How to Measure the Generalizability of Clinical Trials

Abstract

Randomized controlled trials are widely regarded as the gold standard in clinical research and public health. However, they have been criticized for potentially lacking generalizability, as trial participants may not fully represent the target patient population due to the inability to obtain a truly random sample for enrollment. Assessing and evaluating the generalizability of randomized controlled trials is an important issue that has not been addressed adequately in literature. Additionally, although the importance of describing clinical trial generalizability is recognized by clinical trial reporting guidelines (e.g., CONSORT), it provides no clear guidance on statistical tests or estimation procedures. In this paper, we compare five generalizability indexes, including Standardized Mean Difference, C-Statistic, β-Index, Kolmogorov-Smirnov Distance, and Lévy Distance. We simulate a patient population with a treatment effect size of 0.5 (Cohen's d ) and seven covariates that include gender, health insurance, race, baseline symptoms, comorbidity, age, and motivation. We then evaluate the performance of the five generalizability indexes using selected nonrandom and random clinical trial samples under different number of covariates and sample sizes. Our work supports the use of -index and C-statistic due to their strong statistical performance, ease of interpretation and ability to clearly categorize generalizability into levels such as very high, high, medium or low. A -index value between 1 and 0.8 (inclusive) or a C-statistic value between 0.5 and 0.8 (inclusive) indicates that the trail sample is very highly or highly representative of the patient population.