    "Utilisation and payments per participant are both in the top quartile.",
    "Payments per participant are in the top quartile while average support budget is below the top quartile."
  )
) |>
  kable()
signal definition
High budget, low utilisation Average support budget is in the top quartile while utilisation is in the bottom quartile.
High utilisation, high payment Utilisation and payments per participant are both in the top quartile.
High payment, moderate budget Payments per participant are in the top quartile while average support budget is below the top quartile.
Code 
signal_data <- latest_model_data |>
  filter(!is.na(signal))

signal_data |>
  group_by(signal) |>
  summarise(
    service_districts = n_distinct(service_district),
    grouped_cells = n(),
    participants = sum(participant_count, na.rm = TRUE),
    total_payments = sum(payment_amount, na.rm = TRUE),
    median_payment_per_participant = safe_median(payment_per_participant),
    median_utilisation = safe_median(utilisation_rate),
    median_avg_budget = safe_median(avg_support_budget),
    .groups = "drop"
  ) |>
  mutate(
    participants = comma(participants),
    total_payments = dollar(total_payments, accuracy = 1),
    median_payment_per_participant = dollar(median_payment_per_participant, accuracy = 1),
    median_utilisation = percent(median_utilisation, accuracy = 0.1),
    median_avg_budget = dollar(median_avg_budget, accuracy = 1)
  ) |>
  kable()
signal service_districts grouped_cells participants total_payments median_payment_per_participant median_utilisation median_avg_budget
High budget, low utilisation 69 152 5,833 $783,146,000 $111,263 62.0% $159,000
High payment, moderate budget 46 71 5,321 $624,586,000 $111,333 79.0% $127,000
High utilisation, high payment 74 876 60,014 $10,258,221,000 $172,543 88.0% $194,000
Code 
signal_data |>
  group_by(signal, state, service_district, support_class) |>
  summarise(
    grouped_cells = n(),
    participants = sum(participant_count, na.rm = TRUE),
    total_payments = sum(payment_amount, na.rm = TRUE),
    median_utilisation = safe_median(utilisation_rate),
    median_payment_per_participant = safe_median(payment_per_participant),
    .groups = "drop"
  ) |>
  arrange(signal, desc(participants)) |>
  group_by(signal) |>
  slice_head(n = 5) |>
  ungroup() |>
  mutate(
    participants = comma(participants),
    total_payments = dollar(total_payments, accuracy = 1),
    median_utilisation = percent(median_utilisation, accuracy = 0.1),
    median_payment_per_participant = dollar(median_payment_per_participant, accuracy = 1)
  ) |>
  kable()
signal state service_district support_class grouped_cells participants total_payments median_utilisation median_payment_per_participant
High budget, low utilisation QLD Maroochydore Core 4 441 $54,727,000 65.0% $110,685
High budget, low utilisation NSW Murrumbidgee Core 5 332 $37,025,000 62.0% $117,062
High budget, low utilisation NSW North Sydney Core 4 319 $47,653,000 64.5% $143,835
High budget, low utilisation SA Southern Adelaide Core 3 221 $30,030,000 62.0% $130,688
High budget, low utilisation VIC Inner East Melbourne Core 4 220 $26,553,000 60.5% $100,802
High payment, moderate budget NSW Hunter New England Core 2 768 $83,485,000 83.0% $108,090
High payment, moderate budget VIC Bayside Peninsula Core 1 463 $48,582,000 84.0% $107,482
High payment, moderate budget NSW Nepean Blue Mountains Core 3 416 $43,775,000 81.0% $107,688
High payment, moderate budget QLD Beenleigh Core 3 378 $39,623,000 69.0% $108,160
High payment, moderate budget WA South East Metro Core 2 339 $35,984,000 76.0% $110,640
High utilisation, high payment NSW South Western Sydney Core 41 8,489 $1,216,462,000 89.0% $152,151
High utilisation, high payment NSW Western Sydney Core 32 4,943 $781,416,000 89.5% $173,531
High utilisation, high payment NSW Hunter New England Core 14 2,784 $447,264,000 90.0% $185,507
High utilisation, high payment VIC Southern Melbourne Core 25 2,619 $403,094,000 90.0% $170,656
High utilisation, high payment SA Northern Adelaide Core 22 2,402 $392,635,000 88.0% $168,278

Provider Market Context

Provider-market context is one of the few public levers that can plausibly help explain utilisation. The measures below are deliberately rough. participants_per_provider is a demand-pressure proxy, while payment_share_top10 is a concentration proxy. Neither measures provider quality, available hours, wait time, travel burden or workforce depth.

Code 
latest_model_data |>
  group_by(support_class) |>
  summarise(
    grouped_cells = n(),
    participants = sum(participant_count, na.rm = TRUE),
    median_utilisation = safe_median(utilisation_rate),
    median_participants_per_provider = safe_median(participants_per_provider),
    median_payment_share_top10 = safe_median(payment_share_top10),
    median_payment_per_participant = safe_median(payment_per_participant),
    .groups = "drop"
  ) |>
  mutate(
    participants = comma(participants),
    median_utilisation = percent(median_utilisation, accuracy = 0.1),
    median_participants_per_provider = number(median_participants_per_provider, accuracy = 0.1),
    median_payment_share_top10 = percent(median_payment_share_top10, accuracy = 0.1),
    median_payment_per_participant = dollar(median_payment_per_participant, accuracy = 1)
  ) |>
  kable()
support_class grouped_cells participants median_utilisation median_participants_per_provider median_payment_share_top10 median_payment_per_participant
Capital 2475 94,295 72.0% 1.5 76.0% $15,870
Core 5504 521,695 78.0% 1.1 29.0% $77,362
Code 
provider_quadrant_data <- latest_model_data |>
  filter(is.finite(participants_per_provider), is.finite(payment_share_top10)) |>
  mutate(
    provider_pressure = if_else(participants_per_provider >= safe_median(participants_per_provider), "High provider pressure", "Low provider pressure"),
    concentration = if_else(payment_share_top10 >= safe_median(payment_share_top10), "High concentration", "Low concentration"),
    quadrant = paste(provider_pressure, concentration, sep = " / ")
  )

provider_quadrant_data |>
  group_by(quadrant) |>
  summarise(
    service_districts = n_distinct(service_district),
    grouped_cells = n(),
    participants = sum(participant_count, na.rm = TRUE),
    median_utilisation = safe_median(utilisation_rate),
    median_payment_per_participant = safe_median(payment_per_participant),
    median_payment_share_top10 = safe_median(payment_share_top10),
    .groups = "drop"
  ) |>
  mutate(
    participants = comma(participants),
    median_utilisation = percent(median_utilisation, accuracy = 0.1),
    median_payment_per_participant = dollar(median_payment_per_participant, accuracy = 1),
    median_payment_share_top10 = percent(median_payment_share_top10, accuracy = 0.1)
  ) |>
  kable()
quadrant service_districts grouped_cells participants median_utilisation median_payment_per_participant median_payment_share_top10
High provider pressure / High concentration 1 43 5,630 70.0% $22,174 29.0%
Low provider pressure / High concentration 1 43 2,509 80.0% $108,350 29.0%

SIL and SDA Baseline

The utilisation extract includes a SILorSDA field. That makes it possible to set a pre-reform descriptive baseline before mandatory registration for supported independent living and platform providers begins on 1 July 2026. This is not an evaluation of that change; it is a baseline view of utilisation patterns before the reform takes effect.

