SAS Global Forum 2008

Statistics and Data Analysis

Paper 371-2008

Count Data Models in SAS

WenSui Liu, ChoicePoint Precision Marketing, Alpharetta, GA
Jimmy Cela, ChoicePoint Precision Marketing, Alpharetta, GA
ABSTRACT
Poisson regression has been widely used to model count data. However, it is often criticized for its restrictive
assumption of equi-dispersion, meaning equality between the variance and the mean. In real-life applications, count
data often exhibits over-dispersion and excess zeroes. While Negative binomial regression is able to model count
data with over-dispersion, both Hurdle (Mullahy, 1986) and Zero-inflated (Lambert, 1992) regressions address the
issue of excess zeroes in their own rights. Different modeling strategies for count data and various statistical tests for
model evaluation are illustrated through an example of healthcare utilization. The purpose of this paper is to provide
by far the most complete survey of count data modeling strategy in SAS for the user group.

KEYWORDS
Poisson regression, Negative binomial regression, Hurdle regression, Zero-Inflated regression, Overdispersion,
Excess Zeroes, Vuong test.

1. INTRODUCTION
How to model count data as the dependent variable in a regression has become a popular topic in statistics,
econometrics, and epidemiology. Deb and Trivedi (1997) modeled the demand for healthcare utilization by the elderly
using a finite mixture negative binomial regression. Gurmu (1997) evaluated the impact of managed care program on
healthcare utilization using hurdle model. Winkelmann (2004) studied the effect of healthcare reform on the number
of doctor visits in Germany using a number of modified count data models. For more detailed discussions about
recent development in count data models, please refer to Cameron and Trivedi (2001), Winkelmann and
Zimmermann (1995), and Greene (2002).
To illustrate models covered in this paper, we use the same data analyzed by Deb and Trivedi (1997). This data is
originally obtained from National Medical Expenditure Survey (NMES) conducted in 1987 and includes 4406
respondents who were aged 66 or older and covered by Medicare program. In our example, the number of hospital
stays (HOSP) is used as the dependent variable and three types of measures are included in the explanatory
variables, which are self-perceived health status (EXCLHLTH, POORHLTH, and NUMCHRON), demographic data
(AGE and MALE), and socio-economic information (SCHOOL and PRIVINS). The summary statistics of all variables
are given in Table 1.
Table 1.1, Variables Used with Summary Statistics
Mean

Std. Dev.

HOSP

Variable
# of hospital stays

Definition

Obs
4406

0.2960

0.7464

Min
0

Max
8

EXCLHLTH

1 if self-perceived health is excellent

4406

0.0778

0.2680

1
1

POORHLTH

1 if self-perceived health is poor

4406

0.1257

0.3316

NUMCHRON

# of chronic conditions

4406

1.5420

1.3496

AGE

age in years (divided by 10)

4406

7.4024

0.6334

6.6

10.9

MALE

1 if the person is male

4406

0.4035

0.4907

SCHOOL

# of years of education

4406

10.2903

3.7387

18

PRIVINS

1 if the person is covered by private insurance

4406

0.7764

0.4167

As shown in Table 1.1, the variance of HOSP is about two times of the mean, implying the possibility of overdispersion. A further screening on the data also shows that more than 80% of the respondents, 3541 out of 4406,
have no hospital admission, indicating excess zeroes.
A good starting point of count data modeling is to compare the empirical distribution of observed counts to the
univariate Poisson distribution with the mean estimated from the data. Probabilities from two distributions are plotted
in Figure 1.1.

SAS Global Forum 2008

Statistics and Data Analysis

Figure 1.1, Comparison between Observed Probability and Univariate Poisson Probability

The plot in Figure 1.1 clearly shows that univariate Poisson distribution underestimates the probability at 0 and
overestimates the probability at 1. Since Poisson distribution assumes the same mean across the whole sample and
doesnt consider the heterogeneity in each member, it is not surprising to see that the predicted probability does not
fit the observed data well. In the next section, we will allow the observed heterogeneity in the conditional mean of
each sample member by including explanatory variables.

2. POISSON REGRESSION
Poisson regression is the simplest regression model for count data and assumes that each observed count Yi is
drawn from a Poisson distribution with the conditional mean ui on a given vector Xi for case i. Therefore, the density
function of Yi can be expressed as

Exp( ui ) ui
f (Yi | X i ) =
Yi !

Yi

, where ui

= Exp( X i ) .

(2.1)

Given independent observations with the density function in (2.1), the log likelihood function can be obtained by
n

LL = [ ui + Yi Log (ui ) Log (Yi !)] .

