The NLMIXED Procedure

Example 83.4 Poisson-Normal Model with Count Data

(View the complete code for this example.)

This example uses the pump failure data of Gaver and O’Muircheartaigh (1987). The number of failures and the time of operation are recorded for 10 pumps. Each of the pumps is classified into one of two groups corresponding to either continuous or intermittent operation. The data are as follows:

data pump;
   input y t group;
   pump = _n_;
   logtstd = log(t) - 2.4564900;
   datalines;
 5  94.320 1
 1  15.720 2
 5  62.880 1
14 125.760 1
 3   5.240 2
19  31.440 1
 1   1.048 2
 1   1.048 2
 4   2.096 2
22  10.480 2
;

Each row denotes data for a single pump, and the variable logtstd contains the centered operation times.

Letting denote the number of failures for the jth pump in the ith group, Draper (1996) considers the following hierarchical model for these data:

The model specifies different intercepts and slopes for each group, and the random effect is a mechanism for accounting for overdispersion.

The corresponding PROC NLMIXED statements are as follows:

proc nlmixed data=pump;
   parms logsig 0 beta1 1 beta2 1 alpha1 1 alpha2 1;
   if (group = 1) then eta = alpha1 + beta1*logtstd + e;
   else eta = alpha2 + beta2*logtstd + e;
   lambda = exp(eta);
   model y ~ poisson(lambda);
   random e ~ normal(0,exp(2*logsig)) subject=pump;
   estimate 'alpha1-alpha2' alpha1-alpha2;
   estimate 'beta1-beta2' beta1-beta2;
run;

The selected output is as follows.

Output 83.4.1: Dimensions Table for Poisson-Normal Model

The NLMIXED Procedure

Dimensions
Observations Used10
Observations Not Used0
Total Observations10
Subjects10
Max Obs per Subject1
Parameters5
Quadrature Points5


The "Dimensions" table indicates that data for 10 pumps are used with one observation for each (Output 83.4.1).

Output 83.4.2: Iteration History for Poisson-Normal Model

Iteration History
IterationCallsNegative
Log
Likelihood
DifferenceMaximum
Gradient
Slope
1430.69869322.1627685.10725-91.6020
2930.02554680.6731462.76174-11.0489
31229.72632500.2992222.99040-2.36048
41628.73902630.9872992.07443-3.93678
51828.31619330.4228330.61253-0.63084
62128.09564000.2205530.46216-0.52684
72428.04380240.0518380.40505-0.10018
82728.03571340.0080890.13506-0.01875
93028.03392500.0017880.026279-0.00514
103328.03387440.0000510.004020-0.00012
113628.03387271.681E-60.002864-5.09E-6
123928.03387243.199E-70.000147-6.87E-7
134228.03387242.532E-90.000017-5.75E-9

NOTE: GCONV convergence criterion satisfied.


The "Iteration History" table indicates successful convergence in 13 iterations (Output 83.4.2).

Output 83.4.3: Fit Statistics for Poisson-Normal Model

Fit Statistics
-2 Log Likelihood56.1
AIC (smaller is better)66.1
AICC (smaller is better)81.1
BIC (smaller is better)67.6


The "Fit Statistics" table lists the final log likelihood and associated information criteria (Output 83.4.3).

Output 83.4.4: Parameter Estimates and Additional Estimates

Parameter Estimates
ParameterEstimateStandard
Error
DFt ValuePr > |t|95% Confidence LimitsGradient
logsig-0.31610.32139-0.980.3508-1.04290.4107-0.00002
beta1-0.42560.74739-0.570.5829-2.11621.2649-0.00002
beta20.60970.381491.600.1443-0.25301.4725-1.61E-6
alpha12.96441.382692.140.0606-0.16326.0921-5.25E-6
alpha21.79920.549293.280.00960.55683.0415-5.73E-6

Additional Estimates
LabelEstimateStandard
Error
DFt ValuePr > |t|AlphaLowerUpper
alpha1-alpha21.16531.485590.780.45290.05-2.19524.5257
beta1-beta2-1.03540.83899-1.230.24840.05-2.93310.8623


The "Parameter Estimates" and "Additional Estimates" tables list the maximum likelihood estimates for each of the parameters and two differences (Output 83.4.4). The point estimates for the mean parameters agree fairly closely with the Bayesian posterior means reported by Draper (1996); however, the likelihood-based standard errors are roughly half the Bayesian posterior standard deviations. This is most likely due to the fact that the Bayesian standard deviations account for the uncertainty in estimating , whereas the likelihood values plug in its estimated value. This downward bias can be corrected somewhat by using the distribution shown here.