ACADEMIC WRITING GUIDE
How to Conduct and Report Subgroup Analysis and Meta-Regression in a Dissertation (With Worked Examples)
Two studies can estimate different intervention effects without either being wrong. The participant populations may differ; the intervention might be delivered with different intensity; the comparator, follow-up period or risk of bias may also differ. When a meta-analysis displays variation across study results, researchers often ask whether a subgroup analysis or meta-regression can explain that variation.
The attraction is understandable. But subgroup analyses are vulnerable to sparse data, multiple testing, differences between studies and post-hoc storytelling. A statistically attractive difference between subgroups is not automatically a causal explanation or a practical recommendation. This article shows how to plan, conduct, interpret and report investigations of heterogeneity in a dissertation without overstating their evidential value.
It draws on Cochrane's meta-analysis guidance, especially section 10.11, Cochrane's review-reporting guidance and PRISMA 2020.
Teaching-example disclaimer: All studies, estimates, standard errors, model coefficients and scenarios below are fictional. The subgroup summaries and meta-regression coefficients in the numerical examples are assumed illustrative software outputs, not model estimates fitted to real or supplied study-level data. The arithmetic demonstrates how to interpret those hypothetical outputs and does not establish any actual treatment effect, interaction or causal moderator.
First, what are subgroup analysis and meta-regression?
A subgroup analysis divides studies or participants into meaningful groups to examine whether the estimated intervention effect differs according to a characteristic. A review might group studies by intervention delivery mode, participant population or follow-up period. However, subgrouping studies according to a trial-level feature is not equivalent to analysing individual participants' responses to a treatment.
Meta-regression models the association between an effect estimate and one or more study-level variables. A continuous characteristic—such as average participant age or intervention duration—may be considered, subject to available evidence and a justified model. Meta-regression can also represent categorical variables. Both methods investigate possible effect modification across studies; neither automatically demonstrates causality.
Sensitivity analysis asks a different question: how robust is the main result to justified alternative decisions or assumptions, such as excluding a study at critical risk of bias or changing a statistical model. Do not rename every sensitivity check as a subgroup analysis. Acadelyra's sensitivity-analysis guide explains that distinction in more detail.
Heterogeneity assessment asks whether, and how much, results vary across studies. Before exploring reasons for variation, establish what is being summarized and whether the studies are sufficiently comparable. Our I², τ² and prediction-interval guide provides this foundation.
| Method | Core question | Typical output | Main limitation |
|---|---|---|---|
| Heterogeneity assessment | How variable are the effects and how uncertain is that variation? | Forest plot, I², τ², prediction interval | Does not identify the underlying cause |
| Study-level subgroup analysis | Do effect estimates differ across prespecified study groups? | Group summaries and formal between-group comparison | Groups can differ in many uncontrolled ways |
| Meta-regression | Is a study-level characteristic associated with estimated effects? | Model coefficient and uncertainty | Sparse studies, ecological bias, confounding |
| Sensitivity analysis | Do results change under justified alternative review decisions? | Main versus alternative estimate | A changed result does not reveal which analysis is true |
What should come before any subgroup comparison?
Clarify the intervention question and the estimand
Define the population, intervention, comparator, outcome, timing and effect scale. A subgroup comparison is hard to interpret if the studies already address incompatible outcomes, use inconsistent effect directions or blend different causal questions. Do the basic synthesis planning first, as described in Cochrane Chapter 9.
Identify plausible effect modifiers in advance
An effect modifier is a characteristic that plausibly changes the effect of an intervention on the specified outcome. The hypothesis should have a substantive basis, not merely follow a visual pattern in the forest plot. In an educational intervention review, delivery intensity could plausibly matter more than the journal's publication year; in a clinical review, disease severity or treatment dosage may be relevant. Whether those hypotheses can be tested depends on the available studies and their measurement.
