Results from multi-center randomized, observational, and cross-sectional studies involving binary outcomes are often reported in terms that lack meaning and usefulness to an audience of clinicians or health policy analysts. Investigators lack appropriate guidance and software for translating odds ratios generated from generalized linear statistical models into more meaningful measures of treatment effect, such as risk ratios, risk differences, or number needed to treat, along with appropriate confidence intervals. Previously suggested methods for estimating relative risk and risk difference from generalized linear models have been applied mostly to simple cases of single-center studies, or without regard to testing the coverage properties of the resulting confidence intervals. This proposed project seeks to test, and provide for applied statisticians and experienced health services researchers and epidemiologists, a set of well-documented methods and statistical programs to compute unbiased estimates with appropriate confidence intervals (variance) for estimating relative risk, risk difference, and number needed to treat (NNT). The focus will be on estimating effect size rather than on testing hypotheses. To achieve this goal, six specific aims are to: (1) develop and test theoretically sound and generalizable methods for conducting simulations of multi-center data with binary and ordered categorical outcomes, (2) describe, and test bias and coverage of, available multivariable models for binary outcomes with respect to the estimation of clinically meaningful measures of the absolute and relative effect of treatment or exposure, (3) extend available methods for estimating odds ratios for multi-center binary outcomes to estimating risk difference and relative risk, (4) extend the estimation of relative risk and risk difference using models for ordered categorical data: proportional odds and continuation ratio models, (5) extend findings to the estimation of rate differences to generalized linear models for person time data, (6) and extend findings to the estimation of differences in survival rates from multivariable Cox regression models.
These specific aims will serve the overriding goal of the translation of statistical methods for use by non- statistician-investigators for randomized trials, observational studies, and meta analyses, through user- friendly routines available in the public domain for use with commercial software (primarily STATA v 6.0). Publication of findings and programs is proposed to reach clinical investigators, applied statisticians, epidemiologists, and health services researchers.
