Combining Information From Multiple Data Sources to Create Multivariable Risk Models: Illustration and Preliminary Assessment of a New Method

摘要

A common practice of metanalysis is combining the results ofnumerous studies on the effects of a risk factor on a diseaseoutcome. If several of these composite relative risks areestimated from the medical literature for a specific disease, theycannot be combined in a multivariate risk model, as is often donein individual studies, because methods are not available toovercome the issues of risk factor colinearity and heterogeneityof the different cohorts. We propose a solution to these problemsfor general linear regression of continuous outcomes using asimple example of combining two independent variables from twosources in estimating a joint outcome. We demonstrate that whenexplicitly modifying the underlying data characteristics(correlation coefficients, standard deviations, and univariatebetas) over a wide range, the predicted outcomes remain reasonableestimates of empirically derived outcomes (gold standard). Thismethod shows the most promise in situations where the primaryinterest is in generating predicted values as when identifying ahigh-risk group of individuals. The resulting partial regressioncoefficients are less robust than the predicted values.