Cognition in children born preterm: meta-analysis models shed light on differences between studies

A study of how to synthesize results

The central question is methodological: do different ways of combining studies lead to different interpretations of the same data? This work does not evaluate a treatment or a strategy intended to improve cognitive development.

A random-effects meta-analysis allows underlying effects to vary between studies. It generally assumes that their distribution is normal. However, this assumption may be unsuitable when the distribution is skewed or contains several distinct groups.

The researchers compared seven models using approaches including skewed distributions, mixture distributions and Dirichlet processes. They also tested variants of these models in sensitivity analyses. The abstract does not specify the total number of children, their ages or their follow-up durations.

A stable average, with differences better described

Estimates of the average effect were similar across models. The abstract also reports identical estimates of between-study variance. However, it does not provide values that would quantify the cognitive difference between the groups or its uncertainty.

The main contribution lies in how the results are described. Models using skewed distributions flagged studies with extreme effects. Dirichlet process models revealed groups of studies with similar characteristics, while mixture models were less informative.

These clusters suggest that certain clinical or methodological characteristics could explain some of the differences between studies. They do not demonstrate that these characteristics cause those differences.

What this analysis changes about interpreting the evidence

In this example, changing the model did not substantially shift the average estimate. But a stable average does not mean that all studies describe the same situation: examining their distribution can help reveal what an overall synthesis conceals.

This result cannot predict cognitive functioning in an individual child. Nor does it demonstrate that the choice of model is unimportant in all meta-analyses. Its scope remains that of a methodological comparison applied to a specific set of studies.

What we can take away

The average result remains similar across the models compared, but some make differences between studies more visible. This offers insight into evidence synthesis, not a new clinical recommendation.