You have a "large" number of degrees of freedom (92 or 63.3). Plot some t-distributions with that many degrees of freedom, and you will see that they are all very similar (and also very similar to the standard normal distribution). The attached picture shows the quantiles of the t-distributions for a given level of confidence (95% as well as 99%, two-sided) depending on the degrees of freedom. For values greater than 50 there is not much change, as you can see. The horizontal lines show the respective values of the standard normal distribution (= limiting values for d.f. -> infinity).

How heterogeneous are the variances?  And what are the 3 sample sizes?  ANOVA is very robust to heterogeneity of variance when the sample sizes are equal (or nearly so), so I expect the ordinary contrasts (using the MS_within-groups) would also be robust.  If the sample sizes are quite variable, it becomes more important to deal with heterogeneity of variance.  A little Googling took me to p. 245 in the book linked below for some relevant info.  HTH.

https://books.google.ca/books?id=7ep8FjCcloYC&pg=PA785&lpg=PA785&dq=Myers,+Well,+%26+Lorch&source=bl&ots=RCvXYHxtzA&sig=AwZ0l0kOlxyp4gkKDouhpdFAiuE&hl=en&sa=X&redir_esc=y#v=onepage&q=245&f=false

Thanks for the link Bruce! Since the heterogenity is substantial, I will go with adjusted dfs, even though sample sizes are roughly equal (31 - 32 - 32). The book explained how to do it, and I also think I get logic behind the why now:  with homogeneous variances you can pool the condition variances for contrasts (if they are orthogonal), but since the variances may differ substantially for the different conditions included in the different contrasts, pooling could alter the results and is therefore omitted. Consequently, you use different variances for the different contrasts and adjust each dfs to account for heterogenity.

Lenka, given how similar your sample sizes are, I'd be curious to know how the results for the two methods compare--I would expect them not to differ greatly.  If you can, please post both sets of results here when you've finished.  Thanks.

Here are the results for the contrasts. I just ran them through SPSS quickly, but will need to look at them again in detail.  

homogeneous variances assumed

contrast 1 T(92) = 2.391; p = .019; contrast 2 T(92) = -.610, p = .543

no homogeneous variances assumed

contrast 1 T(63.3) = 2.447, p = .017; contrast 2 T(49.4) = -.603, p = .549

The differences in the results are really not that great, but df differs quite noticably. For contrast 2, this is to be expected and probably to a great extend due to the fact that we don't just pool all conditions when variances are inhomogeneous and therefore exclude the first condition and attached df. But even accounted for this, there is quite a gap. 

So if I do understand this correctly, all the advantages of setting up orthogonal contrasts is lost when variances are inhomogeneous - if we can not pool the variances and keep df up, statistically we could just as well run independent t-tests for comparing the conditions - correct? Or did I miss something? 

I would be interested in your comments.

Cheers

You have a "large" number of degrees of freedom (92 or 63.3). Plot some t-distributions with that many degrees of freedom, and you will see that they are all very similar (and also very similar to the standard normal distribution). The attached picture shows the quantiles of the t-distributions for a given level of confidence (95% as well as 99%, two-sided) depending on the degrees of freedom. For values greater than 50 there is not much change, as you can see. The horizontal lines show the respective values of the standard normal distribution (= limiting values for d.f. -> infinity).

Thanks for posting those results, Lenka.  I am not surprised that the results for the two approaches are so similar, especially since the adjusted df are still quite large (as noted by Jochen).  In the old days (before desk-top computers & stats packages), you likely would have actually used a table of the standard normal distribution, and as Jochen's figure shows, you would not have been off by very much in doing so.  HTH.

Hi, thanks Jochen for this illustration! This gives me some background on what I saw in my data. So practically the loss in dfs does not make a lot of difference (not with my sample sizes anyway) and my results support the notion that whichever way I handle this, the results stay put - which is a good thing in my book! From a theoretical viewpoint, I will go with the adjusted values, since the variances are inhomogeneous. Going beyond my original problem, I still would be interested if statistically helmert contrasts and the appropriate t-tests for inhomogeneous variances are identical when we adjust variances and df for both anyway (while they do differ for homogeneous variances since we use pooled variances / dfs - even though this difference might be marginal as in my case). This goes beyond my original question, but now I am curious and will need to follow up on this.