Upending these results entirely comes a subsequent study whose conclusions are completely different.³ In this alternative study, the authors took advantage of a sort of “natural experiment,” a 15-month national piperacillin-tazobactam shortage. Preceding and following the period of shortage, piperacillin-tazobactam was the clear primary option at the authors’ institution. During the shortage period, the prevailing anti-pseudomonal antibiotic coverage for their institution switched, by necessity, to cefepime. The authors make an argument that the shortage exists as a random exogenous factor creating the conditions necessary to interpret their results as a sort of emulated clinical trial.

The results of this emulated clinical trial are starkly different. In their study, comprising over 7,500 patients, these authors demonstrated an opposite, and larger, absolute effect on mortality. Rather than the non-significant difference observed at 14 days in ACORN, here the mortality outcome at 90 days demonstrated a 5.0 percent absolute advantage to cefepime. The authors ultimately conclude, when not otherwise indicated, the use of piperacillin-tazobactam results in potentially “thousands of additional fatalities each year.”

Any conclusion providing evidence of “thousands of additional fatalities” ought give any clinician pause. Clearly, it is always prudent to narrow the spectrum of antibiotic coverage as clinically appropriate. However, each of these antibiotics has historically been used interchangeably for empiric coverage of gram-negative organisms.

There are, of course, concerns regarding the magnitude of this observed effect size. Firstly, and most prominently, this remains a retrospective analysis prone to the effects of uncontrolled biases and confounding. Even a so-called “natural experiment” such as this remains limited in its ability to isolate a causal effect. Of note, despite the proposed disadvantage of piperacillin-tazobactam, crude mortality was unchanged, rather than mproved, during its absence. Most of the observed mortality advantage favoring cefepime emerged following statistical adjustments.

In particular, one specific adjustment, for metronidazole exposure, is suspected to have been problematic. Patients who received metronidazole as extended anaerobic coverage were generally more unwell, leading to commentary suggesting the observations from this analysis are subject to collider bias.⁴ Collider bias is a type of confounding in which an unobserved feature is associated with both an outcome and a covariate used for adjustment. In this case, the concern voiced is high-risk clinical factors independently influence both mortality and metronidazole use, and an adjustment solely for metronidazole neglects this “collider.” The net effect of this adjustment generates an over-correction, after a fashion, bestowing an exaggerated advantage upon cefepime.