2
$\begingroup$

Context:

  • I am running a large survey and am trying to figure out what statistical analyses to perform.

  • The survey asks questions in every format known to man (e.g., likert scales, multi-choice, dichotomous) and is about psychosocial health and work stress.

  • I plan to run a factor analysis to see what demographic questions factor together to effect quality of life and my supervisor has suggested then doing Pearson correlations on the factors that get thrown up. Then running n-way MANOVA to determine % of variance explained. Despite having aced stats up til now, I'm finding my old brain is confused by all this and somehow the order seems backwards.

Question:

  • If you had a huge data set and you expected lots of the questions to come together to be influential factors on a dependent variable (quality of life in this case) what would you do and in what order?
$\endgroup$
5
  • $\begingroup$ can you qualify huge? Different fields have different definitions of this term. $\endgroup$ Sep 29, 2011 at 14:36
  • 1
    $\begingroup$ IMHO that really matters is why do you need either of them? And your factors are constructed not to be orthogonal to each other combinations? Are you trying to construct just a some sort of index for separation purposes or do you need to implement rotation of the factor space to have a nice interpretation? $\endgroup$ Sep 29, 2011 at 22:06
  • $\begingroup$ @richmorrisroe Huge = more than 100 variables altogether $\endgroup$
    – Pundy
    Sep 30, 2011 at 1:50
  • $\begingroup$ @DmitrijCelov I think I need a nice interpretation. I am really lost about this. $\endgroup$
    – Pundy
    Sep 30, 2011 at 1:53
  • $\begingroup$ @rolando2 I'll make two groups from the upper and lower qunitles scorers on my quality of life and psychsocial health measures. Was thinking to see if the factors were different for these two - risk vs protective maybe. $\endgroup$
    – Pundy
    Sep 30, 2011 at 1:53

1 Answer 1

1
$\begingroup$

I don't see the applicability of MANOVA, since it is designed to identify group differences, and I haven't heard you say that you have any groups to compare.

Correlations among factors would be useful as a side analysis to help prepare for the next step. What I mean is, if the factors are essentially uncorrelated with one another (perhaps with *r*s <.3) they could be built into a regression to explain the outcome. If they are more strongly correlated, then you would want to avoid the collinearity problem by simply seeing how strongly the outcome correlates with each factor: no regression, no partial control for one another, because the results would be too muddied.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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