Transparency of clinical trials: Reproducing and extending analyses

Between 2021 and 2023, deidentified patient-level data and associated programming code from the biologics license application for Pfizer's covid-19 vaccine were published on the Internet. The release, which appears to be unprecedented in permitting public access to such granular data, allows new researchers to independently reproduce published analyses and extend published findings by conducting analyses that the original investigators did not report.

We downloaded 2370 files (26 gigabytes), including more than 30 SAS programming files, 2 data reviewer's guides, and 180 SAS Transport datasets from phmpt.org. We documented programming files and adapted them to run in our environment, aiming to reproduce four published tables and a figure reporting efficacy and safety results. We also sought to estimate vaccine efficacy against covid-19 hospitalization, a clinically meaningful and public health relevant measure of potential protection against serious disease that has not been reported.

We were able to reproduce 334 of 352 data elements from four published efficacy and safety tables. Lacking original code, we could not reproduce one key efficacy analysis (a cumulative incidence curve). For our extended analyses, we judged that vaccine efficacy against covid-19 hospitalization could not be reliably calculated using the available datasets due to a paradoxical (and therefore unreliable) finding of more covid-19 hospitalizations (n = 43) than cases of severe covid-19 (n = 31).

The public release of original datasets and patient-level data enabled the exact reproduction of most published efficacy and safety analyses, establishing a firm basis for further analyses. Our experience highlights the feasibility and enormous value of publicly releasing and analyzing patient-level clinical trial datasets.