Specify how each feature will be classified. “Online intervention” may mean live facilitated sessions in one study and fully self-paced content in another. Mixing those designs because a report uses the same label can generate an incoherent subgroup.
Limit the number of questions
When reviewers try many subgroup definitions, effect metrics, outcomes and cut points, interesting results can appear by chance. Prespecify a small set of scientifically credible hypotheses in the review protocol. Distinguish planned analyses from those suggested by observed findings.
Check that enough appropriate studies are available
Subgroup comparisons and meta-regression often have little value when there are very few studies or highly unbalanced groups. A frequently cited minimum of around ten studies per explored characteristic is only a rough warning, not a universal assurance that an analysis with ten studies is sound. More may be needed, especially when precision varies, heterogeneity is large, categories are sparse or multiple predictors are proposed. Sometimes the best analytic decision is not to run a formal subgroup test.
| Proposed modifier | Scientific rationale | Definition to record before analysis | Potential competing explanation |
|---|---|---|---|
| Delivery mode | Facilitated and self-guided interventions may work differently | Prespecified categories based on intervention details | Duration, participant motivation or baseline resources |
| Follow-up timing | Short-term and sustained outcomes differ | Clearly defined outcome windows | Selective outcome availability and attrition |
| Baseline severity | Starting risk may influence absolute or relative benefits | Comparable baseline measure and thresholds | Differences in prognosis or treatment allocation |
| Study design | Methods may produce different estimates | Clearly specified design taxonomy | Confounding and risk of bias rather than true effect modification |
If one variable is essentially a surrogate for another—such as all live courses occurring only in universities and all self-paced courses occurring only in secondary schools—the comparison cannot isolate the delivery mode's contribution.
Subgroup analysis: the crucial difference between within-group and between-group tests
A familiar mistake is to report that the intervention “worked” in one subgroup because that subgroup's estimate was statistically significant, but “did not work” in another because the other P value was above a conventional cutoff. The difference between a statistically significant result and a non-significant result is not itself evidence of a statistically significant difference between the results.
A legitimate comparison asks whether the estimated effects themselves differ, accounting for their uncertainty. Where sufficient suitable studies exist, review authors should use an appropriate formal test for interaction or subgroup differences rather than compare the significance labels of two separate estimates. Cochrane's meta-analysis guidance discusses such methods.
Even a supported between-subgroup difference can be confounded by other study-level differences. An effect modifier is a stronger hypothesis when it was proposed before results were known, has a plausible mechanism, is measured consistently, is supported by an appropriate test, and behaves coherently across related outcomes and settings.
Trial-level versus participant-level subgroups
Suppose some studies enrolled predominantly older adults and others enrolled predominantly younger adults. A between-study comparison using mean age cannot establish how individual age modifies treatment response. Studies can differ simultaneously in location, intervention dose, comparator and methods. This is an ecological-inference problem.
A participant-level subgroup analysis conducted within each randomized trial, followed by appropriate synthesis of the within-trial interaction estimates, is a different design and may offer more credible evidence for participant characteristics. It still needs sound prespecification, compatible data and careful handling of multiplicity. Do not infer individual-level effect modification from a trial-level association.
Meta-regression: what can—and cannot—be learned
In meta-regression, each study contributes an estimated effect and relevant uncertainty, and study-level characteristics may enter a weighted statistical model. A coefficient is an estimated relationship between the covariate and the effect, conditional on the model and available studies. It is not necessarily the change an individual participant would experience if a covariate were altered.
A useful meta-regression requires a defensible effect measure, appropriate weights, sound coding of predictors, explicit treatment of between-study variance, sufficiently many studies and careful uncertainty estimation. Statistical software can produce an apparently precise coefficient even when the data structure does not support a reliable conclusion.
With only a modest evidence base, avoid adding multiple correlated predictors. A model using duration, delivery intensity, setting and participant age may have too little independent information to distinguish their contributions. Study-level covariates can also be measured imprecisely or reported selectively. Missing covariate information should not be silently assigned an arbitrary value.