Code 
sil_sda_utilisation <- utilisation |>
  filter(!is.na(sil_or_sda), sil_or_sda != "ALL", support_class != "ALL") |>
  group_by(quarter_date, sil_or_sda, support_class) |>
  summarise(utilisation = safe_mean(utilisation_rate), rows = n(), .groups = "drop")

sil_sda_utilisation |>
  ggplot(aes(quarter_date, utilisation, colour = sil_or_sda)) +
  geom_line(linewidth = 0.8) +
  geom_point(size = 2) +
  facet_wrap(~ support_class) +
  scale_y_continuous(labels = percent) +
  labs(x = NULL, y = "Average utilisation", colour = "SIL/SDA flag")

Code 
sil_sda_payment_items <- payments_item |>
  filter(str_detect(str_to_lower(paste(support_category, support_item_desc)), "supported independent living|\\bsil\\b|specialist disability accommodation|\\bsda\\b"))

if (nrow(sil_sda_payment_items) > 0) {
  sil_sda_payment_items |>
    group_by(quarter_date, support_class, support_category) |>
    summarise(payment_amount = sum(payment_amount, na.rm = TRUE), participants = sum(payment_participants, na.rm = TRUE), .groups = "drop") |>
    arrange(desc(payment_amount)) |>
    mutate(
      payment_amount = dollar(payment_amount, accuracy = 1),
      participants = comma(participants)
    ) |>
    slice_head(n = 15) |>
    kable()
} else {
  tibble(note = "No payment item descriptions matched SIL/SDA keywords in the national item-level extract.") |>
    kable()
}
note
No payment item descriptions matched SIL/SDA keywords in the national item-level extract.

Regional Context and Join Quality

The regional layer is deliberately conservative. It uses the NDIS service-district-to-LGA mapping, then joins ABS LGA SEIFA, age-structure, population, and remoteness allocation files where the source workbooks resolve cleanly. Remoteness is allocated from ABS mesh-block and SA1 correspondence files up to LGA, then aggregated to service districts with population weights. The table below keeps the join quality visible so the model does not quietly imply more geographic precision than the data supports.

Code 
join_diagnostics |>
  mutate(coverage = percent(coverage, accuracy = 0.1)) |>
  kable()
check rows_or_groups denominator coverage
Budget rows with utilisation 62709 105041 59.7%
Budget rows with active providers 61506 105041 58.6%
Budget rows with grouped payments 63823 105041 60.8%
Budget rows with market concentration 1112 105041 1.1%
Service districts with LGA mapping 80 84 95.2%
Service districts with ABS SEIFA/age context 80 84 95.2%
Service districts with remoteness context 80 84 95.2%
Service districts with LGA participant context 80 84 95.2%
Code 
model_data |>
  filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
  group_by(state, service_district) |>
  summarise(
    participants = sum(participant_count, na.rm = TRUE),
    avg_budget = safe_weighted_mean(avg_support_budget, participant_count),
    utilisation = safe_weighted_mean(utilisation_rate, participant_count),
    payment_per_participant = safe_weighted_mean(payment_per_participant, participant_count),
    participants_per_provider = safe_weighted_mean(participants_per_provider, participant_count),
    payment_share_top10 = safe_mean(payment_share_top10),
    remoteness = safe_mode_character(remoteness),
    irsd_decile = safe_mean(irsd_decile),
    median_age = safe_mean(median_age),
    pct_65_plus = safe_mean(pct_65_plus),
    mapped_lgas = max(mapped_lgas, na.rm = TRUE),
    .groups = "drop"
  ) |>
  arrange(desc(participants)) |>
  slice_head(n = 20) |>
  mutate(
    participants = comma(participants),
    avg_budget = dollar(avg_budget, accuracy = 1),
    utilisation = percent(utilisation, accuracy = 0.1),
    payment_per_participant = dollar(payment_per_participant, accuracy = 1),
    participants_per_provider = number(participants_per_provider, accuracy = 0.1),
    payment_share_top10 = percent(payment_share_top10, accuracy = 0.1),
    irsd_decile = number(irsd_decile, accuracy = 0.1),
    median_age = number(median_age, accuracy = 0.1),
    pct_65_plus = number(pct_65_plus, accuracy = 0.1, suffix = "%"),
    mapped_lgas = comma(mapped_lgas)
  ) |>
  kable()
state service_district participants avg_budget utilisation payment_per_participant participants_per_provider payment_share_top10 remoteness irsd_decile median_age pct_65_plus mapped_lgas
NSW Hunter New England 67,783 $44,696 78.1% $60,500 6.8 NA Outer Regional Australia 5.7 41.3 20.6% 22
NSW South Western Sydney 65,541 $45,184 85.5% $67,085 9.3 NA Major Cities 3.5 36.4 14.1% 7
QLD Brisbane 53,192 $46,187 75.8% $58,075 17.5 NA Major Cities 9.0 35.9 13.2% 1
NSW Western Sydney 49,424 $46,861 85.1% $69,767 8.6 NA Major Cities 6.5 35.2 12.0% 4
VIC Bayside Peninsula 43,744 $42,201 77.1% $52,385 15.5 NA Major Cities 9.2 41.2 18.6% 7
SA Northern Adelaide 41,658 $43,685 75.6% $59,811 48.5 NA Major Cities 3.8 37.1 15.6% 4
VIC Southern Melbourne 36,206 $42,944 82.0% $61,780 14.6 NA Major Cities 6.0 34.7 12.0% 3
VIC Western Melbourne 34,330 $40,186 79.7% $51,890 17.8 NA Major Cities 8.2 34.4 10.3% 5
QLD Beenleigh 33,961 $44,387 76.4% $59,177 19.2 NA Major Cities 4.6 37.2 15.3% 2
VIC North East Melbourne 33,581 $44,364 80.0% $58,280 11.6 NA Major Cities 8.0 37.3 14.3% 5
QLD Caboolture/Strathpine 32,442 $43,681 76.7% $56,970 16.2 NA Major Cities 7.0 39.0 17.3% 1
QLD Robina 29,218 $44,825 77.1% $59,060 13.8 NA Major Cities 8.0 39.2 17.3% 1
SA Southern Adelaide 26,816 $43,756 72.1% $54,782 28.4 NA Major Cities 7.5 41.5 20.1% 4
VIC Brimbank Melton 26,518 $36,866 79.1% $48,447 37.2 NA Major Cities 3.9 35.4 12.5% 2
QLD Maroochydore 26,333 $44,352 75.8% $55,140 12.5 NA Inner Regional Australia 7.5 44.6 22.4% 3
NSW North Sydney 25,978 $49,891 80.9% $72,548 2.6 NA Major Cities 10.0 40.6 17.2% 9
VIC Barwon 25,904 $38,949 71.4% $46,723 13.8 NA Inner Regional Australia 8.0 40.2 19.7% 4
NSW South Eastern Sydney 25,679 $47,941 81.1% $66,487 3.3 NA Major Cities 9.0 38.2 16.4% 6
VIC Outer East Melbourne 25,171 $39,244 75.6% $49,929 10.2 NA Major Cities 9.0 40.0 17.2% 3
NSW Illawarra Shoalhaven 24,407 $43,976 79.7% $58,256 4.1 NA Inner Regional Australia 6.7 41.9 20.9% 4
Code 
latest_service_geo <- model_data |>
  filter(quarter_date == max(quarter_date, na.rm = TRUE)) |>
  group_by(state, service_district) |>
  summarise(
    participants = sum(participant_count, na.rm = TRUE),
    avg_budget = safe_weighted_mean(avg_support_budget, participant_count),
    utilisation = safe_weighted_mean(utilisation_rate, participant_count),
    payment_per_participant = safe_weighted_mean(payment_per_participant, participant_count),
    participants_per_provider = safe_weighted_mean(participants_per_provider, participant_count),
    payment_share_top10 = safe_mean(payment_share_top10),
    remoteness = safe_mode_character(remoteness),
    mapped_lgas = max(mapped_lgas, na.rm = TRUE),
    example_lgas = safe_mode_character(mapped_lga_names),
    irsd_decile = safe_mean(irsd_decile),
    median_age = safe_mean(median_age),
    pct_65_plus = safe_mean(pct_65_plus),
    .groups = "drop"
  ) |>
  mutate(
    mapped_lgas = if_else(is.infinite(mapped_lgas), NA_real_, as.numeric(mapped_lgas)),
    participant_rank = percent_rank(participants),
    budget_rank = percent_rank(avg_budget),
    payment_rank = percent_rank(payment_per_participant),
    utilisation_rank = percent_rank(utilisation),
    provider_pressure_rank = percent_rank(participants_per_provider),
    concentration_rank = percent_rank(payment_share_top10),
    geography_pressure_score = rowMeans(
      cbind(payment_rank, utilisation_rank, provider_pressure_rank, concentration_rank),
      na.rm = TRUE
    ),
    geography_pressure_score = if_else(is.nan(geography_pressure_score), NA_real_, geography_pressure_score)
  )

