engleski

Developing prediction models using a small number of datasets with overlapping variables

Using multiple data sources to develop clinical prediction models increases sample size and precision. However, when some datasets include only a part of the relevant predictors, a common regression analysis cannot be applied unless one part of the data is discarded. To overcome this issue, a recent study proposed to estimate a regression coefficient from a model with all relevant predictors (fully adjusted estimate, available from at least one dataset) using the correlations and conditional independencies between fully and partially adjusted estimates. To validate the proposed method for the prediction of risk of allergic diseases in Croatian population using 4 datasets, we consider the problem of developing a prediction model when the number of datasets is too small to estimate these correlations reliably. The proposed method, modified to include plausible correlation values in advance, was compared to the complete-case estimator in a simulation study. Although both approaches showed similarly low bias, the mean squared error of the complete-case estimator was larger. These results suggest that the proposed method may be better suited for population prediction models even when the number of datasets is small.

prediction models ; meta-analysis ; unmeasured confounding

nije evidentirano