Heparin in hospitalized COVID-19 patients
Spyropoulos and colleagues published randomized clinical trial on therapeutic-dose heparin in hospitalized COVID-19 patients where they concluded that the primary efficacy outcome was reduced in non-intensive care unit (ICU) patients, but not in ICU patients. In the intention-to-treat population, the authors reported a relative risk (RR) between the therapeutic dose and standard dose was 0.46 (95% confidence interval [CI] 0.27 - 0.81; p = 0.004) in the non-ICU stratum and 0.92 (95% CI 0.62 - 1.39; p = 0.71) in ICU stratum. Due to this difference in statistical significance, the authors concluded that the effect was only present in the non-ICU stratum.
Along with colleagues Arthur Albuquerque and Carolina Santolia, we felt the statistical analysis supporting this conclusion was questionable and therefore submitted a letter to the editor which was not accepted for publication as the editors were “unable to assign your letter a sufficient priority for publication in JAMA Internal Medicine”. Consequently in addition to submitting the letter to PubPeer, I have reproduced the letter in this post.
As noted by Altman and Bland, statistical analysis should be targeted to the clinical question: is the effect of therapeutic-dose heparin different between ICU and non-ICU patients? To answer this question, one should directly compare the efficacy of therapeutic-dose heparin between non-ICU and ICU patients using am interaction test, which was not performed in this study. Thus, the authors cannot conclude that the treatment effect differed between these subgroups based on their simple comparison of statistically significant and statistically insignificant p values. We applied the appropriate interaction analysis and found a ratio of relative risks equal to 0.5 [95% CI 0.25, 0.99; p = 0.046] between non-ICU and ICU patients.
Beyond applying the proper statistical test, it must also be properly interpreted. If complete equipoise existed in the null hypothesis of no therapeutic difference between the different hospital populations before this study, then after observing this data there remains a 13% probability that the null hypothesis is true. Of course, if there existed a stronger prior belief in the null hypothesis, then the posterior belief in the null hypothesis of no difference would be even greater than 13%, despite the interaction p value < 0.05.
Highly uncertain times, such as the COVID-19 pandemics, require rapid evidence to aid clinicians in making decisions at the bedside and reliable subgroup analyses are of obvious importance in enhancing personalized clinical decision-making. However, this requires both the appropriate statistical analysis and its proper interpretation.