BMI at baseline & followup with exposure at baseline; model change or BMI at FUP? Control for BMI baseline? For a prospective occupational cohort where everyone is exposed to one or more chemical agents, examining BMI at follow-up compared to a specific chemical exposure at baseline, is it necessary to control for baseline BMI? Is it better to model change in BMI or BMI at followup? There is no unexposed group -- just cohort members unexposed to some agents versus others. All analyses are within-cohort.
 A: To calculate a change you are assuming that BMI is perfectly transformed, and it is only in the very special case of linear models that there is an equivalence between the two methods of analyzing change adjusted for baseline and analyzing follow-up measure adjusted for baseline.  I don't like to use special cases, instead opting for a framework that would allow semi-parametric models to also be used (because they have fewer assumptions.  Example: proportional odds model).  So I prefer models like f(post) ~ g(pre) + other baseline variables.
Also don't take for granted that BMI adequately summarizes height and weight or that height and weight are the correct body size measures to use in your situation.  If height and weight are correct to use, you can check that BMI is a good summary by two approaches.  The first is to put log(weight) and log(height) into the model and see if their coefficients have a ratio of -2.  The second approach checks for nonlinearity of height and weight logs.  In the R rms package you would do
require(rms)
f <- ols(y ~ rcs(log(height), 4) + rcs(log(weight, 4)), data=mydata)
anova(f)  # tests for linearity of log height, log weight, combined

This is for the case where you have a response variable y that is not BMI.  If you have BMI measured at two times and you just want to see that it is properly transformed to meet the Bland-Altman conditions, try different transformations f and plot y=f(post) - f(pre) vs. x=f(post) + f(pre).  Look for flat central tendency at y=0 and equal variation across x.
A: Modeling Change in BMI vs BMI at follow up will likely be determined by the interpretation you plan to use - assuming that you are using linear regression.  One can see that they can be formulated equivalently by noting E[Y2-Y1] = E[Y2] - E[Y1].
To demonstrate equivalent formulations, letting XBeta be the linear predictor of covariates and parameters, E[Y2-Y1] = XBeta can be written equivalently as E[Y2] = E[Y1] + XBeta.  If Y1 (baseline BMI) is among the covariates in the linear predictor X (that is, controlling for baseline BMI), the parameter estimates from the two approaches (modeling change or FUP as the response) will be exactly equivalent for all other X (the only parameter with a different estimate will be the parameter associated with Y1).  If Y1 is not in the covariate list (not controlling for baseline BMI), then the latter equation suggests modeling the change is equivalent to forcing the parameter associated with baseline BMI to be 1 in the model for FUP.
As for controlling for baseline BMI, I believe this is still a debated question that doesn't have a definitive answer and depends on the assumptions you want to make for your data/study.  For a detailed description of some issues, see Allison, P.D. (1990).  Change scores as dependent variables in regression analysis.  Sociological Methodology, 20, 93--114.
