1215. Estimating the Effect of Clostridium difficile Infection on Hospital Length of Stay: How Much Does the Statistical Method Matter?
Session: Oral Abstract Session: New Insights into C. difficile Transmission and Reporting
Saturday, October 5, 2013: 9:00 AM
Room: The Moscone Center: 200-212

Background: Cost estimates for Clostridium difficile infections (CDIs) vary widely. The costs attributable to a hospital associated CDI depend primarily on how CDI increases a patient's length of stay (LOS). Estimating the effect of CDI on LOS is complicated by the relationship between CDI and LOS: longer LOS increases exposure to C. difficile, and C. difficile exposure may increase LOS via CDI.  The purpose of this study is to determine how sensitive the estimates of the effect of CDI on LOS are to different statistical methods: how much of the variation in cost estimates for CDI is attributable to the statistical method used?

Methods: Using data from the 2009 HCUP Nationwide Inpatient Sample (NIS), we estimate the effect of CDI on LOS with each of the following methods: unmatched case control study (CCS), matched CCS, propensity score matched CCS, multivariate OLS regression, propensity score adjusted OLS, and gamma and Poisson based GLM regressions. Next, we perform Monte Carlo simulations to estimate the accuracy of each method: we analyze the ability of each method to recover the "true" LOS using 1200 different randomly generated sets of data based on the NIS.

Results: We find the average number of days added to LOS from CDI varied from 1.76 (SE .006) for GLM-Poisson regression to 8.20 (SE .029) for unmatched CCS. LOS estimates are shown in Figure 1. All estimates are statistically significant (P<.0001).

Our simulations show that the following methods tended to overestimate the effect of CDI on LOS by: 42.03% (SE .064) for unmatched CCS, 33.33% (SE .08) for matched CCS, 32.02% (SE .075) for propensity score matched CCS, 16.48% (SE .065) for OLS regression, and 16.14% (SE .065) for propensity score corrected OLS. The following methods were found to underestimate the effect of CDI on LOS by: 17.3% (SE .03) for GLM-gamma and 33.54% (SE .06) for GLM-Poisson. These estimated errors are depicted in Figure 2.

We estimate that the true effect of CDI on LOS lies somewhere between the propensity score corrected OLS estimate of 4.53 days and the GLM-gamma estimate of 1.99 days.

Conclusion: Estimates of how CDI affects LOS vary widely by the statistical method chosen. The confidence intervals for LOS estimates using different methods did not always overlap, i.e. some methods did not capture the true effect. Our results have broad implications for future CDI cost studies. 

Aaron Miller, MA, Department of Pharmacy Practice and Science, University of Iowa, Iowa City, IA, Linnea Polgreen, PhD, Pharmacy Practice & Science, University of Iowa College of Pharmacy, Iowa City, IA and Philip M. Polgreen, MD, Division of Infectious Diseases, Department of Internal Medicine, University of Iowa Carver College of Medicine, Iowa City, IA

Disclosures:

A. Miller, None

L. Polgreen, None

P. M. Polgreen, None

Findings in the abstracts are embargoed until 12:01 a.m. PST, Oct. 2nd with the exception of research findings presented at the IDWeek press conferences.