SAS Data Analysis Examples
Exact Logistic Regression

Version info: Code for this page was tested in SAS 9.3.

Exact logistic regression is used to model binary outcome variables in which the log odds of the outcome is modeled as a linear combination of the predictor variables.  It is used when the sample size is too small for a regular logistic regression (which uses the standard maximum-likelihood-based estimator) and/or when some of the cells formed by the outcome and categorical predictor variable have no observations.  The estimates given by exact logistic regression do not depend on asymptotic results.

Please note: The purpose of this page is to show how to use various data analysis commands.  It does not cover all aspects of the research process which researchers are expected to do.  In particular, it does not cover data cleaning and checking, verification of assumptions, model diagnostics or potential follow-up analyses.

Example

Suppose that we are interested in the factors that influence whether or not a high school senior is admitted into a very competitive engineering school.  The outcome variable is binary (0/1): admit or not admit.  The predictor variables of interest include student gender and whether or not the student took Advanced Placement calculus in high school.  Because the response variable is binary, we need to use a model that handles 0/1 outcome variables correctly.  Also, because of the number of students involved is small, we will need a procedure that can perform the estimation with a small sample size. 

Description of the data

The data for this exact logistic data analysis include the number of students admitted, the total number of applicants broken down by gender (the variable female), and whether or not they had taken AP calculus (the variable apcalc).  Since the dataset is so small, we will read it in directly. 

options nocenter;

data exlogit;
  input female apcalc admit num;
datalines;
0 0 0 7
0 0 1 1
0 1 0 3
0 1 1 7
1 0 0 5
1 0 1 1
1 1 0 0
1 1 1 6
;
run;

Let's look at some frequency tables.  We will specify the variable num as the frequency weight.

proc freq data = exlogit;
  tables female*(apcalc admit);
  tables apcalc*admit;
  weight num;
run;

Table of female by apcalc

female     apcalc

Frequency|
Percent  |
Row Pct  |
Col Pct  |       0|       1|  Total
---------+--------+--------+
       0 |      8 |     10 |     18
         |  26.67 |  33.33 |  60.00
         |  44.44 |  55.56 |
         |  57.14 |  62.50 |
---------+--------+--------+
       1 |      6 |      6 |     12
         |  20.00 |  20.00 |  40.00
         |  50.00 |  50.00 |
         |  42.86 |  37.50 |
---------+--------+--------+
Total          14       16       30
            46.67    53.33   100.00

Table of female by admit

female     admit

Frequency|
Percent  |
Row Pct  |
Col Pct  |       0|       1|  Total
---------+--------+--------+
       0 |     10 |      8 |     18
         |  33.33 |  26.67 |  60.00
         |  55.56 |  44.44 |
         |  66.67 |  53.33 |
---------+--------+--------+
       1 |      5 |      7 |     12
         |  16.67 |  23.33 |  40.00
         |  41.67 |  58.33 |
         |  33.33 |  46.67 |
---------+--------+--------+
Total          15       15       30
            50.00    50.00   100.00

Table of apcalc by admit

apcalc     admit

Frequency|
Percent  |
Row Pct  |
Col Pct  |       0|       1|  Total
---------+--------+--------+
       0 |     12 |      2 |     14
         |  40.00 |   6.67 |  46.67
         |  85.71 |  14.29 |
         |  80.00 |  13.33 |
---------+--------+--------+
       1 |      3 |     13 |     16
         |  10.00 |  43.33 |  53.33
         |  18.75 |  81.25 |
         |  20.00 |  86.67 |
---------+--------+--------+
Total          15       15       30
            50.00    50.00   100.00

proc tabulate data = exlogit;
  class female apcalc admit;
  tables female='female', admit*apcalc='AP calculus'*F=6. / rts=13.;
  freq num;
run;

-----------------------------------------
|           |           admit           |
|           |---------------------------|
|           |      0      |      1      |
|           |-------------+-------------|
|           | AP calculus | AP calculus |
|           |-------------+-------------|
|           |  0   |  1   |  0   |  1   |
|           |------+------+------+------|
|           |  N   |  N   |  N   |  N   |
|-----------+------+------+------+------|
|female     |      |      |      |      |
|-----------|      |      |      |      |
|0          |     7|     3|     1|     7|
|-----------+------+------+------+------|
|1          |     5|     .|     1|     6|
-----------------------------------------

The tables reveal that 30 students applied for the Engineering program.  Of those, 15 were admitted and 15 were denied admission.  There were 18 male and 12 female applicants.   Sixteen of the applicants had taken AP calculus and 14 had not.  Note that all of the females who took AP calculus were admitted, versus only about half the males. 

Analysis methods you might consider

Below is a list of some analysis methods you may have encountered.  Some of the methods listed are quite reasonable, while others have either fallen out of favor or have limitations. 

• Exact logistic regression - This technique is appropriate because the outcome variable is binary, the sample size is small, and some cells are empty.
• Regular logistic regression - Due to the small sample size and the presence of cells with no subjects, regular logistic regression is not advisable, and it might not even be estimable.
• Two-way contingency tables - You may need to use the fisher or exact with proc freq option to get the Fisher's exact test due to small expected values.

