This page shows an example of a multinomial logistic regression analysis with
footnotes explaining the output. The dataset, mlogit, was collected on
200 high school students and are scores on various tests, including a video game
and a puzzle. The outcome measure in this analysis is the preferred flavor of
ice cream – vanilla, chocolate or strawberry- from which we are going to see
what relationships exists with video game scores (video), puzzle scores (puzzle)
and gender (female). Our response variable, ice_cream, is going to
be treated as categorical under the assumption that the levels of ice_cream
have no natural ordering, and we are going to allow SAS to choose the
referent group. In our example, this will be strawberry. By default, SAS sorts
the outcome variable alphabetically or numerically and selects the last group to
be the referent group. The variable ice_cream is a numeric variable in
SAS, so we will add value labels using proc format.

data mlogit; 
  set "C:mlogit"; 
run;

proc format;
value ice_cream_l 
  1="chocolate"
  2="vanilla" 
  3="strawberry";
run;

Before running the multinomial logistic regression, obtaining a frequency of
the ice cream flavors in the data can inform the selection of a reference group.

proc freq data = mlogit;
  format ice_cream ice_cream_l.;
  table ice_cream;
run;

The FREQ Procedure

                  favorite flavor of ice cream

                                       Cumulative    Cumulative
 ICE_CREAM    Frequency     Percent     Frequency      Percent
chocolate           47       23.50            47        23.50
vanilla             95       47.50           142        71.00
strawberry          58       29.00           200       100.00

We can use proc logistic for this model and indicate that the link
function is a generalized logit. This model allows for more than two categories
in the modeled variable and will compare each category to a reference category.
If we do not specify a reference category, the last ordered category (in this
case, ice_cream = 3) will be considered as the reference.

proc logistic data = mlogit;
  model ice_cream = video puzzle female / link = glogit;
run;

Note that we could also use proc catmod for the multinomial logistic regression.
proc catmod is designed for categorical modeling and multinomial logistic
regression is an example of such a model. The options we would use within proc
catmod would specify that our model is a multinomial logistic regression. On
the direct statement, we can list the continuous predictor variables. On the
response statement, we would specify that the response functions are generalized logits. Finally, on the model
statement, we would indicate our outcome variable ice_cream and the predictor
variables to be included in the model. See the proc catmod code below.
This yields an equivalent model to the proc logistic code above.

proc catmod data = mlogit;
  direct video puzzle female;
  response logits;
  model ice_cream = video puzzle female;
run;

The output annotated on this page will be from the proc logistic commands.
The proc logistic code above generates the following output:

The LOGISTIC Procedure
                               Model Information
Data Set                      WORK.MLOGIT
Response Variable             ICE_CREAM             favorite flavor of ice cream
Number of Response Levels     3
Model                         generalized logit
Optimization Technique        Fisher's scoring

Number of Observations Read         200
Number of Observations Used         200

          Response Profile
 Ordered                       Total
   Value     ICE_CREAM     Frequency

       1            1             47
       2            2             95
       3            3             58

Logits modeled use ICE_CREAM=3 as the reference category.

                    Model Convergence Status
         Convergence criterion (GCONV=1E-8) satisfied.


         Model Fit Statistics
                             Intercept
              Intercept            and
Criterion          Only     Covariates
AIC             425.165        404.070
SC              431.762        430.456
-2 Log L        421.165        388.070


        Testing Global Null Hypothesis: BETA=0
Test                 Chi-Square       DF     Pr > ChiSq
Likelihood Ratio        33.0954        6          ChiSq
VIDEO        2        3.4297        0.1800
PUZZLE       2       11.8188        0.0027
FEMALE       2        4.8352        0.0891


                    Analysis of Maximum Likelihood Estimates
                                            Standard          Wald
Parameter    ICE_CREAM    DF    Estimate       Error    Chi-Square    Pr > ChiSq
Intercept    1             1      5.9691      1.4375       17.2425

