MANOVA, multivariate regression or paired t-tests? help please :)

#1
Hi,

I am exploring how symptom expectations after a mild head injury and behavioral intentions change over the first three months post-injury.

The independent/ predictor variable is time. This is an ordinal measurement (T1= 7 days, T2= 1 month, T3= 3 months)
The dependent variables are:
- Symptom expectations (an interval measurement gained from a symptom questionnaire giving a score from 0 to 72; Likert-scale from "no more of a problem" to "a severe problem")
- behavioural intentions (an interval measurement from a 39 item questionnaire asking for the extent to which participants would intent to engage in various behaviour after experiencing a mild head injury; Likert-scale from "much less than before the injury" to "much more than before the injury")
----The behavioural intentions scale is sub-divided into active behaviours and rest behaviours

I am looking to explore whether behavioural intentions follow a pattern of a gradual return to pre-injury levels across the three time points.

I am also looking to measure whether there is a relationship between expected symptoms and behavioural intentions at the three time points.

Does anyone have any advice on which statistical tests would be appropriate for these research questions and data? I have gotten myself very confused and my Uni's statistical support service is closed until January.
 

Karabiner

TS Contributor
#2
I am looking to explore whether behavioural intentions follow a pattern of a gradual return to pre-injury levels across the three time points.
You can do this using a repeated-measures analysis of variance.
I am also looking to measure whether there is a relationship between expected symptoms and behavioural intentions at the three time points.
Taken literally, you could perform three correlational analyses or three simple linear regressions, one for each time point.
But if you want to analyse thse associations across time points, you should consider multilevel modeling.

With kind regards

Karabiner
 
#3
You can do this using a repeated-measures analysis of variance.

Taken literally, you could perform three correlational analyses or three simple linear regressions, one for each time point.
But if you want to analyse thse associations across time points, you should consider multilevel modeling.

With kind regards

Karabiner
True a repeated ANOVA could work, but it could contribute to inflated Type I errors. Multilevel modeling is most likely the more robust option.
 

fed2

Active Member
#4
Taken literally, you could perform three correlational analyses
I had a similar issue recently dealing with this sort of issue of estimating correlation in a repeated measures setting. I was just going to mention that I found the multilevel modeling approach mentioned above computationally unweildy, at least in SAS. I also saw some references thatsuggested the partial correaltion (adjusted for subject) as being a 'good' notion of repeated measures correlation. Not sure if anyone has a developed preference on that issue?

Computing the correlation per time point is probably the 'kiss' solution though for sure, just my two cents.
 
#5
I had a similar issue recently dealing with this sort of issue of estimating correlation in a repeated measures setting. I was just going to mention that I found the multilevel modeling approach mentioned above computationally unweildy, at least in SAS. I also saw some references thatsuggested the partial correaltion (adjusted for subject) as being a 'good' notion of repeated measures correlation. Not sure if anyone has a developed preference on that issue?

Computing the correlation per time point is probably the 'kiss' solution though for sure, just my two cents.
Good point about SAS; SPSS and Stata do a more efficient job.