(2.2)

i =1

The maximum likelihood estimation of Poisson regression is straightforward using the log likelihood function in (2.2).
In SAS, several procedures in both STAT and ETS modules can be used to estimate Poisson regression. While
GENMOD, GLIMMIX, and COUNTREG are easy to use with standard MODEL statement, NLMIXED, MODEL, NLIN
provide great flexibility to model count data by specifying the log likelihood function explicitly. An illustration of both
NLMIXED and COUNTREG procedures is given below. More detailed examples on how to use all mentioned
procedures can be found on authors blog at statcompute.spaces.live.com.
/* METHOD 1: PROC NLMIXED */
proc nlmixed data = tblNMES;
parms b0 = 0 b1 = 0 b2 = 0 b3 = 0 b4 = 0 b5 = 0 b6 = 0 b7 = 0;
mu = exp(b0 + b1 * EXCLHLTH + b2 * POORHLTH + b3 * NUMCHRON + b4 * AGE +
b5 * MALE + b6 * SCHOOL + b7 * PRIVINS);
ll = -mu + HOSP * log(mu) - log(fact(HOSP));
model HOSP ~ general(ll);
predict mu out = poi_out (rename = (pred = Yhat));
run;
/* METHOD 2: PROC COUNTREG */
proc countreg data = tblNMES type = poisson;
model HOSP = EXCLHLTH POORHLTH NUMCHRON AGE MALE SCHOOL PRIVINS;
run;
/* SAMPLE OUTPUT OF PROC COUNTREG:
Model Fit Summary
Log Likelihood
AIC
SBC
Parameter Estimates
2

-3046
6108
6159

SAS Global Forum 2008

Statistics and Data Analysis

Parameter
Intercept
exclhlth
poorhlth
numchron
age
male
school
privins

Estimate
-3.329044
-0.723412
0.626157
0.264462
0.186406
0.103186
-0.000206
0.108652

Standard
Error
0.339728
0.175644
0.067858
0.018277
0.042014
0.056274
0.007871
0.069251

Approx
Pr > |t|
<.0001
<.0001
<.0001
<.0001
<.0001
0.0667
0.9791
0.1167

t Value
-9.80
-4.12
9.23
14.47
4.44
1.83
-0.03
1.57

*/

While Poisson regression is often used as a baseline model for count data, its assumption of equi-dispersion is too
restrictive for many empirical applications. In practice, the variance of observed count data usually exceeds the mean,
namely over-dispersion, due to the unobserved heterogeneity and/or excess zeroes. With the similar consequences
of heteroskedasticity in the linear regression, over-dispersion in a Poisson regression will lead to deflated standard
errors of parameter estimates and therefore inflated t-statistics. After the development of Poisson regression, it is
always a sound practice to do an additional analysis for over-dispersion. In our example, we will consider two
statistical tests based on the alternative Negative binomial model, which will be covered in our next section.
Cameron and Trivedi (1996) introduced a simple test for over-dispersion based on an auxiliary OLS regression
without the intercept, which can be formulated as

( yi ui )2 yi = u + e , where u = Exp( X ) and e is an error term.

i
i
i
i
i

(2.3)

ui

The significance of t-statistics for the coefficient implies the existence of over-dispersion. Please note that (2.3) is
specific for Negbin 2 form, the most common setting for Negative binomial. For Negbin 1 form, a different formulation
of OLS regression should be used.

( yi ui )2 yi = + e
i

(2.4)

ui

The implementation in SAS with Reg procedure based on (2.3) is given below.
data ols_tmp;
set poi_out;
dep = ((HOSP - Yhat) ** 2 - HOSP) / Yhat;
run;
proc reg data = ols_tmp;
model dep = Yhat / noint;
run; quit;

/* FIT A OLS REGRESSION WITHOUT INTERCEPT */

/* OUTPUT OF AUXILIARY OLS REGRESSION:

Parameter Estimates
Parameter
Standard
Variable
Label
DF
Estimate
Error
Yhat
Predicted Value
1
1.63419
0.22609
*/

t Value
7.23

Pr > |t|
<.0001

The second test for over-dispersion introduced by Greene (2002) is based on the Lagrange multiplier (LM) statistics.
If we consider Poisson regression a parametric restriction of Negative binomial regression with the mean equal to the
variance, the LM statistics can be simply expressed as

(e`e nY )
LM =

2u `u

, where

u = Exp( X ` ) and e = Y u

(2.5)

Under the null hypothesis of Poisson regression, the LM statistics follows the chi-squared distribution with one degree
of freedom. The computation is extremely simple with SAS IML procedure or any other matrix languages.
proc iml;
use poi_out;
read all var {HOSP} into y;
read all var {Yhat} into yhat;
close poisson_out;
e
= (y - yhat);
n
= nrow(y);
ybar
= y`[, :];
LM
= (e` * e - n * ybar) ** 2 / (2 * yhat` * yhat);
Pvalue = 1 - probchi(LM, 1);
print LM Pvalue;
3

