Correct terminology for adding dependant at baseline to model

#1
Hi forum,

I am doing some regression analyses (binomial and multinomial logistic) and I have created a few models. One is corrected for covariates such as demographics, a second is corrected for same covariates and the dependant variable (sick leave frequency) at baseline is included as another covariate in the model. The real determinant is a variable measured a year later. Now, this would create some sort of deviation score, but I cannot find the correct terminology. Is there an existing name for such models? I have to, and find it hard to justify the use of the baseline measurement as a covariate, so I would greately appreciate any help on this matter.

Tim
 

hlsmith

Not a robit
#2
Yes, it is very important to control for the baseline score in non-randomized studies. And even in some randomize studies. The issue may be imbalances in the starting values between subgroups. The example I always give is if I conducted a weight loss study with an implemented intervention and did not randomized subjects. Based on non-ignorability of treatment assignment - I could have more overweight individual in the treatment group. Where these individual have a potential to loss more weight than their peers. Thus you need to control for starting weights otherwise the intervention may appear to have a greater effect than may be true.

Is your study a before and after (e.g., pre-/post-study design) or possible a repeated measures? I am not familiar with a specific term for the controlling for the variable beyond - "controlling for baseline values of the dependent variable" or other general phrasings.
 
#3
Yes, it is very important to control for the baseline score in non-randomized studies. And even in some randomize studies. The issue may be imbalances in the starting values between subgroups. The example I always give is if I conducted a weight loss study with an implemented intervention and did not randomized subjects. Based on non-ignorability of treatment assignment - I could have more overweight individual in the treatment group. Where these individual have a potential to loss more weight than their peers. Thus you need to control for starting weights otherwise the intervention may appear to have a greater effect than may be true.

Is your study a before and after (e.g., pre-/post-study design) or possible a repeated measures? I am not familiar with a specific term for the controlling for the variable beyond - "controlling for baseline values of the dependent variable" or other general phrasings.
Thank you.
I research the association between specific chronic diseases and sick leave. The data come from two questionnaires, one a year after the other. Would I (according to your reasoning) justify using sick leave at baseline by stating how sick leave could be worse (or less severe) for participants with the chronic disease compared to those without it?
 

hlsmith

Not a robit
#4
I guess you also need to establish the ordering of events. Are these new (chronic) conditions or were they present prior to the first survey? If present prior, then you may not need to control for prior sick time when modeling the second survey. You also only have two time points, if nothing has really changed over time, I would fit a model on first instrument data and prior sick use, then apply that model on the more current data to see how well it predicts. There may be a little issue in that since the model would have been built using some of the same peoples' data while others are lost or new. But if this is just for internal use, it should function to supply some association answers.
 
#5
It is not much of a question what model I will use, since that has already been decided. I just require help on the interpretation of adding the baseline outcome measurement to the model, and why that is a good thing to do.

The chronic diseases were present (or not) at baseline. I will still correct for baseline though.