Hydroxychloroquine for prevention of COVID-19 mortality: a population-based cohort study

Background. Hydroxychloroquine has been shown to inhibit severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) in vitro, but early clinical studies found no benefit treating patients with coronavirus disease 2019 (COVID-19). We set out to evaluate the effectiveness of hydroxychloroquine for prevention, as opposed to treatment, of COVID-19 mortality. Methods. We pre-specified and conducted an observational, population-based cohort study using national primary care data and linked death registrations in the OpenSAFELY platform, representing 40% of the general population in England. We used Cox regression to estimate the association between ongoing routine hydroxychloroquine use prior to the COVID-19 outbreak in England and risk of COVID-19 mortality among people with rheumatoid arthritis (RA) or systemic lupus erythematosus (SLE). Model adjustment was informed by a directed acyclic graph. Findings. Of 194,637 patients with RA or SLE, 30,569 (15.7%) received [≥]2 prescriptions of hydroxychloroquine in the six months prior to 1 March 2020. Between 1 March 2020 and 13 July 2020, there were 547 COVID-19 deaths, 70 among hydroxychloroquine users. Estimated standardised cumulative COVID-19 mortality was 0.23% (95% CI 0.18-0.29) among users and 0.22% (95% CI 0.20-0.25) among non-users; an absolute difference of 0.008% (95% CI -0.051-0.066). After accounting for age, sex, ethnicity, use of other immunuosuppressives, and geographic region, no association with COVID-19 mortality was observed (HR 1.03, 95% CI 0.80-1.33). We found no evidence of interactions with age or other immunosuppressives. Quantitative bias analyses indicated observed associations were robust to missing information regarding additional biologic treatments for rheumatological disease. We observed similar associations with the negative control outcome of non-COVID-19 mortality. Interpretation. We found no evidence of a difference in COVID-19 mortality among patients who received hydroxychloroquine for treatment of rheumatological disease prior to the COVID-19 outbreak in England.


Abstract
Background. Hydroxychloroquine has been shown to inhibit severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) in vitro, but early clinical studies found no benefit treating patients with coronavirus disease 2019 . We set out to evaluate the effectiveness of hydroxychloroquine for prevention, as opposed to treatment, of COVID-19 mortality.
Methods. We pre-specified and conducted an observational, population-based cohort study using national primary care data and linked death registrations in the OpenSAFELY platform, representing 40% of the general population in England. We used Cox regression to estimate the association between ongoing routine hydroxychloroquine use prior to the COVID- 19  region, no association with COVID-19 mortality was observed (HR 1.03, 95% CI 0.80-1.33). We found no evidence of interactions with age or other immunosuppressives. Quantitative bias analyses indicated observed associations were robust to missing information regarding additional biologic treatments for rheumatological disease. We observed similar associations with the negative control outcome of non-COVID-19 mortality.

Conclusion.
We found no evidence of a difference in COVID-19 mortality among patients who received hydroxychloroquine for treatment of rheumatological disease prior to the COVID-19 outbreak in England.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
Evaluations of the effectiveness of hydroxychloroquine for prevention, as opposed to treatment, of SARS-CoV-2 infection or severe COVID-19 outcomes are limited. 13 One randomised, controlled trial examining hydroxychloroquine as post-exposure prophylaxis did not demonstrate significant benefit in preventing SARS-CoV-2 infection, though uncertainty in results could not exclude possible benefit. 14 Other trials for the prevention of COVID-19 outcomes are ongoing. 5 In this large population-based cohort study, we examined whether ongoing routine hydroxychloroquine use prior to the outbreak in England was associated with lower risk of COVID-19 mortality.

Study design and population
We conducted an observational cohort study using electronic health record (EHR) data from primary care practices using The Phoenix Partnership (TPP) software linked to Office for National Statistics (ONS) death registrations through OpenSAFELY. This is a data analytics platform developed during the COVID-19 pandemic to allow near real-time analysis of pseudonymised primary care patient records at scale, covering approximately 40% of the population in England, operating within the EHR vendor's highly secure data centre. 15,16 Pseudonymised structured data include demographics, medications prescribed from primary care, diagnoses, and laboratory measures. Details on Information Governance of the OpenSAFELY platform can be found in the Supplementary

