I advice the cooperation with a local statistician.

It should be some kind of hierarchical (multilevel) model that takes into account that values within patients are correlated (and not independent). I don't know what the response variable is, so it's hard to say if the normal assumption makes sense or if one should better analyze odds ratios or something else. The comparisons might then be adjusted for covariables you have, and the relationships might not be linear, and there might be interactions (at least the time:group interaction is your scientific target, so this must be part of the model). There is a lot to think about, too much to be discussed via RG.

You may go for paired t-test. If there is paired matching, then you may go for conditional logistic regressions as well.

I advise a ANCOVA, i.e. a regression:

post = pre + group (intervention/control)

One important assumption in this case is there should not be an interaction between pre*group. This is the equality of slopes assumption.

I advice the cooperation with a local statistician.

It should be some kind of hierarchical (multilevel) model that takes into account that values within patients are correlated (and not independent). I don't know what the response variable is, so it's hard to say if the normal assumption makes sense or if one should better analyze odds ratios or something else. The comparisons might then be adjusted for covariables you have, and the relationships might not be linear, and there might be interactions (at least the time:group interaction is your scientific target, so this must be part of the model). There is a lot to think about, too much to be discussed via RG.

I suggest this well tested method. Fit an ARIMA model to the pre-intervention data. On the basis of this make forecasts for the post-intervention period. The difference between the post-intervention data and forecasts should be modelled for the intervention model transfer function. The intervention model is now obtained. A graph of the intervention model should be superimposed on the time plot. The significance of the model is an indication of the model adequacy.

Geeta Pradeep,

As suggested above, I think that ANCOVA would be most appropriate (more so that any kind of t-tests) because it will permit you to take into account differences in your dependent variable(s) prior to the intervention. As also suggested above, it's important to ensure that the assumptions behind ANCOVA have been satisfied.

Regression analyses would be more robust by using structural equation methods that do not rely on questionable significance that can be limited by power or sample size.

Jochen Wilhelm

Thank you for your advice, I am considering this.

The responses variables are

Clinical -patient HbA1C values, BP, cholesterol values; as well as;

Humanistic - their knowledge of diabetes and satisfaction of treatment; before and after intervention in the intervention group.

Control group is a set of clinical outcomes (as above) for a cohort of patients (n=30) at the same practice with same underlying condition. No pre and post measurements in this case as there was no intervention applied.

The Control group also has no humanistic outcomes measured.

Which is why I was thinking of paired sample t-test between pre ad post-intervention variables and

an independent t-test for comparing intervention and control variables.

The above was based on statistical advice but the expert is no longer involved in the research.

My concern was to identify whether the above options are sufficient to analyses the rate of change between the various arms in the study: pre vs post intervention and intervention vs control/usual care.

Regards

Geeta

Can you say more about how people were chosen for the two groups?

Have you considered propensity matching? And Jochen's suggestion to get a local stats person to help is wise!

If the time span between pre and post is long enough to itself have some effect on the outcomes, lacking pre-post data from the control group can be a problem. Another problem could arise if something the patients encounter during the pre-period affects the outcome during the post-period ("carry over", "traning effect" ect.). Was this discussed with the statistician who was involved initially?

Daniel Wright

Sample was chosen from among people with Type 2 diabetes (T2DM) in a practice site with glycated haemoglobin (HbA1C) >7.5%, using at least one oral antidiabetic medication and who were willing to give consent.

The control group formed people from the same practice, with the same characteristics as above except that they received usual care at the practice. This sample was provided as deidentified sample using the practice database.

The intervention was having a pharmacist involved in diabetes medicines management for the intervention group. N=30 in both groups.

Geeta

Jochen Wilhelm

No, this was not discussed with the statistician..the study duration was 15 months.

The pre and post clinical values were collected using the intervention date as the index date which varied for different people as different people gave consent at different times.

So values taken immediately prior to intervention - pre-intervention data and;

values collected after intervention date- post-intervention.

Control data were collected at only one-time point and as there was no exposure to the intervention, there was no pre and post values.

Hope this makes it clear. The study was designed so as to make it as similar to real world practice and none of the clinical parameters were requested specifically for anyone- values obtained were what their prescribers requested as part of medical care given in both groups.

I must confess that I am a novice at statistics and did not really understand the concerns you have highlighted about the time effect in intervention and 'carry over'

Thanks so much, I sincerely appreciate your advice

Geeta

Jochen Wilhelm

Would it appropriate to consider the clinical values obtained for each case in the control before the actual intervention date as the pre-control and the ones collected after the intervention date as post control even though the control cases were not exposed?

Thanks again

Geeta