if (nrow(latest_service_geo) > 0) {
  latest_service_geo |>
    arrange(desc(geography_pressure_score), desc(participants)) |>
    slice_head(n = 15) |>
    transmute(
      state,
      service_district,
      remoteness,
      mapped_lgas = comma(mapped_lgas),
      participants = comma(participants),
      avg_budget = dollar(avg_budget, accuracy = 1),
      utilisation = percent(utilisation, accuracy = 0.1),
      payment_per_participant = dollar(payment_per_participant, accuracy = 1),
      participants_per_provider = number(participants_per_provider, accuracy = 0.1),
      payment_share_top10 = percent(payment_share_top10, accuracy = 0.1),
      irsd_decile = number(irsd_decile, accuracy = 0.1)
    ) |>
    kable()
} else {
  tibble(note = "No latest-quarter service-district geography profile could be built.") |>
    kable()
}
state service_district remoteness mapped_lgas participants avg_budget utilisation payment_per_participant participants_per_provider payment_share_top10 irsd_decile
NT Darwin Urban Outer Regional Australia 6 4,793 $56,479 81.2% $95,172 97.5 NA 8.7
NT Central Australia Very Remote Australia 3 1,181 $116,076 83.4% $206,752 5.9 NA 5.4
VIC Southern Melbourne Major Cities 3 36,206 $42,944 82.0% $61,780 14.6 NA 6.0
NSW South Western Sydney Major Cities 7 65,541 $45,184 85.5% $67,085 9.3 NA 3.5
NSW Western Sydney Major Cities 4 49,424 $46,861 85.1% $69,767 8.6 NA 6.5
QLD Beenleigh Major Cities 2 33,961 $44,387 76.4% $59,177 19.2 NA 4.6
SA Northern Adelaide Major Cities 4 41,658 $43,685 75.6% $59,811 48.5 NA 3.8
QLD Robina Major Cities 1 29,218 $44,825 77.1% $59,060 13.8 NA 8.0
VIC North East Melbourne Major Cities 5 33,581 $44,364 80.0% $58,280 11.6 NA 8.0
QLD Caboolture/Strathpine Major Cities 1 32,442 $43,681 76.7% $56,970 16.2 NA 7.0
VIC Hume Moreland Major Cities 2 23,331 $40,516 82.6% $55,848 11.2 NA 5.1
QLD Brisbane Major Cities 1 53,192 $46,187 75.8% $58,075 17.5 NA 9.0
VIC Western Melbourne Major Cities 5 34,330 $40,186 79.7% $51,890 17.8 NA 8.2
SA Eastern Adelaide Major Cities 7 7,975 $47,136 75.3% $68,222 8.7 NA 9.0
WA South East Metro Major Cities 4 12,801 $54,248 75.9% $67,261 6.7 NA 6.6
Code 
remoteness_profile <- latest_service_geo |>
  filter(!is.na(remoteness))

