Association vs. causality in transfusion medicine: understanding multivariable analysis in prediction vs. etiologic research

Abstract

In the current medical literature, etiologic and prediction research aims are frequently confused. Investigators tend to use principles from prediction research for their etiologic research questions, which results in misleading interpretation of risk factor findings at hand. We used a questionnaire-based survey to quantify the proportion of International Society of Blood Transfusion (ISBT) 2012, Cancun, visitors who felt confident with a causal interpretation of a stepwise logistic regression model. We designed and distributed a short online questionnaire survey addressing questions about a constructed abstract entitled "Association of transfusion and clinical outcomes in a large cohort" among the participants of ISBT 2012, Cancun. In addition to asking questions about the demographics (age, sex, country of employment, and highest education level) of the participants, we designed 7 statements representing possible interpretations of the findings presented in the abstract and asked the participants to mark Agree, Disagree, or Do Not Know for each statement. Based on the responses to these statements, we quantified the proportion of participants who inferred causality from stepwise multivariable models built to examine a question of association (or prediction).Thirty percent to 40% of the respondents agreed that a stepwise model was a valid method to adjust for confounding, and 60% of them agreed to a causal interpretation of a model built for prediction purposes. These findings suggest that a large proportion of ISBT visitors confuse etiology with prediction in the published transfusion medicine research. Using the results as a platform, we aim to delineate the distinction between etiologic and prediction research, issues of confounding accompanying these research aims and how a multivariable model deals with confounding.