Regarding “Clostridium Difficile Ribotype Does Not Predict Severe Infection”

To the Editor—We read with interest the recent article from Walk et al, “Clostridium difficile Ribotype Does Not Predict Severe Infection” [1]. However, their interpretation makes 2 epidemiologic errors, rendering their conclusions unreliable.

First, they treat “no evidence of difference” as “evidence of no difference”: These are equivalent if, and only if, a study is adequately powered to exclude the possibility of a modest effect, for example, based on a 95% confidence interval (CI) (as in noninferiority trials). The comparison of 027/078 vs other ribotypes for severe disease had an unadjusted odds ratio (OR) of 2.33 (95% CI, 1.03–5.02; P = .035). Adjusting for age, Charlson comorbidity index, hematocrit, or platelets reduced the OR only slightly (still remaining >2.1) with P values of .06–.07 and 95% CIs of .92–4.78 (ie, essentially unchanged). The .05 P value cutoff is an arbitrary threshold from days when significance tables had to be calculated laboriously by hand [2]. Statistically, there is little difference in the strength of evidence provided from findings with P values of .035 and .07, as evidenced by the CIs in Table 3 [1], which all extend above 4, well above any plausible margin that might be considered noninferior. The lowest OR is for tcdC deletion (1.26 [95% CI, .38–4.94]); however, the meaning of this estimate is unclear as this deletion is ubiquitous among ribotype 027 isolates, and hence is completely confounded with the majority of the “hypervirulent” strains included in the dataset. Unfortunately, replicating “no evidence of difference” in a separate validation cohort [1] does not transform it into “evidence of no difference”—it simply means that 2 underpowered studies have been conducted.

Another problem is adjusting for factors (biomarkers at diagnosis) on the causal pathway between an exposure (C. difficile ribotype) and an outcome (disease severity), sometimes denoted mediation or overadjustment bias [3]. Consider a pathogen in which the entire mechanism for causing disease is through raising white blood cell (WBC) count. Suppose there are only 2 strains, 1 of which causes twice the WBC count rise. By definition, all the differential mortality between these 2 strains will be due to their differential WBC count rise—adjusting for this will reduce the WBC count–adjusted effect of a strain to zero. Essentially, if there is any strain effect on a biomarker (not analyzed by Walk et al [1]), then the biomarker-adjusted OR for a strain represents an effect that is not mediated through biomarkers. It cannot be interpreted as the causal effect of strain, because it overadjusts for strain-related biomarker differences.

This study shows that WBC count and albumin are strong predictors of C. difficile severity and that there is an unadjusted association between strain and severity. However, unfortunately it does not address the key question as to the causal impact of strain including biomarker-mediated and biomarker-independent effects. Our larger study of 1893 enzyme immunoassay–positive, culture-positive, multilocus sequence–typed strains shows definitively that strains vary in their overall impact on mortality (adjusted for multiple potential confounders) and in their overall impact on biomarkers (predominantly those associated with inflammatory pathways), and that residual variation in mortality risk remains even after adjusting for biomarker-mediated effects [4].