Dread returns to Mega-Silly One

Click to increase image sizeClick to decrease image size

This article refers to:

Everything should be as simple as possible, but no simpler: towards a protocol for accumulating evidence regarding the active content of health behaviour change interventions

Notes

1. As an aside, the term ‘meta-regression’ is redundant and only serves to confuse. Meta-analysis is always a linear model. In its basic (random effects) form, this model is characterised as:

In which the effect (y) in the ith study is predicted from the average population effect (β₀) and some error: Var(u_i) being the between-study variance (usually denoted as τ²) and Var(e_i) the sampling variance within the ith study. This model can be extended to include fixed effect predictors (a mixed model) by extending the linear model to include one or more predictors (e.g., x_i) and attaching a parameter to them that quantifies their relationship to study effect sizes (e.g., β₁):

As we shall see, you can similarly characterise meta-analysis as a multilevel model in which effect sizes (level 1) are nested within studies (level 2). In as much as ‘regression’ is synonymous with ‘linear model’, the term ‘meta-analysis’ does not need to be distinguished from ‘meta-regression’; all that needs to be declared is what predictor variables (if any) have been added to the model.

2. Lesser known cousin of Dr Spock.

3. There may also be a case for treating something other than the study as the unit of analysis (Cooper, 2010) – for example, the intervention techniques themselves.