Model Information

                    Data Summary

Data Set                      WORK.MLOGIT
Response Variablea            ICE_CREAM             favorite flavor of ice cream
Number of Response Levelsb    3
Model                         generalized logit
Optimization Technique        Fisher's scoring

Number of Observations Readc        200
Number of Observations Usedc        200

a. ResponseVariable – This is the response variable in the model. For this
example, the response variable is
ice_cream.

b. Number of Response Levels – This indicates how many levels exist within the
response variable. It also indicates how many models are fitted in the
multinomial regression. In our dataset, there are three possible values for

ice_cream (chocolate, vanilla and strawberry), so there are three levels to
our response variable. In a multinomial regression, one level of the response
variable is treated as the referent group, and then a model is fit for each of
the remaining levels compared to the referent group. Since we have three levels,
one will be the referent level (strawberry) and we will fit two models: 1)
chocolate relative to strawberry and 2) vanilla relative to strawberry.

c. Number of Observations Read/Used – The first is the number of
observations in the model dataset. The second is the number of observations in the dataset
with valid data in all of the variables needed for the specified model. In this
example, our dataset does not contain any missing values, so the number of
observations used in our model is equal to the number of observations read in
from our dataset.

Response Profiles^d

 Ordered                       Total
   Value     ICE_CREAM     Frequency

       1            1             47
       2            2             95
       3            3             58

Logits modeled use ICE_CREAM=3 as the reference category.

d. Response Profiles – This outlines the order in which the values of our
outcome variable ice_cream
are considered. By default in SAS, the last
value is the referent group in the multinomial logistic regression model. In
this case, the last value corresponds to
ice_cream = 3, which is
strawberry. Additionally, the numbers assigned to the other values of the
outcome variable are useful in interpreting other portions of the multinomial
regression output.

Model Fit Statistics and Overall Tests of Effects

                             Intercept
              Intercept            and
Criterione         Onlyf    Covariatesg
AIC            425.165        404.070
SC             431.762        430.456
-2 Log L       421.165        388.070

        Testing Global Null Hypothesis: BETA=0
Testh                Chi-Squarei       DFj     Pr > ChiSqk
Likelihood Ratio        33.0954        6         <.0001
Score                   30.5499        6         <.0001
Wald                    26.8597        6         0.0002

        Type 3 Analysis of Effects
                        Wald
Effectl     DFm   Chi-Squaren   Pr > ChiSqo
VIDEO        2        3.4297        0.1800
PUZZLE       2       11.8188        0.0027
FEMALE       2        4.8352        0.0891

e. Criterion – These are various measurements used to assess the model
fit. The first two, Akaike Information Criterion (AIC) and Schwarz
Criterion (SC) are deviants of negative two times the Log-Likelihood (-2
Log L). AIC and SC penalize the Log-Likelihood by the number
of predictors in the model.

    AIC – This is the Akaike Information Criterion. It is calculated
as AIC = -2 Log L + 2((k-1) + s), where k is the number of
levels of the dependent variable and s is the number of predictors in the
model. AIC is used for the comparison of models from different samples or
nonnested models. Ultimately, the model with the smallest AIC is
considered the best.

    SC – This is the Schwarz Criterion. It is defined as – 2 Log L +
((k-1) + s)*log(Σ f_i), where f_i‘s
are the frequency values of the i^th observation, and k
and s were defined previously. Like AIC, SC penalizes for
the number of predictors in the model and the smallest SC is most
desireable.

    -2 Log L – This is negative two times the log likelihood. The
-2 Log L is used in hypothesis tests for nested models.

f. Intercept Only – This column lists the values of the specified fit
criteria from a model predicting the response variable without covariates (just
an intercept).

g. Intercept and Covariates – This column lists the values of the
specified fit criteria from a model predicting the response variable with the
covariates indicated in the model statement.

h. Test – This indicates which Chi-Square test statistic is used to
test the global null hypothesis that none of the predictors in either of the
models have non-zero coefficients. The test statistics provided by SAS include
the likelihood ratio, score, and Wald Chi-Square statistics.

