ABSTRACT
In traditional meta-analysis, the data input for each study consist of an estimate of the effect measure and its standard error. It is assumed that given the true value of the effect in a study, the estimate follows a normal distribution with mean equal to the true value and standard deviation equal to the observed standard error, which is assumed to be fixed and known. In this chapter, the potential disadvantages of this approximate normal within-study model are discussed, and it is shown that in many cases it is possible to replace the approximate normal within-study likelihood by the exact likelihood using group based summary data. The resulting methods are often referred to as one-stage methods. Three outcome types are considered: continuous, dichotomous, and number of events/person years. For continuous outcomes, the pseudo IPD method is presented to construct the exact within-study likelihood. The resulting model is a linear mixed model. For dichotomous outcomes, the exact within-study likelihood is based on binomial distributions, leading to a mixed effects logistic regression model. For number of events/person years data per group, the within-study likelihood is a Poisson likelihood, leading to a mixed effects Poisson regression model. All methods can be implemented using standard (generalized) linear mixed model software.
