Meta-Regression

Meta-Regression in Evidence Synthesis and Health Technology Assessment

Meta-regression tests whether the treatment effects found in the studies of a meta-analysis rise or fall with a characteristic of those studies, such as the dose given, the mean age of participants, the length of follow-up or the risk of the outcome in the control group. With subgroup analysis, it is one of the two standard ways to look for the causes of between-study heterogeneity, extending the comparison of subgroups of studies to continuous and multiple characteristics. Health economists meet it when a decision model needs a relative effect for one population, because a meta-regression can show whether that effect depends on who was studied. Its findings are associations between studies, so they carry the weaknesses of observational evidence even when every study is a randomised trial. The same regression idea is used in benefit transfer to predict monetary values, a separate application known as meta-regression transfer.

Regressing study effects on study characteristics

The Cochrane Handbook describes meta-regression as similar in essence to simple regression. The outcome is each study's effect estimate, such as a mean difference, a log odds ratio or a log risk ratio, and the explanatory variables are characteristics of the studies that might influence the size of the effect, often called potential effect modifiers or covariates. It differs in two ways, since studies are weighted by the precision of their estimates and the model normally allows for residual heterogeneity that the covariates leave unexplained.

The NICE Decision Support Unit's technical support document 3 (TSD 3) states the aim as relating the size of the treatment effect to characteristics of the included trials in order to explain some or all of the observed between-study heterogeneity. The characteristics may describe the participants, such as mean age, or the trial itself, such as dose, setting or conduct. Categorical characteristics enter as dummy variables, and the same principles apply to dose-response analysis.

Fixed-effect and random-effects meta-regression models

The two versions differ in whether true effects may vary around the regression line. With one covariate, the random-effects (or mixed-effects) model, in the form Viechtbauer sets out for the metafor package, is written as follows.

$$y_i = \beta_0 + \beta_1 x_i + u_i + e_i, \qquad u_i \sim N(0, \tau^2), \quad e_i \sim N(0, v_i)$$

where $y_i$ is the effect estimate from study $i$ (on the log scale for a ratio measure), $x_i$ is the study's value of the covariate, $\beta_0$ is the expected effect when $x_i = 0$, $\beta_1$ is the change in the expected effect per unit increase in $x_i$, $u_i$ is the study's deviation from the regression line, $\tau^2$ is the residual between-study variance not accounted for by the covariate, $e_i$ is sampling error and $v_i$ is the within-study variance, treated as known.

A fixed-effect meta-regression omits $u_i$, so the only random variation it allows is sampling error. Viechtbauer describes its inference as conditional on the studies included, whereas mixed-effects models address a wider population of possible studies. The Cochrane Handbook generally prefers a random-effects model for testing differences between groups of studies, because a fixed-effect comparison carries a high risk of false-positive results. In a TSD 3 example on BCG vaccine trials and distance from the equator, the fixed-effect model fitted very poorly, so only random-effects results were presented.

For a ratio measure the Handbook advises regressing the log of the effect, so that the exponential of $\beta_1$ is the relative change in the effect per unit of the covariate. Viechtbauer notes that treating the estimate of $\tau^2$ as known makes standard errors too small, and that the Knapp and Hartung adjustment, which bases tests on a t distribution with $k - p$ degrees of freedom for $k$ studies and $p$ coefficients, gives error rates closer to nominal. The fall in $\tau^2$ when the covariate is added, divided by $\tau^2$ without it, estimates the share of heterogeneity the covariate accounts for.

How many studies and which covariates

The Cochrane Handbook says meta-regression should generally not be considered when a meta-analysis has fewer than ten studies. It cites typical advice of at least ten studies for each characteristic modelled and warns that even this will be too few when covariate values are unevenly spread across studies. It also asks that characteristics be pre-specified in the protocol, few in number, since false-positive findings become likelier with each one, and each backed by a scientific rationale. Explanations devised after heterogeneity has been seen can at best generate hypotheses.

Choosing covariates means separating prognostic factors, which predict the outcome, from effect modifiers, which change how well the intervention works. The Handbook's example is smoking, a strong predictor of mortality that need not alter a drug's effect on mortality. Study characteristics also travel together, as when a more intensive regimen is tested in sicker patients, and such confounding, like collinearity between covariates in a meta-regression, can make their influences impossible to disentangle.

