I am struggling to understand what method of statistical analysis would work best for my data. I have a small sample of repeated measures data (15 participants in total) and I've asked them to rate the strength of emotion conveyed by test sentences on a scale from "not at all" to "moderately" to "strongly", i.e. a 3 point scale. Their responses were recorded as numeric scores between 0 and 100, meaning that I have a dependent variable (= strength of emotion) that I think is more or less continuous. The independent variable is sentence type and has 2 levels.
However, I have tested and plotted this data and found that only some repetitions or their residuals are normally distributed. There are also some scores that SPSS reports as potential outliers, but I am reluctant to remove them since I do not think that one person rating the strength of emotion higher than other people means I should discard their data.
My problem is that I don't know what kind of tests I can use for reliable results. Even though some of the data has normally distributed residuals, I am concerned that the potential outliers would render rANOVAs or paired T-tests untrustworthy. Would they also be an issue for non-parametric tests?
Furthermore, I would like a more omnibus analysis than just separate tests for each sentence, AND I would also like to know if another independent variable, e.g. age, interacts with the other IV and has an effect on the dependent. A two-way rANOVA isn't going to work for my data most likely, and I'm not sure if it would be suitable for an ordinal logistic regression either.
Are there tests or methods that could accommodate multiple IVs and interactions, non-normal data, and potential outliers if the DV is considered continuous?
Pay attention to Type I errors and failures to detect effects (Type II errors). Metric models of ordinal data can systematically produce inversions of effects.
Hello Alina,
If the only factor of interest was the repeated measures dimension, then the Friedman two-way anova by ranks would suffice. However, you mention at least one between-subjects factor (age).
For this type of model, you might consider the aligned ranks test (which doesn't impose normality assumption): https://faculty.washington.edu/wobbrock/pubs/chi-11.06.pdf
Good luck with your work.
When theres no normal distribution at residuals, use generalized linear models.
R has glm().
Pay attention to Type I errors and failures to detect effects (Type II errors). Metric models of ordinal data can systematically produce inversions of effects.
- Badges
- Science topic
- Similar topics
- Psychology
- Emotion