Appendix.
We included all adults aged ≥ 18 years registered with a general practice for ≥ 1 year on 1 March 2020 (index date) with information on age, sex, and deprivation. Within this source population, we identified people who were diagnosed with RA or SLE ≥ 6 months prior to the index date and therefore had indication for hydroxychloroquine use prior to the outbreak in England. 17 We studied people with these conditions to minimise the potential for confounding by indication when estimating the effectiveness of hydroxychloroquine use rather than investigate how to prevent severe COVID-19 in this population.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted September 9, 2020. . https://doi.org/10.1101/2020.09.04.20187781 doi: medRxiv preprint

Exposure, outcome and follow-up
The exposure of interest was regular use of hydroxychloroquine (≥2 prescriptions in the 6 months prior to index, "users") compared to no regular use of hydroxychloroquine ("non-users"). The primary outcome was COVID-19 mortality, defined by the presence of ICD-10 codes U07.1 ("COVID-19, virus identified") or U07.2 ("COVID-19, virus not identified") on the death certificate. 18 We followed people from index date until earliest of: date of death, or seven days prior to last date of availability of ONS mortality data to account for reporting lag. People not exposed to hydroxychloroquine before index date were censored if prescribed it during follow-up. Study design is depicted in eFigure 1 in Supplementary Appendix.

Covariates
Potential determinants of regular hydroxychloroquine use and COVID-19 mortality were identified by reviewing existing literature and through discussions with clinicians. As this was a study of prevalent hydroxychloroquine users, we included determinants that may have influenced the initial choice of treatment and whether people remained on treatment. The full list of pre-specified variables included age, sex, ethnicity, index of multiple deprivation quintile, other immunosuppressives (other sDMARDs, oral corticosteroids), smoking status, prescribed non-steroidal anti-inflammatory drugs (NSAIDs), body mass index, hypertension, diabetes severity as measured by diagnostic codes and glycated haemoglobin (HbA1c), heart disease, liver disease, respiratory disease excluding asthma, kidney disease as measured by estimated glomerular filtration rate (eGFR), stroke, dementia, cancer, and influenza vaccination in 2019/20 season. Our methodology for creating codelists has been previously described. 15 We developed a directed acyclic graph (DAG) to identify the minimal set of covariates to adjust for the hypothesised confounding structure, which included age, sex, ethnicity, geographic region, and other immunosuppressives (see eFigure 2 in Supplementary Appendix).

Statistical methods
Patient characteristics were summarised using descriptive statistics, stratified by exposure status.
We used Cox regression models with days since the index date as the timescale to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between regular hydroxychloroquine use and COVID-19 mortality. The competing risk of death from causes other than COVID-19 was accounted for by censoring non COVID-19 deaths; our analysis therefore estimated cause-specific hazards. 19 We sequentially adjusted for sex and age using restricted cubic splines; for the minimal adjustment set informed by the DAG; and finally extended for all extracted covariates listed above. Models were stratified by an indicator variable denoting patient population (i.e., RA or SLE) and geographic region. Multiple imputation (10 imputations) was used to account . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted September 9, 2020. . https://doi.org/10.1101/2020.09.04.20187781 doi: medRxiv preprint for missing ethnicity for 23% of the sample, with the imputation model including all extracted covariates and an indicator for the outcome. Those with missing body mass index were assumed to be non-obese, and those with missing smoking data were assumed to be never-smokers; we did not use multiple imputation for these variables as they are expected to be missing not at random in UK primary care. 20 Proportional hazards were checked by examining the Schoenfeld residuals over time.
We generated cumulative mortality curves standardised to adjust for different covariate distributions in the exposed group. First, a flexible parametric Royston-Parmar model with the same covariates as the DAG-Informed Cox model was fitted, with the baseline hazard modelled using a three degrees-of-freedom spline. The survival function was predicted from this model for every individual with regular hydroxychloroquine use and averaged to produce the curve for the exposed group. To produce the standardised comparison curve, the survival functions were predicted and averaged again for the same individuals, but with exposure set to zero. Patient population was included in the flexible parametric model as a binary indicator variable since the model could not converge with both patient population and geographic region as stratification variables. Comparisons between Cox and Royston-Parmar models can be found in the Supplementary Appendix.
We evaluated pre-specified interactions to determine whether the association between regular hydroxychloroquine use and COVID-19 mortality varied by age, exposure to other sDMARDs, oral corticosteroids, and NSAIDs. Two-sided p-values were calculated from Wald tests on interaction terms.

