The present thesis discusses both the methodology as developed, and applications in practice. Methods of dealing in the statistical analysis with the occurrence of missing data are presented in Chapter 2 as a review of recent literature on this topic. Complete issues of Biometrics, Biometrika, Journal of the American Statistical Association, Statistics in Medicine and the American Journal of Epidemiology have been screened from 1986 up to 1991 on this point. Relevant literature from other sources has also been included. Two important statistics used in occupational health epidemiology are relative risk and prevalence. A study of possible effects on the relative risk in the event of inequality of the fractions of deaths and alives of which vital status is known, is described in Chapter 3. Adjustment procedures are proposed for calculating conservative confidence regions of the 11trueu relative risk, using only little extra information. With respect to prevalence, adjustment procedures have been proposed by several authors. An attempt to combine the imputation method with postulating a missing-data-generating mechanism is presented in Chapter 4, where missing-data~ corrected estimates for the prevalence are presented. These estimates are compared with respect to their likelihood through a likelihood ratio test. Two general approaches concerning small~sample behaviour of statistics are exemplified in Chapters 5 and 6. Many parameter estimates that are used in the analysis of continuous outcomes, have well-known small-sample behaviour. This is not true for the two statistics that are often used in the analysis of categorical variables. Therefore Chapter 5 discusses an approximate Bartlett adjustment for testing the goodness-of-fit of multinomial regression models. This adjustment is applicable to any multinomial regression model, since it only needs estimated category frequencies and does not need any derivatives to be calculated. To illustrate the second approach, the consequences of calculating exact p~values for the Wilcoxon-Mann-Whitney test statistic in the case of ordinal response variables are presented in Chapter 6 with up to ten observations in each sample. From this it appears that the performance of / conditional statistics, based on observed ranks only, is not -so good as that of unconditional statistics based on an underlying, plausible, distribution. The trial protocol of the back school education programme intervention study included an investigation of differences between subgroups of participants and nonparticipants, in order to gain some understanding of the distinctions between those who want to participate and those who do not. If a determinant of participation is related to the primary outcome variable, adjustments have to be made to allow generalization of conclusions. Results obtained are presented in Chapter 7. Chapter 8 describes a cross-sectional study of person-related determinants of GGT levels. Model-based person-related tolerance regions are developed to facilitate detection of potential determinants. The main fmding is that persons with an observed GGT value under 30 U/1 in general do not need further screening, whereas persons with an observed GGT value above 60 U/1 do. For persons with intermediate GGT values person-related determinants should be taken into consideration. Thus, each individual is either "labelled11 to undergo further follow-up or not. Persons who are so labelled are to be compared with persons who are not labelled in the actual liverfunction project. In Chapter 9, a model is postulated for correcting for long-term analytical bias of a specific laboratory compared to a pool of comparable laboratories in an e'-"ternal quality control programme.

, , ,
R. van Strik (Roel)
Erasmus University Rotterdam
hdl.handle.net/1765/39781
Erasmus MC: University Medical Center Rotterdam

Lugtenburg, D. (1992, January 22). Some statistical aspects of the generalizability of occupational health studies. Retrieved from http://hdl.handle.net/1765/39781