SAS Data Analysis Examples
Negative Binomial Regression

Negative binomial regression is for modeling count variables, usually for over-dispersed count outcome variables.

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.

This page was updated using SAS 9.2.

Examples of negative binomial regression

Example 1.  School administrators study the attendance behavior of high school juniors at two schools.  Predictors of the number of days of absence include the type of program in which the student is enrolled and a standardized test in math.

Example 2.  A health-related researcher is studying the number of hospital visits in past 12 months by senior citizens in a community based on the characteristics of the individuals and the types of health plans under which each one is covered.  

Description of the data

Let's pursue Example 1 from above.

We have attendance data on 314 high school juniors from two urban high schools in the file nb_data.sas7bdat. The response variable of interest is days absent, daysabs. The variable math gives the standardized math score for each student.  The variable prog is a three-level nominal variable indicating the type of instructional program in which the student is enrolled.

Let's look at the data.  It is always a good idea to start with descriptive statistics and plots.

proc means data = nb_data;
  var daysabs math;
run;

The MEANS Procedure

Variable    Label                   N            Mean         Std Dev         Minimum         Maximum
-----------------------------------------------------------------------------------------------------
DAYSABS     number days absent    314       5.9554140       7.0369576               0      35.0000000
MATH        ctbs math pct rank    314      48.2675159      25.3623913       1.0000000      99.0000000
-----------------------------------------------------------------------------------------------------

proc univariate data = nb_data noprint;
  histogram daysabs / midpoints = 0 to 50 by 1 vscale = count ;
run;

Each variable has 314 valid observations and their distributions seem quite reasonable. The mean of our outcome variable is much lower than its variance.

Let's continue with our description of the variables in this dataset. The table below shows the average numbers of days absent by program type and seems to suggest that program type is a good candidate for predicting the number of days absent, our outcome variable, because the mean value of the outcome appears to vary by prog. The variances within each level of prog are higher than the means within each level. These are the conditional means and variances. These differences suggest that over-dispersion is present and that a Negative Binomial model would be appropriate.

proc sort data = nb_data;
  by prog;
run;

proc means mean var n data = nb_data;
  by prog;
  var daysabs;
run;

PROG=1

The MEANS Procedure

Analysis Variable : DAYSABS number days absent

        Mean        Variance      N
-----------------------------------
  10.6500000      67.2589744     40
-----------------------------------


PROG=2

Analysis Variable : DAYSABS number days absent

        Mean        Variance      N
-----------------------------------
   6.9341317      55.4474425    167
-----------------------------------


PROG=3

Analysis Variable : DAYSABS number days absent

        Mean        Variance      N
-----------------------------------
   2.6728972      13.9391642    107
-----------------------------------

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. 

• Negative binomial regression - Negative binomial regression can be used for over-dispersed count data, that is when the conditional variance exceeds the conditional mean. It can be considered as a generalization of Poisson regression since it has the same mean structure as Poisson regression and it has an extra parameter to model the over-dispersion. If the conditional distribution of the outcome variable is over-dispersed, the confidence intervals for the Negative binomial regression are likely to be narrower as compared to those from a Poisson regression model.
• Poisson regression - Poisson regression is often used for modeling count data. Poisson regression has a number of extensions useful for count models.
• Zero-inflated regression model - Zero-inflated models attempt to account for excess zeros.  In other words, two kinds of zeros are thought to exist in the data, "true zeros" and "excess zeros".  Zero-inflated models estimate two equations simultaneously, one for the count model and one for the excess zeros.
• OLS regression - Count outcome variables are sometimes log-transformed and analyzed using OLS regression.  Many issues arise with this approach, including loss of data due to undefined values generated by taking the log of zero (which is undefined), as well as the lack of capacity to model the dispersion.

