Risk Factors for Carbapenem-resistant Pseudomonas aeruginosa Infection in Children: Correspondence

To the Editors:

We commend Li et al for their important study evaluating risk factors for carbapenem-resistant Pseudomonas aeruginosa (CRPA) infection among children.¹ As the article emphasizes, multidrug-resistant P. aeruginosa (PA) is a critical pathogen that requires prioritization for new antibiotic research and development.² This study evaluated potential risk factors for CRPA infection with multivariable logistic regression. The authors identified several independent risk factors for CRPA infection, including prior exposure to carbapenems, use of β-lactams/-lactamase inhibitors within the previous 90 days, undergoing a bronchoscopy during time at-risk, and invasive therapy within the last year. Although better understanding of such risk factors is critical, this type of analysis carries inherent barriers to valid inference.

The study analyzed select covariates as risk factors for CRPA infection. Covariates with significant univariate associations (P values < 0.05) were included in a backward stepwise-selected multivariable model. The article’s Table 4 reports effect estimates for 8 potential risk factors from that model. Nevertheless, interpretation of multiple exposure effect estimates from a single model may suffer from the so-called Table 2 Fallacy.³ Studies commonly report multiple adjusted effect estimates—stereotypically reported in an article’s “Table 2”—which may include exposures, mediators, and confounders in a single model.³ The Table 2 Fallacy results from potential confusion/ambiguity between direct- versus total-effects for these covariates.³ Moreover, covariates within the model may be confounded by additional variables not included in the model.³ Interpretation is thus difficult, particularly if mediation or unadjusted confounding exists.

To illustrate, we created a directed acyclic graph (DAG) to display causal relationships between covariates from Table 4 (Fig. 1). The study’s multivariable model did not incorporate previous PA infection nor did univariate/multivariable analyses address previous PA colonization. Either of these findings might predispose to CRPA infection, including via their potential impact on prior carbapenem (or other antibiotic) exposure. Unfortunately, without adjusting for such confounders, the effect estimates for the latter risk factors could be affected.

FIGURE 1.: DAG illustrating relationships with potential risk factors for carbapenem-resistant Pseudomonas aeruginosa (CRPA) infection.

The study was unclear whether “prior carbapenem exposure” and “carbapenem use within the previous 90 days” overlapped. When considering the effect from prior exposure to carbapenems, the multivariable model adjusted for previous carbapenem use in the past 90 days, which is a mediator along the causal pathway (Fig. 1). The total effect that prior exposure has on CRPA infection is not represented because adjusting for mediators removes the effect of the causal pathway mediated through 90-day previous carbapenem use.

We suggest displaying different models for each covariate of interest, with adjustments for relevant covariates.³ This method would allow effect estimates to be displayed that represent the unbiased total effects for each covariate.³ If the authors aim to display direct effects, then we suggest carefully highlighting the appropriate interpretation. Last, it is critical to avoid reliance simply on statistical methods when identifying potential risk factors free from confounding. A nonsignificant variable may still induce bias when evaluating causal relationships. A combination of a priori knowledge should be used with statistical analyses which might add a reliance on DAGs to select covariates for adjustment regardless of statistical significance.