Sensitivity analyses
We adjusted for ethnicity in a model excluding people with missing ethnicity and compared results with those from multiple imputation. In primary analyses, <2% of the comparison group had one prescription of hydroxychloroquine in the six-month exposure window. We re-defined exposure as ≥ 1 prescription in the three months prior to the index date. We compared results from primary models stratified for patient population (RA or SLE) to a model that included patient population as a binary indicator variable as well as modelling each population separately.
We calculated bias-adjusted hazard ratios to evaluate how adjustment for biologic DMARDs (bDMARDs), including targeted synthetic DMARDs (tsDMARDs), which were not available for this analysis, may have produced different results under differing assumptions of prevalence and effect on COVID-19 mortality. 21 Prevalence of bDMARDs in each exposure group was estimated from preexisting literature (18% among users and 21% among non-users); 22,23 however, we also examined . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted September 9, 2020. . https://doi.org/10.1101/2020.09.04.20187781 doi: medRxiv preprint more extreme values of prevalence (see Supplementary Appendix). We assumed a range of potential associations between bDMARDs and COVID-19 mortality from 0.8 to 1.2.
We conducted analyses using non COVID-19 mortality as a negative control outcome, censoring at COVID-19 death. We hypothesized that If associations observed in primary analyses were due to confounding by indication, we would observe a similar association with non COVID-19 mortality.

Software and reproducibility
Data management was performed using Python 3.8 and SQL, and analysis using Stata 16.1. All code for data management and analyses is at: https://github.com/opensafely/hydroxychloroquineresearch. All iterations of the pre-specified protocol are archived with version control at: https://github.com/opensafely/hydroxychloroquine-research/tree/master/protocol.

Patient characteristics
We identified 194,637 people with RA or SLE at least six months prior to 1 March 2020 (i.e., index date) for analysis (Figure 1).
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted September 9, 2020. . https://doi.org/10.1101/2020.09.04.20187781 doi: medRxiv preprint Of these, 30,569 (15.7%) demonstrated regular use of hydroxychloroquine in the six months prior to index date. Hydroxychloroquine users were younger (median: 63 years for users, 66 years for nonusers) and more likely to be female (76% women for users; 70% for non-users); other demographic characteristics between exposure groups were broadly similar ( Table 1). Hydroxychloroquine users were more likely to be on other sDMARDs (52% vs 34%), oral corticosteroids (23% vs 16%), and NSAIDs (22% vs 16%). Distributions of characteristics in RA and SLE populations are shown in eTable 1 and eTable 2, respectively. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted September 9, 2020.  is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
There was no evidence of interaction by age, exposure to other sDMARDs, oral corticosteroids, or NSAIDs (eTable 3).

Sensitivity analyses
Results from all sensitivity analyses provided similar findings to primary analyses (eTable 4). In quantitative bias analyses, values of the bias-adjusted association ranged from HR 0.97 (95% CI 0.75-1.26) to HR 1.09 (95% CI 0.84-1.40) (eTable 5). Hydroxychloroquine use was not associated with the negative control outcome of non COVID-19 mortality after adjustment for age and sex (HR 1.00, 95% CI 0.87-1.15, Figure 3). I . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Key findings
In this national, population-based study of hydroxychloroquine users, we found no evidence that hydroxychloroquine was associated with either a beneficial or harmful impact on COVID-19 mortality. The confidence intervals around the relative risk suggest that we could exclude substantial benefit, though a modest benefit or harm on a relative scale could not be ruled out. However, even if hydroxychloroquine provided a benefit, our results demonstrate a maximum absolute risk reduction of 0.05% in the context of an absolute risk of 0.22% of COVID-19 mortality among non-users. Taken together, our findings do not provide any strong support for a protective effect from ongoing routine hydroxychloroquine use as has been previously hypothesised. Our estimates were robust to multiple sensitivity analyses. We have demonstrated in this study that it is feasible to address specific hypotheses about medicines in a transparent manner in response to speculation, using OpenSAFELY, and to inform regulatory bodies decision making in the absence of high quality, randomised trial data. However, due to the observational nature of the study, a degree of uncertainty persists that can only be addressed through large-scale randomised trials.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Comparison with other evidence
A randomised trial examining hydroxychloroquine for post-exposure prophylaxis did not demonstrate a significant benefit in preventing infection, although the findings were compatible with an absolute risk reduction as much as 7% in the context of an absolute risk of about 14% in the placebo arm. 14 As we await the reporting of ongoing clinical trials of prophylactic use of hydroxychloroquine, related evidence of drug effectiveness among existing users can be generated from observational data. Previous investigations of hydroxychloroquine using observational data 24 had limitations in their design and analysis, 25 including relatively small sample sizes and focusing only on hospitalised patients, which may produce spurious associations. 26 Numerous randomised trials [5][6][7][8][9] have failed to find any clinical benefit of hydroxychloroquine for treatment of COVID-19. Studies to date have also not demonstrated substantial harm, though the RECOVERY trial have shown some evidence of harm looking at death and ventilation as a composite outcome. 27 While hydroxychloroquine has been approved for use in treatment of RA and SLE for many years, recent evidence has suggested potential short-term harms (when coprescribed with azithromycin). 28 While we were underpowered to investigate co-medication with azithromycin, results of our study showed no evidence of an association with mortality from COVID-19 or other causes. This suggests justification to continue trials of hydroxychloroquine for prevention of COVID-19 to confirm our findings from observational data.

