More problems with logistic regression in the medical literature:
(PubMed listing, if you prefer: ncbi.nlm.nih.gov/pubmed/27756470)
Surgical aortic valve replacements from 2008-2014
The median follow-up time was 27 months (IQR 9-46 months, max 70 months)
The long-term outcomes data were analyzed with…logistic regression
But…that’s a lot of variability in the potential time at risk!
Logistic regression only treats the outcome as “Yes” or “No.” It doesn’t account for variability in time at risk or the time of events.
A patient who died one day after surgery counts the same as a patient that died five years after surgery.
Plus, the unequal amount of available follow-up time creates some funky situations. Consider the following hypothetical patients.
Patient A treated in 2008 – died in 2014 – survived 6 years after surgery – in the logistic regression model, simply counts as “dead”
Patient B treated in 2014 – alive in 2017 (goes on to die in 2018, after the paper is published) – survived 4 years after surgery – in the logistic regression model, counts as “alive” (since data were analyzed in 2017)
So…in the logistic regression model, Patient A counts as a “worse” outcome than Patient B (even though they lived longer after surgery)
This could introduce serious biases to the logistic regression model if the patients treated earlier in the study period are different from those treated later in the study period
Because the unequal time at risk means the patients treated in 2008-2009 had a lot longer time at risk (and presumably greater chance to die!) than patients treated in 2013-2014
If a specific profile of patient was more likely to be treated in 2013-2014, their survival might appear to be “better” simply because they had less total time-at-risk after the procedure
The answer (should have been) fairly obvious: use a Cox proportional-hazards regression!
Cox models account for variability in time at risk, and are fairly standard in medical literature.
This should not have been able to slip past journal reviewers and/or editors, and more than likely indicates there was no statistical review.
We can banter about Bayesian statistics versus frequentist statistics and the proper interpretation of confidence intervals all day on #statstwitter.
But the fact that analyzing time-to-event data with logistic regression got past peer review in a relatively respected (IF>3.5) surgery journal is an indication that there’s plenty of “low hanging fruit” that needs to be fixed
I would encourage @annalsthorsurg to consider adding a statistician to their editorial staff, and making sure that future papers have statistical review.
Papers like this continuing to seep into the literature (and exert any influence on guidelines) are ultimately detrimental to health care decision making and patient care
Also worth noting, there are significantly better-quality data from RCT’s published on the subject:
So it’s hard to see what the attraction was in publishing this paper from the journal’s perspective. There should be no "the clinical question was so important we just HAD to let this go and get it through quickly" explanation.