if (nrow(remoteness_profile) > 0) {
  remoteness_profile |>
    group_by(remoteness) |>
    summarise(
      service_districts = n_distinct(service_district),
      participants = sum(participants, na.rm = TRUE),
      avg_budget = safe_weighted_mean(avg_budget, participants),
      utilisation = safe_weighted_mean(utilisation, participants),
      payment_per_participant = safe_weighted_mean(payment_per_participant, participants),
      participants_per_provider = safe_weighted_mean(participants_per_provider, participants),
      payment_share_top10 = safe_mean(payment_share_top10),
      median_irsd_decile = safe_median(irsd_decile),
      .groups = "drop"
    ) |>
    arrange(median_irsd_decile, desc(participants)) |>
    mutate(
      participants = comma(participants),
      avg_budget = dollar(avg_budget, accuracy = 1),
      utilisation = percent(utilisation, accuracy = 0.1),
      payment_per_participant = dollar(payment_per_participant, accuracy = 1),
      participants_per_provider = number(participants_per_provider, accuracy = 0.1),
      payment_share_top10 = percent(payment_share_top10, accuracy = 0.1),
      median_irsd_decile = number(median_irsd_decile, accuracy = 0.1)
    ) |>
    kable()
} else {
  tibble(note = "Remoteness was not available in the downloaded LGA mapping or participants-by-LGA file for this render.") |>
    kable()
}
remoteness service_districts participants avg_budget utilisation payment_per_participant participants_per_provider payment_share_top10 median_irsd_decile
Remote Australia 3 8,446 NA NA NA NA NA 3.5
Outer Regional Australia 17 155,850 NA NA NA NA NA 4.8
Very Remote Australia 9 65,617 NA NA NA NA NA 5.1
Inner Regional Australia 22 308,107 NA NA NA NA 47.5% 6.6
Major Cities 29 726,626 NA NA NA NA NA 7.3

Richer Geography Context

The richer geography layer adds two lenses: socioeconomic decile where the ABS join works, and support-class market concentration. Equity-specific signals are left out of this version so the main post stays focused on spend growth mechanics.

Code 
seifa_summary <- model_data |>
  filter(quarter_date == max(quarter_date, na.rm = TRUE), !is.na(irsd_decile))

if (nrow(seifa_summary) > 0) {
  seifa_summary |>
    mutate(irsd_decile = round(irsd_decile)) |>
    group_by(irsd_decile) |>
    summarise(
      participants = sum(participant_count, na.rm = TRUE),
      avg_budget = safe_weighted_mean(avg_support_budget, participant_count),
      utilisation = safe_weighted_mean(utilisation_rate, participant_count),
      payment_per_participant = safe_weighted_mean(payment_per_participant, participant_count),
      participants_per_provider = safe_weighted_mean(participants_per_provider, participant_count),
      payment_share_top10 = safe_mean(payment_share_top10),
      .groups = "drop"
    ) |>
    arrange(irsd_decile) |>
    mutate(
      participants = comma(participants),
      avg_budget = dollar(avg_budget, accuracy = 1),
      utilisation = percent(utilisation, accuracy = 0.1),
      payment_per_participant = dollar(payment_per_participant, accuracy = 1),
      participants_per_provider = number(participants_per_provider, accuracy = 0.1),
      payment_share_top10 = percent(payment_share_top10, accuracy = 0.1)
    ) |>
    kable()
} else {
  tibble(note = "SEIFA was not joined in this render, so the SEIFA decile summary is skipped rather than showing an NA-heavy table.") |>
    kable()
}
irsd_decile participants avg_budget utilisation payment_per_participant participants_per_provider payment_share_top10
1 971 $41,837 68.9% $32,431 14.6 NA
2 14,228 $48,098 75.7% $66,012 2.4 NA
3 80,224 $44,776 82.9% $65,833 8.6 NA
4 104,874 $42,036 76.3% $56,692 32.6 NA
5 177,249 $44,154 74.9% $58,817 8.4 NA
6 197,676 $42,288 75.6% $57,515 12.3 NA
7 201,112 $44,378 79.0% $60,664 7.7 NA
8 200,022 $42,906 76.0% $54,773 15.1 NA
9 239,409 $44,985 76.1% $58,183 12.2 NA
10 48,881 $44,640 77.6% $64,913 37.9 47.5%
Code 
market_concentration |>
  filter(market_quarter_date == max(market_quarter_date, na.rm = TRUE), support_class != "ALL") |>
  group_by(support_class) |>
  summarise(
    service_districts = n_distinct(service_district),
    median_top10_share = safe_median(payment_share_top10),
    high_concentration_share = safe_mean(as.numeric(payment_share_top10 >= 0.7)),
    .groups = "drop"
  ) |>
  arrange(desc(median_top10_share)) |>
  mutate(
    median_top10_share = percent(median_top10_share, accuracy = 0.1),
    high_concentration_share = percent(high_concentration_share, accuracy = 0.1)
  ) |>
  kable()
support_class service_districts median_top10_share high_concentration_share
Capital 88 62.0% 35.2%
Core 89 34.0% 13.5%
Capacity Building 89 23.0% 4.5%

Outcomes, Spend and Service Access

The outcomes analysis has been moved into a separate companion post: NDIS Outcomes, Spend and Service Access. That page inspects the regional dashboard workbook schema directly, chooses a small set of high-value outcome families, and treats outcomes as possible investment channels rather than only as cost offsets.

Model 1: Average Support Budget

The first model asks whether grouped participant characteristics, region, support class, provider supply, and ABS context explain differences in average support budgets. The expanded model now carries through service district plus joined LGA-level SEIFA, age-structure and population measures where the joins succeed. The test period is the latest available quarter in the study window, so the model is judged on a future quarter rather than a random split.

Code 
budget_model_data <- model_data |>
  filter(!is.na(avg_support_budget), !is.na(participant_count), participant_count > 0) |>
  mutate(
    across(c(state, service_district, age_group, disability_group, support_class, remoteness), \(x) fct_na_value_to_level(fct_lump_n(as.factor(x), n = 20), level = "Missing")),
    log_budget = log1p(avg_support_budget),
    log_participants = log1p(participant_count),
    log_providers = log1p(coalesce(active_providers, 0)),
    log_participants_per_provider = log1p(coalesce(participants_per_provider, 0)),
    log_lga_population = log1p(coalesce(lga_population, 0)),
    payment_share_top10 = coalesce(payment_share_top10, median(payment_share_top10, na.rm = TRUE)),
    payment_share_top10 = ifelse(is.na(payment_share_top10) | is.nan(payment_share_top10), 0, payment_share_top10),
    across(c(irsd_score, irsd_decile, ier_score, ier_decile, ieo_score, ieo_decile, median_age, pct_0_14, pct_15_64, pct_65_plus, remoteness_rank, remoteness_major_cities_share, remoteness_inner_regional_share, remoteness_outer_regional_share, remoteness_remote_share, remoteness_very_remote_share), \(x) replace_na(x, median(x, na.rm = TRUE))),
    across(c(irsd_score, irsd_decile, ier_score, ier_decile, ieo_score, ieo_decile, median_age, pct_0_14, pct_15_64, pct_65_plus, remoteness_rank, remoteness_major_cities_share, remoteness_inner_regional_share, remoteness_outer_regional_share, remoteness_remote_share, remoteness_very_remote_share), \(x) ifelse(is.na(x) | is.nan(x), 0, x))
  )

