Random versus fixed effects: choosing rather than defaulting

Random versus fixed effects: choosing rather than defaulting — setting out what I have, and where I think it stops being reliable.

Session topic: FLOW (N Engl J Med, 2024). Please read it before posting; the discussion is much better when everyone has.

The question I would like us to start with is what the trial set out to estimate, rather than what it found. Once that is on the table we can talk about whether the design could have answered it, and only then about the numbers.

Specific things I would like covered: the population and how far it generalises, how discontinuation was handled, whether the comparator was a fair one, and what the absolute rather than relative effect looks like.

I will summarise at the end and the summary will feed the relevant digest page.

Picking up the opening post: that is the part I would want checked first.

Study quality and weighting: some meta-analyses weight all studies equally; others weight by study size or study quality. The choice affects the result and should be stated and justified.

Fixed-effects versus random-effects models: fixed-effects assumes all studies are estimating the same thing and variation is sampling error. Random-effects assumes studies are estimating effects from different distributions and allows between-study variance. Choice matters if heterogeneity is high.

Inclusion and exclusion criteria: a meta-analysis is only as good as its inclusion criteria. If the criteria are too broad, apples and oranges get pooled. If they are too narrow, the meta-analysis answers a overly specific question.

Sensitivity analysis: the authors re-run the meta-analysis excluding studies one at a time, or by quality, to see whether the pooled estimate changes. Robust results stay similar even when individual studies are excluded.

This follows post #3 rather than contradicting it.

Number needed to treat from a meta-analysis: can be computed from the pooled estimate if the baseline risk is specified. More interpretable than pooled relative effects.

Worth separating two things that post #3 runs together.

Subgroup analysis: sometimes a meta-analysis reports separate pooled estimates for different subgroups (e.g., by baseline body mass index or by trial duration). Be cautious — many subgroup analyses are exploratory and less reliable than the main analysis.

For anyone arriving from a search: the marked solution above is the direct answer, and the replies underneath it add the caveats that make it safe to use.

When a meta-analysis is unhelpful: if the included studies are heterogeneous in population, intervention, or outcome, pooling them produces a number that represents nothing in particular. Reading the individual studies is more useful than reading the pooled estimate.

Pooled estimates and heterogeneity: when trials differ in population, duration, or comparator, a pooled estimate answers a question that no individual trial asked. High heterogeneity means effects genuinely differ across studies. The pooled number is an average of things that should not have been averaged.

Picking up post #8: that is the part I would want checked first.

I disagree with the reply above, and I think the disagreement is substantive rather than terminological.

The distinction being drawn does not survive when you look at the published data for this specific question. I would be glad to be shown wrong on this, because the version I am arguing against is more convenient.

Publication bias: what did not get published? Small studies with negative results are less likely to be published than large studies with positive results. A forest plot with only large studies on the positive end is a red flag for unpublished small negative studies.

Why forest plots are more informative than pooled numbers: they show the variation across studies, which tells you whether the effect is consistent or heterogeneous. A narrow confidence interval around a meaningless centre is less useful than a wider interval that shows real differences.

Why forest plots are more informative than pooled numbers: they show the variation across studies, which tells you whether the effect is consistent or heterogeneous. A narrow confidence interval around a meaningless centre is less useful than a wider interval that shows real differences.