written by Eric J. Ma on 2016-12-20 | tags: data science statistics
For real. Read on for the statistical perspective on why.
I saw this article published by JAMA, posted by a number of my friends on Facebook. The claim here was that patients treated by women doctors show lower readmission rates than patients treated by men, controlling for as many other factors as can be plausibly controlled for.
When I saw the clickbait-y headlines and posts by my mostly left-leaning friends basically going "women are better than men", my first instinct was "go check the stats". (I'm probably best labelled as centrist.) Indeed, I think I would have made Prof. Allen Downey proud - I didn't buy the p-values at face value. Rather, I was most interested in the effect size.
If you check the main figure on the paper, it'll look like this:
Peeking into the text for a bit more context for mortality (it's roughly the same magnitude difference elsewhere).
Physician Sex and Patient Mortality
The final sample for the analyses of 30-day mortality rates included 1 583 028 hospitalizations treated by 57 896 physicians. Overall 30-day mortality for the entire sample was 179 162 (11.32%). Patients cared for by female physicians had lower 30-day mortality than did patients treated by male physicians (10.82% vs 11.49%; risk difference [RD], –0.67%; 95% CI, –0.80% to –0.54%; P < .001; number needed to treat [NNT] to prevent 1 death, 149) after accounting for patient characteristics (Table 2). The difference in mortality persisted after adjustment for hospital fixed effects (female physicians, 10.91% vs male physicians, 11.46%; adjusted RD, –0.55%; 95% CI, –0.67% to –0.42%; P < .001; NNT, 182). Further adjusting for physician characteristics had a limited effect on these results (female physicians, 11.07% vs male physicians, 11.49%; adjusted RD, –0.43%; 95% CI –0.57% to –0.28%; P < .001; NNT, 233).
If you look at just the difference between the two 'treatment' groups (male-treated vs. female-treated patients), there only seems to be a 0.42% difference in mortality rates after controlling for hospital patient characteristics, hospital fixed effects, physician characteristics, length of stay, use of care, discharge location, patient volume, and physicians’ years of practice. Damn, that's a really small difference! The effect size (here improperly defined as ratio of larger to smaller) is only 1.03. Hardly an effect worth mentioning, isn't it?
That 0.42% difference occurred from a sample size of more than ~1.5 million hospitalizations. This is mortality that we're speaking of here, so in context, we're talking about ~6,600+ fewer deaths when women treat patients. 6,600+! Hospitalizations happen frequently and in large numbers, so a small percentage difference still means large numbers of people. If you happen to be pro-life, this is something that you really should care about.
I still take issue with the article's emphasis on "statistically significant or not" method of reporting the results, but I think that's a flaw of the medical literature publishing culture, and something that the authors probably had to conform to in order to successfully publish the results. Nonetheless, if you even just skim the text of the paper, the data are consistent across all comparisons made, regardless of whether the results were "statistically significant" or not. That's pretty strong evidence, in my opinion.
I send out a newsletter with tips and tools for data scientists. Come check it out at Substack.
If you would like to receive deeper, in-depth content as an early subscriber, come support me on Patreon!