Study of critically ill patients with COVID-19 in New York City – Authors' reply

We agree with Daniele Piovani and Stephanos Bonovas that informative censoring, if present, could represent a potential source of bias in the survival analyses in our Article.¹ However, sensitivity analysis suggests that any such bias is likely to be minimal.

To evaluate the effect of assigning different observation times on our regression estimates, we reconstructed our primary Cox model with patients discharged from hospital alive considered event-free throughout the study period, as suggested by Piovani and Bonovas. The generated hazard ratios were consistent with those we previously reported, with older age (adjusted hazard ratio 1·31 [95% CI 1·10–1·56] per 10-year increase), chronic cardiac disease (1·71 [1·05–2·78]), chronic pulmonary disease (3·12 [1·58–6·19]), and higher concentrations of interleukin-6 (1·13 [1·04–1·23] per decile increase), and D-dimer (1·10 [1·01–1·20] per decile increase) associated with mortality in the multivariable model. Regarding the cumulative incidence of hospital mortality at 28 days, reconstruction of this function yielded an estimate of approximately 40%.

In addition, more definitive in-hospital outcomes for the patients included in our cohort are now available. As of July 2, 2020, by which time all patients had at least 90 days of observation, a final in-hospital outcome was known for 250 (97%) of 257 patients. 113 (44%) patients had died (including 96 [47%] of 203 patients who received invasive mechanical ventilation), 133 (52%) patients were discharged alive, four (2%) were transferred to another hospital, and seven (3%) remained hospitalised.