If a regression includes ratio effects, the appropriate analysis often models log-transformed effects rather than raw ratios, with interpretation translated back carefully. Report the precise model, effect scale, estimation method, study count, covariate coding, uncertainty and software version. Statistical sophistication is not a substitute for substantive credibility.
| Decision | What to document | Why it matters |
|---|---|---|
| Outcome and effect scale | Compatible outcome, direction, transformations | Prevents meaningless comparisons |
| Candidate modifier | Scientific rationale and prespecification | Reduces result-driven hypothesis fishing |
| Predictor definition | Unit, categories, source and missing values | Makes studies comparable |
| Model | Weighting, heterogeneity estimator, uncertainty method | Supports reproducibility and interpretation |
| Evidence size | Number of studies and distribution of covariates | Shows whether a relationship is estimable |
| Bias and confounding | Correlated study characteristics and methodological differences | Prevents causal overreach |
| Exploratory status | Planned versus post-hoc, multiplicity and limitations | Calibrates confidence in findings |
Worked example: delivery mode in a fictional dissertation review
Imagine a systematic review of academic writing-support interventions. The student plans to synthesize effects on a comparable writing-performance outcome. Some eligible studies examine facilitated workshops, while others examine self-paced modules.
Decision A: inspect eligibility and measurement. Are the outcomes genuinely comparable? Do both formats target the same kind of improvement, use similar scoring direction and report compatible follow-up windows? If not, separate syntheses or structured narrative comparison may be preferable to a common subgroup model.
Decision B: define groups from study reports. The student uses the prespecified rule that facilitated sessions must include scheduled interaction with an instructor. A course that merely provides optional help may not meet that definition. Classification decisions are recorded with citations to source reports.
Decision C: examine correlated characteristics. Suppose all workshops occurred at postgraduate institutions, while most self-paced modules were delivered to undergraduates. Then the apparent delivery-mode difference is entangled with learner level and setting. The comparison cannot isolate format as the reason for different estimated effects.
Decision D: assess whether formal comparison is defensible. If the evidence is sparse or unbalanced, the student does not run a subgroup-difference test merely to populate a Results table. Instead, the pattern is described as exploratory, and the limitations are explicit.
Decision E: distinguish interpretation from confirmation. The student may note that effects appear different across delivery formats but cannot conclude that facilitated sessions causally outperform self-paced modules without stronger supporting evidence.
The scenario above is a qualitative decision example. The separate numerical demonstrations below use explicitly fictional, assumed model summaries to show how an appropriate analysis would be interpreted—not to imply that a formal subgroup model could be fitted to the sparse scenario above. For a real dissertation, the Results section must report actual subgroup estimates, intervals and interaction statistics when justified; otherwise document why the analysis was not conducted.
| Question | Recorded decision | Reporting consequence |
|---|---|---|
| Are the outcome scales aligned? | Check scoring and direction before synthesis | Do not pool incompatible measures |
| Are groups defined a priori? | Use the protocol's delivery-mode criteria | Disclose post-hoc reclassification |
| Is evidence sufficient and balanced? | Review study counts and subgroup spread | Do not force an unreliable test |
| Are groups confounded by setting? | Compare learner level and institution types | Avoid attributing association solely to delivery mode |
| Is there a valid interaction test? | Use only when justified by the data | Do not compare two separate P-value labels |
| What is the inference strength? | Classify findings as exploratory or more credible | Avoid causal claims based on trial-level patterns |
Worked numerical example: compare subgroup effects, not separate significance labels
Imagine that, in a separate fictional meta-analysis with adequate studies and a prespecified delivery-mode comparison, software reports the following two independently estimated subgroup summaries on the log risk-ratio scale. This simplified example assumes the subgroups contain non-overlapping studies; actual analyses must account for the fitted model, between-study heterogeneity, dependencies and appropriate uncertainty estimation.
| Prespecified subgroup | Hypothetical pooled log(RR) | Hypothetical SE | Interpretation of direction |
|---|---|---|---|
| Facilitated support | −0.30 | 0.12 | More negative, lower event risk on this effect scale |
| Self-paced support | −0.10 | 0.10 | Less negative, lower event risk on this effect scale |
Step 1 — compare estimates directly. The between-subgroup difference is −0.20 log units: −0.30 − (−0.10). The relevant question is whether that difference is estimated precisely, not whether each subgroup's separate P value crosses a significance threshold.