SAS Global Forum 2008

Statistics and Data Analysis

quit;
/* OUTPUT OF LM STATISTICS:
LM
794.14707

PVALUE
0

*/

Both tests come up with the same conclusion and confirm our suspicion of over-dispersion.
To evaluate the goodness-of-fit of a regression for count data, the most popular but somewhat problematic practice is
to compare the predicted and observed values of the dependent variable. However, a measure of goodness-of-fit
solely based upon the expected value is unable to address the improvement achieved by a model with less restrictive
variance assumption. A better alternative is to compare the predicted and observed probabilities of each count
outcome by taking the probability distribution into consideration. In Figure 2.1 below, a plot comparing observed
probabilities to predicted probabilities side by side is given.
Figure 2.1, Comparison between Observed and Predicted Probability from Poisson Regression

Compared with Figure1.1, we can see a moderate improvement shown in Figure 2.1 after considering the observed
heterogeneity in Poisson regression. However, the under-prediction at 0 and over-prediction at 1 suggest that a
further improvement is still possible. In the next section, we will discuss an alternative model with less restrictive
assumption, namely Negative binomial regression.

3. NEGATIVE BINOMIAL REGRESSION

As the most common alternative to Poisson regression, Negative binomial regression addresses the issue of overdispersion by including a dispersion parameter to accommodate the unobserved heterogeneity in the count data.
While there are many variants of Negative binomial, we will only focus on the Negbin 2 form in our paper.
Negative binomial regression can be considered a generalization of Poisson regression and assumes that the
conditional mean ui of Yi is not only determined by Xi but also a heterogeneity component ei unrelated to Xi. The
formulation can be expressed as

ui = Exp( X i + ei ) = Exp( X i )Exp(ei ) , where Exp(ei ) ~ Gamma( 1 , 1 )

(3.1)

As a result, the density function of Yi can be derived as

Yi + 1
f (Yi | X i ) =
(Yi + 1) 1

( )

1
1

+ ui

Yi

ui
1

+ ui

(3.2)

And the corresponding log likelihood function becomes

Yi + 1
LL = Log
1
i =1
(Yi + 1)
n

Yi + 1 Log (1 + ui ) + Yi Log (ui ) .

( ) (

(3.3)

Similar to Poisson regression, Negative binomial regression can be modeled by SAS either directly with GENMOD,
GLIMMIXED, and COUNTREG procedures or through log likelihood function in (3.3) with NLMIXED, MODEL, and
NLIN procedures. A brief example of SAS code and related output is given below.
4

SAS Global Forum 2008

Statistics and Data Analysis

/* METHOD 1: PROC NLMIXED */

proc nlmixed data = tblNMES;
parms b0 = 0 b1 = 0 b2 = 0 b3 = 0 b4 = 0 b5 = 0 b6 = 0 b7 = 0;
mu = exp(b0 + b1 * EXCLHLTH + b2 * POORHLTH + b3 * NUMCHRON + b4 * AGE +
b5 * MALE + b6 * SCHOOL + b7 * PRIVINS);
ll = lgamma(HOSP + 1 / alpha) - lgamma(HOSP + 1) - lgamma(1 / alpha) +
HOSP * log(alpha * mu) (HOSP + 1 / alpha) * log(1 + alpha * mu);
model HOSP ~ general(ll);
predict mu out = nb_out (rename = (pred = Yhat));
run;
/* METHOD 2: PROC COUNTREG
*/
proc countreg data = tblNMES type = negativebinom method = qn;
model HOSP = EXCLHLTH POORHLTH NUMCHRON AGE MALE SCHOOL PRIVINS;
run;
/* SAMPLE OUTPUT OF PROC COUNTREG:
Model Fit Summary
Log Likelihood
AIC
SBC

Parameter
Intercept
exclhlth
poorhlth
numchron
age
male
school
privins
_Alpha

Parameter Estimates
Standard
Estimate
Error
-3.752640
0.446835
-0.697875
0.193318
0.613926
0.095392
0.289418
0.026470
0.238444
0.055265
0.153862
0.073033
-0.002271
0.010203
0.093922
0.090494
1.766727
0.160492

