1
$\begingroup$

In Andy Field's example for one-way ANOVA with repeated measures, he measures "retching time" for eight different celebrities eating four different gross foods.

He tests things like sphericity, runs one-way ANOVA and post hoc tests.

I have a similar set up but with multiple measures for each subject-factor combination. Imagine in Andy's example that for each celebrity-food pair I'd measured 50 different retching times.

I've seen it suggested, on this site even, to just take the average across these 50 trials. Is this "legal". Shouldn't I have to verify that the variance for each celebrity-food pair is small? If so how? Should these trials go into one large analysis or do I just need to preprocess my data?

As a last cry for help. How is the best accomplished with a program like SPSS?

$\endgroup$
1
$\begingroup$

Not only is the averaging "legal", it's mandatory. Keeping the degrees of freedom that you get from not averaging wretching times is what is illegal.

Why do you think that the variance for each celebrity-food pair needs to be small?

$\endgroup$
  • $\begingroup$ I would think the variance needs to be small for the mean to be a meaningful representative. $\endgroup$ – Alec Jacobson Dec 21 '13 at 6:25
  • $\begingroup$ The variance is the variance and the mean is the mean. The variance of your measurement of the mean should be small but that's strongly determined by how many measures you make (standard error) and, with repeated measures, the strength of the correlation of the effect across subjects. A strong correlation gives you lots of power. Typically you don't run a large enough N in repeated measures designs to get good estimates of raw values, only effects. $\endgroup$ – John Dec 21 '13 at 15:47
  • $\begingroup$ So, let me see if I understand. In this example, we should average across trials and then analyze variances of these means because we're trying to compare effects of different gross foods. This ok because we're not trying to report average retching times for each gross food individually. $\endgroup$ – Alec Jacobson Dec 21 '13 at 18:13
  • $\begingroup$ If you wanted to keep all 50 measurements I think I'm correct in sayings that you can use a linear mixed effects model and include subject as a random effect, as well as the subject x condition term as a random effect. This would allow you to estimate the between subject error, as well as the between subject error within each condition. $\endgroup$ – Martyn Feb 23 '14 at 22:15

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.