In the absence of a large number of contributing studies, we may instead prefer a model-based approach in which study membership is treated as a fixed (rather than a random) factor. In fixed-effects IDA the set of available studies is not construed as a random sample of a broader population of studies but instead constitutes the universe of all of the studies of interest. Following Fisher's (1922) guideline to include all measures associated with the sampling framework when fitting models to the data, we include study membership as an explanatory variable in all analyses. For example, in our work we pool data drawn from three studies each of which used a high-risk sampling design to oversample children of alcoholic parents. We controlled for selection criteria used to create our fixed sample of studies by creating two dummy coded variables to capture variance associated with the three contributing studies and these variables are exogenous predictors in all of our fitted models. This accounts for the sampling framework at the level of the pooled analysis. However, we also implement Fisher's recommendation within each individual study. To do so, we control for factors that influence selection into the individual studies, namely parent alcoholism.
