Lawrence H. Moulton

Analyzing repeated measures on generalized linear models via the bootstrap

The analysis of longitudinal data for which the response variables have nonnormal error distributions previously has been complex and/or dependent on restrictive assumptions. In this paper simple methods are introduced for the class of generalized linear models (GLMs). Regressions are fit to the data at each observation time; functions of the resulting coefficients may be bootstrapped, or the coefficients combined through closed-form estimation of their covariances. Application is made to a data set on xerophthalmia in Indonesian children.

Bootstrapping generalized linear models

Abstract Several methods for bootstrapping generalized linear regression models are introduced. One-step techniques, both conditional and unconditional on the covariates, are examined with respect to robustness and coverage properties.

ASSESSMENT OF POTENTIAL BIASES IN THE APPLICATION OF MSHA RESPIRABLE COAL MINE DUST DATA TO AN EPIDEMIOLOGIC STUDY

Systematic errors in exposure data will result in biased estimates of the exposure-response relationship derived from epidemiologic analyses. Thus, adjustment of exposure data to account for identified errors may provide for a more accurate assessment of effect. In preparing to apply respirable coal mine dust exposure data collected by the Mine Safety and Health Administration (MSHA) to a study of the pulmonary status of underground coal miners, an assessment of potential systematic errors was undertaken. Potential errors stemming from adjustment of controls during sampling, concentration-dependent sampling, truncation of sampling results, identified sampling equipment problems, and a disproportionate number of low concentration samples in mine operator-collected samples were identified and evaluated. Methods to account for these errors and adjust mean exposures by mine, occupation, and year are given.

Bartlett correction factors in logistic regression models

Bartlett correction factors for likelihood ratio tests of parameters in conditional and unconditional logistic regression models are calculated. The resulting tests are compared to the Wald, likelihood ratio, and score tests, and a test proposed by Moolgavkar and Venzon in Modern Statistical Methods in Chronic Disease Epidemiology, (Wiley, New York, 1986).

Latency and time-dependent exposure in a case-control study

Detailed historical data are elicited often from subjects in retrospective studies, yielding time-dependent measures of exposures. Investigation of a hypothesized period of latency can be made by examining disease/exposure relationships in multiple time windows, either along the age or time-before diagnosis axes. We suggest splitting the data into many time intervals and separately fitting regression models to the available data in each interval. Covariances between estimated coefficients from different intervals are empirically estimated, and used for assessing variability of specified functions of the time-specific coefficients. Alternative methods of interval formation and their consequences are discussed. We apply these methods to a French case-control study of oral contraceptive use and cervical cancer incidence, and compare the results to those of standard analyses.

ALCOHOL AND TRAFFIC SAFETY: A SENSITIVITY ANALYSIS OF DATA FROM COMPOSITE SOURCES

Risk factors associated with single-vehicle driver fatalities are explored in a sensitivity analysis of data from composite sources. Information on fatalities was taken from the Federal Accident Reporting System data base for 1976-1981. Characteristics of the driving population were given by the 1973 National Roadside Breath Testing Survey (Wolfe 1974). Using Bayes theorem and logistic regression analysis, the effect of changing driver characteristics on the probability of a fatality was explored. The method used is proposed for a case-control study in which the controls may not accurately represent the population from which the cases were drawn. Risk factors identified are generally in agreement with previous reports.

Confidence intervals for the comparison of two poisson distributions

Confidence interval construction the difference in mean event rates for two Index independent , Poisson samples is discussed. Intervals are derived by considering Bayes estimates of the mean event rates using a family of noninformative priors. The coverage probabilities of the proposed are compared to those of the standard Wald interval for of observed events. A compromise method of constructing interval based on the data is suggested and its properties are evaluated. The method is illustrated in several examples.