Significance tests for binary logistic regressions

#1
Hi everyone,
I am currently working on my master thesis in finance and I am using a binary logistic regression model to analyse my data.

I have read through a lot of guides and manuals on how and when to use it and also how to interpret the findings. However there is one thing I have not quite understood so far:
There are several significance tests and best fit models that available for logistic regressions (Hosmer and Lemeshow, Omnibus, Pearson, Deviance, likelihood ratio, AIC, BIC, Wald).
SPSS does not provide AIC, BIC and likelihood ratio test for BINARY logistic regressions, only for MULTINOMIAL regression models. However I have read several articles which also use a BINARY model but do report one or more of those three statistics.

Now my question is: Are these three statistics (AIC, BIC, likelihood ratio) suitable for a BINARY logistic model or not? Or are they only valid for multinomial logistic models?

I hope this was not too confusing and I would be very glad if someone could clarify this for me.

Thank you very much,
Captain
 
#2
Hm...no answer so far.

Is it that no one knows an answer or was my question unclear?
I would really appreciate some help.

So if you know anything about this, please let me know. It would be great help!

Cheers,
Captain
 
#3
constuct

Hi,
I try to help u but I not strong in English language.Please attention to read.
Logistic regression use when Outcome variable is Dichotomous or binary analyze with several independent variable both type(continuous, categorical)
The logistic regression consist of several step.
1.Univariate -> analyze each independent var with Outcome var
- To find Confounding effect and Interaction effect if both case is occur and how to handle please find it on internet
2.in step one u will know variable that u will use to fit model next can fit model and this step have several sub-step u can find in "hosmer and lemeshow (Applied logistic regression)
3.When u report does not serious about numerical about statistic but u should
explain relation of each variable and method that u use such as OR.When u have equation to predict u shoud simplify it
The STATA program present the AIC BIC Wald and Ratio,I suggest u to use STATA
 
Last edited:
#4
alamp,
thank you for your answer. But I know how the regression works and how to interpret it. My problem is that I am not sure if the tests are valid for BOTH, binary AND mutlinomial regressions. Especially since researchers use different tests in different articles...(the probably use the one with the most supportive results).

Anyways...thank you for your effort, I appreciate it.

Captain
 

terzi

TS Contributor
#5
This is an old post but the question is quite interesting.

First of all, Logistic Regression is a model estimated via Maximum Likelihood, therefore you can forget about R^2, F ratio and all those beautiful measures from OLS regression.

As you may know, several measures exist for this kind of models, and the one you use will depend on the comparisons you want to make. In general, the most reliable measure for comparing models is the Likelihood ratio test. Most works will use this value when comparing two models. The sad part of likelihood ratio test is that it will only be valid for nested models and exactly equal data sets (i.e. you cannot compare a model for men against another for women). Also if you estimated a model with Restricted Maximum Likelihood, the likelihood ratio may not be valid. In case you want to compare non nested models, you use AIC or BIC and its respective ratio test. Wald test is a more general approach although it is only reliable with larger samples.

Finally, let me remark that the above is valid for any model estimated through maximum likelihood.