-2857
5731
5789

t Value
-8.40
-3.61
6.44
10.93
4.31
2.11
-0.22
1.04
11.01

Approx
Pr > |t|
<.0001
0.0003
<.0001
<.0001
<.0001
0.0351
0.8238
0.2993
<.0001

*/

Please note that Negative binomial regression is the extension of Poisson with a more liberal variance assumption
and could collapsed into Poisson regression with the dispersion parameter equal to 0. This important fact provides a
possibility to do the model comparison between Poisson and Negative binomial regressions. First of all, we can
looked at the reported t-statistics of dispersion parameter, Alpha, to assess the significance of over-dispersion. Then
a likelihood ratio (LR) test, which follows Chi-square distribution with 1 degree of freedom, between 2 regressions can
be used to determine the preferred model for the data. In our example, the t-statistics of Alpha is 11.01 and the LR
test is -2(LLPoisson - LLNegbin) = -2[-3046 (-2857)] = 378, both of which are highly significant and indicate that Negative
binomial regression is preferred over Poisson regression. This result is also consistent with our findings of overdispersion in Section 2.
The goodness-of-fit of Negative binomial regression can be visualized in the similar way to Figure 2.1, as shown in
Figure 3.1.
Figure 3.1, Comparison between Observed and Predicted Probability from Negative Binomial Regression

SAS Global Forum 2008

Statistics and Data Analysis

Compared with Figure 1.1 and 2.1, we can clearly see the significant improvement made by Negative binomial
regression in Figure 3.1, a nearly complete overlap between predicted and observed probabilities. However, Negative
binomial regression is not without criticism. The inclusion of unobserved heterogeneity will increase the probabilities
of both zero counts and high counts but might not yield a good fit for the distribution of count outcome with excess
zeroes. In the next two sections, we will introduce two alternative models to handle excess zeroes, which are Hurdle
regression (Mullahy 1986) and Zero-inflated regression (Lambert 1992).

4. HURDLE REGRESSION
Originally developed by Mullahy (1986), Hurdle regression is also known as two-part model. Instead of assuming that
count outcome comes from a single data generating process, Hurdle regression considers count outcome generated
by two systematically different statistical processes, a binomial distribution determining if a count outcome is zero or
nonzero and a truncated-at-zero distribution for count data governing all positive counts conditional on nonzero
outcomes. The attraction of Hurdle regression is that it reflects a two-stage decision-making process in most human
behaviors and therefore has an appealing interpretation. For instance, it is patients decision whether to contact the
doctors office and to make the initial visit. However, after the patients first visit, doctor plays a more important role in
determining if the patient needs to make follow-up visits. Therefore, in a regression setting, the first decision might be
reflected by a Logit or Probit regression, while the second one can be analyzed by a truncated Poisson or Negative
binomial regression. Moreover, different explanatory variables are allowed to have different impacts at each decision
process.
The most popular formulation of a Hurdle regression is called Logit-Poisson model, which is the combination of a
Logit regression modeling zero vs. nonzero outcomes and a truncated Poisson regression modeling positive counts
conditional on nonzero outcomes. Its probability density function is given as

Y
f (Yi | X i ) = (1 i ) Exp( ui ) ui i
(1 Exp( ui )) Yi !

for Yi = 0
for Yi > 0

, where

i = P(Yi = 0 )

and

ui = Exp( X i )

(4.1)

The log-likelihood function of a Logit-Poisson regression therefore can be expressed as the sum of log-likelihood
functions of two components as below
n
I (Yi = 0)Log ( i ) +

LL =

i =1 I (Yi > 0 )(Log (1 i ) ui + Yi Log (ui ) Log (1 Exp( u i )) Log (Yi !))

(4.2).

Unlike Poisson and Negative binomial regressions, Hurdle regression can only be modeled through log-likelihood
function with NLMIXED, MODEL, and NLIN procedures in SAS. For the simplicity, we use the same explanatory
variables in both components of Logit-Poisson regression. However, in practice, two sets of explanatory variables do
not have to coincide. An example of NLMIXED is given below.
/* METHOD 1: PROC NLMIXED */
proc nlmixed data = tblNMES tech = dbldog;
parms a0 = 0 a1 = 0 a2 = 0 a3 = 0 a4 = 0 a5 = 0 a6 = 0 a7 = 0
b0 = 0 b1 = 0 b2 = 0 b3 = 0 b4 = 0 b5 = 0 b6 = 0 b7 = 0;
eta0 = a0 + a1 * EXCLHLTH + a2 * POORHLTH + a3 * NUMCHRON + a4 * AGE +
a5 * MALE + a6 * SCHOOL + a7 * PRIVINS;
exp_eta0 = exp(eta0);
p0 = exp_eta0 / (1 + exp_eta0);
etap = b0 + b1 * EXCLHLTH + b2 * POORHLTH + b3 * NUMCHRON + b4 * AGE +
b5 * MALE + b6 * SCHOOL + b7 * PRIVINS;
exp_etap = exp(etap);
if HOSP = 0 then ll = log(p0);
else ll = log(1 - p0) - exp_etap + HOSP * etap - lgamma(HOSP + 1)
- log(1 - exp(-exp_etap));
model HOSP ~ general(ll);
predict exp_etap out = hdl_out1 (keep = pred HOSP rename = (pred = Yhat));
predict p0 out = hdl_out2 (keep = pred rename = (pred = p0));
run;
/* SAMPLE OUTPUT OF PROC NLMIXED:
Fit Statistics
-2 Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)

