Repeated Measures ANOVA (Split-Plot ANOVA) analysis of residuals

[Martin Marko]

Hi,

how to process decent residual analysis in repeated measures? I know that the residuals should have normal distribution (altogether an in every level of repeated factor too), but is there any other way to examine the residuals?

I am considering Levene's test to chcek if the variances among factor levels (particular repeated observations) are similar + plot it.

In linear regressions, I used to plot residuals to predicted values, but since in ANOVA the "predicted" are repeated means, how to adjust the analysis? I think about making absolute values of residuals (or not absolute?) and test all pairs by t-test (or other ANOVA for residuals).

Any help would be perfect, any idea how to enhance the analysis would be priceless!
Thank you,

M.

PS: would be the analysis the same in between-subject design (the same as repeated).

 

[Miner]

Depending on what your particular SW package will allow, try plotting the residuals by each time period of the repeated measures and by subject. If the SW will not do it directly, try storing the residuals first.

 

[Martin Marko]

So.. plot Factor levels (X axis) and Residuals (Y axis) and chceck if the error bars are equal? Is it okey to use Levene's test testing this (One-way ANOVA + Homogenity)? I have acess to SPSS and posibly Statistica 10 (bud I work in SPSS).
Thank you

 

[Miner]

Yes. You would typically use Bartletts test unless the distributions are nonnormal. Then use Levene's.

 

[Martin Marko]

Yes. You would typically use Bartletts test unless the distributions are nonnormal. Then use Levene's.

Okay
... normal data = Bartletts
... non-normal is for Levenes.

Just last one question, may I use Bartletts test from Factor Analysis, since (as I know) there is no other accessable Bartletts test. (Put repeated factor variables into as "items" and check the Bartletss Test chi statistics).

Thank you for your help.
Best regards,

M.

 

[Miner]

I am not famililiar with SPSS or Statistica, so I cannot answer that question though I am doubtful about it.

You can use Levene's test on normally distributed data, so I recommend using it rather than try the other approach. Be aware that Levene's is more conservative than Bartlett's, so it might reject the null when Bartlett's would not reject it. The math is not overly complex, so you might want to run Bartlett's in a spreadsheet to verify if Levene's rejects the null. If Levene's fails to reject, there is no need.