Using the exact logistic model

Let's run the exact logistic analysis using proc logistic with the exact statement.  We will include the option estimate = both on the exact statement so that we obtain both the point estimates and the odds ratios in the output.  We will also need to use the freq statement, for which we will specify the frequency weight variable num. 

proc logistic data = exlogit desc;
  freq num;
  model admit = female apcalc;
  exact female apcalc / estimate = both;
run;

The LOGISTIC Procedure

              Model Information

Data Set                      WORK.EXLOGIT
Response Variable             admit
Number of Response Levels     2
Frequency Variable            num
Model                         binary logit
Optimization Technique        Fisher's scoring

Number of Observations Read           8
Number of Observations Used           7
Sum of Frequencies Read              30
Sum of Frequencies Used              30

          Response Profile

 Ordered                      Total
   Value        admit     Frequency

       1            1            15
       2            0            15

Probability modeled is admit=1.

NOTE: 1 observation having nonpositive frequency or weight was excluded since it does not
      contribute to the analysis.

                    Model Convergence Status

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


         Model Fit Statistics

                             Intercept
              Intercept            and
Criterion          Only     Covariates

AIC              43.589         31.194
SC               44.990         35.398
-2 Log L         41.589         25.194

        Testing Global Null Hypothesis: BETA=0

Test                 Chi-Square       DF     Pr > ChiSq

Likelihood Ratio        16.3947        2         0.0003
Score                   14.2886        2         0.0008
Wald                     9.6706        2         0.0079

             Analysis of Maximum Likelihood Estimates

                               Standard          Wald
Parameter    DF    Estimate       Error    Chi-Square    Pr > ChiSq

Intercept     1     -2.5984      1.1361        5.2310        0.0222
female        1      1.4513      1.2037        1.4537        0.2279
apcalc        1      3.6685      1.1904        9.4973        0.0021

           Odds Ratio Estimates

             Point          95% Wald
Effect    Estimate      Confidence Limits

female       4.269       0.403      45.179
apcalc      39.193       3.801     404.075

Association of Predicted Probabilities and Observed Responses

Percent Concordant     80.4    Somers' D    0.756
Percent Discordant      4.9    Gamma        0.885
Percent Tied           14.7    Tau-a        0.391
Pairs                   225    c            0.878

Exact Conditional Analysis

             Conditional Exact Tests

                                   --- p-Value ---
Effect   Test          Statistic    Exact      Mid

female   Score            1.5143   0.3401   0.2438
         Probability      0.1925   0.3401   0.2438
apcalc   Score           13.0574   0.0003   0.0002
         Probability    0.000283   0.0003   0.0002

                     Exact Parameter Estimates

                         Standard       95% Confidence
Parameter    Estimate       Error           Limits           p-Value

female         1.3605      1.1698     -1.1290      5.3680     0.4557
apcalc         3.3387      1.1251      1.1017      7.2659     0.0006

                  Exact Odds Ratios

                          95% Confidence
Parameter   Estimate          Limits          p-Value

female         3.898      0.323    214.433     0.4557
apcalc        28.182      3.009   >999.999     0.0006

• The output begins with information about the dataset used and the model run.  Next, we see information about the response variable, including the number of 0s and 1s.  We see a note indicating that the 1s are being modeled (because we used the desc option on the proc logistic statement), and a note warning us about the 0 count for one of the lines of data. 
• We next see model fit statistics, which can be used to compare models, and tests of the overall model.  We see that the overall model is statistically significant. 
• Next, we have tables giving us the maximum likelihood estimates.  After the table giving the association between the predicted probabilities and the observed responses, we see the results of the exact conditional analysis.  Both the score test and the probability test are given.  The variable female is not statistically significant, but the variable apcalc is.  For every one unit change in apcalc, the expected log odds of admission (admit) increases by 3.34.  The intercept is not included in the output because its sufficient statistic was conditioned out when creating the joint distribution of female and apcalc. 
• The final table in the output is table of exact odds ratios.  The odds for an applicant who had taken AP calculus was about 28.2 times greater than for one who had not taken the course.

We can also graph the predicted probabilities.  To do this, we will create a new variable called p using the output statement.  Then we will use proc gplot to graph p.

proc logistic data = exlogit desc;
  freq num;
  model admit = female apcalc;
  exact female apcalc / estimate = both;
  output out = pred predicted = p;
run;

symbol1 c=blue v=circle i=join;
symbol2 c=red  v=plus i=join; 
symbol3 c=black v=square i=join;
axis1 label=(r=0 a=90) minor=none;
axis2 minor=none order=(0 1); 
proc gplot data= pred;
  plot p*female=apcalc / vaxis=axis1 haxis=axis2;
run;
quit;

Things to consider

• Exact logistic regression is a very memory intensive procedure, and it is relatively easy to exceed the memory capacity of a given computer. 
• Firth logit may be helpful if you have separation in your data.  You can use the firth option on the model statement to run a Firth logit.  This option was added in SAS version 9.2.
• Exact logistic regression is an alternative to conditional logistic regression if you have stratification, since both condition on the number of positive outcomes within each stratum.  The estimates from these two analyses will be different because conditional logit conditions only on the intercept term, while exact logistic regression conditions on the sufficient statistics of the other regression parameters as well as the intercept term.

See also

• SAS documentation for proc logistic

References

• Collett, D.  Modeling Binary Data, Second Edition.  Boca Raton:  Chapman and Hall.
• Cox, D. R. and Snell, E. J. (1989).  Analysis of Binary Data, Second Edition. Boca Raton: Chapman and Hall.
• Hirji, K. F. (2005).  Exact Analysis of Discrete Data. Boca Raton: Chapman and Hall.