# Should the outcome variable be measured at least twice for a longitudinal study?

I am trying to find the association between BMI and onset age of a condition with linear regression model. I have multiple records of BMI measurement. But the outcome variable, onset age of condition, obviously, only once.

It seems outcome variable should be measured at least twice for a longitudinal study.

Is this correct? If yes, then how do I make use of this multiple BMI observation in my linear regression model?

-

Is the BMI measured before or after the onset of the condition?

If it is measured before the onset you could use survival analysis. These models often allow you to introduce time varying covariates.

If they are measured after the onset, you will have to ask yourself how you think some future event will affect the condition now. In that case I would use the BMI closest to the event onset and consider that an (imperfect) proxy for the BMI at the onset of the condition.

-
BMIs are measured before the onset but not on regular intervals. The data is from PHC. So the BMI readings are taken whenever each patients visit the hospitals. I was doing a normal linear regression with onset age as dependent variable and BMI closest to the onset of disease as independent variable. But I was wondering if I can use all the other measurements of BMI. Please suggest the best technique to model this. – arshad Mar 20 '14 at 8:55
There is no such thing as the best technique, there are always multiple trade-offs that only you can make. I gave my suggestion in the answer. – Maarten Buis Mar 20 '14 at 9:29

The first thing I would do is make some graphs of BMI over time per person and color them as to whether they were pre-onset or post-onset.

Then I'd think about whether the trend in BMI pre-onset is an interesting predictor. If it is, there are some methods that allow time varying covariates that are not regular; or, you could try some sort of cluster analysis (either formal or not) on the trends and then see if those were related to onset.

Survival analysis - time to event - may be more powerful than methods that simply look at onset vs. not (logistic regression). It may also be a more interesting question, substantively

-