Selection bias in clinical trials.

TLDR

Randomized trials are essential for determining treatment differences, yet selection biases can limit generalizability, making it hard to extrapolate population prognosis from the selected study cohort. In a sarcoma trial, 48 of 90 eligible patients were not randomized (24 by physician choice, 24 refusing), lower‑stage patients were more likely to be randomized (65% vs 37% for higher stage), and nonrandomized patients had poorer disease‑free survival, resulting in divergent efficacy conclusions between randomized and nonrandomized groups even after adjustment.

Abstract

Of 90 patients with intermediate or high-grade sarcoma eligible for a randomized trial of adjuvant doxorubicin (Adriamycin, Adria Laboratories, Columbus, Ohio), 48 were not entered: 24 (27%) by physician's choice and 24 refused randomization. Sixty-five percent of lower stage patients were randomized compared with 37% of those with higher stage (P = .02). Patients with extremity lesions were more frequently offered participation in the study (P = .07). Patients with lower stage lesions accepted randomization more readily than those with higher stage lesions (P = .01). As predicted by the higher stage and percentage of central lesions, the disease-free survival of nonrandomized patients was inferior to that of randomized patients (P = .15). Thus, patients at high risk appeared to avoid randomization and adjuvant doxorubicin in this trial, resulting in an inferior disease-free survival for the nonrandomized control group. Important questions generally require randomized trials that reliably determine relative treatment differences. If, however, the patients in a clinical trial are not representative of the entire patient population because of patient and physician selection biases, the generalizability of the results to the entire patient population may be compromised. For example, the prognosis of the general population cannot necessarily be inferred from the selected group in the study. In this study, the randomized and nonrandomized series yielded differing conclusions regarding treatment efficacy, even when an adjustment was made for known prognostic facts.