Negative binomial regression analysis

Negative binomial models can be estimated in SAS using proc genmod. On the class statement we list the variable prog.  After prog, we use two options, which are given in parentheses.  The param=ref option changes the coding of prog from effect coding, which is the default, to reference coding.  The ref=first option changes the reference group to the first level of prog.  We have used two options on the model statement.  The type3 option is used to get the multi-degree-of-freedom test of the categorical variables listed on the class statement, and the dist = negbin option is used to indicate that a negative binomial distribution should be used. 

proc genmod data = nb_data;
  class prog (param=ref ref=first);
  model daysabs = math prog / type3 dist=negbin;
run;

The GENMOD Procedure

                         Model Information

Data Set                   WORK.NB_DATA
Distribution          Negative Binomial
Link Function                       Log
Dependent Variable              DAYSABS    number days absent


Number of Observations Read         314
Number of Observations Used         314


   Class Level Information

                      Design
Class     Value     Variables

PROG      1          0      0
          2          1      0
          3          0      1


             Criteria For Assessing Goodness Of Fit

Criterion                     DF           Value        Value/DF

Deviance                     310        358.5193          1.1565
Scaled Deviance              310        358.5193          1.1565
Pearson Chi-Square           310        339.8771          1.0964
Scaled Pearson X2            310        339.8771          1.0964
Log Likelihood                         2151.5227
Full Log Likelihood                    -865.6289
AIC (smaller is better)                1741.2578
AICC (smaller is better)               1741.4526
BIC (smaller is better)                1760.0048


Algorithm converged.


                       Analysis Of Maximum Likelihood Parameter Estimates

                                     Standard     Wald 95% Confidence          Wald
Parameter          DF    Estimate       Error           Limits           Chi-Square    Pr > ChiSq

Intercept           1      2.6153      0.1964      2.2304      3.0001        177.40        <.0001
MATH                1     -0.0060      0.0025     -0.0109     -0.0011          5.71        0.0168
PROG          2     1     -0.4408      0.1826     -0.7986     -0.0829          5.83        0.0158
PROG          3     1     -1.2787      0.2020     -1.6745     -0.8828         40.08        <.0001
Dispersion          1      0.9683      0.0995      0.7916      1.1844

NOTE: The negative binomial dispersion parameter was estimated by maximum likelihood.

     LR Statistics For Type 3 Analysis

                          Chi-
Source           DF     Square    Pr > ChiSq

MATH              1       5.61        0.0179
PROG              2      45.05        <.0001

• The output begins the Model Information table and the Criteria for Assessing Goodness of Fit table.  The number of observations read and used is given.  In this example, we have no missing data, so all 314 observations that are read in are used in the analysis.  In the Criteria for Assessing Goodness of Fit table, we see the Pearson Chi-Square of 339.88. This is not a test of the model coefficients (which we saw in the header information), but a test of the model form: are the data overdispersed when modeled with a negative binomial distribution? A low p-value from this test suggests misspecification or other problems with the model. We can get the p-value of
this test.  The non-significant p-value suggests that the negative binomial model is a good fit for the data.

data test; 
  pval = 1 - probchi(339.8771, 310);
run;

proc print data = test; run;

Obs      pval
 1     0.11703

• The Analysis of Maximum Likelihood Parameter Estimates table is presented next, which gives the regression coefficients, standard errors, the Wald 95% confidence intervals for the coefficients, chi-square tests and p-values for each of the model variables.  In this example, the variable math has a coefficient of -0.006, which is statistically significant.  This means that for each one-unit increase in math, the expected log count of the days absent decreases by .0006.  The indicator for prog=2 is the expected difference in log count between group 2 and the reference group (prog=1).  The expected log count for level 2 of prog is 0.44 lower than the expected log count for level 1. The indicator variable prog=3 is the expected difference in log count between group 3 and the reference group. The expected log count for level 3 of prog is 1.28 lower than the expected log count for level 1.  To determine if prog itself, overall, is statistically significant, we can look at the LR Statistics for Type 3 Analysis table that includes the two degrees-of-freedom test of this variable. The two degree-of-freedom chi-square test indicates that prog is a statistically significant predictor of daysabs. The chi-square value for this test is 45.05 with a p-value of .0001.  This indicates that the variable prog is a statistically significant predictor of daysabs.
• Additionally, there is an estimate of the dispersion coefficient (often called alpha).  A Poisson model is one in which this alpha value is constrained to zero. In this example, the estimated alpha has a 95% confidence interval that does not include zero, suggesting that the negative binomial model form is more appropriate than the Poisson. An estimate greater than zero suggests over-dispersion (variance greater than mean). An estimate less than zero suggests under-dispersion, which is very rare.

