# To chi square or not to chi square?

I am running a frequency count on the instances of powerless of language in the language of prosecuting attorneys and the same frequency count of the defendants language.

I am asking if the frequency of the powerless language by women makes a favorable verdict and less powerless language by men creates a favorable verdict. I have 400 cases/transcripts.

Below is my hypothesis: Is the use of powerless language used by defense attorneys in arguments independent of the verdict?

a. Is the use of powerless language used by female defense
attorneys in closing    arguments independent of the verdict?
Ha) The use of powerless language by a female defense attorney is
not independent of the verdict.
Ho) The use of powerless language by a female defense attorney is
independent of the verdict
b. Is the use of powerless language used by male defense attorneys
in closing  arguments independent of the verdict?
Ha) The use of powerless language by a male defense attorney is
not independent of the verdict.
Ho) The use of powerless language by a male defense attorney is
independent of the verdict.


Would I run two separate chi squares: one on all of the prosecuting attorneys and one on all the defending attorneys with the verdicts?

-
I don't think a chi-squared analysis is appropriate here. 1st, your data aren't really separate, so you don't want to treat them that way. Prima fascia, you want to use a logistic regression model that can take all variables into consideration simultaneously. However, the key issue in your situation is that for every favorable verdict for the prosecution, the verdict for the defense is necessarily unfavorable. IE you don't have independence here, which makes your situation more difficult / interesting. –  gung Oct 17 '12 at 18:31

I think you should use logistic regression with verdict as the dependent variable and sex of attorney and powerless language as independent variables.

First, you need to consider how to operationalize 'powerless language'. You say you have a frequency count, but it might be better to scale this by the amount of language. Trials last very different times, so perhaps your measure should be "powerless statements per hour" or some such. Call this variable PL (you can call it whatever you'd like).

All your hypotheses are about language of the defense attorney, which simplifies things a bit.

The next independent variable is sex of the defense attorney. Code this, e.g., 1 for female and 0 for male.

Suppose your dependent variable is labeled "V" and can be G or NG (guilty or not guilty).

P(G) ~ PL + S + PL*S

where ~ means "is related to"

Then you use logistic regression.

In R

m1 <- glm(V~PL + Sex + PL*Sex, family = 'binomial')
summary(m1)


In SAS

proc logistic data = mydata;
class sex;
model V = PL|Sex;
run;

-
I've been meaning to provide an answer to this Q, but maybe I should let you since you've already started. LR is clearly the right answer here, but I wonder if you'd expand on this a bit (I suspect the OP isn't very statistically sophisticated). Can you talk a little about the interaction terms she'll need to include, given her data & her goals, & how she'd interpret (eg the signs of) the coefficients that result? –  gung Oct 21 '12 at 0:35
+1, thanks Peter, I think this will be helpful for her. –  gung Oct 21 '12 at 1:06
You're welcome. I hope you are right. –  Peter Flom Oct 21 '12 at 1:11