Hello,
I want to examine whether a set of categorical variables (scores from a questionnaire, organised into categories) are associated with a set of continuous variables (BMI and waist circumference). I am thinking of using polychoric correlation, but not sure whether this is the most appropriate statistical test.
I would appreciate any thoughts and ideas. Thank you in advance.
Neli
Have you consider combining your ordinal-scored variables into a scale (or more than one scale) that you could use a continuous variable (or variables)?
If you are interested in this option, I have compiled a set of relevant resources:
https://www.researchgate.net/post/How_can_one_help_people_create_simple_scales_from_Likert-scored_items
Hi Neli Pavlova if you collapse your categories you can use point biserial correlation. If you want to see the differences in your anthropometric data for each category, then you can use a Kruskal-Wallis test. If you wish to control for other variables affecting your anthropometric data, you can use a regression approach with one-hot encoding of your categories and the control covariates. (On the other hand, if you want to see how your anthropometric data is associated with your categories, then multinominal/ordinal regression is suitable).
POLYCHORIC CORRELATON is used only for ordinal/ordinal crossing data sets, i.e. Y = ordinal and X = ordinal.
In the present case, you have nominal and quantitative, the appropriate correlation coefficient is point biserial.
Dear Rolando Gonzales and Paul Louangrath ,
Thank you for your suggestions on using point biseral correlation. Since my categorical variable is not binary how would you approach it? It has four ordinal categories.
Thank you in advance for your time.
Best wishes,
Neli
Hi Neli Pavlova , you can
1) binarize your categories,
or
2) Use a Kruskal-Wallis test.
If you wish to control for other variables affecting your anthropometric data, you can use a regression approach with one-hot encoding of your categories and the control covariates.
On the other hand, if you want to see how your anthropometric data is associated with your categories, then multinominal/ordinal regression is suitable.
Converting your ordinal variables to binary would not only be a loss of information but would also require a potentially arbitrary decision about where to set the cut-point(s). I personally would not recommend making choices like that just so your data fit into one particular statistical test.
Hi, Neli,
You describe a ubiquitous measurement and statistical analysis issue in current empirical science!
Sometimes a scale which is ordinal by design is actually binary in practice--but one has to use the correct analysis to see through the fog. Here is a brief paper which clearly focuses on this issue: it has a surprise ending!
- https://odajournal.com/2016/11/11/novometric-vs-oda-reliability-analysis-vs-polychoric-correlation-with-relaxed-distributional-assumptions-inter-rater-reliability-of-independent-ratings-of-plant-health/
There exists a superior, modern option to point-biserial. Here is a short report which makes this fact utterly obvious...
- https://odajournal.com/2015/12/12/1418/
Optimal wishes,
Paul
- Badges
- Science topic
- Similar topics
- Cognitive Science
- Thinking