We can also see the results as incident rate ratios by using estimate statements with the exp option. 

proc genmod data = nb_data;
  class prog (param=ref ref=first);
  model daysabs = math prog / type3 dist=negbin;
  estimate 'prog 2' prog 1 0 / exp;
  estimate 'prog 3' prog 0 1 / exp;
  estimate 'math'   math 1   / exp;
run;

< - some output omitted - >

                                    Contrast Estimate Results

                 Mean         Mean           L'Beta  Standard                L'Beta          Chi-
Label        Estimate   Confidence Limits  Estimate     Error   Alpha   Confidence Limits  Square

prog 2         0.6435    0.4500    0.9204   -0.4408    0.1826    0.05   -0.7986   -0.0829    5.83
Exp(prog 2)                                  0.6435    0.1175    0.05    0.4500    0.9204
prog 3         0.2784    0.1874    0.4136   -1.2787    0.2020    0.05   -1.6745   -0.8828   40.08
Exp(prog 3)                                  0.2784    0.0562    0.05    0.1874    0.4136
math           0.9940    0.9892    0.9989   -0.0060    0.0025    0.05   -0.0109   -0.0011    5.71
Exp(math)                                    0.9940    0.0025    0.05    0.9892    0.9989

The output above indicates that the incident rate for prog=1 is 0.64 times the incident rate for the reference group (prog=1). Likewise, the incident rate for prog=3 is 0.28 times the incident rate for the reference group holding the other variables constant. The percent change in the incident rate of daysabs is a 1% decrease (1 - .99) for every unit increase in math.

The form of the model equation for negative binomial regression is the same as that for Poisson regression. The log of the outcome is predicted with a linear combination of the predictors:

log(daysabs) = Intercept + b₁(prog=2) + b₂(prog=3) + b₃math.

This implies:

daysabs = exp(Intercept + b₁(prog=2) + b₂(prog=3)+ b₃math) = exp(Intercept) * exp(b₁(prog=2)) * exp(b₂(prog=3)) * exp(b₃math)

The coefficients have an additive effect in the log(y) scale and the IRR have a multiplicative effect in the y scale. The dispersion parameter in negative binomial regression does not effect the expected counts, but it does effect the estimated variance of the expected counts.

For additional information on the various metrics in which the results can be presented, and the interpretation of such, please see Regression Models for Categorical Dependent Variables Using Stata, Second Edition by J. Scott Long and Jeremy Freese (2006).

Below we use estimate statements to calculate the predicted number of events at each level of prog, holding all other variables (in this example, math) in the model at their means.

proc genmod data = nb_data;
  class prog (param=ref ref=first);
  model daysabs = math prog / type3 dist=negbin;
  estimate 'prog 1' intercept 1 prog 0 0 math 48.2675 / exp;
  estimate 'prog 2' intercept 1 prog 1 0 math 48.2675 / exp;
  estimate 'prog 3' intercept 1 prog 0 1 math 48.2675 / exp;
run;

< - some output omitted - >

                                    Contrast Estimate Results

                 Mean         Mean           L'Beta  Standard                L'Beta          Chi-
Label        Estimate   Confidence Limits  Estimate     Error   Alpha   Confidence Limits  Square

prog 1        10.2369    7.4291   14.1058    2.3260    0.1636    0.05    2.0054    2.6466  202.22
Exp(prog 1)                                 10.2369    1.6744    0.05    7.4291   14.1058
prog 2         6.5879    5.5916    7.7618    1.8852    0.0837    0.05    1.7213    2.0492  507.76
Exp(prog 2)                                  6.5879    0.5512    0.05    5.5916    7.7618
prog 3         2.8501    2.2720    3.5753    1.0473    0.1157    0.05    0.8207    1.2740   82.00
Exp(prog 3)                                  2.8501    0.3296    0.05    2.2720    3.5753

In the output above, we see that the predicted number of events for level 1 of prog is about 10.24, holding math at its mean.  The predicted number of events for level 2 of prog is lower at 6.59, and the predicted number of events for level 3 of prog is about 2.85. Note that the predicted count of level 2 of prog is (6.5879/10.2369) = 0.64 times the predicted count for level 1 of prog. This matches what we saw in the after in the incident rate ratio output table.

