Methodological progress note: Descriptive, predictive, and causal observational studies

Observational data can be used to answer descriptive, predictive, and causal questions. Analysis of observational data can be challenging and may require complex analytical methods to produce meaningful, unbiased results; for example, when addressing confounding in causal studies. In recent years, there has been a growing appreciation that these different questions should be answered with different analytical approaches. This Methodological Progress Note introduces the key differences between these types of research questions and analytical approaches. In descriptive research, additional variables may be incorporated, for example, via stratification or statistical adjustment, to further characterize the outcome distribution with respect to those variables. In predictor research, the researcher may adjust for established predictors to assess the extent to which the candidate predictor may be a predictor of the outcome, over and above the existing predictors. In prediction modeling, a set of variables is selected based on their collective ability to most accurately predict the outcome within all populations and settings for which the model will be used. In causal research, additional variables are included to account for confounding bias. Examples relating to central venous access device failure are provided.