Since you want to see if patients develop diabetes, I suppose that at time 1 all patients did not have diabetes (and that they were then randomized to the treatment or control group). If this is so, then time-point 1 is not informative for disease at all and it is sufficient to simply analyze whether or not a patient has diabetes at time 2. (it means that a "change" can only go from "no" to "yes" and not the other way around)

This can be analyzed using a standard logistic model with the response being "diabetes at time 2 (yes/no)".

If the patients could have diabetes at time 1, then their status can change from "no" to "yes" or from "yes" to "no", and it can also remain the same at "no" as well as at "yes". So there are 4 possibilities. I am not sure how this could be analyzed in a smart way. My first idea is to split the analysis into two groups (patients with diabetes at time 1 and patients without diabites at time 1) and procees like described above. The results of these parts will be "comparable" when the covariates in the two groups are similar. If this is not the case, each analysis will stand independently for itself and refer to the respective values of the covariates.If you look for a way to statistically adjust for the covariates between the groups, I would suggest you really go to a local statistician and discuss your experiment and your problem in detail.

Hi Jiali,

I am assuming the "Change Model" you are referring to is the "Pre and Post intervention" kind of scenario. If this is the case, you can try ANCOVA (which tends to investigate whether the post-test means, adjusted for pre-test scores, differ between the two groups) or you can perform a simple t-test with the treatment variable as your independent variable and pre-post change scores as the dependent variables, since this cannot be treated as repeated measures or a longitudinal study.

You can also check out what others had to say in the forums in the links:

I have a document that might help you.

I hope this helps.

Regards

http://stats.stackexchange.com/questions/3466/best-practice-when-analysing-pre-post-treatment-control-designs

http://www.theanalysisfactor.com/pre-post-data-repeated-measures/

Hi,

I guess the answer depends on what exactly your treatment is. Is it a cure or a pre-treatment to avoid one to develop diabetes ? (I admit I don't know much about diabetes but hopefully a little more about statistics !)

If first, I woud perform a logisitc regression on diabetes-stricken people at time t = 0, include control variables and of course your dummy indicating whether one received the treatment between t = 0 and t = 1.

If second, the same but on non diabetes-stricken people at time t = 0.

You can also have a look at this (http://www.stat.columbia.edu/~gelman/arm/chap9.pdf) and adapt the general theory concerning linear regression model to your issue.

I guess what is causing you troubles is that you want to study several relations at the same time (0 -> 0, 0 -> 1, 1 ->0, 1->1). 

Hope this helps !

Since you want to see if patients develop diabetes, I suppose that at time 1 all patients did not have diabetes (and that they were then randomized to the treatment or control group). If this is so, then time-point 1 is not informative for disease at all and it is sufficient to simply analyze whether or not a patient has diabetes at time 2. (it means that a "change" can only go from "no" to "yes" and not the other way around)

This can be analyzed using a standard logistic model with the response being "diabetes at time 2 (yes/no)".

If the patients could have diabetes at time 1, then their status can change from "no" to "yes" or from "yes" to "no", and it can also remain the same at "no" as well as at "yes". So there are 4 possibilities. I am not sure how this could be analyzed in a smart way. My first idea is to split the analysis into two groups (patients with diabetes at time 1 and patients without diabites at time 1) and procees like described above. The results of these parts will be "comparable" when the covariates in the two groups are similar. If this is not the case, each analysis will stand independently for itself and refer to the respective values of the covariates.If you look for a way to statistically adjust for the covariates between the groups, I would suggest you really go to a local statistician and discuss your experiment and your problem in detail.

Normally it would be straightforward to do this as a generalized linear mixed-effects model with time*treatment as independent variables. (e.g., if the outcome variable were instead correct/incorrect responses on some questions, and you wanted to see if the treatment made people perform more accurately.) But in this case that is probably not appropriate, because there is an extra dependency in the data: if someone already has diabetes at time=0, then they are guaranteed to still have it (I assume?) at time=1. So I think Geoffrey's suggestion is probably the best. I'm not very familiar with this research area, but my naive impression is that it seems like the best option is to only analyze people who did not already have diabetes at time=0, and then just use logistic regression on their outcomes at time=1, with independent variables being Treatment and whatever other control variables/covariates are already known to affect their likelihood of contracting diabetes.

Above I was assuming that time 1 is a "pre-treatment" and time 2 a "post-treatment" time. But maybe I've misunderstood, maybe time 1 is like a year after the treatment and time 2 6 years after, or something. In such a case, maybe another option (again, I'm not familiar with this research area, so maybe this is crazy) would be to instead treat the outcome as an ordinal variable (1=never developed diabetes; 2=developed diabetes by the second time point; 3=developed diabetes already by the first time point) and use ordinal regression with Treatment (and other covariates) as the IV; then a negative coefficient for Treatment would indicate that treatment lowered risk of diabetes (and looking at the threshhold coefficients would tell you more specifically if it delayed diabetes from t1 to t2, or from t2 to never).

The best option, of course, is to check out related studies in this research area and see what they did. This particular kind of study design is a bit unfamiliar to me, but I assume it is probably not unusual in your research area, and maybe techniques have been developed to analyze these kinds of data.

If i understand it correctly, you wanted to know the chances of developing diabetes prospectively. If this is your case, at time t0 you drop study participants who have already diabetes and start with the participants that are free from diabetes, and look at time t1, whether they are cases or non cases. For modelling this you can either use logistic regression model or a modified poisson model. For adjustment (identification of confounders) you can use graphical methods like DAGs (Directled acyclic Graphs) or you can base on your background knowledge and select dempgraphic, behavioural and socio-economic factors that could be associated with your response and primary exposure variable. These factors you can determine from former similar studies (literatures that are already existing).