I am having the following data (as an example). I am interested in measuring the difference in sentences between several texts in English and in Spanish. I can do that either with Mann-Whitney U and then calculate the R (effect size) or use biserial correlation. The differences between the text however are probably affected by two other variables related to the language. Can I implement these variables in a partial correlation test? Is it even acceptable that the controls are dichotomous in the form that they are in my example?
I managed to use a pcor script with a slight modification (instead of using scale I use pseudoinverse). I do receive a result but I wonder, if there is another method that can give me an effect size while controlling for these variables?
df = data.frame(case = c(1,1,2,2,3,3), language = c("en","es","en","es","en","es"),sents = c(100,200,250,300,350,340),
control_sents = c(10,12,10,12,10,12), control_userpower = c(5,10,5,10,5,10))
source("pcor.R")
pcor.test(df$language,df$sents,df[,c("control_userpower","control_sents")],use="mat")
I think in two AF(finite automaton) deterministic o non deterministc for soluction the problem
How would a finite automaton help me calculate the effect size for a correlation?
the computational power of Fahlman's Recurrent
Cascade Correlation (RCC) architecture to that of finite state automata
I agree Irfan, and that is why I wanted to use r for having the effect size. The issue is that in terms of a partial correlation, it doesn't work when you have two dichotomous variables and one interval. For example
r y z
1 100 55
2 200 58
1 300 55
2 400 58
1 500 55
2 600 58
pearson correlation will give
ry = a number
zy = a number (exactly the same to ry)
rz = 1
So mathematically a partial correlation cannot work.
Is there a way I can implement z in y and then run a correlation?
To give more information. y is the number of words between different texts and z is the number of words between english and spanish that I know should be identical. So z could be perceived as the natural variation between languages and the number of words needed to construct identical sentences.
In order to get the true difference for y I need to adjust it by z(the natural variance between the two languages).
Can I create a metric y.new = y / z and then run a correlation between r and y.new? Or is there another metric meant to adjust a number based on another?
Use codings 0 and 1 for the dichotomous variable, then you may use pearson as well.
- Badges
- Science topic