# Multinomial logistic regression

With SPSS procedure NOMREG (available in versions 9 or higher), you may estimate the influence of variables on a dependent variable with several categories (such as "Brand A", "Brand B", "Brand C", "Brand D"). Note that if these categories are ordered (such as in statements like "agree strongly" ... "disagree strongly"), an ordered logistic regression model usually should be preferred.

*Example*

NOMREG whiskey WITH age BY socclass | |

/ PRINT CPS MFI PARAMETER SUMMARY LRT. |

Here, the choice of a certain brand of whiskey is regressed on a metric variable, age (hence keyword WITH) and a categorical variable socclass (i.e. social class) (hence keyword BY). Note that the very useful keyword CONTRAST, available in procedures such as LOGISTIC REGRESSION or COXREG, cannot be used here. Rather, the last category of the categorical variable is used as a reference category. Therefore, you are advised to code (or recode) your categorical variables according to your needs.

Additional subcommands are available, such as the SAVE subcommand with exactly the same keywords as in the PLUM procedure for ordinal logistic regression. If you wish to include interaction effects in your model specification, subcommand / MODEL (followed by, for instance, age socclass age*socclass) has to be used. Note that other links functions than the logit link are not available. (There is a procedure to estimate probit models, but I have not yet tried this - usually I do this advanced stuff with other software.)

By default, the last category of the dependent variable is used as the base (or reference) category. You may chance this as follows:

NOMREG whiskey (BASE = FIRST) WITH age BY socclass | |

/ PRINT CPS MFI PARAMETER SUMMARY LRT. |

Instead of keyword FIRST, you may also indicate a particular value of the dependent variable. e.g. the number 3. If the dependent variable is a string variable or a formatteed variable (such as date, currency), the value must be put between quotes.

Results of multinomial logistic regression are not always easy to interpret. A clearer interpretation can be derived from the so-called "marginal effects" (on the probabilities), which are not available in the SPSS standard output.

© W. Ludwig-Mayerhofer, IGSW | Last update: 26 Jun 2007