5758.4
5790.4
5790.6
5892.7

Parameter Estimates
6

SAS Global Forum 2008

Parameter
a0
a1
a2
a3
a4
a5
a6
a7
b0
b1
b2
b3
b4
b5
b6
b7

Statistics and Data Analysis

Estimate
4.2311
0.5826
-0.6953
-0.3078
-0.2752
-0.1948
-0.00593
-0.01924
-0.4693
-0.9422
0.3373
0.1426
-0.01229
-0.03854
-0.01815
0.2589

Standard
Error
0.4889
0.1991
0.1073
0.02890
0.06061
0.08008
0.01126
0.09944
0.5627
0.4949
0.1008
0.02967
0.06834
0.09227
0.01290
0.1139

DF
4406
4406
4406
4406
4406
4406
4406
4406
4406
4406
4406
4406
4406
4406
4406
4406

t Value
8.65
2.93
-6.48
-10.65
-4.54
-2.43
-0.53
-0.19
-0.83
-1.90
3.35
4.81
-0.18
-0.42
-1.41
2.27

Pr > |t|
<.0001
0.0035
<.0001
<.0001
<.0001
0.0151
0.5982
0.8466
0.4043
0.0570
0.0008
<.0001
0.8573
0.6762
0.1597
0.0231

Alpha
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05

*/

Similarly to Negative binomial regression, Hurdle regression might become Poisson regression with the restriction of
parameters and therefore they can be considered nested models. Thus, it is straightforward to use the Likelihood
Ratio (LR) test discussed in the previous section to compare Hurdle regression and Poisson regression. In our case,
the LR test is given as -2(LLPoisson - LLHurdle) = -2[-3046 - (-2879)] = 334, which is highly significant and suggests that
Hurdle regression is preferred to Poisson regression.
While Hurdle regression and Poisson regression are nested, Hurdle regression and Negative binomial regression are
not. As a result, the LR test cannot be used to compare these non-nested models. In statistics literature, two methods
are generally used to compare non-nested models. The first approach is to use information criteria such as AIC or
BIC. However, due to its more parsimonious parameterization and higher log likelihood function, Negative binomial
regression is often reported to be favored over Hurdle regression. The second one is to use Vuong test, as proposed
by Greene (1994). Since more comprehensive introduction about Vuong test is given in the next section, we will skip
the detailed discussion here.
Again, Figure 4.1 visualizes the goodness-of-fit of Hurdle regression, which looks as good as the one provided by
Negative binomial regression.
Figure 4.1, Comparison between Observed and Predicted Probability from Hurdle Regression

5. ZERO-INFLATED REGRESSION
Introduced by Lambert (1992), Zero-inflated regression is another way to model count data with excess zeros. Similar
to Hurdle regression, Zero-inflated regression can also be considered a mixture of two statistical processes, one
always generating zero counts and the other generating both zero and nonzero counts. However, it is slightly different
from Hurdle regression with all zero counts from a single statistical process and assumes that zero counts might
come from two different sources. More specifically, in a Zero-inflated regression, a Logit model with binomial
assumption is used to determine if an individual count outcome is from the always-zero or the not-always-zero group
and then a model for count data, either Poisson or Negative binomial, to model outcomes in the not-always-zero
group. In the paper, we will limit our discussion to Zero-inflated Poisson (ZIP) regression with formulation

SAS Global Forum 2008

Statistics and Data Analysis

Log i = Z i
1 i

and Log

(ui ) = X i ,

(5.1)

where Zi and Xi are covariate matrix. However, the same idea can be easily generalized to Zero-inflated Negative
Binomial (ZINB) regression.
The density function of a ZIP model is given as

i + (1 i ) Exp( ui )

Y
f (Yi | X i ) =
Exp( ui ) ui i
1 i )
(

Yi !

for Yi = 0
for Yi > 0

, where 1 i

= P(Yi ~ Poisson (ui ))

(5.2)

And its log-likelihood function is expressed as

n
I (Yi = 0)Log (i + (1 i )Exp( ui )) +

LL =

i =1 I (Yi > 0 )(Log (1 i ) + Yi Log (ui ) u i Log (Yi !))

