Comparison of multistate model, survival regression, and matched case–control methods for estimating excess length of stay due to healthcare-associated infections

Tools

Pan, Jiafeng and Kavanagh, Kimberley and Stewart, Sally and Robertson, Chris and Kennedy, Sharon and Manoukian, Sarkis and Haahr, Lynne and Graves, Nicholas and Reilly, Jacqui (2022) Comparison of multistate model, survival regression, and matched case–control methods for estimating excess length of stay due to healthcare-associated infections. Journal of Hospital Infection, 126. pp. 44-51. ISSN 1532-2939 (https://doi.org/10.1016/j.jhin.2022.04.010)

[thumbnail of Pan-etal-2022-Comparison-of-multistate-model-survival-regression-and-matched-case-control-methods]
Preview
 Text. Filename: Pan-etal-2022-Comparison-of-multistate-model-survival-regression-and-matched-case-control-methods.pdf
Accepted Author Manuscript
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo
Download (278kB)| Preview

Abstract

Background: A recent systematic review recommended time-varying methods for minimizing bias when estimating the excess length of stay (LOS) for healthcare-associated infections (HAIs); however, little evidence exists concerning which time-varying method is best used for HAI incidence studies. Aim: To undertake a retrospective analysis of data from a one-year prospective incidence study of HAIs, in one teaching hospital and one general hospital in NHS Scotland. Methods: Three time-varying methods – multistate model, multivariable adjusted survival regression, and matched case–control approach – were applied to the data to estimate excess LOS and compared. Findings: The unadjusted excess LOS estimated from the multistate model was 7.8 (95% confidence interval: 5.7–9.9) days, being shorter than the excess LOS estimated from survival regression adjusting for the admission characteristics (9.9; 8.4–11.7) days, and the adjusted estimates from matched case–control approach (10; 8.5–11.5) days. All estimates from the time-varying methods were much lower than the crude time-fixed estimates of 27 days. Conclusion: Studies examining LOS associated with HAI should consider a design which addresses time-dependent bias as a minimum. If there is an imbalance in patient characteristics between the HAI and non-HAI groups, then adjustment for patient characteristics is also important, where survival regression with time-dependent covariates is likely to provide the most flexible approach. Matched design is more likely to result in data loss, whereas a multistate model is limited by the challenge in adjusting for covariates. These findings have important implications for future cost-effectiveness studies of infection prevention and control programmes.

ORCID iDs

Pan, Jiafeng ORCID logoORCID: https://orcid.org/0000-0001-5993-3209, Kavanagh, Kimberley ORCID logoORCID: https://orcid.org/0000-0002-2679-5409, Stewart, Sally, Robertson, Chris, Kennedy, Sharon, Manoukian, Sarkis, Haahr, Lynne, Graves, Nicholas and Reilly, Jacqui;
CORE (COnnecting REpositories)