Modelling the effect of a change in socioeconomic status on changes in health outcomes

[John Adler]

I have panel data on an individuals health status (both self reported and objective) and their socioeconomic status at baseline and ten years later. I would like to investigate how the health status of these individuals changes, ten years later, based on changes in this socioeconomic status and the unemployment rate in their area of living. Initially I had thought a standard logistic fixed effects regression would work, but I'm not sure how to incorporate the baseline and final health and socioeconomic status into this model. Particularly I don't think that the standard model allows me to consider the effect of a change in socioeconomic status on changes in health outcomes in the same individuals measured at two time points. If possible it would be great if someone could make some suggestion as to the best methodological approach (particularly in stata) to this problem.

[Clyde Schechter]

I would think the answer would depend on the kind of data that the health status and SES variables provide you with, and you don't describe that. Are these categorical variables? Ordinal? Some kind of quasi-continuous rating scales? What?

[Roman Mostazir]

Could you please show us some data example using -dtaex-. Please read the FAQ Section on how to make a useful post. It would make things easier for others to help you.

On your point, it is not clear how many years of data you are calling 'panel data'. If you have two time points of data i.e. baseline and at 10yrs, you can take the delta score of socio economic status (new value-baseline value) and regress (linear regression/logit or whatever suits the data) the outcome variable (health status, time point-10) on that change score having the base line health status (time point-1) included in the model as a covariate i.e. the model should look like:

health outcome at year 10 =  Intercept + baseline health outcome + delta score of socio economic status

If you have more data points for health outcomes and only two time points (time-1 & 10) for socio-economic status, then it could be tricky. You need to decide whether it makes sence to use a delta score as a covariate because being a hypothesised cause (using the term loosely), socio-status should change before any event (change in health outcome) took place. Thus, health outcome at year-5 should not be attributed to a change in socio-status that took place at year 10.

[daniel klein]

If you have two time points of data i.e. baseline and at 10yrs, you can take the delta score of socio economic status (new value-baseline value) and regress (linear regression/logit or whatever suits the data) the outcome variable (health status, time point-10) on that change score having the base line health status (time point-1) included in the model as a covariate i.e. the model should look like:

health outcome at year 10 = Intercept + baseline health outcome + delta score of socio economic status

Note that the coefficients estimated from this model will be biased because, the time-invariant part of the error-term (that is not shown here) affects both, health at baseline and health 10 years after. Thus, there is a correlation of unobserved factors and the predictors that is implied by the model formulation and that renders the estimated coefficients biased.

I do not see why you think a "standard" fixed-effects (within variance estimator) does not give you the answer you are looking for.

Best
Daniel