(5.3)

In SAS, ZIP can be fitted either through log-likelihood function or directly with experimental COUNTREG procedure in
ETS module.
/* METHOD 1: PROC COUNTREG */
proc countreg data = tblNMES type = zip;
model HOSP = EXCLHLTH POORHLTH NUMCHRON AGE MALE SCHOOL PRIVINS
/ zi(link = logistic, var = EXCLHLTH POORHLTH NUMCHRON AGE MALE SCHOOL PRIVINS);
run;
/* METHOD 2: PROC NLMIXED */
proc nlmixed data = tblNMES tech = dbldog;
parms a0 = 0 a1 = 0 a2 = 0 a3 = 0 a4 = 0 a5 = 0 a6 = 0 a7 = 0
b0 = 0 b1 = 0 b2 = 0 b3 = 0 b4 = 0 b5 = 0 b6 = 0 b7 = 0;
eta0 = a0 + a1 * EXCLHLTH + a2 * POORHLTH + a3 * NUMCHRON + a4 * AGE +
a5 * MALE + a6 * SCHOOL + a7 * PRIVINS;
exp_eta0 = exp(eta0);
p0 = exp_eta0 / (1 + exp_eta0);
etap = b0 + b1 * EXCLHLTH + b2 * POORHLTH + b3 * NUMCHRON + b4 * AGE +
b5 * MALE + b6 * SCHOOL + b7 * PRIVINS;
exp_etap = exp(etap);
if HOSP = 0 then ll = log(p0 + (1 - p0) * exp(-exp_etap));
else ll = log(1 - p0) + HOSP * etap - exp_etap - lgamma(HOSP + 1);
model HOSP ~ general(ll);
predict exp_etap out = zip_out1 (keep = pred HOSP rename = (pred = Yhat));
predict p0 out = zip_out2 (keep = pred rename = (pred = p0));
run;
/* SAMPLE OUTPUT OF PROC COUNTREG:
Model Fit Summary
Log Likelihood
AIC
SBC

Parameter
Intercept
exclhlth
poorhlth
numchron
age
male
school
privins
Inf_Intercept
Inf_exclhlth
Inf_poorhlth
Inf_numchron
Inf_age
Inf_male
Inf_school
Inf_privins
*/

-2878
5788
5890

Parameter Estimates
Standard
Estimate
Error
-0.366506
0.572032
-0.919990
0.458460
0.324926
0.101157
0.127746
0.033867
-0.024359
0.068806
-0.059629
0.099133
-0.012473
0.013520
0.229208
0.114004
4.265976
0.971218
-0.369944
0.717395
-0.589745
0.195174
-0.280116
0.062396
-0.405962
0.119765
-0.334773
0.162429
-0.019390
0.022126
0.224859
0.196133

t Value
-0.64
-2.01
3.21
3.77
-0.35
-0.60
-0.92
2.01
4.39
-0.52
-3.02
-4.49
-3.39
-2.06
-0.88
1.15

Approx
Pr > |t|
0.5217
0.0448
0.0013
0.0002
0.7233
0.5475
0.3562
0.0444
<.0001
0.6061
0.0025
<.0001
0.0007
0.0393
0.3808
0.2516

SAS Global Forum 2008

Statistics and Data Analysis

Please note that in a ZIP regression, explanatory variables used in two components do not need to be the same.
However, when all covariates in both sub-models are identical, ZIP regression can become a more parsimonious
model by assuming that the coefficient vector in Logit component are the product between the coefficient vector in

Poisson component and a scalar parameter (tau), namely ZIP(tau) model. Its formulation can be expressed as

Log i = X i
1 i

and Log

(ui ) = X i .

(5.4)

And the estimation of ZIP(tau) in SAS is straightforward with NLMIXED procedure given below.
/* METHOD 1:
proc nlmixed
parms b0 =
eta0 = tau

PROC NLMIXED */
data = tblNMES;
0 b1 = 0 b2 = 0 b3 = 0 b4 = 0 b5 = 0 b6 = 0 b7 = 0 tau = 1;
* (b0 + b1 * EXCLHLTH + b2 * POORHLTH + b3 * NUMCHRON + b4 * AGE +
b5 * MALE + b6 * SCHOOL + b7 * PRIVINS);
exp_eta0 = exp(eta0);
p0 = exp_eta0 / (1 + exp_eta0);
etap = b0 + b1 * EXCLHLTH + b2 * POORHLTH + b3 * NUMCHRON + b4 * AGE +
b5 * MALE + b6 * SCHOOL + b7 * PRIVINS;
exp_etap = exp(etap);
if HOSP = 0 then ll = log(p0 + (1 - p0) * exp(-exp_etap));
else ll = log(1 - p0) + HOSP * etap - exp_etap - lgamma(HOSP + 1);
model HOSP ~ general(ll);
predict exp_etap out = zip_out1 (keep = pred HOSP rename = (pred = Yhat));
predict p0 out = zip_out2 (keep = pred rename = (pred = p0));
run;
/* SAMPLE OUTPUT OF PROC NLMIXED:
Fit Statistics
-2 Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)