Ecological bias and aggregation bias in trial-level covariates

In an aggregate-data meta-regression, a patient characteristic such as age enters only as a study mean. The Cochrane Handbook explains that a strong relationship between age and treatment effect within each trial can be invisible across trials with similar mean ages, and that the direction across studies can even be the opposite of the direction within them. This aggregation bias, also called ecological bias or the ecological fallacy, is covered under confounding by aggregation.

TSD 3 adds that with few but precise trials a highly significant relationship between effect and covariate can be entirely spurious. With individual patient data and a continuous covariate, a within-trial comparison avoids ecological bias and has far greater power, because patients' covariate values vary much more than trial means. TSD 3 also describes aggregation bias without any effect modification. When a covariate changes baseline risk, pooling over it biases the treatment effect towards the null in non-linear models, a particular danger in survival analysis.

Treatment effect and baseline risk

A common question for decision models is whether the relative effect varies with the underlying risk of the event. The Cochrane Handbook notes that underlying risk is usually measured as the observed risk in each study's comparator group and summarises known and unknown case-mix factors. Its stroke illustration shows why the scale matters, since a treatment that cuts risk to 80% of baseline removes 10 percentage points from a 50% risk but only 4 from a 20% risk.

Regressing effects on the observed comparator group risk invites regression to the mean, because that risk is part of the effect estimate, and a comparator risk that is high by chance tends to produce a larger effect. The Handbook calls the result a false correlation and recommends correction methods and statistical expertise. TSD 3 instead uses the true baseline, estimated with its uncertainty in a Bayesian model, as the covariate, and warns that naive regression on observed baseline risk ignores this correlation.

Use in NICE appraisals and economic models

The NICE technology appraisal and highly specialised technologies manual (PMG36, section 3.4.10) says that known clinical heterogeneity, for example from patient characteristics, may be explored using subgroup analyses and meta-regression. When event risks differ substantially between the control groups of the studies, it asks whether the relative effect is constant over different baseline risks, especially when the model's comparator-arm event rate is very different from the rates in the meta-analysis.

A fitted meta-regression lets a model use the effect expected for the decision population instead of the trial average. TSD 3 centres the covariate at its mean, so the estimated effect applies at that mean and the effect at any other value is recovered as follows.

$$\theta(z) = d + b\,(z - \bar{x})$$

where $\theta(z)$ is the relative effect (for example a log odds ratio) at covariate value $z$, $d$ is the effect at the mean covariate value $\bar{x}$ and $b$ is the interaction coefficient.

TSD 3 also describes meta-regression as a way to borrow strength from trials in a related population when direct evidence is thin, while adjusting for a systematic difference between the populations. In a network meta-analysis it favours a single interaction term common to all treatments relative to the reference, since separate terms could lead a strict cost-effectiveness analysis to recommend different treatments for different subgroups.

Worked example adjusting a treatment effect to a lower-risk population

All numbers in this example are illustrative. Fourteen placebo-controlled trials of a preventive drug report one-year event rates, with control-group risks between 4% and 35%. A random-effects meta-analysis gives a pooled log odds ratio of -0.40 (odds ratio 0.670) with $\hat{\tau}^2 = 0.09$. A Bayesian random-effects meta-regression on the true baseline log-odds, centred at the trials' mean of $\bar{x} = -1.70$ (a risk of about 15.4%), gives $d = -0.40$, $b = 0.30$ and a residual $\hat{\tau}^2 = 0.03$. The covariate accounts for $(0.09 - 0.03)/0.09 = 67\%$ of the heterogeneity, and the relative effect weakens as baseline risk rises. The decision model's population has a one-year baseline risk of 5%.

Step 1 (baseline on the model scale). $z = \ln(0.05/0.95) = -2.944$.

Step 2 (adjusted relative effect). $\theta(z) = -0.40 + 0.30 \times (-2.944 + 1.70) = -0.40 - 0.373 = -0.773$, an odds ratio of $e^{-0.773} = 0.462$.

Step 3 (absolute effect). Baseline odds are 0.05 / 0.95 = 0.0526. With the adjusted odds ratio, treated odds are 0.0526 × 0.462 = 0.0243 and treated risk is 0.0243 / 1.0243 = 0.0237, so 26.3 events are avoided per 1,000 treated. With the pooled odds ratio, treated odds are 0.0526 × 0.6703 = 0.0353, treated risk is 0.0353 / 1.0353 = 0.0341, and 15.9 events are avoided per 1,000.

