Correcting for Non-adherence in a Randomized Study of Hip Protectors to Prevent Fractures Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 20, 2019
  • Gill, Karminder S.
    • Affiliation: Gillings School of Global Public Health, Department of Epidemiology
  • Between October 2002 and October 2004 the Hip Impact Protection Project (HIP PRO) cluster randomized 1042 nursing home residents to wear hip protectors on either the left or right hip; residents were followed for 676 person-years of observation. The intent-to-treat (ITT) incidence rate ratio, comparing protected to unprotected hips, was 1.23 (95% confidence limit (CL): 0.65, 2.34); overall adherence was 74%. When non-adherence is substantial an ITT analysis estimates the effectiveness of treatment in a mixed population comprised of both compliers and non-compliers and, therefore, under-estimates the etiologic effect of treatment to the extent that the study population is comprised of non-compliers. Because of the problems inherent to ITT analyses of studies with non-trivial amounts of non-adherence, there have been calls to supplement the ITT effect estimate with adherence corrected effect estimates. Three relatively new methods in the epidemiology literature correct for non-adherence in randomized studies, and provide unbiased effect estimates: marginal structural models (using inverse probability-of-censoring weights (IPCWs)), structural nested models, and instrumental variable analysis. We employed IPCWs to correct for non-adherence in the HIP PRO study under an assumption that we had measured and correctly modeled all important joint determinants of adherence and hip fracture, and obtained a hazard ratio of 0.55 (95% CL: 0.13, 2.40). Under a structural nested modeling approach, we employed a rank-preserving structural failure time model to identify the survival difference that would have been observed had all participants adhered to their assigned treatment. The factor by which time to a hip fracture was expanded under continuous exposure to hip protectors was 2.41 (95% CL: 0.31, 18.7), assuming a Weibull distribution for time to hip fracture. The estimated hazard ratio under constant exposure was 0.46 (95% CL: 0.07, 2.84). Using data from the HIP PRO study, we found apparent differences in results between the ITT analysis and analyses correcting for non-adherence. We do not take the adherence corrected results as a complete reversal of the prior analysis; rather, we see these results as supplementing the ITT analysis.
Date of publication
Resource type
Rights statement
  • In Copyright
  • "... in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Epidemiology of the School of Public Health."
  • Cole, Stephen
Place of publication
  • Chapel Hill, NC
  • Open access

This work has no parents.