Toolkit · 06 of 13

Subgroup analysis

Pre-specified subgroups and tests for subgroup differences.

What it is

A subgroup analysis splits the studies by a characteristic fixed at the study level, such as design, type of procedure or length of follow-up, pools each group separately, and asks whether the effect differs between them.

The answer comes from the test for subgroup differences, not from comparing P values: a significant effect in one subgroup and a non-significant one in the other does not mean that the effect differs.

Subgroups are made of studies, not of randomized patients, so the comparison is observational: a difference may come from anything else those studies share. A few pre-specified subgroups, each with a clinical rationale, give credible answers; many post hoc ones mostly produce chance findings.

Worked exampleCompensatory sweating, by study design

	Studies	Odds ratio (95% CI)	I²
Observational	7	0.34 (0.16-0.73)	78%
RCT	4	0.42 (0.19-0.94)	0%

Both subgroups favor T2-sparingsimilar effects in RCTs and observational studiesTest for subgroup differences: P = 0.73no evidence that design changes the effect

The heterogeneity sits within the observational studies, not between the two designs.

When to use it

Use it when

The subgroups were pre-specified in the protocol, each with a clinical reason to expect a different effect.
The characteristic is categorical and defined at the study level: design, type of procedure, technique, follow-up period.
Each subgroup has at least two studies, and preferably the analysis has 10 or more studies overall.
There is heterogeneity you want to explain, or a clinical question about one group of patients or procedures.

Choose another tool when

The characteristic is continuous, such as mean age or dose: use meta-regression rather than cutting it into categories.
You want to know whether a single study drives the result: use sensitivity analysis.
The characteristic varies between patients within each study, such as sex or age: study-level subgroups cannot answer it; it needs individual patient data or trials reporting results by subgroup.

The template, parameter by parameter

Click a highlighted argument to see what it does and which values to use. Picking a value updates the code.

# m: the model from the binary or continuous tool
m_sg <- update(m,
               subgroup = design,
                = FALSE,
                = TRUE)

# forest plot: see section 4

The forest plot, parameter by parameter

These arguments change only how the result is shown, not the result itself. Click one to see the options.

png("subgroups.png",
    width = 5500, height = 1400 + 200 * m_sg$k, res = 500)
forest(m_sg,
        = studlab,
        = TRUE,
        = FALSE,
        = "grey",
        = TRUE,
        = TRUE,
        = "navyblue",  = "black",
       label.left = "Favors intervention",
       label.right = "Favors control")
dev.off()

How to read the results

Test for subgroup differences

This is the result to report: Q between subgroups, its degrees of freedom and P value.
The test has low power, so P < 0.10 is often taken as evidence of a difference.
Never compare the P values of the subgroups with each other.

Within the subgroups

Compare the subgroup estimates and their intervals with the overall result.
Check whether I² falls within the subgroups: if it does, the characteristic explains part of the heterogeneity.
A subgroup with one or two studies gives an unstable estimate.

Credibility

Was the subgroup pre-specified, with the direction of the difference predicted?
Is the difference consistent across related outcomes, and is the number of subgroup analyses small?
Could another study characteristic explain it? Tools such as ICEMAN formalize these questions.

In my work

Subgroup analyses from two published reviews, next to the call behind them. The first figure was redrawn from the published study data and reproduces the reported results. Click a figure to enlarge it.

Compensatory sweating, by study designOdds ratio, 11 studies (4 RCTs, 7 observational); test for subgroup differences P = 0.73. From J Cardiothorac Surg, 2026 ↗

m_sg <- update(m,                   # main model
               subgroup = design,   # RCT or observational
               tau.common = FALSE)  # tau² per subgroup

forest(m_sg,
       col.subgroup = "grey",       # subtotals in grey
       print.subgroup.name = FALSE, # heading "RCT" only
       sort.subgroup = TRUE)        # alphabetical

Need for transfusion, by type of surgeryRisk ratio, 12 trials. Neither subgroup is significant on its own, yet the overall effect is, and the test for subgroup differences is P = 0.64: the effect does not differ by type of surgery. From JAMA Surg, 2026 ↗

m_sg <- update(m,                   # main model
               subgroup = type)     # type of surgery

forest(m_sg,
       col.subgroup = "grey",       # subtotals in grey
       print.subgroup.name = FALSE) # heading only

References

Higgins JPT, Thomas J, Chandler J, et al. Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Cochrane; 2024. Chapter 10.11 (investigating heterogeneity).
Borenstein M, Higgins JPT. Meta-analysis and subgroups. Prev Sci. 2013;14:134-143.
Richardson M, Garner P, Donegan S. Interpretation of subgroup analyses in systematic reviews: a tutorial. Clin Epidemiol Glob Health. 2019;7:192-198.
Sun X, Ioannidis JPA, Agoritsas T, Alba AC, Guyatt G. How to use a subgroup analysis: users' guide to the medical literature. JAMA. 2014;311:405-411.
Schandelmaier S, Briel M, Varadhan R, et al. Development of the Instrument to assess the Credibility of Effect Modification Analyses (ICEMAN) in randomized controlled trials and meta-analyses. CMAJ. 2020;192:E901-E906.
Thompson SG, Higgins JPT. How should meta-regression analyses be undertaken and interpreted? Stat Med. 2002;21:1559-1573.