Step 2 — calculate an illustrative uncertainty interval. Under the simplified independence assumption, the difference has SE = √(0.12² + 0.10²) ≈ 0.1562. The approximate normal 95% interval is −0.20 ± (1.96 × 0.1562), or −0.506 to 0.106 after rounding. Because the interval includes 0 on the log scale, this hypothetical output does not clearly establish a between-subgroup difference. This hand calculation is a pedagogical approximation; a valid meta-analysis must use the interaction or subgroup-difference method appropriate to its model rather than defaulting to this approximation.
Step 3 — interpret without causal overreach. Even if a formal between-subgroup test were persuasive, it would not prove that facilitation itself caused a difference: the two sets of studies could also differ in setting, follow-up, participant characteristics and bias. Report whether the comparison was planned, how groups were defined, how many studies contributed and the uncertainty around the interaction.
A dissertation-ready illustration would read: “The fictional pooled log risk ratios were −0.30 and −0.10 across facilitated and self-paced formats. Their illustrative between-group contrast was −0.20 (approximate 95% CI −0.506 to 0.106). The data do not establish a reliable difference between delivery formats, and residual study-level confounding remains possible.” Replace every number with correctly derived output from the actual review.
Worked numerical example: interpret a meta-regression coefficient
A different fictional review considers whether intervention duration is associated with effect size across studies. Suppose a univariable meta-regression fitted in appropriate software reports a log(RR) intercept of −0.10, a duration slope of −0.025 log(RR) per additional week, and a slope SE of 0.020. These are given model outputs, not estimates we have fitted from a dataset; the following arithmetic shows their interpretation, not the model-fitting process.
The assumed fitted line is:
predicted log(RR) = −0.10 − (0.025 × duration in weeks)
For a duration of 4 weeks, the predicted log(RR) is −0.20: −0.10 − (0.025 × 4). At 8 weeks, it is −0.30: −0.10 − (0.025 × 8). The predicted difference is −0.10 log units over four additional weeks. A more negative log(RR) is associated with a lower event-risk ratio under this hypothetical fitted model, not with an established within-person or causal benefit of extending treatment.
Using a simple normal approximation for illustration, the slope's 95% confidence interval is −0.0642 to 0.0142: −0.025 ± (1.96 × 0.020). It includes zero. Therefore, even taken at face value, these invented model outputs are compatible with no study-level duration association. Actual meta-regression inference may require different small-sample or heterogeneity-aware uncertainty procedures, so the software's correctly specified confidence interval should take precedence over this teaching approximation.
| Quantity | Fictional output or calculation | Responsible interpretation |
|---|---|---|
| Intercept | −0.10 log(RR) at zero weeks | A mathematical model reference; may lie outside observed durations |
| Slope | −0.025 log(RR) per week | Across-study association, not individual causal response |
| Predicted value at 4 weeks | −0.20 log(RR) | Model-based fitted value, not an observed study effect |
| Predicted value at 8 weeks | −0.30 log(RR) | Compare with 4-week prediction cautiously |
| Approximate slope interval | −0.0642 to 0.0142 | Includes zero; no clear association established |
The model could be misleading if longer interventions also occur in older populations, are delivered by different institutions or are studied with different comparators. A dissertation should therefore report study counts, duration range, covariate coding, weighting, variance estimator, missing covariates, model diagnostics and whether the moderator was prespecified. Cochrane's investigation-of-heterogeneity guidance and the trial-level tutorial explain these limitations.
