1

I have 4 groups based on BPD symptoms at baseline and follow-up. (BPD T1 and T2; BPD T1 not T2; no BPD T1 but BPD T2; Never BPD). I am examining whether psychosocial functioning outcomes (SOFAS score) differ amongst the groups.

As I wish to control for Axis I disorders (mood, anxiety, and substance use) and treatment (service 1 or 2) I am using an ANCOVA.

I would also like to control for the baseline measure of the outcome variable (SOFAS at T1).

The Axis I disorders and the treatment are all dichotomous variables; SOFAS scores (both T1 and T2) are continuous.

When using SPSS is it correct to put the dichotomous variables in the "fixed factors"box?

ANCOVA Fixed factors = Groups; mood; anxiety; substance use; service Covariate = SOFAS T1 DV = SOFAS T2.

Does this seem reasonable?

From my readings it seems OK but my supervisor is unsure and wants me to get further advice. So I would appreciate any comments. Thank you

Sarah
  • 323

1 Answers1

1

Your readings are correct.

The only additional thing you need is that you should be making sure that the correlation between SOFAS T1 and the fixed factors is minimal. An ANCOVA is inappropriate when the covariate is correlated with the predictor variables.

Also, you really might want to consider taking those variables that you've currently dichotomized, make them continuous, and do multiple regression instead of ANCOVA. It's unlikely that anxiety is measured as a factor. All of the scales I know of it are minimally ordinal, likely the sum of ordinals (substance and mood may or may not be a factor but it can still be entered in multiple regression).

Community
  • 1
John
  • 22,838
  • 9
  • 53
  • 87
  • I'm not clear on why a lack of correlation is needed for ANCOVA to work. Great point about dichotomization; this will likely ruin the analysis. – Frank Harrell Dec 29 '13 at 14:18
  • The math will "work" but in order for a strict ANCOVA interpretation of the model to be taken (and it's really just an interpretation of the model) one of the assumptions is that the covariate is independent of the predictors. You can examine a model where they correlate but you need to look at substantially more complex interpretation other than this is the effect of x with the covariate removed. Miller and Chapman 2001 is a readable treatment of this (if somewhat flawed). -http://apsychoserver.psychofizz.psych.arizona.edu/JJBAReprints/PSYC501A/Readings/Miller_Chapman_JAP_2001.pdf – John Jan 05 '14 at 16:44