Have some beneficial effects of hydroxychloroquine been overestimated? Potential biases in observational studies of drug effects: Comment on the article by Pons-Estel et al

Authors


Have Some Beneficial Effects of Hydroxychloroquine been Overestimated? Potential Biases in Observational Studies of Drug Effects: Comment on the Article by Pons-Estel et al

To the Editors:

The published medical literature contains scores of observational studies that estimate amazingly beneficial effects for drugs. Recent articles have highlighted a tremendous benefit from exposure to antimalarial drugs, such as an 87% improvement in survival in systemic lupus erythematosus (1, 2). In the setting of an observational study, such surprising beneficial drug effects should be reassessed for the possibility of immortal time bias (3, 4).

As an example, we refer to the observational study recently published in Arthritis Care & Research, in which Pons-Estel et al report that hydroxychloroquine (HCQ) use in patients with lupus nephritis is associated with an astonishing 88% reduction in renal damage (2). This result is even more surprising when we consider that cyclophosphamide did not achieve an effect of that magnitude on renal damage in randomized controlled trials of lupus nephritis (5).

In their Cox proportional hazards model, with renal damage as the dependent variable, Pons-Estel et al classified HCQ exposure as any use during the followup period (i.e., ever/never used) (2). They did not consider when this exposure started. By doing so, any unexposed person-time from cohort entry (T0) to the start of actual exposure in subjects is misclassified as exposed (3, 4). This time period is “immortal,” since the outcome under study could not have occurred during this time, and therefore confers an undue advantage to the exposed group. As a result, the estimated hazard ratio (HR) of the exposure gives the impression that it is much more effective in preventing renal damage than it actually is. When the size of the unexposed group is small compared with the exposed group, the impact of this bias on the estimated HR can be particularly large (3).

To rule out this bias, a time-dependent Cox proportional hazards model, which correctly classifies the immortal person-time preceding exposure as unexposed and the person-time following exposure as exposed, could be used (3). This approach should be considered the standard.

In conclusion, immortal time bias is, unfortunately, a common phenomenon. To avoid this bias, careful consideration of exposures with respect to person-time must be performed. The occurrence of this bias extends to various observational designs, and its presence has been identified in the study of many other drugs (4).

Evelyne Vinet MD*, Sasha Bernatsky MD, PhD*, Samy Suissa PhD†, * McGill University Health Center, Montreal, Quebec, Canada, † McGill University, Montreal, Quebec, Canada.

Ancillary