test_quarter <- max(budget_model_data$quarter_date, na.rm = TRUE)
budget_train <- filter(budget_model_data, quarter_date < test_quarter)
budget_test <- filter(budget_model_data, quarter_date == test_quarter)

budget_baseline <- lm(log_budget ~ quarter_index + log_participants, data = budget_train)

factor_terms <- c("state", "service_district", "age_group", "disability_group", "support_class", "remoteness")
factor_terms <- factor_terms[map_int(budget_train[factor_terms], \(x) n_distinct(x, na.rm = TRUE)) > 1]
numeric_terms <- c("quarter_index", "log_participants", "log_providers", "log_participants_per_provider", "payment_share_top10", "irsd_score", "irsd_decile", "ier_score", "ier_decile", "ieo_score", "ieo_decile", "median_age", "pct_0_14", "pct_15_64", "pct_65_plus", "log_lga_population", "remoteness_rank", "remoteness_major_cities_share", "remoteness_inner_regional_share", "remoteness_outer_regional_share", "remoteness_remote_share", "remoteness_very_remote_share")
numeric_terms <- numeric_terms[map_lgl(budget_train[numeric_terms], \(x) isTRUE(sd(x, na.rm = TRUE) > 0))]
budget_formula <- as.formula(paste("log_budget ~", paste(c(numeric_terms, factor_terms), collapse = " + ")))

budget_expanded <- lm(budget_formula, data = budget_train)

set.seed(20260526)
budget_rf_train <- budget_train |>
  drop_na(log_budget) |>
  slice_sample(n = min(nrow(drop_na(budget_train, log_budget)), 8000))

budget_rf <- randomForest(
  budget_formula,
  data = budget_rf_train,
  ntree = 60,
  importance = TRUE
)

score_regression <- function(actual, predicted) {
  tibble(
    rmse = sqrt(mean((actual - predicted)^2, na.rm = TRUE)),
    mae = mean(abs(actual - predicted), na.rm = TRUE)
  )
}

partial_dependence_data <- function(model, data, features, response_transform = identity,
                                    x_transforms = list(), feature_labels = NULL,
                                    grid_n = 25, sample_n = 1500) {
  pd_sample <- data |>
    slice_sample(n = min(nrow(data), sample_n))

  map_dfr(features, \(feature) {
    if (!feature %in% names(pd_sample) || !is.numeric(pd_sample[[feature]])) {
      return(tibble())
    }

    x <- pd_sample[[feature]]
    x <- x[is.finite(x)]
    if (length(unique(x)) < 2) return(tibble())

    if (length(unique(x)) <= grid_n) {
      grid <- sort(unique(x))
    } else {
      grid <- quantile(x, probs = seq(0.02, 0.98, length.out = grid_n), na.rm = TRUE)
      grid <- sort(unique(as.numeric(grid)))
    }

    x_transform <- x_transforms[[feature]]
    if (is.null(x_transform)) x_transform <- identity
    feature_label <- feature_labels[[feature]]
    if (is.null(feature_label)) feature_label <- feature

    map_dfr(grid, \(value) {
      newdata <- pd_sample
      newdata[[feature]] <- value
      tibble(
        feature = feature,
        feature_label = feature_label,
        feature_value = x_transform(value),
        predicted_response = mean(response_transform(predict(model, newdata = newdata)), na.rm = TRUE)
      )
    })
  })
}

plot_partial_dependence <- function(pdp_data, y_label, line_colour) {
  if (nrow(pdp_data) == 0) {
    return(tibble(note = "No partial dependence data could be computed for the selected features.") |> kable())
  }

  ggplot(pdp_data, aes(feature_value, predicted_response)) +
    geom_line(linewidth = 0.8, colour = line_colour) +
    facet_wrap(~ feature_label, scales = "free_x", labeller = label_wrap_gen(width = 26)) +
    theme(axis.text.x = element_text(size = 8)) +
    labs(x = NULL, y = y_label)
}

budget_scores <- bind_rows(
  score_regression(budget_test$log_budget, predict(budget_baseline, newdata = budget_test)) |> mutate(model = "Baseline linear"),
  score_regression(budget_test$log_budget, predict(budget_expanded, newdata = budget_test)) |> mutate(model = "Expanded linear"),
  score_regression(budget_test$log_budget, predict(budget_rf, newdata = budget_test)) |> mutate(model = "Random forest")
) |>
  select(model, rmse, mae)

budget_scores |> kable(digits = 3)
model rmse mae
Baseline linear 1.053 0.843
Expanded linear 0.539 0.398
Random forest 0.289 0.187
Code 
budget_predictions <- budget_test |>
  mutate(
    pred_log_budget = predict(budget_rf, newdata = budget_test),
    pred_budget = expm1(pred_log_budget),
    budget_residual = avg_support_budget - pred_budget
  )

ggplot(budget_predictions, aes(pred_budget, avg_support_budget)) +
  geom_abline(linetype = "dashed", colour = "grey50") +
  geom_point(aes(size = participant_count), alpha = 0.35, colour = "#1b6ca8") +
  scale_x_continuous(labels = dollar) +
  scale_y_continuous(labels = dollar) +
  scale_size_area(labels = comma) +
  labs(x = "Predicted average support budget", y = "Observed average support budget", size = "Participants")

Predicted versus observed average support budgets in the holdout quarter.
Code 
importance(budget_rf) |>
  as.data.frame() |>
  tibble::rownames_to_column("feature") |>
  arrange(desc(IncNodePurity)) |>
  slice_head(n = 12) |>
  ggplot(aes(reorder(feature, IncNodePurity), IncNodePurity)) +
  geom_col(fill = "#1b6ca8") +
  coord_flip() +
  labs(x = NULL, y = "Random forest importance", title = "Most useful predictors for average support budget")

The clearest partial-dependence shape is provider pressure. In this fitted model, higher participants per active provider is associated with lower predicted average support budgets after a small low-pressure bump. The age-structure curves are weaker, but older local age profiles also lean slightly lower after the rest of the grouped data is averaged out.

Partial dependence: average support budget random forest

These curves hold the rest of the training distribution in place and vary one predictor at a time. They are useful for shape, not causality. Provider pressure is shown on its original scale even though the model uses log1p(participants_per_provider).

