3
$\begingroup$

I have an experiment with 10 subjects.

Each of them has to exert force and real-time feedback is received. Each subject experiences all 3 types of feedback (within-subject factor).

Moreover, each type of feedback may be filtered either at 2 or 5 Hz. Again, each subject experiences both filtering frequencies (within-subject factor).

Finally, for each feedback type and filtering frequency, the participant must follow a trace with one of two mean force levels. Once more, both mean force levels are presented to every subject for every feedback type and frequency combination (within-subject factor).

Should I implement a 3-way ANOVA or a Linear Mixel Model (e.g. outcome ~ feedback * frequency * force + (1|subject))? Or something else?

$\endgroup$
4
  • $\begingroup$ Do the subjects all receive the same combinations of feedback, frequency and force, and each subject only once for each combination? $\endgroup$ Dec 6, 2019 at 17:53
  • $\begingroup$ @RobertLong Do the subjects all receive the same combinations of fedback, frequency and force? -> Yes. All of them (3 feedbacks * 2 frequencies * 2 forces = 12 combinations for each subject). BUT each subject receives 2 times each combinations, so there are in total 24 trials per subject. $\endgroup$
    – Luisda
    Dec 6, 2019 at 18:04
  • $\begingroup$ OK, and are you interested in any change between measurement occasions ? $\endgroup$ Dec 6, 2019 at 18:47
  • 1
    $\begingroup$ @RobertLong No. Different trials are just meant as a mean to increase statistical power. $\endgroup$
    – Luisda
    Dec 6, 2019 at 19:15

1 Answer 1

3
$\begingroup$

Since the measurements are repeated for each subject, there is likely to be correlation within subjects because the measurements are not independent. A mixed effect model with random intercepts for subjects is a good way to model these data, whereas a 3-way ANOVA would lead to biased results.

$\endgroup$
5
  • $\begingroup$ The model you described would be coded as error ~ feedback * frequency * force + (1|subject)? I say this because of the doubt about the variables being nested $\endgroup$
    – Luisda
    Dec 6, 2019 at 23:04
  • $\begingroup$ If error is your outcome variable then yes. These are no nested random effects since you have only 1 grouping variable, but you have clustered data because of the repeated measures. $\endgroup$ Dec 7, 2019 at 3:54
  • $\begingroup$ I fitted the LME model as recommended by you. When checking for model soundness, I found the residuals to be really large and also their variance depends linearly on the x values. Because error is the outcome variable and is a RMSE, it follows a Chi-squared distribution. Is that affecting my residuals? Thanks for any help $\endgroup$
    – Luisda
    Dec 9, 2019 at 17:40
  • $\begingroup$ That's a completely different question to do with model fit, so please describe the experimental setup explaining how all the variables including the response arise, in a new question. $\endgroup$ Dec 9, 2019 at 18:26
  • $\begingroup$ I posted a new question treating the topic here $\endgroup$
    – Luisda
    Dec 10, 2019 at 2:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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