Fixed effect or random effects in a quitting review: read the assumption

Start with the analysis that supports the headline

Find the specific comparison and its stated model in the methods or results. A review can contain several analyses with different outcomes and study sets. The model label attached to one result should not be borrowed for another simply because they appear in the same paper.

Cochrane explains common-effect assumptions and random-effects summaries that allow a distribution of underlying effects. This introductory article helps a reader retain that distinction. It does not choose a statistical model for a new analysis, judge a named review, or establish a treatment effect. Statistical appraisal requires more than identifying a phrase in the methods.

Cochrane: Cochrane Handbook, Chapter 10: Analysing data and undertaking meta-analyses

Keep allocation and pooling separate

A randomized trial allocates an intervention using a random process. A random-effects meta-analysis is a way of combining study estimates. It can include different underlying study designs when the review permits them. The word “random” in the model name does not repair nonrandom assignment or establish that every included study was a trial.

For a hypothetical reading example, a review combines trials of several support formats and reports a random-effects result. That label does not show that every format had the same effect, or that a new tracking app reproduces their intervention. You still need the comparison map: participants, support, comparator, outcome, and follow-up.

Understand why the summary may change

Different pooling assumptions can change study weights and uncertainty. A random-effects average is not a promise that the average effect will occur in every setting. A fixed-effect result should not be called more trustworthy simply because its interval is narrower. Precision must be interpreted with the assumptions that produced it.

If a report presents both models, check why it did so and whether the interpretation addresses any material difference. Do not select whichever estimate looks more favorable. The question is whether the conclusion depends on a modeling decision and whether the studies can answer a meaningful common question, not which label sounds sophisticated.

Write a bounded interpretation

A useful note states: “This analysis used this model to combine these study results for this outcome; its summary does not remove these setting differences.” Keep important unexplained variation visible. A statistical method cannot make smoking and vaping outcomes identical or turn an untested product into the studied intervention.

If the assumptions remain unclear, record a question for the review authors or a qualified professional rather than recreating the analysis from a screenshot. You can still make practical preparations while the evidence question is open. The reading task ends when the model label no longer functions as a quality badge and you can explain what remains uncertain about the pooled claim.