Inverse-Variance Weighting in Meta-Analysis: DerSimonian-Laird, REML and Hartung-Knapp

Every pooled effect estimate in a meta-analysis is a weighted average, and inverse-variance weighting is the default rule for setting those weights: a study’s weight is w = 1/v, the inverse of its variance. Precise studies (large samples, tight confidence intervals, small variance) pull the pooled estimate toward themselves; imprecise studies contribute little. This is the same mechanical principle behind most continuous-outcome meta-analyses, and it is what a forest plot’s box sizes are actually showing. This guide covers what changes between a fixed-effect and a random-effects application of that principle, the two competing ways software estimates the between-study variance term a random-effects model needs (DerSimonian-Laird and REML), and the Hartung-Knapp adjustment that corrects the confidence interval those weights produce.

For the broader question of whether to run a fixed-effect or random-effects model at all, see Fixed-Effect vs. Random-Effects Meta-Analysis. This guide assumes that choice is already made and focuses on how the random-effects weights themselves get computed, and why two analysts running the same dataset through different software defaults can get different confidence intervals from identical data.

The inverse-variance principle in a fixed-effect model

Under a fixed-effect model, every included study is assumed to be estimating the same true underlying effect, and any difference between studies is attributed entirely to sampling error. The pooled estimate is the variance-weighted average of the individual study estimates:

θ̂ = Σ(wᵢθᵢ) / Σ(wᵢ), where wᵢ = 1/vᵢ

vᵢ is the squared standard error of study i’s own effect estimate. A study with a standard error half as large gets four times the weight. This is why a single very large trial can dominate a fixed-effect pooled result almost entirely — the model has no mechanism for saying “the trials disagree because they were run in different populations,” it can only attribute disagreement to chance.

Inverse-variance weighting is the general-purpose method, used for continuous outcomes (mean differences, standardized mean differences) and any effect measure with a known standard error. For binary outcomes specifically (odds ratios, risk ratios), the Mantel-Haenszel method is often preferred instead, particularly when event rates are low or the number of studies is small — it behaves better than inverse-variance weighting in those conditions because it does not require large-sample approximations of each study’s variance. The two methods usually converge closely when data are plentiful; they can diverge with sparse events.

Random-effects models add a between-study variance term

A random-effects model relaxes the fixed-effect assumption: it treats each study’s true effect as a draw from a distribution of true effects, not a single fixed value. That distribution’s spread is the between-study variance, denoted τ² (tau-squared). The random-effects weight for study i becomes:

wᵢ* = 1 / (vᵢ + τ²)

Adding τ² to every study’s variance has two consequences worth understanding, not just accepting: it widens the confidence interval around the pooled estimate (extra uncertainty from between-study variation is now represented), and it flattens the weight distribution — large, precise studies still get more weight than small ones, but the gap shrinks, because τ² is added to every study’s denominator regardless of its individual precision. A very large trial no longer dominates the pooled estimate to the same degree it would under a fixed-effect model. For the full mechanics of what heterogeneity means and how τ²/I² relate to prediction intervals, see Heterogeneity in Meta-Analysis.

The random-effects model requires an estimate of τ² before it can compute anything — and unlike the individual-study variances vᵢ (which come directly from each study’s reported standard error), τ² has to be estimated from the pattern of disagreement across the included studies themselves. That estimation step is where DerSimonian-Laird and REML diverge.

DerSimonian-Laird: the classic method-of-moments estimator

DerSimonian and Laird’s 1986 method estimates τ² directly from Cochran’s Q statistic (the standard test for heterogeneity) in a single closed-form calculation — no iteration required:

τ̂²_DL = max(0, (Q − (k − 1)) / C)

k is the number of studies and C is a scaling constant built from the fixed-effect weights. Because it resolves in one step, DL was for decades the default (and, in many packages, the only) option for estimating τ² in random-effects meta-analysis software, including Cochrane’s own RevMan.

DL’s practical weakness is well documented: it tends to underestimate τ² when the number of studies is small or when study sizes vary substantially, which in turn produces confidence intervals for the pooled effect that are narrower than they should be. In a specialist field like evidence synthesis, where reviews commonly pool a handful of trials rather than dozens, this is not a corner case — it is the typical situation the estimator is weakest in.

REML: now the default in RevMan

Restricted maximum likelihood (REML) estimates τ² iteratively, and — unlike ordinary maximum likelihood, which treats the pooled mean effect as known when estimating τ² and so tends to underestimate it — REML accounts for the degrees of freedom used up in estimating that mean, giving a less biased estimate of the between-study variance. It generally performs better than DL across small-to-moderate numbers of studies and extends naturally to meta-regression models with covariates (see Meta-Regression in Meta-Analysis).

This is not just a textbook preference. The current Cochrane Handbook (Chapter 10, §10.10.4.4) documents a real, recent software change: as of 2024, RevMan added REML as an available τ² estimator alongside the original DerSimonian-Laird method, and REML is now RevMan’s default option, with DerSimonian-Laird retained as an alternative. Other software has generally defaulted to REML for longer — R’s widely used metafor package and Stata’s meta suite both use REML by default — so this brings RevMan in line with where the rest of the evidence-synthesis toolchain already was. A review’s methods section citing “REML” for τ² estimation is describing current standard practice, not an unusual choice; one citing plain DerSimonian-Laird without qualification may be running an older analysis or an explicit, defensible preference for the classic method (large k, replicating an earlier review’s exact numbers) — it is not automatically wrong, but it is worth noticing.

