Cancer trials per cancer case: a 68-fold income gap, but the 50-fold test is inconclusive

The question

Most cancer cases and deaths occur in low- and middle-income countries. Earlier studies report that most cancer trials run in high-income countries, mostly as shares (for example, "75% of trials are high-income-only"). They do not report trials per unit of burden. We pre-registered one question: how steep is the income gradient? We also pre-registered one hypothesis: high-income countries have more than 50 times as many trials per unit of cancer burden as low-income countries.

Method

• Trials. We took all interventional studies on ClinicalTrials.gov that matched "cancer" and started between 2015 and 2024: 48,736 records, equal to the API's total count. A trial counts once for every country where it lists at least one site.
• Burden. We used GLOBOCAN 2022 incident cancer cases, both sexes, for 185 countries (18.7 million cases worldwide), with the World Bank income groups that GLOBOCAN ships.
• Metric. For each income group we summed trials and cases across its countries, keeping countries with zero trials. We did not average country ratios.

$R_g = 100{,}000 \times \frac{\sum_{c \in g} T_c}{\sum_{c \in g} B_c}, \qquad \text{Ratio} = \frac{R_{\text{HIC}}}{R_{\text{LIC}}}$

• Decision rule, fixed in advance. We built a 95% interval by resampling countries within each group 2,000 times. The hypothesis is supported only if the whole interval lies above 50, refuted if it lies wholly below 50, and inconclusive otherwise.

Results

Trial density falls steadily with each step down in income. The high-to-low ratio is 68.0, with a 95% bootstrap interval of 33.9 to 170.8. Because the interval includes 50, the pre-registered verdict is inconclusive. The gradient is clearly very steep, but these data cannot tell whether it exceeds 50-fold.

The low-income figure rests on very few trials. 15 of 27 low-income countries had no registered trial in the window, and five countries account for 52 of the 64 trials: Uganda 21, Malawi 11, Syria 10, Ethiopia 6, Sudan 4. Dropping one low-income country at a time moves the ratio between 59.5 and 93.9.

Sensitivity analyses. 7 of 9 alternative definitions gave a ratio above 50:

• cancer deaths instead of cases: 114.9
• treatment trials only: 104.6
• phase 3 trials only: 90.6, but with just 13 low-income trials
• high-income group without the USA: 64.0
• current World Bank income groups: 60.9
• trials starting 2020–2024: 59.2
• counting site rows instead of trials: 432.9

The two definitions that stop multinational trials from being counted once per country both fell below 50:

• single-country trials only: 38.1
• each trial counted once per income group: 36.0

How multinational trials are counted therefore matters as much as anything else in the analysis.

By cancer site. Trials were matched to sites by keywords in their condition text:

• breast: 82.5
• cervix: 53.5
• liver: 79.0
• stomach: undefined, because no low-income trials were found

Low-income counts for each site are in single or low double digits, so these site figures are fragile.

Caveats

• Registry coverage. ClinicalTrials.gov is a US registry. It under-covers trials registered only in Indian, Chinese, Pan-African or other registries, so the gap is probably overstated by an unknown amount.
• Burden measurement. Many low-income cancer estimates are modelled from neighbouring countries because local registries are sparse.
• Exclusions. Taiwan (1,660 trials) and Hong Kong (574) have no separate GLOBOCAN estimate, and five territories have no income label. All were excluded.
• Ecological description. This compares country-level registry counts only. It says nothing about any individual's access to trials and carries no clinical implication.

What runs next

The next run will repeat the analysis on the WHO's multi-registry trial platform (ICTRP), with deduplication. We will pre-register "each trial counted once per income group" as the primary metric. Pooling low-income countries with their neighbours, or using a count model, could narrow an interval that is currently too wide to decide the 50-fold question.