Code 
budget_pdp <- partial_dependence_data(
  budget_rf,
  budget_rf_train,
  features = c("log_participants_per_provider", "median_age", "pct_65_plus", "irsd_decile", "payment_share_top10"),
  response_transform = expm1,
  x_transforms = list(
    log_participants_per_provider = expm1,
    pct_65_plus = \(x) x * 100,
    payment_share_top10 = \(x) x * 100
  ),
  feature_labels = list(
    log_participants_per_provider = "Participants per active provider",
    median_age = "Median LGA age",
    pct_65_plus = "LGA population aged 65+ (%)",
    irsd_decile = "SEIFA disadvantage decile",
    payment_share_top10 = "Top 10 provider payment share (%)"
  )
)

plot_partial_dependence(budget_pdp, y_label = "Predicted average support budget", line_colour = "#1b6ca8") +
  scale_y_continuous(labels = dollar)

Partial dependence curves for selected average support budget predictors.

Model 2: Plan Utilisation

Utilisation is a bounded rate, so this first model keeps the interpretation simple: it predicts the observed utilisation rate for grouped cells and checks where predictions differ most from the latest quarter. The same joined geography fields used in the budget model are available here when they have non-missing variation.

Code 
util_model_data <- model_data |>
  filter(!is.na(utilisation_rate), !is.na(avg_support_budget), utilisation_rate >= 0, utilisation_rate <= 1.5) |>
  mutate(
    utilisation_rate = pmin(utilisation_rate, 1),
    across(c(state, service_district, age_group, disability_group, support_class, remoteness), \(x) fct_na_value_to_level(fct_lump_n(as.factor(x), n = 20), level = "Missing")),
    log_budget = log1p(avg_support_budget),
    log_participants = log1p(coalesce(participant_count, 0)),
    log_providers = log1p(coalesce(active_providers, 0)),
    log_participants_per_provider = log1p(coalesce(participants_per_provider, 0)),
    log_lga_population = log1p(coalesce(lga_population, 0)),
    payment_share_top10 = coalesce(payment_share_top10, median(payment_share_top10, na.rm = TRUE)),
    payment_share_top10 = ifelse(is.na(payment_share_top10) | is.nan(payment_share_top10), 0, payment_share_top10),
    across(c(irsd_score, irsd_decile, ier_score, ier_decile, ieo_score, ieo_decile, median_age, pct_0_14, pct_15_64, pct_65_plus, remoteness_rank, remoteness_major_cities_share, remoteness_inner_regional_share, remoteness_outer_regional_share, remoteness_remote_share, remoteness_very_remote_share), \(x) replace_na(x, median(x, na.rm = TRUE))),
    across(c(irsd_score, irsd_decile, ier_score, ier_decile, ieo_score, ieo_decile, median_age, pct_0_14, pct_15_64, pct_65_plus, remoteness_rank, remoteness_major_cities_share, remoteness_inner_regional_share, remoteness_outer_regional_share, remoteness_remote_share, remoteness_very_remote_share), \(x) ifelse(is.na(x) | is.nan(x), 0, x))
  )

util_test_quarter <- max(util_model_data$quarter_date, na.rm = TRUE)
util_train <- filter(util_model_data, quarter_date < util_test_quarter)
util_test <- filter(util_model_data, quarter_date == util_test_quarter)

util_baseline <- lm(utilisation_rate ~ quarter_index + log_budget, data = util_train)
util_factor_terms <- c("state", "service_district", "age_group", "disability_group", "support_class", "remoteness")
util_factor_terms <- util_factor_terms[map_int(util_train[util_factor_terms], \(x) n_distinct(x, na.rm = TRUE)) > 1]
util_numeric_terms <- c("quarter_index", "log_budget", "log_participants", "log_providers", "log_participants_per_provider", "payment_share_top10", "irsd_score", "irsd_decile", "ier_score", "ier_decile", "ieo_score", "ieo_decile", "median_age", "pct_0_14", "pct_15_64", "pct_65_plus", "log_lga_population", "remoteness_rank", "remoteness_major_cities_share", "remoteness_inner_regional_share", "remoteness_outer_regional_share", "remoteness_remote_share", "remoteness_very_remote_share")
util_numeric_terms <- util_numeric_terms[map_lgl(util_train[util_numeric_terms], \(x) isTRUE(sd(x, na.rm = TRUE) > 0))]
util_formula <- as.formula(paste("utilisation_rate ~", paste(c(util_numeric_terms, util_factor_terms), collapse = " + ")))

util_expanded <- lm(util_formula, data = util_train)

set.seed(20260526)
util_rf_train <- util_train |>
  drop_na(utilisation_rate) |>
  slice_sample(n = min(nrow(drop_na(util_train, utilisation_rate)), 8000))

util_rf <- randomForest(
  util_formula,
  data = util_rf_train,
  ntree = 60,
  importance = TRUE
)

util_scores <- bind_rows(
  score_regression(util_test$utilisation_rate, predict(util_baseline, newdata = util_test)) |> mutate(model = "Baseline linear"),
  score_regression(util_test$utilisation_rate, predict(util_expanded, newdata = util_test)) |> mutate(model = "Expanded linear"),
  score_regression(util_test$utilisation_rate, predict(util_rf, newdata = util_test)) |> mutate(model = "Random forest")
) |>
  select(model, rmse, mae)

util_scores |> kable(digits = 3)
model rmse mae
Baseline linear 0.176 0.131
Expanded linear 0.172 0.127
Random forest 0.163 0.114
Code 
util_predictions <- util_test |>
  mutate(
    pred_utilisation = pmin(pmax(predict(util_rf, newdata = util_test), 0), 1),
    utilisation_residual = utilisation_rate - pred_utilisation
  )

ggplot(util_predictions, aes(pred_utilisation, utilisation_rate)) +
  geom_abline(linetype = "dashed", colour = "grey50") +
  geom_point(aes(size = participant_count), alpha = 0.35, colour = "#7a4b9d") +
  scale_x_continuous(labels = percent) +
  scale_y_continuous(labels = percent) +
  scale_size_area(labels = comma) +
  labs(x = "Predicted utilisation", y = "Observed utilisation", size = "Participants")

Predicted versus observed utilisation in the holdout quarter.
Code 
importance(util_rf) |>
  as.data.frame() |>
  tibble::rownames_to_column("feature") |>
  arrange(desc(IncNodePurity)) |>
  slice_head(n = 12) |>
  ggplot(aes(reorder(feature, IncNodePurity), IncNodePurity)) +
  geom_col(fill = "#7a4b9d") +
  coord_flip() +
  labs(x = NULL, y = "Random forest importance", title = "Most useful predictors for plan utilisation")

The utilisation PDPs tell a different story from the budget model. Average support budget has the strongest visible positive shape, while participants per provider is close to flat after the rest of the model context is averaged over. That is a useful caution: provider pressure may matter in particular regions or supports, but it is not a simple national monotonic utilisation effect in this grouped model.

