Solved – When to use correlation as effect size when performing meta-analysis with diverse predictors

Question

I am a PhD student and am conducting a literature review on musculoskeletal physiotherapy and am struggling to get guidance on this subject as most health researchers conduct meta-analyses of intervention effects/ effectiveness and not of correlations.

I am looking at the predictors of attendance and adherence to treatment recommendations in musculoskeletal physiotherapy.

• The dependent variable is adherence operationalised into 4 aspects – attendance, in clinic adherence, home exercise adherence and long term adherence.
• The independent variables range from nominal ones like gender, ordinal like socio economic status as well as interval ratio measurements.

Having looked at literature and discussed with a couple of people, I feel that correlation 'r' is the effect size that we would need to use, and for nominal or ordinal data, we can use 2×2 contigency tables, calculate t or chi square and then convert it to 'r'. Then we apply an 'r' to 'z' transformation to combine results and then reconvert it to 'r' for interpretation.

Is this the right approach? What other factors should be considered?

Another issue is that I am taking several different diseases like neck pain, low back pain, ankle injury etc. Is it too heterogeneous to conduct meta analysis even though they have a common symptomatic presentation?

Answer 1

Just a few thoughts:

• If you have multiple predictors then you will probably want to perform multiple meta-analyses; one for each predictor.
• When performing the meta-analysis, you will need to think about what properties of the studies you record. If you record the type of disease and aspect of adherence, then you can perform moderator analysis to see whether the size of predictor-outcome correlations vary by the moderator. This should partially resolve the issue of combining research from multiple diseases. It becomes an empirical question as to whether the size of correlations vary by disease type.
• In general, Pearson's correlation seems like a reasonable approach to aggregating relationships. You should be able to extract information from most reports that describe bivariate relationships that allow calculation of a correlation. Dedicated meta-analysis software generally makes this task easier.
• You might want to think about the effect of measurement and scaling on obtained sizes to correlations (e.g., adherence could presumably be measured in a binary (yes, no) way or in a continuous way (e.g., degree of adherence). Correlations are typically larger when scales are either more nuanced or more reliable.