Who pays? Three ways to fund a health system

TidyTuesday 2026-04-21 · WHO Global Health Expenditure Database, 195 countries

NoteSession 3 · autonomously developed · Claude Opus 5

Dataset choice, analytical angle, figures and prose are Claude Opus 5’s, produced working autonomously with no human steering during the session. How the sessions differ.

Every country’s health bill is paid from exactly three pockets. Government raises it through taxes and social insurance. Private spending comes out of households and employers, most of it at the point of care. External spending comes from abroad — aid, development banks, global health funds.

The WHO tracks all three for 195 countries from 2000 to 2023, and — usefully — as shares that sum to exactly 100. How much a country spends is largely a story about how rich it is. Which pocket it comes from is a story about politics, and it is the number that decides whether a hospital visit bankrupts you.

TipReproducing this page

Everything runs from data/health_spending.csv, committed in this repository. To refresh:

u <- paste0("https://raw.githubusercontent.com/rfordatascience/tidytuesday/",
            "main/data/2026/2026-04-21/health_spending.csv")
download.file(u, "data/health_spending.csv")

Packages: tidyverse only. Source: WHO Global Health Expenditure Database, via TidyTuesday 2026-04-21.

library(tidyverse)

theme_set(theme_minimal(base_size = 13))
pal_src <- c("government" = "#1a7fa3", "private" = "#c75146", "external" = "#8a6fb5")

raw <- read_csv("data/health_spending.csv", show_col_types = FALSE)

shares <- raw |>
  filter(str_detect(indicator_code, "_che$")) |>
  select(country = country_name, iso3 = iso3_code, year, indicator_code, value) |>
  pivot_wider(names_from = indicator_code, values_from = value) |>
  rename(government = gghed_che, private = pvtd_che, external = ext_che)

totals <- raw |>
  filter(indicator_code == "che_usd2023") |>
  select(country = country_name, year, che = value)

spend <- shares |>
  left_join(totals, by = c("country", "year")) |>
  filter(!is.na(government), !is.na(private), !is.na(external))

n_countries <- n_distinct(spend$country)
yr_range    <- range(spend$year)

183 countries, 2000 to 2023.

Where the world’s health money comes from

global <- spend |>
  filter(!is.na(che)) |>
  summarise(across(c(government, private, external),
                   \(x) weighted.mean(x, che)), .by = year) |>
  pivot_longer(-year, names_to = "source", values_to = "share") |>
  mutate(source = factor(source, levels = c("external", "private", "government")))

ggplot(global, aes(year, share / 100, fill = source)) +
  geom_area(colour = "white", linewidth = 0.3) +
  annotate("segment", x = 2020, xend = 2020, y = 1.02, yend = 0.98,
           colour = "grey25", arrow = arrow(length = unit(0.14, "cm"))) +
  annotate("text", x = 2020, y = 1.06, label = "2020", size = 3.6, colour = "grey25") +
  scale_fill_manual(values = pal_src, name = NULL,
                    breaks = c("government", "private", "external")) +
  scale_y_continuous(labels = scales::percent, expand = expansion(mult = c(0, 0.10))) +
  labs(
    title = "Governments pay for six dollars in ten of the world's health care",
    subtitle = "Share of global health expenditure by source, weighted by spending",
    x = NULL, y = "share of global health expenditure"
  ) +
  theme(legend.position = "top",
        plot.title = element_text(face = "bold"),
        panel.grid.minor = element_blank())

Composition of global health expenditure, weighting each country by what it actually spends.

Weighted by what countries actually spend, the government share of world health financing rose from 50.4% in 2000 to 57.5% on the eve of the pandemic, spiked to 60.8% in 2020, and has drifted back to 58.7%. External funding is invisible at this scale — 0.3% of global health spending — because the countries that depend on it do not spend much in dollar terms.

That last clause should make you suspicious of the whole chart.

The same question, asked the other way

A spending-weighted average tells you where the world’s money comes from. It is dominated by a handful of large, rich health systems: the United States alone is a substantial fraction of the total. It says almost nothing about the financing arrangements a randomly chosen country actually has.

both <- bind_rows(
  spend |> filter(!is.na(che)) |>
    summarise(share = weighted.mean(government, che), .by = year) |>
    mutate(basis = "weighted by spending  (“where the world's money comes from”)"),
  spend |>
    summarise(share = mean(government), .by = year) |>
    mutate(basis = "each country counted once  (“what the typical country does”)")
)

ggplot(both, aes(year, share / 100, colour = basis)) +
  geom_line(linewidth = 1.1) +
  scale_colour_manual(
    values = c("weighted by spending  (“where the world's money comes from”)" = "#1a7fa3",
               "each country counted once  (“what the typical country does”)"  = "#c75146"),
    name = NULL) +
  scale_y_continuous(labels = scales::percent, limits = c(0.4, 0.7)) +
  labs(
    title = "The rise of public health financing is a big-spender story",
    subtitle = "Government share of health expenditure, two ways of averaging",
    x = NULL, y = "government share"
  ) +
  theme(legend.position = "top",
        legend.text = element_text(size = 9.5),
        plot.title = element_text(face = "bold"),
        panel.grid.minor = element_blank())

Government share of health spending, computed two ways: weighting countries by their spending, and treating each country equally.

Two lines, one underlying question, and they disagree about the headline.

Weighted by spending, the government share climbed 8.4 points over the period — a real and substantial shift toward public financing. Counting each country once, it moved 3.2 points, from 47.3% to 50.5%. Essentially flat.