Partial dependence: plan utilisation random forest

These plots are most useful for asking access questions. For example, if predicted utilisation falls as participants per provider rises, that is consistent with provider pressure constraining use of plans. It is still not enough to prove access failure, because participant mix and support needs are only partly observed in the public files.

Code 
util_pdp <- partial_dependence_data(
  util_rf,
  util_rf_train,
  features = c("log_participants_per_provider", "median_age", "log_budget", "pct_0_14", "irsd_decile", "payment_share_top10"),
  response_transform = \(x) pmin(pmax(x, 0), 1),
  x_transforms = list(
    log_participants_per_provider = expm1,
    log_budget = \(x) expm1(x) / 1000,
    pct_0_14 = \(x) x * 100,
    payment_share_top10 = \(x) x * 100
  ),
  feature_labels = list(
    log_participants_per_provider = "Participants per active provider",
    median_age = "Median LGA age",
    log_budget = "Average support budget ($000s)",
    pct_0_14 = "LGA population aged 0-14 (%)",
    irsd_decile = "SEIFA disadvantage decile",
    payment_share_top10 = "Top 10 provider payment share (%)"
  )
)

plot_partial_dependence(util_pdp, y_label = "Predicted utilisation", line_colour = "#7a4b9d") +
  scale_y_continuous(labels = percent)

Partial dependence curves for selected plan utilisation predictors.

The supporting models help separate the mechanics behind spend growth:

  • Average support budget: the strongest random-forest predictors are support class, disability group, age group, active providers, participants per active provider.
  • Plan utilisation: the strongest random-forest predictors are average support budget, disability group, support class, age group, active providers.

Model 3: Overall Spend

The main model asks how well public explanatory factors describe actual payment dollars for each grouped quarterly cell. I exclude lagged payments and lagged budgets, and I also keep current average budget and utilisation out of this model. Those variables are useful components of spend, but including them here would make the model less useful for describing upstream drivers such as participant scale, disability mix, age mix, geography and provider-market context.

Code 
spend_model_data <- model_data |>
  filter(
    !is.na(payment_amount), payment_amount > 0,
    !is.na(participant_count), participant_count > 0
  ) |>
  mutate(
    across(c(state, service_district, age_group, disability_group, support_class, remoteness), \(x) fct_na_value_to_level(fct_lump_n(as.factor(x), n = 25), level = "Missing")),
    log_payment = log1p(payment_amount),
    log_participants = log1p(participant_count),
    log_providers = log1p(coalesce(active_providers, 0)),
    log_participants_per_provider = log1p(coalesce(participants_per_provider, 0)),
    log_lga_population = log1p(coalesce(lga_population, 0)),
    payment_share_top10 = coalesce(payment_share_top10, median(payment_share_top10, na.rm = TRUE)),
    payment_share_top10 = ifelse(is.na(payment_share_top10) | is.nan(payment_share_top10), 0, payment_share_top10),
    across(c(irsd_score, irsd_decile, ier_score, ier_decile, ieo_score, ieo_decile, median_age, pct_0_14, pct_15_64, pct_65_plus, remoteness_rank, remoteness_major_cities_share, remoteness_inner_regional_share, remoteness_outer_regional_share, remoteness_remote_share, remoteness_very_remote_share), \(x) replace_na(x, median(x, na.rm = TRUE))),
    across(c(irsd_score, irsd_decile, ier_score, ier_decile, ieo_score, ieo_decile, median_age, pct_0_14, pct_15_64, pct_65_plus, remoteness_rank, remoteness_major_cities_share, remoteness_inner_regional_share, remoteness_outer_regional_share, remoteness_remote_share, remoteness_very_remote_share), \(x) ifelse(is.na(x) | is.nan(x), 0, x))
  )

spend_test_quarter <- max(spend_model_data$quarter_date, na.rm = TRUE)
spend_train <- filter(spend_model_data, quarter_date < spend_test_quarter)
spend_test <- filter(spend_model_data, quarter_date == spend_test_quarter)

spend_baseline <- lm(log_payment ~ quarter_index + log_participants, data = spend_train)

spend_factor_terms <- c("state", "service_district", "age_group", "disability_group", "support_class", "remoteness")
spend_factor_terms <- spend_factor_terms[map_int(spend_train[spend_factor_terms], \(x) n_distinct(x, na.rm = TRUE)) > 1]
spend_numeric_terms <- c(
  "quarter_index", "log_participants",
  "log_providers", "log_participants_per_provider", "payment_share_top10",
  "irsd_score", "irsd_decile", "ier_score", "ier_decile", "ieo_score", "ieo_decile",
  "median_age", "pct_0_14", "pct_15_64", "pct_65_plus", "log_lga_population",
  "remoteness_rank", "remoteness_major_cities_share", "remoteness_inner_regional_share",
  "remoteness_outer_regional_share", "remoteness_remote_share", "remoteness_very_remote_share"
)
spend_numeric_terms <- spend_numeric_terms[map_lgl(spend_train[spend_numeric_terms], \(x) isTRUE(sd(x, na.rm = TRUE) > 0))]
spend_expanded_formula <- as.formula(paste("log_payment ~", paste(c(spend_numeric_terms, spend_factor_terms), collapse = " + ")))
spend_expanded <- lm(spend_expanded_formula, data = spend_train)

set.seed(20260526)
spend_rf_train <- spend_train |>
  drop_na(log_payment) |>
  slice_sample(n = min(nrow(drop_na(spend_train, log_payment)), 10000))

spend_rf <- randomForest(
  spend_expanded_formula,
  data = spend_rf_train,
  ntree = 80,
  importance = TRUE
)

spend_scores <- bind_rows(
  score_regression(spend_test$log_payment, predict(spend_baseline, newdata = spend_test)) |> mutate(model = "Participant-scale linear"),
  score_regression(spend_test$log_payment, predict(spend_expanded, newdata = spend_test)) |> mutate(model = "Explanatory linear"),
  score_regression(spend_test$log_payment, predict(spend_rf, newdata = spend_test)) |> mutate(model = "Explanatory random forest")
) |>
  select(model, rmse, mae)

spend_scores |> kable(digits = 3)
model rmse mae
Participant-scale linear 1.355 1.176
Explanatory linear 0.501 0.361
Explanatory random forest 0.366 0.270
Code 
spend_predictions <- spend_test |>
  mutate(
    pred_log_payment = predict(spend_rf, newdata = spend_test),
    pred_payment = expm1(pred_log_payment),
    payment_residual = payment_amount - pred_payment
  )

ggplot(spend_predictions, aes(pred_payment, payment_amount)) +
  geom_abline(linetype = "dashed", colour = "grey50") +
  geom_point(aes(size = participant_count), alpha = 0.35, colour = "#0b6b57") +
  scale_x_continuous(labels = dollar) +
  scale_y_continuous(labels = dollar) +
  scale_size_area(labels = comma) +
  labs(x = "Predicted grouped payments", y = "Observed grouped payments", size = "Participants")