How to write the Methods section
Your Methods should explain what you intended to investigate before inspecting results, how the modifier was defined, which eligible studies contributed, and how subgroup differences or meta-regression were evaluated. If no formal analysis was justified, state that decision rather than imply that an absent table reflects an oversight.
Model Methods paragraph
We identified potential effect modifiers before synthesis based on the review question and the intervention literature. Planned subgroup comparisons considered delivery format using explicit classifications derived from trial reports. We assessed whether outcomes and effect measures were compatible and whether enough studies contributed to each category for an interpretable comparison. Formal between-subgroup comparisons were planned only where the evidence structure supported them. Unplanned patterns were labelled exploratory, and we did not interpret between-study associations as individual-level treatment interactions.
This is model language, not evidence that those methods were actually applied in a particular dissertation. Replace it with the analysis actually performed and disclose protocol deviations.
How to write the Results and Discussion sections
Results should describe the studies in each group, actual estimates and their uncertainty, the between-group interaction assessment when justified, relevant heterogeneity and any missing covariate data. Avoid reporting only the group that appears most favorable. If the subgroup analysis was not feasible, say why.
Model Results paragraph when a test is not justified
The eligible studies varied in delivery format, participant level and follow-up. Although delivery mode was identified as a potential effect modifier, the available studies were too sparse and their characteristics too closely correlated for a reliable subgroup comparison. We therefore did not estimate a formal interaction or meta-regression coefficient. Descriptive patterns are presented without a causal interpretation.
Model Discussion paragraph
Differences in intervention delivery offer one plausible explanation for variation in reported effects, but the evidence does not isolate delivery format from participant background, institutional setting or methodological quality. The available results therefore do not establish a differential treatment effect across formats. Additional primary studies with comparable outcome measures and prespecified within-study interactions would be needed to evaluate that hypothesis more convincingly.
Where valid interactions have been estimated, discuss their magnitude, precision, direction and credibility, rather than relying on a single statistical threshold. Cochrane's reporting chapter recommends specifying potential modifiers, definitions, prespecification and the methods used to compare groups.
What about multiple comparisons and post-hoc findings?
A review may offer many apparent choices for splitting studies: region, intervention duration, baseline severity, setting, outcome scale, year, sample size and design. Trying each split can produce eye-catching results even when no real modifier exists. The more hypotheses explored after inspecting the data, the stronger the need for transparent exploratory labelling and independent confirmation.
Do not conceal analyses that failed to produce a notable result while highlighting only the attractive ones. Keep a decision log of all prespecified and additional comparisons. A protocol amendment may explain legitimate changes, but should not rewrite history by pretending a post-hoc idea was planned from the start. See Acadelyra's systematic-review protocol guide.
If your evidence is unsuitable for meta-regression, a carefully structured narrative synthesis can compare contexts, designs and outcomes transparently without claiming to quantify effect modification.
How subgroup analysis relates to risk of bias and GRADE
Risk-of-bias judgments can inform planned sensitivity analyses and the interpretation of study differences, but a subgroup defined by study quality may also reflect other correlated differences. A finding driven by high-risk-of-bias studies should be described cautiously rather than celebrated as a new effect modifier.
GRADE certainty is generally judged for a specified outcome and comparison; exploratory subgroup findings do not automatically justify a new high-certainty conclusion. If important inconsistency remains unexplained, consider what that means for the certainty of the overall evidence, without a mechanical downgrade based only on a subgroup P value.
Common mistakes to avoid
Comparing two significance labels. “Significant here, not significant there” is not a formal interaction test.
Inventing groups after seeing effects. Data-driven categories can overfit. Disclose unplanned analyses.
Overfitting meta-regression. Too few studies or too many correlated predictors make coefficients unstable and easily misinterpreted.
Treating between-study associations as individual-level effects. Ecological relationships cannot substitute for participant-level interaction analyses.
