Safety of hydroxychloroquine, alone and in combination with azithromycin, in light of rapid wide-spread use for COVID-19: a multinational, network cohort and self-controlled case series study

Jennifer C.E Lane; James Weaver; Kristin Kostka; Talita Duarte-Salles; Maria Tereza F. Abrahao; Heba Alghoul; Osaid Alser; Thamir M Alshammari; Patricia Biedermann; Edward Burn; Paula Casajust; Mitch Conover; Aedin C. Culhane; Alexander Davydov; Scott L. DuVall; Dmitry Dymshyts; Sergio Fernández Bertolín; Kristina Fišter; Jill Hardin; Laura Hester; George Hripcsak; Seamus Kent; Sajan Khosla; Spyros Kolovos; Christophe G. Lambert; Johan ver der Lei; Ajit A. Londhe; Kristine E. Lynch; Rupa Makadia; Andrea V. Margulis; Michael E. Matheny; Paras Mehta; Daniel R. Morales; Henry Morgan-Stewart; Mees Mosseveld; Danielle Newby; Fredrik Nyberg; Anna Ostropolets; Rae Woong Park; Albert Prats-Uribe; Gowtham A. Rao; Christian Reich; Jenna Reps; Peter Rijnbeek; Selva Muthu Kumaran Sathappan; Martijn Schuemie; Sarah Seager; Anthony Sena; Azza Shoaibi; Matthew Spotnitz; Marc A. Suchard; Joel Swerdel; Carmen Olga Torre; David Vizcaya; Haini Wen; Marcel de Wilde; Seng Chan You; Lin Zhang; Oleg Zhuk; Patrick Ryan; Daniel Prieto-Alhambra

is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10.1101/2020.04.08.20054551 doi: medRxiv preprint further minimise potential unresolved confounding with calibrated HRs (CalHRs) and 95% confidence intervals estimated. 44, 45 For SCCS, safety of HCQ therapy was assessed separately as a secondary analysis, regardless of indication, comparing exposed and unexposed time periods within the same individuals. The method is self-controlled in that it makes within-person comparisons of event rates during periods of hypothesized increase risk with other periods of baseline risk, with eliminates all time-invariant confounding. Because we do not compare between persons, the SCCS is robust to between-person differences, even including unmeasured differences (like genetics). However, the method is vulnerable to time-varying confounders: the time of exposure may be incomparable to the time when not exposed. To adjust for this, we included many time-varying co-variates in the models, including age, season, and other drug exposures. The effects of age and season were assumed constant within each calendar month and were modelled using bicubic splines with 5 knots. A conditional Poisson regression was used to fit the outcome model using the Cyclops package, with a hyperparameter selected through 10-fold crossvalidation. 46 Study diagnostics (power, propensity score distribution, covariate balance, empirical null distribution) were evaluated by clinicians and epidemiologists to determine which database-target-comparatoroutcome-analysis variants could produce unbiased estimates. Database-target-comparator-analysis variants with zero event outcomes in the time-at-risk window or contained analyses with baseline covariate with standardized mean difference>0.1 after stratification were excluded from analysis. Study diagnostics for all database-target-comparator-outcome-analysis will be provided as part of study, regardless of which effect estimation results are unblinded. All the proposed analyses were conducted for each database separately, with estimates combined in fixed effects meta-analysis methods where I2 . CC-BY-NC 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.