Step 4 (cost-effectiveness). The drug costs GBP 300 per person per year, and each event avoided saves GBP 2,000 and gains 0.5 QALYs (one year, no discounting). Per 1,000 treated, the adjusted effect gives a net cost in GBP of $C = 300{,}000 - 26.3 \times 2{,}000 = 247{,}400$ for 26.3 × 0.5 = 13.15 QALYs, an ICER of about GBP 18,800 per QALY. The pooled effect gives $C = 300{,}000 - 15.9 \times 2{,}000 = 268{,}200$ for 7.95 QALYs, about GBP 33,700 per QALY.

Using the effect at the model's own baseline risk lowers the ICER by about 44%, from about GBP 33,700 to about GBP 18,800 per QALY. With fourteen trials and one covariate, the adjustment would normally be a scenario unless the interaction was pre-specified, plausible and precise, and modelling the true baseline protects the slope from regression to the mean.

Meta-regression compared with nearby methods

Subgroup analysis of studies handles one categorical characteristic, while meta-regression adds continuous and multiple covariates within one model. TSD 3 notes that network meta-analysis can be understood as a form of meta-regression but rests on randomised comparisons, so it does not inherit the biases of observational studies.

Meta-regression transfer belongs to benefit transfer. Johnston and colleagues describe benefit-function transfers that estimate a meta-regression across valuation studies and use it to predict values for a new policy site. There the dependent variable is a monetary value and the aim is prediction for an unstudied setting, whereas clinical meta-regression explains variation in relative effects between trials.

Limitations and common misreadings

Geissbühler and colleagues examined 81 meta-analyses published in 2002 and 2012 that included aggregate-data meta-regression. Of these, 57 (70%) had at least one of three pitfalls, with 53 (65%) susceptible to the ecological fallacy, 14 (17%) at risk of overfitting with fewer than five studies per covariate, and 5 (6%) regressing effects on the observed risk of the outcome by inappropriate methods. Only two of the 57 discussed these limitations frankly.

1. A significant slope shows cause. The Cochrane Handbook stresses that participants are randomised within trials, not between them, so a characteristic may only proxy for another.
2. A non-significant slope shows no relationship. Many meta-regressions have low power, and the Handbook urges particular caution before concluding that no dose-response relationship exists.
3. A heterogeneity statistic decides whether to run one. Geissbühler and colleagues argue that the decision should follow clinically important variation seen graphically or indicated by $\tau^2$, not heterogeneity statistics such as Cochran's Q or I-squared.

Sources

• Deeks JJ, Higgins JPT, Altman DG, McKenzie JE, Veroniki AA (editors). Chapter 10: Analysing data and undertaking meta-analyses (last updated November 2024). In: Higgins JPT, Thomas J, Chandler J, Cumpston M, Li T, Page MJ, et al (editors). Cochrane Handbook for Systematic Reviews of Interventions version 6.5. Cochrane; 2024. https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-10
• Dias S, Sutton AJ, Welton NJ, Ades AE. NICE DSU Technical Support Document 3: Heterogeneity: subgroups, meta-regression, bias and bias-adjustment. Decision Support Unit; 2011, last updated April 2012. https://www.sheffield.ac.uk/sites/default/files/2022-02/TSD3-Heterogeneity.final-report.08.05.12.pdf
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). 2022, last updated 13 October 2026. Section 3.4. https://www.nice.org.uk/process/pmg36/chapter/evidence-2
• Viechtbauer W. Conducting meta-analyses in R with the metafor package. Journal of Statistical Software. 2010;36(3):1-48. https://doi.org/10.18637/jss.v036.i03
• Geissbühler M, Hincapié CA, Aghlmandi S, Zwahlen M, Jüni P, da Costa BR. Most published meta-regression analyses based on aggregate data suffer from methodological pitfalls: a meta-epidemiological study. BMC Medical Research Methodology. 2021;21:123. https://doi.org/10.1186/s12874-021-01310-0
• Johnston RJ, Boyle KJ, Loureiro ML, Navrud S, Rolfe J. Guidance to enhance the validity and credibility of environmental benefit transfers. Environmental and Resource Economics. 2021;79:575-624. https://doi.org/10.1007/s10640-021-00574-w