Choosing the correct analysis ANCOVA/MLM

[joebacon]

I apologize in advance for the simplicity of this question.

My question at it's essence is "what is the correct test to run in this scenario?"

I have been collecting data of the following nature: 

Insufficiently active adults (N = 116, M age=40.1±10.3 years; 78% female; M daily steps = 4854.99 ± 1332.75) were recruited in self-selected teams (N = 24) to receive a web-based intervention with gamification component (TECH+C) or without (TECH). PA and SB were measured on 7 days at baseline and after the intervention (12 weeks) using an accelerometer.

My questions that I would like answered are:

1. How many insufficiently active individuals who were recruited for this study were overly sedentary?
   1. Will report mean, median SD and range of daily sedentary minutes for the total sample, and by group.
   2. Will report the N for a split by 1) median and by 2) 8 hours as cut-off (Loprinski et al., 2016)
2. Did TECH or TECH+C lead to an incidental change in average sedentary time in minutes per day (i.e. group X time interaction)?
   1. Will examine correlations between demographic variables and sedentary time and include all significant correlations as covariates in the ANCOVA.
   2. Normality will be examined and transformations will be done if needed

• Will run a 2 (time) X 2 (group) mixed ANCOVA with sedentary time in min/day as outcome.

1. Did both interventions lead to an incidental change in average sedentary time in minutes per day (i.e. main effect of time, controlling for group)?
   1. An ANCOVA with time as factor and group as covariate.
      1. 2x2 ANCOVA, looking at the main effect of time OR
      2. ANCOVA with only time as factor, and group as covariate.
   2. What are moderators of the interventions?
      1. Does baseline sedentary time (above vs. below the median) moderate the intervention effect (both groups) on average MVPA min/day?
         1. Normality will be examined and transformations will be done if needed.
         2. Sedentary time will be coded as 1 for above the median and 0 for below the median.

• Will determine the interaction effect using regression: MVPAmin = a + b*SedTime + c*group + d*SedTimeXgroup + e

1. Does baseline PA level (above vs. below the median) moderate the intervention effect (both groups) on average MVPA min/day?
   1. Normality will be examined and transformations will be done if needed.
   2. PA level (i.e. step count/day) will be coded as 1 for above the median and 0 for below the median.

• Will determine the interaction effect using regression: MVPAmin = a + b*StepCount + c*group + d*StepCountXgroup + e

I have highlighted my analysis plan questions with how I would like to answer them. First and foremost, if anyone has any recommendations that appear better, please help me learn! Mostly, I am worried about the ANCOVA. is that appropriate here? It seems I have leveled data of the two groups split by intervention measured at two time points. However, people enrolled with sort of a “team” for the program, so we do expect there could be differences amongst these teams for PA, though less potential for sedentary time and it doesn’t seem like there are differences here… but,there were a number of these “teams” within each of the intervention conditions.

Is a HLM more appropriate here to capture the baseline and group differences? Any resources or discourse would help straighten out my mind, but I needed someone to talk to before proceeding which brought me to you bright individuals.

[PaigeMiller]

You don't state what your concerns are about using ANCOVA, but from what I can see, it ought to work properly.

[joebacon]

My biggest concern about using ANCOVA is that there are two time points and two groups (one with the intervention and one without) but within these two groups there were "teams". How do i control for those teams that could potentionally lead to difference within the groups? Or do I just examine their means and collapse them between groups?

[PaigeMiller]

---

@joebacon wrote:
My biggest concern about using ANCOVA is that there are two time points and two groups (one with the intervention and one without) but within these two groups there were "teams". How do i control for those teams that could potentionally lead to difference within the groups? Or do I just examine their means and collapse them between groups?

---

If you think teams have an effect, and you have the team information recorded in the data, then you add the variable that represents team into the model.

[joebacon]

I checked SB min/day by team. The mean at baseline was 597.46 +- 44.17 min/day, ranging from 520.46 to 696.33 min/day. I checked for differences by running an ANOVA (probably not be the optimal approach?). There is a main effect of team, but I am not sure if that has to do with alpha inflation (this ANOVA is comparing 24 groups…)

Df Sum Sq Mean Sq F value Pr(>F)
Team 23 103447 4498 2.774 0.00887 **
TimePoint 1 1053 1053 0.649 0.42861
Residuals 23 37295 1622

If I stay with the ANOVA approach, I will update the results by controlling for baseline PA and team. Any thoughts on this Paige?

[PaigeMiller]

Normality will be examined and transformations will be done if needed

You say this several times, and I just want to point out that it is the residuals after the model fit that must be normally distributed. It doesn't matter whether anything else is normally distributed when you fit such a model.