i. Chi-Square – These are the values of the specified Chi-Square test
statistics.

j. DF – These are the degrees of freedom for each of the tests three
global tests.  Since all three are testing the same hypothesis, the degrees
of freedom is the same for all three.  There are a total of six parameters
(two models with three parameters each) compared to zero, so the degrees of
freedom is 6.

k. Pr > ChiSq – This is the p-value associated with the specified Chi-Square
statistic. Here, the null hypothesis is that there is no relationship between
the any of the predictor variable and the outcome,
ice_cream (i.e., the estimates of
the all of the predictors in both of the fitted models is zero). If the p-value is less than
the specified alpha (usually .05 or .01), then this null hypothesis can be
rejected. In this example, all three tests indicate that we can reject the null
hypothesis.

l.
Effect – Here, we are interested in the effect of of each predictor on the
outcome variable considering both of the fitted models at once.

m. DF –
The degrees of freedom for this analysis refers to the two
fitted models, so DF=2 for all of the variables.

n. Wald Chi-Square –
This is the post-estimation test statistic of the
parameter across both models.

o. Pr > ChiSq – This is the p-value associated with the Wald Chi-Square
statistic. Here, the null hypothesis is that there is no relationship between
the predictor variable and the outcome,
ice_cream (i.e., the estimates of
the predictor in both of the fitted models are zero). If the p-value is less than
the specified alpha (usually .05 or .01), then this null hypothesis can be
rejected.

Analysis of Maximum Likelihood Estimates

                    Analysis of Maximum Likelihood Estimates
                                            Standard          Wald
Parameterp   ICE_CREAMq   DFr   Estimates      Errort   Chi-Squareu   Pr > ChiSqv
Intercept    1             1      5.9691      1.4375       17.2425        <.0001
Intercept    2             1      4.0572      1.2229       11.0065        0.0009
VIDEO        1             1     -0.0465      0.0251        3.4296        0.0640
VIDEO        2             1     -0.0229      0.0209        1.2060        0.2721
PUZZLE       1             1     -0.0819      0.0238       11.8149        0.0006
PUZZLE       2             1     -0.0430      0.0199        4.6746        0.0306
FEMALE       1             1      0.8494      0.4482        3.5913        0.0581
FEMALE       2             1      0.0328      0.3500        0.0088        0.9252

                 Odds Ratio Estimates
                          Point          95% Wald
Effect    ICE_CREAM    Estimatew     Confidence Limitsx
VIDEO     1               0.955       0.909       1.003
VIDEO     2               0.977       0.938       1.018
PUZZLE    1               0.921       0.879       0.965
PUZZLE    2               0.958       0.921       0.996
FEMALE    1               2.338       0.971       5.628
FEMALE    2               1.033       0.520       2.052

p. Parameter – This columns lists the predictor values and the
intercept–the parameters that were estimated in the model. The intercept and
each predictor appears twice because two models were fitted.

q. ICE_CREAM – Two models were defined in this multinomial
regression: one relating chocolate to the referent category, strawberry, and
another model relating vanilla to strawberry. The ice_cream number indicates to
which model an estimate, standard error, chi-square, and p-value refer. We can
refer to the response profiles to determine which response corresponds to which
model. Our ice_cream categories 1 and 2 are chocolate and vanilla,
respectively, so values of 1 correspond to
the chocolate relative to strawberry model and values of 2 correspond to the
vanilla relative to strawberry model.

r. DF – These are the degrees of freedom for parameter in the
specified model.  Since our predictors are continuous variables, they all
have one degree of freedom in each model.

s.
Estimate –
These are the estimated multinomial logistic regression
coefficients for the models. An important feature of the multinomial logit model
is that it estimates k-1 models, where
k is the number of levels
of the outcome variable. SAS treats strawberry as the referent group and
estimates a model for chocolate relative to strawberry and a model for vanilla
relative to strawberry. Therefore, each estimate listed in this column must be
considered in terms both the parameter it corresponds to and the model to which
it belongs. The standard interpretation of the multinomial logit is that for a
unit change in the predictor variable, the logit of outcome
m relative to
the referent group is expected to change by its respective parameter estimate
(which is in log-odds units) given the other variables in the model are held
constant.