Predicted versus observed grouped quarterly payments in the holdout quarter.
Code 
importance(spend_rf) |>
  as.data.frame() |>
  tibble::rownames_to_column("feature") |>
  arrange(desc(IncNodePurity)) |>
  slice_head(n = 15) |>
  ggplot(aes(reorder(feature, IncNodePurity), IncNodePurity)) +
  geom_col(fill = "#0b6b57") +
  coord_flip() +
  labs(x = NULL, y = "Random forest importance", title = "Most useful predictors for grouped payment spend")

For overall spend, the strongest predictors in the random forest are support class, participant count, disability group, active providers, age group, participants per active provider. Fit is expected to be weaker than a lagged model, but the trade-off is interpretability: the model is forced to lean on observable mix, scale, geography and market variables rather than yesterday’s spend.

The partial dependence curves below are descriptive shape checks. They hold the rest of the training distribution in place and vary one numeric predictor at a time, so they are useful for asking “what does the fitted model tend to do as this factor changes?”, not for claiming causality.

Code 
spend_pdp <- partial_dependence_data(
  spend_rf,
  spend_rf_train,
  features = c("log_participants", "log_providers", "log_participants_per_provider", "median_age", "pct_65_plus", "payment_share_top10"),
  response_transform = expm1,
  x_transforms = list(
    log_participants = expm1,
    log_providers = expm1,
    log_participants_per_provider = expm1,
    pct_65_plus = \(x) x * 100,
    payment_share_top10 = \(x) x * 100
  ),
  feature_labels = list(
    log_participants = "Participants",
    log_providers = "Active providers",
    log_participants_per_provider = "Participants per active provider",
    pct_65_plus = "LGA population aged 65+ (%)",
    payment_share_top10 = "Top 10 provider payment share (%)",
    median_age = "Median LGA age"
  )
)

plot_partial_dependence(spend_pdp, y_label = "Predicted grouped payments", line_colour = "#0b6b57") +
  scale_y_continuous(labels = dollar)

Partial dependence curves for selected grouped payment spend predictors.

Reading the Top Predictors

The top categorical predictors are not just statistical conveniences. support_class tells the model what kind of support is being paid for, disability_group captures broad differences in support needs and service patterns, and age_group captures lifecycle differences in both participant numbers and support intensity. The tables below are latest-quarter grouped payment summaries, so they should be read as model-cell summaries rather than unique participant counts.

For numeric predictors, the partial-dependence chart gives the main shape. Participant count is the scale variable: larger grouped cells have higher predicted payments. Active providers also tends to act as a market-size proxy, while participants per provider is more about local pressure or thinness. The age-structure and provider-concentration curves are weaker, but they help the model distinguish otherwise similar regions and support markets.

Support class

Code 
format_spend_factor_summary(spend_support_summary) |>
  kable()
category total_payments paid_participants payment_per_paid_participant grouped_rows median_cell_payment
Core $33,376,937,000 445,707 $74,885 4,383 $3,661,000
Capital $1,110,695,000 65,554 $16,943 2,363 $353,000

Disability group

Code 
format_spend_factor_summary(spend_disability_summary) |>
  kable()
category total_payments paid_participants payment_per_paid_participant grouped_rows median_cell_payment
Intellectual Disability $9,013,455,000 89,121 $101,137 875 $2,755,000
Autism $5,670,942,000 186,734 $30,369 535 $2,952,000
Psychosocial disability $3,806,680,000 47,455 $80,217 345 $4,756,000
Other Neurological $3,531,125,000 32,773 $107,745 647 $1,911,000
ABI $2,885,983,000 22,634 $127,507 505 $2,748,000
Cerebral Palsy $2,528,100,000 23,565 $107,282 755 $1,239,000
Stroke $1,444,029,000 13,869 $104,119 337 $1,895,000
Other Physical $1,239,948,000 24,950 $49,697 584 $964,000
Down Syndrome $1,056,436,000 8,893 $118,794 312 $2,877,500
Multiple Sclerosis $1,019,682,000 14,974 $68,097 375 $1,350,000

Age group

Code 
format_spend_factor_summary(spend_age_summary) |>
  kable()
category total_payments paid_participants payment_per_paid_participant grouped_rows median_cell_payment
55 to 64 $8,055,101,000 85,518 $94,192 1,337 $1,891,000
45 to 54 $5,436,150,000 57,052 $95,284 1,124 $1,563,000
65+ $5,167,903,000 53,947 $95,796 1,087 $1,701,000
25 to 34 $4,821,438,000 51,027 $94,488 678 $1,742,000
35 to 44 $4,459,941,000 44,441 $100,356 873 $1,730,000
19 to 24 $3,322,372,000 47,501 $69,943 451 $1,614,000
9 to 14 $1,485,000,000 86,229 $17,222 466 $603,500
15 to 18 $1,199,646,000 38,795 $30,923 308 $701,500
0 to 8 $540,081,000 46,751 $11,552 422 $355,000

Useful next explanatory data would include functional-capacity measures, plan duration and review timing, informal-support availability, support hours, provider wait times, workforce depth, plan-management type, local therapy availability, and direct measures of unmet demand. Those would help distinguish genuine support need from access friction and administrative timing.

Caveats

  • These are aggregate public datasets. They do not contain participant-level need, functional capacity, goals, informal supports or local service availability.
  • A plan budget is not the same thing as a payment, and a payment is not the same thing as a good outcome.
  • Disability-group payment trends use public aggregate rows. Where a national disability total is not exposed, support-class rows are summed and treated as a directional spend signal rather than a unique participant count.
  • Utilisation can be low because a participant does not need a support, cannot find a provider, faces administrative barriers, or has timing issues inside a plan. The public data cannot separate all of these explanations.
  • Regional ABS joins are approximate because NDIS service districts and ABS geographies do not always align perfectly.
  • Outcomes are handled in the companion post because the dashboard workbook structure and investment interpretation need their own workflow.

Next Steps

The next useful extension would be to connect NDIS service access with MBS/PBS or hospital-pressure indicators, especially for regions where utilisation, provider concentration and outcomes all point in the same direction.

]]> ndis health-policy modelling australia https://blog.aydins-workbench.com/posts/ndis-budget-utilisation-prediction/ Tue, 26 May 2026 00:00:00 GMT