Multiple linear regression vs logistic regression

#1
I am not a statistician myself, and haven't done regressions for very long time and have a basic question.

I need to do multiple regression for my little research project. I have 4 continuous variables as x variables that are supposed to predict customer's satisfaction, which is measured on the scale 1 to 5: Very happy, Happy, Neither, Unhappy, Very unhappy.

I have records of about 10000 customers to use. I was going to do multiple linear regression, however, I am reading a research project that is almost identical to what I want to do, the y variable is measured on a similar scale, but the statistician there (and I know he is a statistician with a degree in statistics) used "multiple logistic regression"

I was under the impression that logistic regression is used only when 'y' variable is binary (yes or no, alive or dead...etc). Is it actually possible to use logistic regression for something like this??? if so, how???

Many thanks
 

hlsmith

Not a robit
#4
Your data are ordinal/discrete they are bound and don't take values between the integers. Ordinal logistic regression is the appropriate method for ordinal data.
 

noetsi

Fortran must die
#5
With only five distinct levels linear regression almost certainly will generate violations of assumptions such as homoskedacity so ordinal logistic regression is your only real choice. Well you might collapse the categories into happy/not happy (removing the neither responses) and run binomial logistic regression instead. This has the advantage of being easier to interpret in some cases and may be neccessary as ordinal regression makes an assumption that logistic regression does not [commerical software test for this if you ask for it].
 

CB

Super Moderator
#6
the statistician there (and I know he is a statistician with a degree in statistics) used "multiple logistic regression"
It may be that you are thinking of multinomial logistic regression. This is logistic regression with a categorical dependent variable that takes more than two values. (Conceptually it's very similar to binary logistic regression). It differs from ordinal logistic regression in that it doesn't take into account your knowledge that the dependent variable is ordered. The resulting model is more complex than an ordered logistic model, but with weaker assumptions. I agree that ordered logistic regression makes sense here, though make sure you do some reading about its assumptions and how to test them.