Subgroup Analysis

Subgroup Analysis

Subgroup analysis estimates and compares an association or treatment effect across predefined or exploratory categories of participants. In health economics it can reveal decision-relevant variation in baseline risk, benefit, harm, cost or cost-effectiveness, but smaller groups and multiple comparisons make attractive-looking differences unreliable without careful design. This page explains what a subgroup comparison asks, how an interaction is calculated, and how uncertain heterogeneity should affect a decision.

Specify subgroups and the estimand

Define a subgroup using characteristics known before treatment whenever the aim is to compare treatment effects, such as baseline risk, age category or biomarker status. Prespecify the rationale, boundaries, primary outcome, effect scale and statistical comparison before viewing results when feasible. A subgroup defined by a post-treatment response may reflect the treatment itself and need different causal methods; it should not be treated as an ordinary baseline modifier.

Separate prognostic factors, which predict outcome regardless of treatment, from effect modifiers, which change the treatment-versus-comparator effect on a stated scale. A modifier on the absolute risk-difference scale need not be a modifier on the risk-ratio scale. Report whether groups are mutually exclusive and exhaustive, and whether results apply to the whole population or only a clinically defined stratum.

Work through a risk-difference interaction

Suppose a fictional randomized trial reports a 12-month event risk of 20% with standard care and 10% with a new intervention in a higher-risk subgroup. In a lower-risk subgroup the corresponding risks are 8% and 6%. Define the effect as intervention minus standard care: $0.10-0.20=-0.10$ in the higher-risk group and $0.06-0.08=-0.02$ in the lower-risk group. The difference between those two risk differences is $-0.10-(-0.02)=-0.08$, an eight-percentage-point interaction contrast on the additive scale.

In a spreadsheet, =10%-20% and =6%-8% produce the two subgroup risk differences; =(10%-20%)-(6%-8%) gives -8%, displayed as minus eight percentage points. The risk ratios are $0.10/0.20=0.50$ and $0.06/0.08=0.75$, which give a different kind of contrast. These are illustrative risks, not observed sample data; without sample sizes, variances and study design details, one cannot compute a valid confidence interval or p-value for the interaction. Nor does this arithmetic establish that treatment truly works differently between groups.

Judge whether heterogeneity is credible

A subgroup pattern is stronger when it follows a prespecified biological or decision rationale, is measured reliably at baseline, uses an appropriate interaction test, has sufficient precision, and is consistent across related outcomes and credible studies. Randomization protects treatment comparisons within baseline-defined groups on average, but subgroup estimates are often imprecise. A significant treatment result within only one group is not itself a significant difference between groups.

Searching many cut points, outcomes and categories creates multiplicity: some striking differences will occur by chance. Selective reporting compounds the problem. Display all prespecified comparisons, estimates and uncertainty intervals, distinguish exploratory from confirmatory findings, and avoid treating a p-value as the size or importance of an effect. An interaction test can be underpowered, so an inconclusive test does not prove effects are identical. Validation in independent evidence may be important before restricting access.

For meta-analysis, distinguish within-trial subgroup comparisons from across-study comparisons. A comparison of trials that enrolled mostly older patients with trials that enrolled mostly younger patients can reflect differences in setting or treatment, an ecological problem, rather than patient-level effect modification. Cochrane provides specific guidance on tests for subgroup differences and meta-regression; the chosen model and number of studies matter.

Use subgroup findings in economic decisions

Economic value can vary even without a true relative treatment-effect interaction. If two groups share the same relative risk reduction but have different baseline risks, their absolute avoided events, QALYs and downstream costs may differ. Intervention delivery costs, uptake and adverse effects can also vary. Estimate incremental costs and health effects for each policy-relevant subgroup with appropriately aligned inputs and time horizons, while retaining uncertainty and shared parameter dependencies.

For a decision rule, define the population eligible for each option and compare subgroup-specific expected net benefit at a stated threshold or opportunity cost. Do not choose a subgroup solely because its observed ICER is below a cutoff in a small sample. Restricting treatment may create implementation costs, misclassification and equity concerns, particularly when subgroup status is hard to observe or access to testing differs. If evidence does not credibly support different policies, show the uncertainty rather than claiming precision from post hoc splits.

Report the analysis so it can be checked

State the subgroup definition and measurement timing; whether it was prespecified; the number of participants and events in each treatment-by-subgroup cell; the outcome, effect scale and estimates; interaction result with uncertainty; and the full set of comparisons considered. Explain missing data, overlap between groups, adjustment, and any sensitivity or external validation. For economic analyses also show costs, health effects, subgroup sizes and how an eligibility rule would operate in practice.

Sources and further reading

• Cochrane Handbook, Chapter 10, on subgroup comparisons and meta-regression in evidence synthesis: https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-10
• SPIRIT–CONSORT, reporting standards for randomized trials, including planned and exploratory analyses: https://www.consort-spirit.org/
• Espinoza and colleagues, “The value of heterogeneity for cost-effectiveness subgroup analysis” (2014), on decision use of heterogeneity: https://pmc.ncbi.nlm.nih.gov/articles/PMC4232328/
• NICE, Health technology evaluations: the manual, for economic evaluation in defined decision populations: https://www.nice.org.uk/process/pmg36/chapter/economic-evaluation-2