Strengths and limitations
The greatest strengths of this study were the statistical power and the detailed longitudinal, routinely collected data to ascertain routine hydroxychloroquine use prior to the outbreak of COVID-19 in England. We were able to focus analyses on patients with indications for the use of hydroxychloroquine, a key component to mitigate confounding by indication in pharmacoepidemiological research of real-world data. Prior to starting analysis, we developed a DAG to identify a minimal set of covariates to adjust for the hypothesised confounding structure. We also fitted models adjusting for additional characteristics suggested as potentially important in consultation with clinicians. We performed informative sensitivity analyses including quantitative bias analyses to test key assumptions about missing data on bDMARD treatments. Lastly, the optimal timing 29 and dose 30 of hydroxychloroquine for therapeutic and prophylactic use for COVID-19 has been debated. Our population included regular users of hydroxychloroquine, in doses routinely used in clinical practice, with clarity that hydroxychloroquine administration occurred before exposure to SARS-CoV-2. Finally, we pre-specified our study protocol and have shared all analytical code.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted September 9, 2020. . https://doi.org/10.1101/2020.09.04.20187781 doi: medRxiv preprint We also recognise possible limitations. One is the risk of residual confounding by use of medications not prescribed in general practice, namely bDMARDs. While the majority of prescriptions in England are supplied by general practitioners in primary care, some medicines, including biologics, may be supplied by hospitals for various reasons including cost. We have advocated for these data to be more widely shared but, at present, they are not available. 31,32 Although recent nationwide information on the prevalence of concomitant use of bDMARDs and hydroxychloroquine were unavailable, we demonstrated that our results were robust to a wide range of plausible assumptions about the use of these drugs and their potential relationship with COVID-19 mortality in quantitative bias analysis. A further potential source of confounding is severity of rheumatological disease, which is not captured in primary care records. However, the addition of a number of chronic comorbidities and proxies for health status in extended adjustment did not alter our findings. Another important consideration is the potential for exposure misclassification, whereby people prescribed hydroxychloroquine were not taking it as directed. In some reportedly rare cases in March, local shortages of hydroxychloroquine may have occurred due to inappropriate stockpiling; however, the United Kingdom has not suffered major shortages during the COVID-19 outbreak. 33 An additional limitation in primary care prescribing data is drug exposure misclassification, whereby people may not adhere to medications as directed. Finally, COVID-19 mortality as an outcome reflects the risk of exposure to and acquiring SARS-CoV-2 infection, as well as the risk of developing severe disease and subsequent death. We were not able to explore the risk of SARS-CoV-2 infection in the current study due to the lack of complete or representative testing data. However, if hydroxychloroquine had a strong protective effect on the risk of SARS-CoV-2 infection, we would have expected to see this reflected in lower risk of COVID-19 mortality.

Summary
We carried out a large study of patients who were prescribed hydroxychloroquine for its licensed purpose and followed them up to look for clear signals of benefit in mortality from COVID-19 and other causes. We found no evidence of benefit after adjusting for important differences in those who had received hydroxychloroquine. Completion of randomised trials for prevention of severe outcomes is warranted to confirm these observational findings. The use of hydroxychloroquine for prevention of COVID-19 mortality outside trial settings is currently not justified.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted September 9, 2020. Funders had no role in the study design, collection, analysis, and interpretation of data; in the writing of the report; and in the decision to submit the article for publication. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

Guarantor
The copyright holder for this preprint this version posted September 9, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted September 9, 2020. . https://doi.org/10.1101/2020.09.04.20187781 doi: medRxiv preprint