Read on Twitter
1/Hydroxychloroquine (HCQ) is looking worse&worse as an option for care of #COVID19 The best new study offers a *superb* example of a statistical approach to data where treatment wasn't randomly assigned. I'll summarize the finding and "propensity scores" https://www.nejm.org/doi/full/10.1056/NEJMoa2012410?query=featured_home
2/In this NY hospital (March 7-April 8), docs were permitted to offer HCQ but didn’t have to. And, the 811 who got HC were sicker than 565 who didn’t. If we compare outcomes with vs without HCQ, it might make HCQ look bad, because of who got it (“confounding by indication”).
3/To address such imbalances, you can do 2 things. Either do a large randomized trial (hard to do fast!), or- as was done here- attempt to analyze in a way that corrects for imbalances. The latter is never as definitive as the former, but it’s faster.
4/This paper devised a model to predict “who got HCQ”. This study used a SH*T-TON of variables: demographics, clinical factors, lab tests, & many specific meds And their model did well in assessing who was most likely to have received HCQ (c=0.81, far better than flipping a coin)
5/Once each patient has a modeled "propensity" to get HCQ, you (a) match each person who got it to one who didn’t (but had similar likelihood of getting it) or (b) weight the data so someone “less likely to get HCQ” (but DID get it) carries more weight
https://www.ncbi.nlm.nih.gov/pubmed/21818162
https://www.ncbi.nlm.nih.gov/pubmed/21818162
6/Had the researchers not addressed the imbalance in who got HCQ, the chance of intubation or death seemed 2x higher in persons who got it. After making the appropriate adjustment, there’s no difference (Hazard Ratio, 1.04 with a 95% confidence interval 0.8-1.3). HCQ didn't help
9/This differs from an earlier study of 368 Veterans, comparing outcomes (death, death after ventilation) with HCQ, HCQ + Azithro, or No HCQ. They also balanced dissimilar groups with propensity scores, and found 2.6x higher risk of death in the HC group. https://www.medrxiv.org/content/10.1101/2020.04.16.20065920v2
10/Before anyone concludes a “harm was caused” by the VA study (2.6x higher death), let’s emphasize the study was smaller & restricted to males. Although it used similar statistical techniques, they couldn't control for as many variables. It was less authoritative evidence
11/Also, let's note the early, tantalizing trial from China (n=62), showing earlier recovery with HCQ, but: the study carried out differed greatly from the study that was planned; recovery indicators were debatable and statistical concerns were major: https://www.cebm.net/covid-19/hydroxychloroquine-for-covid-19-what-do-the-clinical-trials-tell-us/
12/And let's also note a randomized trial of Chinese adults (n=150) where HC did not result in superior clearance of the SARS-CoV-2 virus at 28 days. Doesn’t look good: https://www.medrxiv.org/content/10.1101/2020.04.10.20060558v2
13/In sum:
*So far, no benefit for HCQ if one is hospitalized with #COVID19
*One can’t ever fully eliminate the problems of deducing cause vs effect in retrospective data
*But, one can do a better by having a ton of data and applying “propensities” https://www.bmj.com/content/367/bmj.l5657
*So far, no benefit for HCQ if one is hospitalized with #COVID19
*One can’t ever fully eliminate the problems of deducing cause vs effect in retrospective data
*But, one can do a better by having a ton of data and applying “propensities” https://www.bmj.com/content/367/bmj.l5657
