Estimating excess length of stay (LoS) attributable to MRSA using an instrumental variable approach

Background:  Matched cohort studies and multivariate analyses have estimated excess LoS associated with MRSA infection, but few studies controlled for the simultaneous relationship between MRSA infection and time in hospital, potentially yielding upwardly biased estimates. Objective: To derive a less biased estimate and assess the excess LoS attributable to MRSA infection using an instrumental variable approach to control for endogeneity bias.

Methods: Using a sample of 797 MRSA-colonized and 164 MRSA-infected surgical patients taken from a prospective cohort study at the University of Geneva Hospitals, two-stage regression techniques were used to predict LoS as a function of MRSA infection, while controlling for patient-level confounders and endogeneity. The exogenous relationship between MRSA infection and LoS was estimated using receipt of rapid PCR screening on admission as the instrumental variable. Receipt of rapid PCR was strongly correlated with the MRSA covariate, yet uncorrelated with the LoS outcome variable, affording an opportunity to control for the simultaneous relationship between MRSA infection and LoS.

Results: The median LoS of MRSA-infected surgical patients was 48 days (IQR 26-79) compared to 13 (IQR 6-24) for MRSA carriers (P<.001). Significant differences between infected and colonized patients were seen with diabetes (32% vs 17%), the proportion transferred from another facility (61% vs 43%), occurrence of pressure ulcers (5% vs 1%), CVL-catheterization (37% vs 17%), urinary catheterization (73% vs 41%), and immunosuppressive therapy (12% vs 6%), respectively (all P<.01). When controlling for patient-level variables, co-morbidities, and in-hospital events, the excess LoS attributable to MRSA infection from the multivariate analysis was 13 days (95%CI 11-15). When using receipt of rapid PCR as an instrument to decrease endogeneity bias, the first-stage regression coefficient for the instrument was of the expected sign (negative) and statistically significant (P<.001), suggesting that receipt of rapid PCR had a protective effect on LoS.  In the second-stage regression, the coefficient for infection suggested an excess attributable LoS of only 4.6 days, however without statistical significance (P>.5).   The loss of significance could be explained by the small portion of MRSA variance explained by receipt of PCR (R²=.08).  Other strong predictors (P<.01) of excess LoS included time in ICU (16d, 95%CI 8-26), CVL-catheterization (12d, 95%CI 4-20), and transfer from another facility (8d, 95%CI 4-12).

Conclusions: Results shown here are indicative of an endogeneous relationship between MRSA and LoS, suggesting that existing estimates of excess LoS associated with MRSA infection are potentially biased due to confounding and endogeneity bias, as infection risk increases with time in hospital yet time in hospital increases with infection.