Hydroxychloroquine in RECOVERY

Why the Hydroxychloroquine Comparison Matters

A large randomized trial can answer a practical question quickly: should a treatment continue to be tested or used in a particular patient group? The hydroxychloroquine comparison in RECOVERY addressed this question among patients hospitalized with COVID-19. The result was not a reduction in the primary outcome of 28-day mortality. Some secondary outcomes also raised concern, but they require careful interpretation rather than a claim that harm was proved on every measure.

The key is to keep three things distinct: the treatment assignment being compared, the particular outcome and time horizon, and the measure used to summarize the difference. In particular, a crude ratio of the proportions who died is not identical to the trial’s reported rate ratio.

Trial Comparison and Analysis

The published comparison included 1,561 participants assigned to hydroxychloroquine and 3,155 assigned to usual care. The study was open-label: patients and clinicians knew the assigned treatment. Randomization helps make the groups comparable, on average, at the time of assignment; it does not ensure that every measured or unmeasured baseline characteristic is identical in this particular sample. The intention-to-treat comparison preserves the benefits of randomization by retaining participants in their assigned groups.

Mortality is an objective outcome, although follow-up and outcome ascertainment still need to be sufficiently complete. The primary comparison concerns the effect of assignment to the hydroxychloroquine strategy in this hospitalized population and setting. It does not, by itself, estimate the effect of starting the drug earlier in infection, in a different population, or under a different treatment strategy.

For a fixed follow-up horizon, a simple descriptive analysis can compare the observed risks. If \(a\) of \(n_H\) participants assigned hydroxychloroquine die by day 28, and \(c\) of \(n_C\) participants assigned usual care die by day 28, then:

Here, \(RD\) is the risk difference, in percentage points when the risks are expressed as percentages. \(RR=1\) corresponds to equal observed risks; a value above 1 indicates a higher observed risk in the hydroxychloroquine group. These crude summaries are useful, but they are not automatically the same as the trial’s reported time-to-event estimate.

Primary Outcome: No Evidence of a Mortality Benefit

RECOVERY reported a rate ratio for 28-day mortality of \(1.09\), with a 95% confidence interval from \(0.97\) to \(1.23\) and \(P=0.15\). The estimate is above 1, but the interval includes 1. The results therefore did not provide statistically significant evidence of a difference in mortality at the conventional 5% level, and they did not show a mortality benefit from assignment to hydroxychloroquine.

The reported rate ratio should not be substituted for the crude risk ratio calculated from the two proportions. The crude risk ratio from the rounded observed risks is approximately \(0.2697/0.2504=1.077\), or \(1.08\). The trial’s reported rate ratio uses its time-to-event analysis; it is a different summary. Both point in the direction of a higher outcome frequency in the hydroxychloroquine group, but the reported inferential result is the one to quote when describing the trial analysis.

The 95% confidence interval is compatible with effects ranging from a modestly lower mortality rate to a meaningfully higher one under the analysis. It is not a 95% probability statement about the unknown effect. The \(P\)-value is the probability, assuming the null model and the test’s assumptions, of obtaining a result at least as incompatible with that model as the observed result. It is not the probability that the null hypothesis is true.

Secondary Outcomes and Their Interpretation

The secondary results help explain why the overall evidence was concerning even though the primary mortality comparison was not statistically significant. Among participants not receiving invasive mechanical ventilation at randomization, the reported outcome of invasive mechanical ventilation or death occurred in \(30.7\%\) of those assigned hydroxychloroquine and \(26.9\%\) of those assigned usual care. The reported rate ratio was \(1.14\) (95% CI \(1.03\) to \(1.27\); \(P=0.03\)).

This is evidence of a worse result on that composite outcome in the hydroxychloroquine group under the trial analysis. The endpoint combines two events; it does not establish that each component separately increased. Also, it applies to the subgroup not already receiving invasive ventilation at randomization, not to every randomized participant.

The proportion discharged alive within 28 days was \(59.6\%\) in the hydroxychloroquine group and \(62.9\%\) in the usual-care group. The reported rate ratio was \(0.90\) (95% CI \(0.83\) to \(0.98\); \(P=0.02\)), indicating a lower rate of discharge alive under the trial’s analysis. This outcome is clinically relevant, but it is not interchangeable with mortality: a participant who dies cannot be discharged alive, and discharge decisions in an open-label trial may also be influenced by clinical practice.

Why the Treatment Arm Was Stopped

RECOVERY’s hydroxychloroquine arm stopped recruiting after review of accumulating evidence showed no beneficial effect on mortality. The decision was made after the independent data monitoring process reviewed the evidence and recommended stopping enrollment to this comparison. The central practical conclusion was that continuing to assign hospitalized patients to hydroxychloroquine in this trial was not justified by evidence of mortality benefit, with secondary findings adding concern.

A stopping decision is not the same as proof that the treatment has exactly zero effect. Nor does the decision mean the primary outcome crossed a conventional significance threshold in the direction of harm: the reported mortality confidence interval included \(1\). Rather, an interim decision considers whether the evidence and the expected value of further recruitment justify continuing that arm. In a platform trial, a particular treatment comparison can close while the broader trial continues to evaluate other questions.

Diagnostics, Pitfalls, and Limits

When reading a result like this, check the outcome definition, analysis population, time horizon, and effect measure before interpreting a number. Verify whether a percentage is a fixed-time risk or whether a reported estimate summarizes event rates over follow-up. Check that the confidence interval and \(P\)-value refer to the same outcome and analysis. For secondary or subgroup outcomes, note whether the result is exploratory and whether multiple comparisons may affect how much weight to place on a nominal \(P\)-value.

• Pitfall: treating the crude risk ratio and reported rate ratio as identical. The crude ratio \(1.08\) comes from the two observed proportions; the reported time-to-event rate ratio was \(1.09\). Use the label and estimate reported for the relevant analysis.
• Pitfall: calling a non-significant result proof of no effect. The mortality interval includes both a modest reduction and an increase. The correct conclusion is no demonstrated mortality benefit, not established equivalence.
• Pitfall: calling a significant secondary endpoint proof of mortality harm. The ventilation-or-death composite is a different outcome in a defined subgroup. Report it as such and do not replace the primary mortality result with it.
• Pitfall: assuming randomization removes every concern. Randomization supports an unbiased comparison of assigned strategies, but open-label care can affect decisions such as discharge, and results from hospitalized patients in the early pandemic do not automatically apply to other settings or treatment timing.

Compared with observational comparisons, the randomized assignment makes the treatment groups more credible counterfactual comparators, subject to adherence, follow-up, and analysis considerations. Compared with relying only on a single primary endpoint, secondary outcomes can reveal clinically important patterns, but they should be interpreted with their definitions and multiplicity in mind. The case therefore supports a firm practical conclusion about the lack of demonstrated mortality benefit in this population, while remaining appropriately cautious about the exact magnitude and mechanism of any harm.