Software commonly offers other τ² estimators beyond these two — Paule-Mandel, Sidik-Jonkman, and empirical Bayes variants among them — each with its own bias/variance tradeoffs under different study-count and heterogeneity conditions. REML and DerSimonian-Laird remain the two an editorial reviewer or research-support office is most likely to encounter in a submitted manuscript, which is why they are the focus here.

Why the confidence interval needs its own fix: the Hartung-Knapp adjustment

Estimating τ² from data, rather than knowing its true value, introduces an extra layer of uncertainty that the standard random-effects calculation does not account for. The conventional approach treats τ² as if it were known exactly once estimated, then computes the confidence interval around the pooled effect using a standard normal (z) distribution. When the number of studies is small, this understates real uncertainty and can produce confidence intervals that are too narrow — increasing the chance of a false-positive conclusion.

The Hartung-Knapp-Sidik-Jonkman (HKSJ, commonly shortened to Hartung-Knapp or HK) adjustment is not an alternative τ² estimator — it is a correction applied after pooling, to the standard error and confidence interval of the summary effect. It rescales the standard error using the variability of individual study estimates around the pooled effect, and replaces the z-distribution with a t-distribution on k − 1 degrees of freedom, which is wider and better reflects the genuine uncertainty when few studies are available.

The current Cochrane Handbook is explicit about when to apply it (§10.10.4.4): RevMan prompts reviewers to use the HKSJ method whenever the estimated between-study variance is greater than zero and more than two studies are included, specifically to avoid two failure modes — confidence intervals that are too narrow when τ² is estimated at exactly zero, and confidence intervals that are unstable when only two studies are pooled. The Handbook (§10.10.4.5) separately cautions that with only two or three studies, HKSJ’s own performance is limited, and recommends a sensitivity analysis (comparing the HKSJ interval against the standard approach) rather than treating either as automatically authoritative in that narrow-k range.

One caveat worth knowing before reading a Hartung-Knapp interval at face value: when study sizes are very unequal and the pooled result is dominated by one large study, an unadjusted HKSJ calculation can occasionally produce a confidence interval narrower than the standard method — the opposite of its intended effect. Most current software implements a modified/truncated version of HKSJ that prevents this by never returning an interval narrower than the standard Wald-type interval would give.

Reading it on a forest plot

A forest plot’s box sizes are a direct visualization of the weights discussed above. Two things worth checking when reading one, or when reviewing a manuscript that includes one: whether the model is fixed-effect or random-effects (stated in the plot or its caption, and confirmable by whether box sizes look proportional to raw sample size, which is more characteristic of a fixed-effect model, versus more evenly sized, which is more typical under substantial heterogeneity in a random-effects model), and whether the reported confidence interval on the summary diamond used a standard or Hartung-Knapp calculation — the methods section, not the plot itself, is where that is usually stated.

Practical guidance by study count

• Large sets of studies (roughly 30 or more): DerSimonian-Laird and REML tend to converge; either is defensible, and DL’s simplicity/reproducibility can be a reasonable choice for k this large.
• Small to moderate sets (roughly under 30): REML is the better-performing default and matches current software defaults across RevMan, metafor, and Stata’s meta suite.
• Any random-effects analysis with more than two studies and τ² estimated above zero: apply the Hartung-Knapp adjustment to the confidence interval, per current Cochrane Handbook guidance.
• Exactly two or three studies: report a sensitivity analysis comparing standard and Hartung-Knapp intervals rather than presenting either alone as definitive — per Cochrane Handbook §10.10.4.5.

Frequently asked questions

What is inverse-variance weighting in meta-analysis?

It is the standard method for combining effect estimates across studies by weighting each one inversely to its variance — more precise (larger, lower-variance) studies contribute more to the pooled result. Under a random-effects model, the between-study variance (τ²) is added to every study’s variance before weighting, which narrows the gap between large and small studies’ influence compared with a fixed-effect model.

Is REML always better than DerSimonian-Laird?

REML generally estimates τ² with less bias, especially with fewer studies, which is why it is now the default in RevMan, metafor, and Stata’s meta suite. DerSimonian-Laird remains a reasonable, simpler choice for very large numbers of studies, or when reproducing an earlier analysis that specifically used it.

Should Hartung-Knapp always be used instead of the standard confidence interval?

Current Cochrane Handbook guidance recommends it whenever τ² is estimated above zero and more than two studies are pooled, because it better reflects the extra uncertainty in an estimated τ². With only two or three studies, its own performance is limited enough that a sensitivity analysis comparing both approaches is recommended rather than relying on either alone.

Which pooling method should be used for binary outcomes like odds ratios?

Mantel-Haenszel weighting is commonly preferred over inverse-variance weighting for binary outcomes, particularly with low event rates or few studies, because it avoids some of the large-sample approximations inverse-variance weighting relies on. The two methods typically converge when event data are plentiful.

Related CASRAI coverage

This guide sits alongside CASRAI’s other evidence-synthesis method pages: Fixed-Effect vs. Random-Effects Meta-Analysis for the underlying model choice, Heterogeneity in Meta-Analysis for I², τ², and prediction intervals, Meta-Regression in Meta-Analysis for extending these same τ²-estimation methods to models with covariates, Standardized Mean Difference (SMD) and Hedges’ g in Meta-Analysis for the effect-size measure these weights are typically applied to, and the Cochrane Handbook chapter guide for a map of where each of these topics sits in the Handbook itself. See also the dictionary entries for meta-analysis, systematic review, network meta-analysis, the Cochrane Handbook, effect size, and funnel plot.