Is there any downside to using a repeated-measure ANOVA when you have a within-subjects 2-level factor where 1 level is represented by more trials than the other. I know in ANOVA we are comparing means, but does the the tighter standard error on the one level throw off the ANOVA? For example, if my IV is stimulus congruency and I present congruent trials 80 times while only presenting incongruent trials 20 times, can I just run the the ANOVA including all of the trials or should I run the analysis by selecting the same amount of trials for each factor level (i.e. select only 20 of the congruent trials to pair with the 20 incongruent trials)? If this is OK for a strictly within-subjects design, does it cause problems if a balanced or unbalanced between-subjects variable is added to the design?

  • $\begingroup$ What software are you using? There are different types of 'balance' in ANOVA, and the precise requirements can get rather complicated. Some packages may be able to deal with designs that others can't, but any decent package should give an error, or at least a warning, if the data don't fulfil its requirements. $\endgroup$ – onestop Jan 25 '11 at 19:50
  • $\begingroup$ @onestop In the past I have used SAS or SPSS. Currently, I'm trying to transition to mostly using R $\endgroup$ – Matt Jan 25 '11 at 19:54

I think there are ways to do unbalanced repeated measures ANOVA; maybe look at the books "Variance Components" by Searle, Casella, and McCulloch, and "Components of Variance" by Cox and Solomon. It also seems to be in some books about multilevel / hierarchical / mixed effect models (e.g. Raudenbush and Bryk), where it is treated as a special case of a hierarchical / multilevel / mixed effect model.

I'm not sure what programs support unbalanced repeated measures anova ("Unbalanced designs create special difficulties for the analysis of variance."), but most statistical programs support mixed effect models. Is there any reason that would be insufficient?

  • $\begingroup$ @B R I don't know if it will change your answer, but I have edited the question to be more specific. $\endgroup$ – Matt Jan 25 '11 at 19:57
  • $\begingroup$ It won't! I don't know enough about advanced ANOVA techniques to say if that changes anything. My answer could really be summarized as "There may be ways to overcome these difficulties, but why not just use mixed models?" $\endgroup$ – B R Jan 26 '11 at 0:49
  • $\begingroup$ Generally I would say a reason to not just used mixed models is that others are not familiar with the method and its outputs. $\endgroup$ – russellpierce Apr 13 '11 at 14:24
  • $\begingroup$ P.S. @BR, the link you provide is talking about unbalanced mixed designs on the between subjects factor, I think that is a slightly different answer than the question being asked. $\endgroup$ – russellpierce Apr 13 '11 at 14:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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