Parameter
b0
b1
b2
b3
b4
b5
b6
b7
tau

Estimate
-1.3944
-0.2685
0.3223
0.1391
0.1040
0.07254
-0.00039
0.04292
-1.8406

Parameter
Standard
Error
0.2698
0.09606
0.05980
0.02195
0.02789
0.03383
0.004641
0.04216
0.4585

5768.7
5786.7
5786.7
5844.2

Estimates
DF
4406
4406
4406
4406
4406
4406
4406
4406
4406

t Value
-5.17
-2.80
5.39
6.34
3.73
2.14
-0.08
1.02
-4.01

Pr > |t|
<.0001
0.0052
<.0001
<.0001
0.0002
0.0321
0.9331
0.3087
<.0001

Alpha
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05

*/

In Figure 5.1 and 5.2 below, plots between observed probability and predicted probability are used to visualize the
goodness-of-fit of ZIP and ZIP(tau) regressions. It is clear that both models fit the observed count outcomes as well
as Negative binomial and Hurdle regression.
Figure 5.1, Comparison between Observed and Predicted Probability from ZIP Regression

SAS Global Forum 2008

Statistics and Data Analysis

Figure 5.2, Comparison between Observed and Predicted Probability from ZIP( ) Regression

While plotting the prediction can be used as an informal way to assess goodness-of-fit, Vuong test is considered a
better method to compare ZIP regression to other non-nested models for count data, such as Poisson regression,
Negative Binomial regression, or Hurdle regression. If we define

P (Y | X )
mi = Log 1 i i
P2 (Yi | X i )

(4.3)

where PN(Yi|Xi) is the predicted probability of observed count for case i from model N, then Vuong statistic to test the
hypothesis E(mi = 0) is expressed as

V=

1 n

n i=1 mi
n

(4.4)

1 n
(mi m )2

i =1
n

If V > 1.96, the first model is preferred. If V < -1.96, then the second one is preferred. SAS implementation of Vuong
test to compare ZIP regression to Poisson regression is given below.
data poi_pred (keep = poi_prob);
set poi_out;
/* OUTPUT FROM POISSON REGRESSION */
do i = 0 to 8;
poi_prob = pdf('poisson', i , Yhat);
if hosp = i then output;
end;
run;
data zip_pred (keep = zip_prob);
merge zip_out1 zip_out2;
/* OUTPUT FROM ZIP REGRESSION */
do i = 0 to 8;
if i = 0 then zip_prob = p0 + (1 - p0) * pdf('poisson', i, Yhat);
else zip_prob = (1 - p0) * pdf('poisson', i, Yhat);
if hosp = i then output;
end;
run;
data compare;
merge poi_pred zip_pred;
m = log(zip_prob / poi_prob);
run;
proc sql;
select
mean(m)
as mbar,
std(m)
as s,
sqrt(count(*)) * mean(m) / std(m) as v
from
compare;
quit;
10

SAS Global Forum 2008

Statistics and Data Analysis

/* RESULT OF VUONG TEST:

mbar
0.038138

s
0.375956

v
6.733444

*/

From the above result of Vuong test, it is clearly shown that ZIP regression fit the data better than Poisson regression.

6. MODELS EVALUATION
In previous sections, five models for count data have been built with the healthcare utilization data: Poisson
regression, Negative binomial regression, Hurdle regression, ZIP regression, and ZIP(tau) regression. In practice, it is
often an interest to compare these models both in statistical sense and in business sense.
Table 6.1, Estimated Coefficients of Four Models
Poisson
Regresson
INTERCEPT
EXCLHLTH
POORHLTH
NUMCHRON
AGE
MALE
SCHOOL
PRIVINS
alpha
tau
Log Likelihood
# of Parameters
AIC
BIC
Vuong Test

Neg Bin
Regresson

-3.3290
-0.7234
0.6262
0.2645
0.1864
0.1032
-0.0002
0.1087

-3.7526
-0.6979
0.6139
0.2894
0.2384
0.1539
-0.0023
0.0939
1.7667

-3046
8
6108
6159

-2857
9
5732
5790

Hurdle
Logit
Poisson
4.2311
-0.4693
0.5826
-0.9422
-0.6953
0.3373
-0.3078
0.1426
-0.2752
-0.0123
-0.1948
-0.0385
-0.0059
-0.0182
0.2589
-0.0192

Zero-Inflated
Logit
Poisson
4.2660
-0.3665
-0.9200
-0.3700
-0.5897
0.3249
-0.2801
0.1277
-0.4060
-0.0243
-0.3348
-0.0596
-0.0194
-0.0125
0.2292
0.2249

-2879
16
5790
5892

-2878
16
5788
5890
6.73

ZIP (tau)
-1.3944
-0.2685
0.3223
0.1931
0.1040
0.0725
-0.0004
0.0429
-1.8406

-2887
9
5792
5850
6.61

Highlighted Coefficients are significant at 5%.

