2
$\begingroup$

I am working on a research project to determine the impact of several different interventions on a dependent variable. Over the last couple days, I have been reading about possible ways to approach the problem, but I haven't found anything that exactly fits how I have conceptualized the problem (though perhaps I'm thinking about it in the wrong way).

The easiest way to explain the question is with an example. Suppose I want to determine the longitudinal effectiveness of two new procedures intended to lower blood pressure. I want to know not only whether they're effective, but also which is most effective and if other factors impact the outcome (e.g. covariates like age). Considering the following:

  1. Each patient undergoes only one intervention.
  2. I have 5-10 blood pressure measurements for each patient pre- and post-intervention at uneven time periods.
  3. Since blood pressure fluctuates naturally based on a variety of environmental factors (which I can't control for), I would like to take all of my measurements into consideration rather than using a simple point estimate (e.g. mean, median, etc.).

So far, I have investigated these options, but none of them seem to take into account everything I would like to:

1. Mixed-design ANOVA:

  In this case, I would use the intervention (and potentially, some other elements like patient sex, age, etc.) as the between-subjects factors and time as the within-subjects factor. The issue here is that I would not be able to take all of my measurements into account, I would have to use 1-2 point estimates at different time periods.

2. Interrupted Time Series:

  In this method, the measurements from each individual could be view as a time series. I could use interrupted time series analysis to test if there is a change at or close to the time of intervention. The problem is that this can only be used on a single time series (or possibly one time series from each intervention group), so it would be an incomplete picture of the data.

3. Multiple Analyses:

  I could perform multiple analyses to answer the different portions of my question. In this method, I could potentially use a t-test or non-parametric test (e.g. Wilcoxon) to determine separately if there is a difference pre- and post-intervention for each treatment. I could then use ANCOVA on the post-treatment data (with the pre-treatment mean/median of each sample as a covariate) to determine which of the two treatments was most successful. However, both of these would fail to take into account the variation within each of the subjects.

Note:

Previous literature on this subject has just used point estimates (e.g. median of 3 measurements) for pre-intervention and difference at pre-determined time points (e.g. 1 day, 1 week, 4 weeks, etc.) post-intervention

Is there some other type of analysis/model/test that I haven't found that could answer these questions? Am I thinking about the problem in the wrong way? Perhaps it is okay to use the point estimates and I am just overly worried about the natural fluctuation in the measurement?

Update - here's an example of the kind of data I am dealing with:

╔═══════════╦════════════╦══════════════╦═══════════╦══════════╗
║  Subject  ║ Treatment  ║ Measurement  ║   Date    ║ Pre/Post ║
╠═══════════╬════════════╬══════════════╬═══════════╬══════════╣
║ Patient 1 ║ A          ║           12 ║ 1/1/2018  ║ Pre      ║
║ Patient 1 ║ A          ║           16 ║ 1/23/2018 ║ Pre      ║
║ Patient 1 ║ A          ║           13 ║ 2/4/2018  ║ Pre      ║
║ Patient 1 ║ A          ║           14 ║ 2/7/2018  ║ Pre      ║
║ Patient 1 ║ A          ║           12 ║ 2/21/2018 ║ Post     ║
║ Patient 1 ║ A          ║           10 ║ 3/5/2018  ║ Post     ║
║ Patient 1 ║ A          ║            8 ║ 3/12/2018 ║ Post     ║
║ Patient 1 ║ A          ║            9 ║ 4/15/2018 ║ Post     ║
║ Patient 1 ║ A          ║            7 ║ 5/12/2018 ║ Post     ║
║ Patient 2 ║ B          ║           21 ║ 1/1/2018  ║ Pre      ║
║ Patient 2 ║ B          ║           22 ║ 1/23/2018 ║ Pre      ║
║ Patient 2 ║ B          ║           19 ║ 2/4/2018  ║ Pre      ║
║ Patient 2 ║ B          ║           17 ║ 2/7/2018  ║ Pre      ║
║ Patient 2 ║ B          ║           24 ║ 2/21/2018 ║ Pre      ║
║ Patient 2 ║ B          ║           21 ║ 3/5/2018  ║ Pre      ║
║ Patient 2 ║ B          ║           10 ║ 3/12/2018 ║ Post     ║
║ Patient 2 ║ B          ║           15 ║ 4/15/2018 ║ Post     ║
║ Patient 2 ║ B          ║           13 ║ 5/12/2018 ║ Post     ║
╚═══════════╩════════════╩══════════════╩═══════════╩══════════╝
$\endgroup$
-1
$\begingroup$

I don't agree with your reflection about Interrupted Time Series as there may be multiple time point of Intervention and multiple ( waiting to be discovered) time points of intervention via Intervention Detection schemes here https://docplayer.net/12080848-Outliers-level-shifts-and-variance-changes-in-time-series.html

$\endgroup$
  • $\begingroup$ Do you mind elaborating on where you disagree? ITS would definitely allow me to find the intervention point(s) as you say in your answer. My issue with it is that I haven't found a way to include the data from all of my samples into the ITS analysis. Is there a way to do this? Thanks! $\endgroup$ – gaw89 Jul 3 '18 at 13:45
  • $\begingroup$ It would help me if you actually posted all of your data so I can more fully understand your issue. By the way I am in the Philadelphia area also. $\endgroup$ – IrishStat Jul 3 '18 at 13:47
  • $\begingroup$ "I have 5-10 blood pressure measurements for each patient pre- and post-intervention at uneven time periods." .... This being so vitiates standard time series analysis which requires equally spaced observations. $\endgroup$ – IrishStat Jul 3 '18 at 13:55
  • $\begingroup$ I updated with an example of the data. Obviously, this is not real data, but it approximates the kind of thing that I am dealing with. Also, I am aware that time series requires equally spaced observations, which is precisely why I mentioned that my observations are not evenly spaced. Sorry for not making that more clear in my question. $\endgroup$ – gaw89 Jul 3 '18 at 13:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.