Model Number 1: chocolate relative to strawberry

Intercept – This is the multinomial logit estimate for chocolate
relative to strawberry when the predictor variables in the model are evaluated
at zero. For males (the variable
female evaluated at zero) with zero

video and
puzzle scores, the logit for preferring chocolate to
strawberry is 5.9696. Note that evaluating
video and
puzzle at
zero is out of the range of plausible scores. If the scores were mean-centered,
the intercept would have a natural interpretation: log odds of preferring
chocolate to strawberry for a male with average
video and
puzzle
scores.

video – This is the multinomial logit estimate for a one unit increase
in video score for chocolate relative to strawberry, given the other
variables in the model are held constant. If a subject were to increase his

video score by one point, the multinomial log-odds for preferring chocolate
to strawberry would be expected to decrease by 0.0465 unit while holding all
other variables in the model constant.

puzzle – This is the multinomial logit estimate for a one unit
increase in puzzle score for chocolate relative to strawberry, given the
other variables in the model are held constant. If a subject were to increase
his puzzle score by one point, the multinomial log-odds for preferring
chocolate to strawberry would be expected to decrease by 0.0819 unit while
holding all other variables in the model constant.

female – This is the multinomial logit estimate comparing females to
males for chocolate relative to strawberry, given the other variables in the
model are held constant. The multinomial logit for females relative to males is
0.8495 unit higher for preferring chocolate to strawberry, given all other
predictor variables in the model are held constant. In other words, females are
more likely than males to prefer chocolate to strawberry.

Model 2: vanilla relative to strawberry

Intercept – This is the multinomial logit estimate for vanilla
relative to strawberry when the other predictor variables in the model are
evaluated at zero. For males (the variable
female evaluated at zero) with
zero video and
puzzle scores, the logit for preferring vanilla to
strawberry is 4.0572.

video – This is the multinomial logit estimate for a one unit increase
in video score for vanilla relative to strawberry, given the other
variables in the model are held constant. If a subject were to increase his

video score by one point, the multinomial log-odds for preferring vanilla to
strawberry would be expected to decrease by 0.0229 unit while holding all other
variables in the model constant.

puzzle – This is the multinomial logit estimate for a one unit
increase in puzzle score for vanilla relative to strawberry, given the
other variables in the model are held constant. If a subject were to increase
his puzzle score by one point, the multinomial log-odds for preferring
vanilla to strawberry would be expected to decrease by 0.0430 unit while holding
all other variables in the model constant.

female – This is the multinomial logit estimate comparing females to
males for vanilla relative to strawberry, given the other variables in the model
are held constant. The multinomial logit for females relative to males is 0.0328
unit higher for preferring vanilla to strawberry, given all other predictor
variables in the model are held constant. In other words, males are less likely
than females to prefer vanilla ice cream to strawberry ice cream.

t.
Standard Error – These are the standard errors of the individual
regression coefficients for the two respective models estimated.

u.
Chi-Square –
This column lists the Chi-Square test statistic of the
given parameter and model.

v.
Pr > Chi-Square – This is the p-value used to determine whether or
not the null hypothesis that a particular predictor’s regression coefficient is
zero, given that the rest of the predictors are in the model, can be rejected.
If the p-value less than alpha, then the null hypothesis can be rejected and the
parameter estimate is considered to be statistically significant at that alpha
level. The Chi-Square
test statistic values follows a Chi-Square
distribution which is used to test against the alternative hypothesis that the
estimate is not equal to zero. In multinomial logistic regression, the
interpretation of a parameter estimate’s significance is limited to the model in
which the parameter estimate was calculated. For example, the significance of a
parameter estimate in the chocolate relative to strawberry model cannot be
assumed to hold in the vanilla relative to strawberry model.

