Analysing outcomes over varying time periods

[Rob Wood]

Hi There,

I would like to assess outcomes (resource use and costs) in subjects following an intervention, whilst adjusting for confounding factors. However, the data I am are using are longitudinal in nature and therefore I have varying time periods over which to assess these outcomes across my subjects e.g. Subject 1 has 10 months of follow-up, Subject 2 has 6 months, Subject 3 has 15 months etc.

I would like advice on which model I should be constructing to analyse these data given the varying follow-up times and to allow for a meaningful interpretation of the resultant coefficient/estimate.

I thought about using a standard regression technique e.g. Poission model, and including follow-up time and the interaction between follow-up time and each confounding factor, but this seemed too simple for this task.

I would appreciate any advice on this topic.

Thanks in advance,

Rob.

[Carlo Lazzaro]

Rob:
see -xtpoisson-.

[Rob Wood]

Hi Carlo,

Thank you for responding. It looks as though -xtpoisson- will do the trick for the resource use aspect of my request i.e. count data, but it won't work for the cost outcomes. Any further suggestions for modelling the cost data? Should I linearly transform the cost data to a standard period of time e.g. 12 months, or is that too basic an assumption?

Kind Regards,

Rob.

[Carlo Lazzaro]

Rob:
resources consumed by patients are usually expressed as counts (say, full blood counts 0,1,2...N) and -xtpoisson- will fit them.
Costs are continuous and you should consider -xtreg-.
Hence, the choice between these panel data regression models heavily depends on your research goal.
In health economics, longitudinal studies have costs as dependent variable and some predictor may well be a (health care) resource that is known (from previous literature) to play a relevant role as a cost driver.

[Rob Wood]

Hi Carlo,

That makes complete sense. Having looked at both -xtpoisson- and -xtreg- it appears that within -xtpoisson- the user can specify the variable containing the follow-up time using the available options; however, having looked into -xtreg- there doesn't appear to be a similar option. How do I specify the follow-up time within -xtreg-?

Thanks again for your help on this.

Kind Regards,

Rob.

[Carlo Lazzaro]

Rob:
the way I see the issue is that you should have data in -long- format (which the best for almost all the Stata-based procedures; see -help reshape- if your data are still in -wide- format), so that for each -panelid- (ie, patient) you have multiple observation (one for each follow-up measurement, with follow-up being your -panelid-). You might be interested in having an example of panel data in long- format just taking a look at the following example dataset (use http://www.stata-press.com/data/r15/nlswork.dta).
Do not worry if you end up with an unbalanced panel dataset, because Stata can handle both balanced and unbalanced panels without any problem.