Meta-Analysis

Refers to the statistical integration, or synthesis, of data from more than one study for the purpose of analysis. The findings from any one study become the unit of analysis within the meta-analysis framework. The studies will usually have been conducted by different researchers at different points in time and in different locations, but will have enough methodological similarity to warrant integration. Meta-analysis has been used in a wide range of scholarly areas, including education research, and is gaining increasing importance in epidemiology and related clinical fields.

By combining data from myriad investigations, a meta-analysis may result in increased precision of estimates and a greater confidence in hypothesis testing. This is particularly important when single studies have utilized relatively small sample sizes, and as such would be hindered by low statistical power and a high likelihood of type II errors. Meta-analysis usually involves the comparison of odds (and odds ratios) or risk (and relative risk ratios). In clinical research, meta-analysis has been a very powerful technique for addressing issues of interpretation when the results of studies contradict one another; by combining the data in one analysis, more robust and conclusive interpretations may be generated.

The studies included in a meta-analysis need to have similar measures and methodologies. As such, meta-analysis has been particularly useful when analyzing the results of randomized clinical trials, rather than quasi-experimental or observational designs. Additionally, problems associated with Simpson's paradox and publication bias may limit the usefulness of meta-analysis. Simpson's paradox refers to a situation wherein a statistical association observed in separate analyses may be lost when the studies are combined. This concern is particularly relevant for meta-analyses that combine data from studies with very large differences in sample sizes. Publication bias may be a problem for meta-analysis because it represents a selection bias. That is, studies that find a statistically significant association are more likely to be published than studies that find no significant differences in outcome; as such, the source studies for a data analysis will be filtered through the publication bias. This may result in false-positive findings. However, the inclusion of non-published studies in a meta-analysis also brings to light potential problems, given that they have not passed a peer-review process.

The Cochrane Database of Systematic Reviews is a well-known effort that prepares, disseminates, and periodically updates meta-analyses of randomized [Page 1118]controlled trials on health interventions. Contemporary textbooks in biostatistics typically include discussion of techniques associated with meta-analysis, including discussion of techniques for standardizing outcome measures, conducting sensitivity analysis, and presenting results (e.g., box-plots).