1

I am trying to check the significance of counselling in the technique for inhaler usage (rotahaler only) at different points of time in same subjects.

The subjects' baseline technique will be assessed using a checklist scoring : zero for incorrect/ missed step and 1 for correct step. The total score is 8 (i.e., the checklist has 8 steps). Then subject will be counselled and an immediate post intervention score is taken followed by another assessment after 2 weeks.

I want to compare the means at 3 points in time. Can I use a repeated measures ANOVA here? If not, which test is suitable?

jonsca
  • 1,772
Pari
  • 91

1 Answers1

3

Repeated measures ANOVA makes assumptions that may not be met; multi-level modeling makes fewer assumptions. In particular, RM-ANOVA assumes sphericity.

There was a thread on this

There is also a nice section on RM-ANOVA vs. MLM and other methods in Hedeker and Gibbons. They write "ANOVA for repeated measures assumes compound symmetry which implies constant variances and covariances over time. Clearly, such an assumption has little, if any, validity for longitudinal data".

Community
  • 1
Peter Flom
  • 119,535
  • 36
  • 175
  • 383
  • That is not true. Repeated measures Anova requires specification of a covariance structure. Sphericity is only one possible choice. In SAS there are many options including AR(1), ARMA(1,1), compound symmetry, Toeplitz and even unstructured if there is enough data to allow it. – Michael R. Chernick Aug 14 '12 at 21:30
  • Wolfinger and Cheng of SAS seem to indicate otherwise. In GLM, there is no "select covariance structure" in MIXED there is. On p. 3 they say:

    PROC MIXED provides you with a variety of possible structures to choose from in addition to the Type H and unstructured matrices used by PROC GLM. These include compound symmetry, autoregressive andother time series structures, random coefficients models,and spatial correlations.

    It also points out other advantages of MIXED, such as using all the data even if some is missing.

     – Peter Flom Aug 14 '12 at 21:47
  • That is exactly my point for repeated measures in the mixed model there are all these possibilities for covariance structure and not just sphericity. Proc GLM has nothing to do with this. – Michael R. Chernick Aug 14 '12 at 21:56
  • PROC MIXED doesn't do repeated measures ANOVA, it does MLM. PROC GLM does repeated measures ANOVA. – Peter Flom Aug 14 '12 at 22:14
  • Check again proc mixed does repeated measures anova for mixed models. – Michael R. Chernick Aug 14 '12 at 22:31
  • Admittedly proc glm does repeated measures for fixed effect models. But where is it required that the covariance structure for the repeated measures must be spherical? – Michael R. Chernick Aug 14 '12 at 22:40
  • 3
    Go argue with the SAS authors and with Hedeker and Gibbons, who I cited. I'm done. – Peter Flom Aug 14 '12 at 22:40
  • OK, the site is asking us to stop. So, I'll stop. – Peter Flom Aug 14 '12 at 22:41
  • 2
    Before I stop and possibly take this to discussion I must state that proc mixed does repeated meausres ANOVA for the mixed model. You can check this in SAS Help. – Michael R. Chernick Aug 14 '12 at 22:45
  • 2
    (+1) I don't know anything about SAS but I think the contended portion of this answer is true - see here. – Macro Aug 14 '12 at 23:12
  • @PeterFlom The article linked compare the way proc glm and proc mixed do repeated measures ANOVA. So the paper confirms what I said the proc mixed does repeated measures ANOVA for mixed models. Regarding sphericity the authors suggest testing for sphericity. They do not say that sphericity nis a requirement for repeated measures ANOVA. – Michael R. Chernick Aug 15 '12 at 01:09
  • I have done more reading on this subject and the Wolfinger and Cheng article provides an excellent illustration of how the glm and mixed procedures deal differently with repeated measures. For univariate repeated measures models it appears that sphericity is required and tested for. But there are adjustments that can be made when the test for sphericity fails. So I still think that repeated measures ANOVA can be applied in this case regardless of the covariance structure. But I was wrong about not needing the assumption and also have chosen to delete my answer. – Michael R. Chernick Aug 15 '12 at 15:35
  • I think we are just arguing about terms, then. To me (and Hedeker, and others I've seen) the models fitted by MIXED should not be called ANOVA. ANOVA is part of the general linear model, while MIXED is part of the generalized linear model (although that terminology also varies). – Peter Flom Aug 16 '12 at 05:59