Both numbers are correct. They answer different questions, and the difference between them is the finding: over twenty-three years, health financing became more public in the places that already spend the most, and barely changed in the median country. A reader shown only the first line would conclude the world had moved decisively toward public health financing. A reader shown only the second would conclude nothing had happened. Neither should be shown alone.

The third pocket

External financing is a rounding error globally and the dominant fact of life in a few dozen countries.

aid_now <- spend |> filter(year == 2023, external >= 25) |> select(country, external)
aid_then <- spend |> filter(year == 2000) |> select(country, external_2000 = external)

aid <- aid_now |>
  left_join(aid_then, by = "country") |>
  mutate(country = fct_reorder(country, external))

n25 <- spend |> summarise(n = sum(external > 25), .by = year)

ggplot(aid, aes(y = country)) +
  geom_segment(aes(x = external_2000, xend = external, yend = country),
               colour = "grey78", linewidth = 1.3, lineend = "round") +
  geom_point(aes(x = external_2000), colour = "grey55", size = 2.6) +
  geom_point(aes(x = external), colour = pal_src[["external"]], size = 3) +
  scale_x_continuous(labels = \(x) paste0(x, "%")) +
  labs(
    title = "Where health care is funded from abroad",
    subtitle = paste0("External share, 2000 (grey) to 2023 (purple).\n",
                      "Countries funding 25% or more from abroad in 2023."),
    x = "share of health spending from external sources", y = NULL
  ) +
  theme(panel.grid.major.y = element_blank(),
        plot.title = element_text(face = "bold"))

Countries where external sources funded at least a quarter of health spending in 2023, with their 2000 position.

The number of countries funding a quarter or more of their health system from abroad went from 10 in 2000 to 31 in 2010, and stands at 29 today. The grey-to-purple movement on that chart is almost entirely rightward: for 24 of the 26 countries shown, external dependence is higher now than at the start of the century. Only Rwanda and Micronesia have moved the other way. (3 countries — Somalia, South Sudan, and Zimbabwe — have no grey dot at all, having no 2000 observation; South Sudan did not exist. A missing grey dot means no measurement, not a starting value of zero.)

In Micronesia (Federated States of), 71% of all health spending is externally financed. Seven countries are above half. These are health systems whose funding is decided, to a first approximation, somewhere else — which makes them acutely exposed to decisions taken in donor capitals for reasons that have nothing to do with them.

What the pandemic did

covid <- spend |>
  filter(year %in% c(2019, 2020)) |>
  select(country, year, government) |>
  pivot_wider(names_from = year, values_from = government, names_prefix = "y") |>
  filter(!is.na(y2019), !is.na(y2020)) |>
  mutate(change = y2020 - y2019)

med <- median(covid$change)

ggplot(covid, aes(change)) +
  geom_vline(xintercept = 0, colour = "grey45", linewidth = 0.5) +
  geom_histogram(binwidth = 1, fill = pal_src[["government"]], colour = "white",
                 linewidth = 0.3) +
  geom_vline(xintercept = med, colour = "#c75146", linewidth = 1) +
  annotate("text", x = med + 0.8, y = Inf, vjust = 1.8, hjust = 0, size = 3.9,
           colour = "#c75146", fontface = "bold",
           label = sprintf("median %+.1f points", med)) +
  labs(
    title = sprintf("In 2020, %.0f%% of countries shifted health spending onto the state",
                    100 * mean(covid$change > 0)),
    subtitle = "Change in government share of health expenditure, 2019 to 2020",
    x = "percentage-point change in government share", y = "countries"
  ) +
  theme(plot.title = element_text(face = "bold"),
        panel.grid.minor = element_blank())

Change in the government share of health spending between 2019 and 2020, across all countries with both years recorded.

70% of countries increased the government share of health spending in 2020, by a median of 1.9 percentage points. This is the clearest thing in the dataset: a global shock, an almost universal response in the same direction, visible in nearly every country at once.

It is also the clearest illustration of what these shares do and don’t mean. A rising government share can mean the state spent more. It can equally mean households spent less — and in 2020, households spent dramatically less, because elective care stopped and people stayed away from clinics. A share is a ratio, and a ratio moves when either part moves. Separating those two stories needs the constant-dollar series, and even then needs a view about what would have happened otherwise.

NoteWhat this can and cannot say

These are financing sources, not out-of-pocket costs. “Private” (PVT-D) bundles household out-of-pocket payments together with private insurance and employer schemes. A country with a 40% private share where all of it is prepaid insurance is a completely different place to live than one where all of it is cash at the counter. This dataset, as published here, cannot tell those two apart, and no claim on this page rests on the private share alone.

No population, so no per-capita anything. The file carries total expenditure in constant 2023 US$ but no population, so every level comparison would be a comparison of country size. That is why this page is about composition throughout: shares are comparable across countries in a way that totals are not.

Shares are ratios and move for two reasons. The COVID section above is the sharpest case, but it applies everywhere: a rising external share can mean more aid or a collapsing domestic budget. Directional readings of share changes should be treated as questions, not answers.

Health accounts are estimates, and unevenly good ones. WHO harmonises national health accounts of very different quality; external financing in particular is often better recorded by donors than by recipients. The countries with the largest external shares are generally those with the weakest statistical systems, which is precisely where the numbers should be trusted least.

2023 is provisional and coverage is incomplete. 183 of 195 countries have all three shares in at least one year, but coverage varies year to year; the country composition of the unweighted average is therefore not perfectly constant over time, which adds a small amount of noise to the flatter of the two lines above.