We conducted this study at New York–Presbyterian Hospital (NYP)–Columbia University Irving Medical Center (CUIMC), a quaternary, acute care hospital in northern Manhattan. We obtained samples from all admitted adults who had a positive test result for the virus SARS-CoV-2 from analysis of nasopharyngeal or oropharyngeal swab specimens obtained at any point during their hospitalization from March 7 to April 8, 2020. Follow-up continued through April 25, 2020. These tests were conducted by the New York State Department of Health until the NYP–CUIMC laboratory developed internal testing capability with a reverse-transcriptase–polymerase-chain-reaction assay on March 11, 2020. Patients who were intubated, who died, or who were transferred to another facility within 24 hours after presentation to the emergency department were excluded from the analysis. The institutional review board at CUIMC approved this analysis under an expedited review.
A guidance developed by the Department of Medicine and distributed to all the house staff and attending staff at our medical center suggested hydroxychloroquine as a therapeutic option for patients with Covid-19 who presented with moderate-to-severe respiratory illness, which was defined as a resting oxygen saturation of less than 94% while they were breathing ambient air. The suggested hydroxychloroquine regimen was a loading dose of 600 mg twice on day 1, followed by 400 mg daily for 4 additional days. Azithromycin at a dose of 500 mg on day 1 and then 250 mg daily for 4 more days in combination with hydroxychloroquine was an additional suggested therapeutic option. The azithromycin suggestion was removed on April 12, 2020, and the hydroxychloroquine suggestion was removed on April 29, 2020. The decision to prescribe either or both medications was left to the discretion of the treating team for each individual patient.
Patients receiving sarilumab were allowed to continue hydroxychloroquine. Patients receiving remdesivir as part of a randomized trial either did not receive or had completed a course of treatment with hydroxychloroquine.
We obtained data from the NYP–CUIMC clinical data warehouse. This warehouse contains all the clinical data available on all inpatient and outpatient visits to one of the CUIMC facilities (see the Data Extraction section in the Supplementary Appendix, available with the full text of this article at NEJM.org). No data were manually abstracted from the electronic medical record or charts. The data obtained included patients’ demographic details, insurance status, vital signs, laboratory test results, medication administration data, historical and current medication lists, historical and current diagnoses, clinical notes, historical discharge disposition for previous inpatient hospitalizations, and ventilator use data.
From the clinical data warehouse, we obtained the following data elements for each patient: age; sex; patient-reported race and ethnic group; current insurance carrier; the first recorded vital signs on presentation; the ratio of the partial pressure of arterial oxygen to the fraction of inspired oxygen (Pao2:Fio2) at admission, estimated with the use of methods developed by Brown and colleagues8,9 (see the Data Extraction section in the Supplementary Appendix); the first recorded body-mass index as calculated for measured height and weight (the body-mass index is the weight in kilograms divided by the square of the height in meters), grouped on the basis of the Centers for Disease Control and Prevention guidelines for adults; the first recorded inpatient laboratory tests; past and current diagnoses; patient-reported smoking status; and medication administration at baseline. Details of the variables assessed are provided in the Supplementary Appendix.
Patients were defined as receiving hydroxychloroquine if they were receiving it at study baseline or received it during the follow-up period before intubation or death. Study baseline was defined as 24 hours after arrival at the emergency department.
The primary end point was the time from study baseline to intubation or death. For patients who died after intubation, the timing of the primary end point was defined as the time of intubation.
We calculated bivariate frequencies to examine the associations among the preadmission variables described above. Patients without a primary end-point event had their data censored on April 25, 2020.
Cox proportional-hazards regression models were used to estimate the association between hydroxychloroquine use and the composite end point of intubation or death. An initial multivariable Cox regression model included demographic factors, clinical factors, laboratory tests, and medications. In addition, to help account for the nonrandomized treatment administration of hydroxychloroquine, we used propensity-score methods to reduce the effects of confounding. The individual propensities for receipt of hydroxychloroquine treatment were estimated with the use of a multivariable logistic-regression model that included the same covariates as the Cox regression model. Associations between hydroxychloroquine use and respiratory failure were then estimated by multivariable Cox regression models with the use of three propensity-score methods.
The primary analysis used inverse probability weighting. In the inverse-probability-weighted analysis, the predicted probabilities from the propensity-score model were used to calculate the stabilized inverse-probability-weighting weight.10 Kaplan–Meier curves and Cox models that used the inverse-probability-weighting weights were reported.
We conducted a secondary analysis that used propensity-score matching and another that included the propensity score as an additional covariate. In the propensity-score matching analysis, the nearest-neighbor method was applied to create a matched control sample. Additional sensitivity analyses included the same set of analyses with the use of a different study baseline of 48 hours after arrival to the emergency department as well as analyses that defined the exposure as receipt of the first dose of hydroxychloroquine before study baseline only. Multiple imputation was used to handle missing data, and model estimates and standard errors were calculated with Rubin’s rules.11 The nonparametric bootstrap method was used to obtain 95% pointwise confidence intervals for the inverse-probability-weighted Kaplan–Meier curves. The statistical analyses were performed with the use of R software, version 3.6.1 (R Project for Statistical Computing).