Table 2.

Three major sources of systematic bias in RWE studies

Bias	Description	Example
Selection bias	Bias in an effect estimate arises when there are differences in how patients enter or exit the study population in ways that are related to both the exposure and outcome.	1. The study selects initiators of drug X and prevalent users of drug Y as comparison groups. Prevalent users are patients who have “survived” without developing the outcome. This imbalance in susceptibility to develop the outcome will be reflected in the effect estimate. 2. About 40% of patients exposed to drug X and drug Y are lost to follow-up. Twice as many outcomes are missed due to censoring in drug X exposed patients than drug Y.
Information (measurement) bias	Bias in an effect estimate arises when there is incorrect classification of a key study parameter such as exposure or outcome.	The code algorithm used to measure an outcome of myocardial infarction has a PPV of 90%. This means that 10% of patients classified as having the outcome did not actually have the outcome. The PPV is the same across compared exposure groups, thus the estimated effect will be biased toward the null.
Confounding	Bias in an effect estimate arises when risk factors for the outcome are imbalanced between compared groups.	Patients who initiate treatment with drug X are more likely to be over 65, women, and of low socioeconomic status than patients who initiate treatment with drug Y. These 3 factors are risk factors for the outcome, thus the apparent risks of drug X will be inflated unless the imbalance in these baseline confounding factors are addressed in design or analysis.

PPV, positive predictive value; RWE, real-world evidence.