We can similarly obtain the predicted number of events for values of math while holding prog constant.

proc genmod data = nb_data;
  class prog (param=ref ref=first);
  model daysabs = math prog / type3 dist=negbin;
  estimate 'math 20' intercept 1 prog 0 0 math 20 / exp;
  estimate 'math 40' intercept 1 prog 0 0 math 40 / exp;
run;

                                    Contrast Estimate Results

                  Mean         Mean           L'Beta  Standard                L'Beta          Chi-
Label         Estimate   Confidence Limits  Estimate     Error   Alpha   Confidence Limits  Square

math 20        12.1267    8.6305   17.0391    2.4954    0.1735    0.05    2.1553    2.8355  206.80
Exp(math 20)                                 12.1267    2.1043    0.05    8.6305   17.0391
math 40        10.7569    7.8092   14.8172    2.3755    0.1634    0.05    2.0553    2.6958  211.38
Exp(math 40)                                 10.7569    1.7576    0.05    7.8092   14.8172

The table above shows that when prog held at its reference level and math at 20, the predicted count (or average number of days absent) is about 12.13;  when prog held at its reference level and math at 40, the predicted count is about 10.76. If we compare the predicted counts at these two levels of math, we can see that the ratio is (10.7569/12.1267) = 0.887. This matches the IRR of 0.994 for a 20 unit change: 0.994^20 = 0.887.

You can graph the predicted number of events using the commands below.  Proc genmod must be run with the output statement to obtain the predicted values in a dataset we called pred1.  We then sorted our data by the predicted values and created a graph with proc sgplot.

The graph indicates that the most days absent are predicted for those in program 1.  The lowest number of predicted days absent is for those students in program 3.

proc genmod data = nb_data;
  class prog (param=ref ref=first);
  model daysabs = math prog / type3 dist=negbin;
  output out = nb_pred predicted = pred1;
run;

proc sort data = nb_pred;
  by pred1;
run;

proc sgplot data = nb_pred;
  series x=math y=pred1 / group = prog;
run;

Things to consider

• It is not recommended that negative binomial models be applied to small samples.
• Negative binomial models assume that only one process generates the data.  If more than one process generates the data, then it is possible to have more 0s than expected by the negative binomial model; in this case, a zero-inflated model (either zero-inflated Poisson or zero-inflated negative binomial) may be more appropriate.
• If the data generating process does not allow for any 0s (such as the number of days spent in the hospital), then a zero-truncated model may be more appropriate.  Such models can be estimated with proc countreg.
• Count data often have an exposure variable, which indicates the number of times the event could have happened.  This variable should be incorporated into your negative binomial model with the use of the offset option on the model statement.
• The outcome variable in a negative binomial regression cannot have negative numbers.

References

• Long, J. S. 1997. Regression Models for Categorical and Limited Dependent Variables. Thousand Oaks, CA: Sage Publications.
• Long, J. S. and Freese, J.  2006.  Regression Models for Categorical Dependent Variables Using Stata, Second Edition.  College Station, TX:  Stata Press.
• Cameron, A. C. and  Trivedi, P. K.  2009.  Microeconometrics Using Stata.  College Station, TX:  Stata Press.
• Cameron, A. C. and  Trivedi, P. K.  1998.  Regression Analysis of Count Data.  New York:  Cambridge Press.
• Cameron, A. C.  Advances in Count Data Regression Talk for the Applied Statistics Workshop, March 28, 2009. http://cameron.econ.ucdavis.edu/racd/count.html .
• Dupont, W. D.  2002.  Statistical Modeling for Biomedical Researchers:  A Simple Introduction to the Analysis of Complex Data.  New York:  Cambridge Press.

See also

• Annotated output for negative binomial regression
• SAS Online Manual
  • proc genmod