Hey all researchers,
I've a little query, how can one find the association between the nominal (say as case & control, dependent variable) and different continuous variables (independent variable).
All I can think is binomial regression? Is there any other possibility?
Can Spearman's correlation be used ?
Thank you
If your nominal variable has only two levels, you can use a traditional correlation statistic and test (Pearson, Spearman Kendall, as appropriate).
If yo have one dichotomous variable (case or control) and another continuous variable, you can use the Point-biserial correlation to assess the correlation of these two variables. Logic regression (binary logistic or multiple logistic) can also be used and in fact can help you to control the effect of the other independent variables on your outcome while investigating how one of your independent variables is associated with the outcome. Another simple option would be to use a two sample independent t-test to compare if the mean of your independent variable is different for the two groups( cases vs controls) provided that the test assumption are not violated.
My idea is to use a nonlinear regression model.
Of course there are different types.
Thank you Henok G. Woldu for suggesting the options to exhibit the data,
I've already shown the differences in the mean values of the continuous variables among the cases and controls.
But I am also intended to show the strength of association of continuous variables (there are 7 continuous variables) with case and control.
Thanks
You might also think about a measure of effect size. Traditionally matched with t-tests is Cohen's d. It is essentially the difference in means divided by the pooled standard deviation. (There are a few different variants). This can be related to Pearson's r in cases like yours.
Hello Sal Mangiafico , Thanks for the input sir.
I got few variables with a non-normal distribution (Mann Whitney's test applied), so I understand the effect size is Z square divided by n ?? ? Is that correct??
Thanks
There are a couple of different effect size statistics used for a Mann-Whitney test.
One, as you mention, is called r, and is the z value divided by the square root of the total sample size. This statistic tracks pretty well with Spearman's rho (but with the opposite sign). I probably don't like this measure as an effect size statistic very much. It is true that 0 represents no effect. But the statistic doesn't reach a value of 1, and its maximum depends on how equal the sample sizes are. It's instructive to play with some toy data to see this behavior.
Another two are Vargha and Delaney's A and Cliff's delta. What's nice about VDA is that it is easy to explain. It's the probability of an observation of one group being larger than an observation in the other group. Because of this, 0.5 represents no effect and 0 or 1 represents total stochastic dominance. Because this can be a little confusing to some readers, a variant is called Cliff's delta, with 0 representing no effect, and -1 or 1 representing complete stochastic dominance.
thank you so much Sal Mangiafico .
I need to go through VDA and Cliff's delta.
I may bother you again in case of some queries/doubts. :)
Thank you all :)
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers