Dear Forum Users,
I am currently performing analysis using a multinomial logit model and would like to perform tests in respect of certain criteria, below. I would be very grateful if anyone was able to suggest solutions to the difficulties that I have encountered in performing such test, and to suggest any relevant tests that they believe I may have missed.
The key issue appears to be finding an alternative to the -mlogtest- command, which does not function with factor variables.
1. Validating the use of the multinomial logit model
I have performed a Brant Test and a likelihood ratio test to prove that my model violates the proportional odds assumption, meaning that an ordered logit model is not appropriate. I have considered the use of a generalised ordered logit model using -gologit2-, which may be more parsimonious and interpretable. However, I am encountering difficult with the application of the command.
2. Testing the independence irrelevant alternatives assumption (IIA) of the multinomial logit model
My research implies that a test of the IIA assumption can be performed by using the Stata command mlogtest, iia. However, as of April 23, 2010, -mlogtest, iia- does not work with factor variables.
Similarly, the Small-Hsiao and Hausman tests, -mlogtest, smhsiao- and mlogtest, hausman-, do not function with factor variables.
3. Testing whether the dependent variable categories can be combined
Similarly this test requires the command -mlogtest, combine- or -mlogtest, lrcom-, which does not function.
4. Model fit
The model fit can be observed using the -fitstat- command. However, the output is particularly muddled with several values provided, see example below. Do any Users have experience in which measure might be preferable for the purposes of interpreting the model fit? I understand based on online commentary that this type of analysis for multinomial logit models is somewhat lacking.
5. Independent variable significance
I intend to test for this using a Wald Test or likelihood ratio test.
I would be grateful for any comments or advice.
Many thanks,
Harrison
I am currently performing analysis using a multinomial logit model and would like to perform tests in respect of certain criteria, below. I would be very grateful if anyone was able to suggest solutions to the difficulties that I have encountered in performing such test, and to suggest any relevant tests that they believe I may have missed.
The key issue appears to be finding an alternative to the -mlogtest- command, which does not function with factor variables.
1. Validating the use of the multinomial logit model
I have performed a Brant Test and a likelihood ratio test to prove that my model violates the proportional odds assumption, meaning that an ordered logit model is not appropriate. I have considered the use of a generalised ordered logit model using -gologit2-, which may be more parsimonious and interpretable. However, I am encountering difficult with the application of the command.
2. Testing the independence irrelevant alternatives assumption (IIA) of the multinomial logit model
My research implies that a test of the IIA assumption can be performed by using the Stata command mlogtest, iia. However, as of April 23, 2010, -mlogtest, iia- does not work with factor variables.
Similarly, the Small-Hsiao and Hausman tests, -mlogtest, smhsiao- and mlogtest, hausman-, do not function with factor variables.
3. Testing whether the dependent variable categories can be combined
Similarly this test requires the command -mlogtest, combine- or -mlogtest, lrcom-, which does not function.
4. Model fit
The model fit can be observed using the -fitstat- command. However, the output is particularly muddled with several values provided, see example below. Do any Users have experience in which measure might be preferable for the purposes of interpreting the model fit? I understand based on online commentary that this type of analysis for multinomial logit models is somewhat lacking.
5. Independent variable significance
I intend to test for this using a Wald Test or likelihood ratio test.
I would be grateful for any comments or advice.
Many thanks,
Harrison
Comment