A downside of being an epidemiologist who thinks about antibiotic resistance 24-7 is that eventually it's hard to read the literature for fear of seeing another paper where authors make a critical mistake. It's unfortunate, because the review process should catch these common errors and at least make the authors mention limitations in their discussion sections. A case in point is a recent paper by Chen et al. in CID that sought to assess the impact of MRSA SCCmec type and vancomycin MIC on treatment failure. In this case, treatment failure was defined as all-cause 30-day mortality, persistent bacteremia, or recurrent bacteremia. They provide nice graphs showing higher mortality for hospital associated-MRSA strains (SCCmec type I, II, III) and for strains with vancomycin MICs > 2. So far, so good. But then, I looked at their multivariable model.
As expected, they controlled for septic shock or severe sepsis in their model. They should not have done this. If it were a comparative effectiveness study assessing various treatment outcomes, it would be appropriate to collect severity of illness including evidence of sepsis before treatment initiation and control for it in the model. However, Chen's study was only looking at outcomes based solely on strain differences. Thus, they should never have controlled for shock or sepsis. How else would a bacteremic patient die if not through sepsis? For a longer description of why controlling for factors in the causal pathway is suboptimal, see my post from 3/19/2012. I have pasted the key sections below. Note: We have been writing letters and review articles in CID pointing this issue out since the year 2000.
March 19, 2012: My first ever publication, and in some ways still my favorite, was a letter to the editor of CID that I wrote in 2000 pointing out a common flaw in outcome studies of infectious diseases. In the letter, I discussed a paper that looked at the outcomes (death) associated with methicillin-resistance in patients with S. aureus bacteremia. In the analysis, the authors controlled for septic shock in their regression model. I pointed out that shock is in the causal pathway between infection and death and, therefore, should not be controlled for in regression in models. This would be like controlling for car accidents when looking at the association between cell phone use and death. In infectious diseases, if you remove shock from the causal pathway, it is hard to see how you might otherwise die.
The error of controlling for intermediates is frequently repeated in ID outcome studies when, for example, authors control for illness severity using the APACHE score. If the APACHE is measured after the infection manifests, this variable would be in the causal pathway and should not be controlled for in the regression model. The APACHE should be measured before the infection manifests, as we did here. Jessina McGregor and JJ Furuno (both now at Oregon State) published a nice systematic review on optimal methods for ID outcome studies in CID back in 2007. Wouter Rottier (with Marc Bonten) just published a meta-analysis looking at the impact of confounders and intermediates (factors in the causal pathway) on ESBL-bacteremia outcomes. (JAC, March 5, 2012) I highly recommend that you read these studies prior to undertaking an ID outcome study.