Using the same studies repeatedly as if independent. Shared participants, multiple outcomes and multi-arm trials require appropriate dependence handling.
Ignoring covariate missingness. Incomplete study descriptions can bias which studies enter a modifier analysis.
Claiming heterogeneity is 'explained' solely because a P value is small. Statistical association and causal explanation are different standards.
Reporting only the attractive subgroup. Readers need all prespecified comparisons, estimates and uncertainty.
Changing the outcome or effect scale for convenience. Review compatibility and measurement first.
Treating a post-hoc finding as confirmatory. Exploration can motivate future research, not automatically confirm a mechanism.
Frequently asked questions
Is subgroup analysis mandatory if I² is high?
No. High observed heterogeneity is a reason to examine context and model assumptions, not an instruction to run unsupported subgroup tests. Some evidence bases are too small or incompatible for reliable effect-modifier analysis.
Can I use a subgroup analysis with only a few studies?
You may describe the available study groups, but formal inference is often unreliable with sparse evidence. A small P value from a poorly supported test does not overcome insufficient data.
What is the difference between a subgroup and a sensitivity analysis?
Subgroups ask whether effects differ across a characteristic; sensitivity analyses ask whether conclusions change when a defensible analytic decision changes. The same study feature may motivate both, but the research questions differ.
Does a significant effect in one subgroup and a nonsignificant effect in another prove an interaction?
No. Compare effect estimates directly using an appropriate interaction or between-subgroup test when feasible.
Can meta-regression prove the intervention works better in older participants?
Not from trial-level average age alone. Differences between study populations can confound that association. Stronger individual-level effect-modification evidence requires appropriate within-study comparisons and compatible data.
How many studies do I need for meta-regression?
There is no guaranteed safe minimum. Around ten studies per characteristic is a common cautionary benchmark, not a sufficient condition. More may be required depending on design, covariate distribution, precision and model complexity.
Should I include an exploratory meta-regression in the abstract?
Only when it is relevant and responsibly qualified. Avoid foregrounding a fragile post-hoc association in a way that implies the analysis was prespecified or causal.
What if the protocol did not mention a modifier I now think is important?
Report the analysis as post-hoc or exploratory, document the rationale and any amendment, and avoid presenting it as a planned confirmatory test.
Final quality checklist
Before submitting, confirm that your review question and effect scale are clear; effect modifiers are scientifically justified; group definitions are reproducible; planned and post-hoc analyses are distinguished; sparse evidence and multiple testing are addressed; subgroup effects are compared directly when appropriate; meta-regression methods and limitations are disclosed; trial-level and individual-level inference are not confused; all eligible results are reported; and uncertainty is integrated into the Discussion.
The standard is not whether your software produced a subgroup coefficient. The standard is whether the question, evidence and method support the interpretation you give that coefficient.
Final takeaway
Subgroup analysis and meta-regression can reveal useful hypotheses about why study effects vary, but their credibility depends on prior rationale, compatible data, suitable study counts, appropriate comparisons and transparent reporting. A dissertation is stronger when it declines an unsupported subgroup test than when it reports a confident-looking result that the evidence cannot sustain.
For help refining your own systematic-review structure, methods narrative or presentation of results, explore Acadelyra's Dissertation & Thesis Support and Research Support services. Your analytical decisions and claims must remain grounded in your actual data and academic requirements.
Authoritative references and further reading
Cochrane. Handbook for Systematic Reviews of Interventions, Chapter III: Reporting the review.
Page, M. J., et al. (2021). The PRISMA 2020 statement: An updated guideline for reporting systematic reviews.
Godolphin, P. J., Tweed, C. D., & Fisher, D. J. (2026). Exploring heterogeneity in meta-analysis using trial-level characteristics: A tutorial. Cochrane Evidence Synthesis and Methods, 4(5), e70099. https://doi.org/10.1002/cesm.70099.