Plan and run a meta-analysis

A diagnostic test is one test or scale that detects a condition the person has now. A prediction model combines several factors to predict a condition or a future event.

Differences in people, settings or methods don’t automatically rule out pooling. If you accept them, MetaLab uses random effects and records them.

This goes in your methods output, so you can discuss it with your supervisor or statistician.

Answer No if two rows share a cohort or a control group, or come from several models in one dataset.

Decide from how the studies were done, not from their results. When populations or settings differ, the answer is usually Yes.

List the adjustment variables in the CSV. Adjusted estimates from different models may still not estimate the same effect.

For example, depression. Different ways of measuring it, such as different questionnaires, can go under the differences answer above. Different outcomes, such as depression and anxiety, should be analysed separately.

Use one time point from each study. Different time points across studies can be pooled, for example 3, 5 and 7 years.

For example, every row is treatment vs control, and a higher score always means worse symptoms.

Hedges’ g is an SMD with a small-sample correction; Cohen’s d has none.

Usually 95%. Every row must use the same level.

Adding 0.5 lets inverse-variance and random-effects models run, but can bias results when events are rare. Mantel–Haenszel (common-effect) needs no correction for studies with a zero in one group. Studies with no events (or only events) in both groups cannot be analysed here.