Model 1: chocolate relative to strawberry

For chocolaterelative to strawberry, the Chi-Square test statistic
for the intercept
is 17.2425 with an associated p-value of <0.0001. With an
alpha level of 0.05, we would reject the null hypothesis and conclude that the
multinomial logit for males (the variable
female evaluated at zero) and
with zero video and
puzzle scores in chocolaterelative to
strawberry are found to be statistically different from zero.

For chocolate relative to strawberry, the Chi-Square test statistic for the
predictor video is 3.4296 with an associated p-value of 0.0640. If we set
our alpha level to 0.05, we would fail to reject the null hypothesis and
conclude that for chocolate relative to strawberry, the regression coefficient
for video has not been found to be statistically different from zero
given puzzle and
female are in the model.

For chocolate
relative to strawberry, the Chi-Square test statistic for
the predictor puzzle is 11.8149 with an associated p-value of 0.0006. If we
again set our alpha level to 0.05, we would reject the null hypothesis and
conclude that the regression coefficient for
puzzle has been found to be
statistically different from zero for chocolaterelative to strawberry
given that video and
female are in the model.

For chocolate
relative to strawberry, the Chi-Square test statistic for
the predictor female is 3.5913 with an associated p-value of 0.0581. If we
again set our alpha level to 0.05, we would fail to reject the null hypothesis
and conclude that the difference between males and females has not been found to
be statistically different for chocolate relative to strawberry given that
video and
puzzle are in the model.

Model 2: vanilla relative to strawberry

For vanilla relative to strawberry, the Chi-Square test statistic for the
intercept is 11.0065 with an associated p-value of 0.0009. With an alpha level of
0.05, we would reject the null hypothesis and conclude that a) the multinomial logit for males (the variable
female evaluated at zero) and with zero

video and
puzzle scores in vanilla relative to strawberry are
statistically different from zero; or b) for males with zero
video and

puzzle scores, there is a statistically significant difference between the
likelihood of being classified as preferring vanilla or preferring strawberry.
Such a male would be more likely to be classified as preferring vanilla to
strawberry. We can make the second interpretation when we view the interceptas a specific covariate profile (males with zero
video and
puzzle
scores). Based on the direction and significance of the coefficient, the
intercept
indicates whether the profile would have a greater propensity
to be classified in one level of the outcome variable than the other level.

For vanillarelative to strawberry, the Chi-Square test statistic for
the predictor video is 1.2060 with an associated p-value of 0.2721. If we
set our alpha level to 0.05, we would fail to reject the null hypothesis and
conclude that for vanilla relative to strawberry, the regression coefficient for
video has not been found to be statistically different from zero given

puzzle and
female are in the model.

For vanilla relative to strawberry, the Chi-Square test statistic for the
predictor puzzle is 4.6746 with an associated p-value of 0.0306. If we
again set our alpha level to 0.05, we would reject the null hypothesis and
conclude that the regression coefficient for
puzzle has been found to be
statistically different from zero for vanillarelative to strawberry
given that video and
female are in the model.

For vanilla relative to strawberry, the Chi-Squaretest statistic for the
predictor female is 0.0088 with an associated p-value of 0.9252. If we
again set our alpha level to 0.05, we would fail to reject the null hypothesis
and conclude that for vanilla relative to strawberry, the regression coefficient
for female has not been found to be statistically different from zero
given puzzle and
video are in the model.

w. Odds Ratio Point Estimate – These are the proportional odds ratios.
They can be obtained by exponentiating the estimate, e^estimate.

x. 95% Wald Confidence Limits – This is the Confidence Interval (CI)
for the proportional odds ratio given the other predictors are in the model. For
a given predictor with a level of 95% confidence, we say that we are 95%
confident that the “true” population proportional odds ratio lies between the
lower and upper limit of the interval.  The CI is equivalent to the Wald
Chi-Square test statistic; if the CI includes 1, we would fail to reject the
null hypothesis that a particular ordered logit regression coefficient is zero
given the other predictors are in the model at an alpha level of 0.05. The CI is
more illustrative than the Wald Chi-Square test statistic.