| Premises |
The CD-based mapping method can be used for both aggregate data and individual participant data. This method can be applied to both fixed- and random-effects meta-analysis models. Flexible model-based inference, which includes confidence intervals or regions at all levels for all testing questions (i.e., above and beyond a few isolated point estimates), can be obtained. Data can be combined from exploratory studies (e.g., to generate hypothesis testing questions) as well as from confirmatory studies (e.g., large-scale clinical trials). |
| Assumptions |
A full model is assumed to be shared by all studies. Study-level systematic missing data are assumed to be missing at random. Appropriate mapping matrices can be identified for all studies to validly link the expectation of the study-specific parameters to the vector of hyperparameters of the full model. For studies with missing numerical covariates, their population means (e.g., zero) can be reasonably assumed and accommodated in mapping matrices. |
| Challenges |
A true full model may be difficult to identify or justify, especially when combining data from highly heterogeneous studies. Appropriate mapping matrices can be challenging to identify as the dimensions of data (i.e., covariate by study) increase. When item-level IPD are available, an additional analysis may be needed to retrospectively establish a commensurate metric. When individual studies have small samples and the number of studies in a meta-analysis is small, it may be desirable to accommodate uncertainty surrounding covariance estimators. |
