Stata Technical Bulletin
19
Saved results
koopman and koopmani save in r():
Scalars
r(χ)
r(m)
r(y)
r(n)
r (theta)
r(theta-l)
r(theta-u)
number of events in group 1
total number of subjects in group 1
number of events in group 2
total number of subjects in group 2
odds ratio between group 1 and 2
lower bound of CI for theta
upper bound of CI for theta
References
Bishop, Y. M. M., S. E. Fienberg, and P. W. Holland. 1977. Discrete Multivariate Analysis. Cambridge, MA: MIT Press.
Koopman, P. A. R. 1984. Confidence intervals for the ratio of two binomial proportions. Biometrics 40: 513-517.
sg155 Tests for the multinomial logit model
Jeremy Freese, University of Wisconsin-Madison, [email protected]
J. Scott Long, Indiana University, [email protected]
Abstract: The command mlogtest designed to simplify the use of several tests associated with the multinomial logit model is
introduced and illustrated.
Keywords: multinomial logit model, Hausman, likelihood-ratio test, Wald test.
Introduction
There are several tests that are commonly used in association with the multinomial logit model (MNLM hereafter). First,
we can test that all of the coefficients associated with an independent variable are simultaneously equal to zero (that is, test that
a variable has no effect). Second, we can test whether the independent variables differentiate between two outcomes; this test
is commonly used to determine if two outcomes can be combined. Third, we can assess the assumption of the independence of
irrelevant alternatives (IIA) using either a Hausman test or the LR test proposed by McFadden et al. (1976) and improved by
Small and Hsiao (1985). While each of these tests can be computed using either the test, Irtest, or hausman commands in
Stata or the smhsiao command of Nick Winter (available at the SSC-IDEAS archive), in practice computing these tests can be
awkward and/or tedious. The mlogtest command is designed to simplify the use of these tests. mlogtest is a post-estimation
command that requires that mlogit is the last model estimated.
Given the difficulties of interpretation associated with the MNLM, it is tempting to search for a more parsimonious model by
excluding variables or combining outcome categories based on a series of statistical tests. While mlogtest facilitates computing
tests that can be used in a specification search, great care is required. First, these tests all involve multiple coefficients. While
the overall test might indicate that as a group the parameters are not significantly different from zero, an individual parameter
can still be substantively and statistically significant. Accordingly, one needs to carefully examine the individual coefficients
involved in each test before deciding to revise a model. Second, as with all searches that use repeated, sequential tests, there is
a danger of overfitting the data. When models are constructed based on prior testing using the same data, significance levels
should only be used as rough guidelines.
Syntax
mlogtest [, detail iia hausman Ir wald combine Ircomb smhsiao set Varrlstt [∖ varlist...] ) all base ]
Options
detail reports the full hausman output for the IIA test. The default is to provide only a summary of the results.
iia specifies that both tests of the IIA assumption should be performed.
hausman requests Hausman tests of the IIA assumption.
Ir requests that LR tests for each independent variable should be performed.
wald requests that Wald tests for each independent variable should be performed.
combine requests Wald tests of whether dependent categories can be combined.