You are here

Article Text

Article menu
Download PDFPDF
Electronic pages
Exposure assessment 2
  1. Y. Jin,
  2. J. A. Deddens
  1. University of Cincinnati

    Statistics from Altmetric.com

      Request Permissions

      If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

      084 ANALYSIS OF REPEATED MEASURE EXPOSURE DATA WITH VALUES BELOW THE LIMIT OF DETECTION

      Objectives:

      Occupational studies often involve repeated measures of log-normally distributed exposure data that contain values below the limit of detection (LOD). We will illustrate new methods for dealing with such data. Results will be contrasted with the traditional methods of substituting LOD/2 or LOD/sqrt(2) for values below the LOD. We will also illustrate new methods for computing the correlation of two variables both of which have values below the LOD. Simulations results will be given as well as examples from occupational studies (eg, pesticide exposure).

      Methods:

      Traditionally when analysing exposure data with values below the LOD, one substitutes LOD/2 or LOD/sqrt(2) for values below the LOD and then uses standard methods for computing correlations, and uses usual mixed linear models for repeated measures exposure data. For data without repeated measures it is possible to obtain maximum likelihood estimates using PROC LIFEREG by treating the values below the LOD as being left censored. Recently Lyles et al (Biometrics 2001;57:1238–44) showed how to find the maximum likelihood of the correlation between two left censored variables. Theibaut et al (Comput Methods Programs Biomed 74;2004:255–260) has showed how to use PROC NLMIXED to perform maximum likelihood estimation of repeated left censored data with one random effect. These methods were developed for AIDS/HIV data and do not seem to be widely known in the occupational health literature. We have also developed SAS MACROs for performing Bayesian estimation of random effect linear models with values below the LOD.

      Results:

      The NLMIXED method yields maximum likelihood estimates of parameters in linear mixed models. Simulations show that maximum likelihood methods provide excellent parameter estimates, correlation estimates, and tests of hypotheses. Bayesian methods also provide excellent estimates and tests of hypotheses. In contrast, the traditional methods of substituting LOD/2 provide very biased estimates of both geometric means and …

      View Full Text