AIC = -2 * LL + 2 * # of Parameters, BIC = -2 * LL + Log(# of cases) * # of Parameters

Estimated coefficients of all five models together with related statistics are listed in Table 6.1. While Poisson
regression provides a baseline model for count data, the other four demonstrate the better fit than the basic Poisson
regression. It is interesting to see that although Negative binomial and ZIP(tau) regressions have very different
assumption and specification, they all indicate that the information related health status and demographic determines
the frequency of hospital admissions but socio-economic variables do not. But if we take a look at Hurdle and ZIP
regressions, we should see a different story from these composite models. First of all, the coefficient significance in
the Logit component suggests that whether an individual is admitted in the hospital depends on the health status and
demographics information. However, the Poisson component indicates that the frequencies of hospital admissions
are related to both the heath status and insurance status. A potential interpretation is that if the individual with health
problem is covered by the private insurance, he/she might be admitted into a hospital more often than the one without
the coverage of private insurance.
If we make the justification of best model solely based on the statistical tests reported in Table 6.1, it is very tempting
to conclude that Negative binomial regression out-performs its counterparts for the lower AIC and BIC. On the other
hand, composite models such as Hurdle and ZIP regressions provide a greater flexibility of modeling zero outcomes
and a more intuitive interpretation. The major difference between Hurdle and ZIP regressions is that the Logit
component in Hurdle regression describes the probability of a zero count, while the Logit component in ZIP
regression estimates the probability of a zero count from the always-zero or the not-always-zero group. In general,
these two models might lead to the similar goodness-of-fit and close interpretation. However, due to the complex
parameterization, they share the same limitation and would often suffer from the over-fitting.

7. CONCLUSION
In this paper, we have reviewed several modeling strategies for count data and their implementations in SAS. Basic
Poisson models with and without the consideration of observed heterogeneity is a good starting point for count data
modeling. For count data with the evidence of over-dispersion, Negative Binomial regression with a more liberal
assumption on variance is able to provide a better solution. If the over-dispersion results from a high frequency of
zero counts, advanced composite models such as Hurdle regression and ZIP regression might give more satisfactory
fit to the data. An example in healthcare utilization has been used in our paper to demonstrate the usage of various
11

SAS Global Forum 2008

Statistics and Data Analysis

models for count data and related statistical tests. However, successfully applications can also be extended to other
business problems, such as database marketing, credit risk, and quality control.

REFERENCES
Cameron, A. C. and Trivedi, P. K. (1996), Count Data Models for Financial Data, Handbook of Statistics, Vol. 14,
Statistical Methods in Finance, 363-392, Amsterdam, North-Holland.
Cameron, A. C. and Trivedi, P. K. (2001), Essentials of Count Data Regression, A Companion to Theoretical
Econometrics, 331-348, Blackwell.
Deb, P. and Trivedi, P. (1997), Demand for Medical Care by the Elderly: A Finite Mixture Approach, Journal of
Applied Econometrics, Vol. 12, No. 3, 313-336.
Greene, W. (2002), Econometric Analysis, Prentice Hall.
Greene, W. (1994), Accounting for Excess Zeros and Sample Selection in Poisson and Negative Binomial
Regression Models, Working Paper, Department of Economics, New York University
Gurmu, S. (1997), Semi-Parametric Estimation of Hurdle Regression Models With an Application to Medicaid
Utilization, Journal of Applied Econometrics, Vol. 12, No. 3, 225-242.
Lambert, D (1992), Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing, Technometrics,
Vol. 34, No. 1, 1 14.
Mullahy, J. (1986), Specification and Testing of Some Modified Count Data Models, Journal of Econometrics, 33,
341-365
Winkelmann, R. and Zimmermann, K. F. (1995), Recent Developments in Count Data Modeling: Theory and
Application, Theory and Applications, Journal of Economic Surveys, 9, 1-24.
Winkelmann, R. (2004), Health Care Reform and The Number of Doctor Visits An Econometric Analysis, Journal of
Applied Econometrics, 19, 455 - 472.

CONTACT INFORMATION
Wensui Liu, Statistical Project Manager
ChoicePoint Precision Marketing, Alpharetta, GA
Email: [email protected]
Jimmy Cela, AVP
ChoicePoint Precision Marketing, Alpharetta, GA
Email: [email protected]
SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS
Institute Inc. in the USA and other countries. indicates USA registration.
Other brand and product names are trademarks of their respective companies.

12