0
$\begingroup$

I am doing my dissertation on the relationship between schizotypy and cognitive functioning. For my main research question, I want to understand if certain subscales of the STA (schizotypy questionnaire) are related to aspects of cognitive dysfunction. There are 3 IVs - the 3 subscales of the STA: Unusual Perceptual Experiences (UPE), Magical Thinking (MT) and Paranoid Ideation (PI) and 3 DVs - 3 measures of cognitive function: Trails B Test, Visual Attention Task, Rhyming Task. All of the IVs and DVs are continuous variables. I also want to test this in both low schizotypes and high schizotypes to compare (I have already identified and coded this aspect).

My hypothesis is that Magical Thinking will predict cognitive dysfunction and the others I am unsure of.

What I am stuck on is whether to run a MANOVA, individual ANOVAs for each, or a regression? I have read through Andy Field's Statistics book but I feel unsure of what to do next.

Improve this question
$\endgroup$
1
  • $\begingroup$ Updated my answer with a couple of links for more reading about truly multivariate regressions and how to implement them. $\endgroup$
    – EdM
    Commented Jun 29, 2021 at 16:02

1 Answer 1

Reset to default
1
$\begingroup$

ANOVA/MANOVA are for categorical predictors. With continuous predictors you need to use a regression model.* Consider whether you need to evaluate interactions among the predictors to test your hypothesis correctly (something that is pretty much built into a multi-way ANOVA). For example: might the association of Magical Thinking with the outcomes depend on the level of Paranoid Ideation?

What you then have is best analyzed as a multivariate (multiple outcome) regression model. The regression coefficient estimates will be the same as for separate regressions for each outcome, but handling the outcomes properly together allows for evaluation of correlations among outcomes and corresponding adjustments of the error estimates for the coefficients.

Issues specific to multivariate regressions are explained in this R Journal article and these notes.

*ANOVA/MANOVA are essentially just regression models with categorical predictors that also include their interactions. There is no fundamental distinction of them from regression, and the same software accomplishes both (perhaps with different displays of results).

Improve this answer
$\endgroup$
5
  • $\begingroup$ Thank you! One more question because I am having trouble clarifying it online/in my textbook. Is a multivariate regression the same as a multiple linear regression? And if not, are they still similar enough that the assumptions tests are the same? I appreciate it. $\endgroup$
    – Emma
    Commented Jun 29, 2021 at 15:25
  • 1
    $\begingroup$ @Emma the terminology is used ambiguously. Strictly, "multivariate" should apply to multiple outcomes, distinguished from multiple regression (more than one predictor variable). In practice, many use "multivariate" when they mean multiple regression with only a single outcome. Significance tests for true "multivariate" regressions take into account the correlations among outcomes (unlike a set of individual-outcome multiple regressions), but otherwise they are like other multiple regressions. See this page for example. $\endgroup$
    – EdM
    Commented Jun 29, 2021 at 15:48
  • $\begingroup$ Thank you again. Do you know of any resources for interpreting the SPSS output for a multivariate regression? I was able to run the test but can't seem to find a good guide on interpreting - particularly where you are looking at the interactions among predictors. Thank you! $\endgroup$
    – Emma
    Commented Jun 29, 2021 at 16:08
  • $\begingroup$ @Emma closest I know is this page for SPSS MANOVA output. That might help with the outcome-correlation part of things (what seems to be new for you), as MANOVA is just multivariate outcomes with categorical predictors. Like other predictors, interaction terms in multivariate multiple regression have the same point estimates as in single-outcome regressions, but the standard errors, confidence intervals, p-values etc are adjusted for outcome correlations. $\endgroup$
    – EdM
    Commented Jun 29, 2021 at 17:33
  • $\begingroup$ Thank you sincerely for all your help, it has made this all so much clearer! $\endgroup$
    – Emma
    Commented Jun 30, 2021 at 11:55

Your Answer

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

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