Our working group has a data set that contains about 10 measured urinary pesticides or their metabolites in a sample of about 1000 women. We want to see if we can describe exposure patterns among this sample to different socioeconomic variables, for example state of residence, age, parity, body mass index, diet, etc. Some of these explanatory variables are continuous, some ordinal and some categorical.

The dependent variables (the pesticides) are basically censored continuous variables. Many have only 10-30% of the measurements above the detection limit of the apparatus and within the group of measurements for each pesticide the distribution of measurements is highly skewed.

We are looking for help in designing a procedure that takes into account the multiple measurements of pesticides within each subject, the censored nature of the dependent variables and covariance among the various pesticides to see if we can detect patterns of pesticide exposure that might be associated with the personal characteristics of the subjects.

We’ve thought of using principal components or factor analysis but the lack of multinormality of the group of outcome measurements and the censored nature of the measurements suggests that the results of that analysis will be suspect. We are looking at “gsem” with “intreg” for solutions but aren’t sure how to specify the model.

Can any one provide pointers that would help us solve this thorny analysis problem?