Meta-Regression and Ecological Bias: Explaining Heterogeneity Without Being Misled

Meta-regression tries to explain why trials in a meta-analysis disagree by relating each trial's effect to a trial-level characteristic such as average age, dose, or baseline risk. It is useful for generating hypotheses, but the relationships it finds are observational and can be distorted by ecological bias, where a pattern seen across trial averages does not hold, or even reverses, inside individual patients. Trust it more when the covariate was prespecified, when there are enough trials, and when the claim stays at the trial level rather than the individual one.

From subgroup analysis to meta-regression

When trials in a meta-analysis disagree, the natural next question is why. Maybe the studies with older patients, or higher doses, or sicker populations, saw bigger effects. Subgroup analysis splits the trials into a few bins and compares them. Meta-regression is the smoother, more formal version: it fits a line relating each trial's estimated effect to a characteristic of that trial.

Done properly, it is a random-effects regression. Each trial is weighted by its own precision and by the residual heterogeneity that the covariate fails to explain. The output tells you how much of the scatter between trials a given factor accounts for, and whether the slope is distinguishable from flat.

Why the association is observational, not causal

Here is the trap that catches careful readers. The trials inside a meta-analysis may each be a randomized controlled trial, so it is tempting to treat the whole synthesis as if it carried the causal authority of randomization. Meta-regression does not.

What gets randomized inside a trial is treatment assignment. What varies across trials is everything else: their populations, settings, and eras. The covariate in a meta-regression, say mean age, was not assigned at random across studies, so a trial with older patients may also differ in dose, follow-up, or outcome definition. Any of those could be the real driver behind an apparent age effect. A meta-regression association is therefore an observational finding sitting on top of randomized parts, and it can be confounded like any observational result.

Ecological bias: the trap of trial-level averages

The deepest pitfall is ecological bias, also called aggregation bias. A meta-regression usually uses a trial-level summary, such as the average age of everyone in the trial. A relationship between those averages and the treatment effect describes trials, not people.

That distinction is not pedantic. A pattern across trial averages can be weaker than, unrelated to, or even opposite to the pattern within individuals. Average age might track with larger effects across trials simply because the older-skewing trials also happened to use a more effective protocol. To learn how age modifies the effect inside a patient, you generally need the patient-level data, which is exactly what an individual participant data meta-analysis recovers by estimating the relationship within each trial before combining.

Too few trials, too many covariates

Meta-regression is also chronically underpowered. A meta-analysis might pool ten or fifteen trials, which is a tiny sample when the unit of analysis is the trial rather than the patient. Standard guidance suggests you need something on the order of ten trials for each covariate you want to examine, and most reviews fall short.

Compounding this is the temptation to test many covariates and report whichever turns up significant. With enough candidate explanations and a handful of trials, chance alone will hand you a convincing-looking slope. This is data dredging, and the defense against it is prespecification: naming, before the analysis, the small number of factors you believe could genuinely modify the effect and why.

What meta-regression can and cannot tell you

Used well, meta-regression is a hypothesis-generating tool. It can flag that dose, or baseline risk, or trial quality might explain some of the disagreement between studies, and it can guide the design of future trials that test those ideas directly.

Used carelessly, it manufactures individualized advice out of trial averages. When you read a claim that a treatment works better in older or higher-risk patients, drawn from a meta-regression, ask three things. Was the covariate prespecified, or fished from many? Were there enough trials to support the analysis? And is the conclusion being stated at the level the data actually live, the trial, rather than being quietly promoted into a statement about individual patients? Baseline risk deserves extra caution, since the outcome sits on both sides of the regression and creates spurious correlation unless special methods are used.