2

I'm looking to model the effectiveness of an intervention on human behavior. I have control groups and treatments groups for the outcome of interest and pre and post measurements (just 1 for each) for the outcome. The outcome can be calculated as a rate or count variable. I have specific features of the person that may be of interest, so I'm assuming I need to do some regression to hold those constant, but I don't know which regression to use, how to incorporate the treatment/control variable or how to incorporate the post and pre measurements when I look at the regression. Any help with understanding which power analysis to do would be greatly appreciated.

Statsnoob
  • 23
  • 2
  • 1
    Was treatment randomized? If not, what assumptions are you willing to make about the data? See Matthay et al (202) for a guide on which methods correspond to which assumptions. If treatment was randomized, your options open up a lot because you can focus on decreasing variance rather than removing bias. – Noah Mar 02 '24 at 17:09
  • @Noah thank you for your response! I'm not sure if treatment type was randomized, it's an assumption I am willing to make. Is there anything else you'd look for? I'm analyzing this after the experiment was already conducted and have little information on how it was carried out. – Statsnoob Mar 04 '24 at 00:36

1 Answers1

1

Beyond first looking at your data descriptively, you would want to do a type of mixed-effects regression. If your outcome is count, mixed-effects Poisson/negative binomial regression are options (glmer and glmer.nb functions from lme4 package if using R).

In your model, participant ID would be a random effects term, treatment type (treat vs. control) and time period (pre. vs. post.) would be fixed effects. It would be useful also to include an interaction term between treatment-type*type period to assess if any changes in pre to post differ between treatment groups.

In most statistical software to do mixed-effects regression, your data will need to be in 'long' format as opposed to the conventional 'wide'.

Jack
  • 334
  • Poisson/negative binomial can both be used for rates as well correct? Oh so you would have two rows for each participant? And then have time period as an independent variable? Also, what kind of things would you look for descriptively? I've looked at correlations, general distributions, etc. – Statsnoob Mar 04 '24 at 00:35
  • @Statsnoobl Yes, both can be used for that type of outcome. Negative binomial would be used when there is overdispersion (variance>mean). See this post for more info on checking for overdispersion: https://stats.stackexchange.com/questions/66586/is-there-a-test-to-determine-whether-glm-overdispersion-is-significant. – Jack Mar 05 '24 at 04:38
  • The problem I am looking at also has other important variables change from pre intervention to post. One in particular measures the amount of times the subjects had a chance to do the behavior we are attempting to change and another looks at the time frame each subject was measured over. How would you adjust for these variables? @Jack – Statsnoob Mar 07 '24 at 03:41