The Power of Proc Nlmixed
Introduction
• Proc Nlmixed fits nonlinear mixed-effects models
(NLMMs) – models in which the fixed and random effects
have a nonlinear relationship
• NLMMs are widespread in pharmacokinetics – ? genesis
of procedure
• Nlmixed was first available in Version 7 (experimental)
and then in Version 8 (production)
Introduction (cont.)
• Nlmixed is similar to the Nlinmix and Glimmix macros but
uses a different estimation method, and is much easier to use
• Macros iteratively fit a set of GEEs, whereas Nlmixed directly
maximizes an approximation of the likelihood, integrated over
the random effects
Example: logistic regression with
residual error
• Say you design an experiment with a single treatment that
has i 1,2,3 levels
• Each treatment level is randomly assigned to a bunch of
plots, and each plot contains m = 30 trees
• Replication is balanced, so that the same number of plots
j 1,2, ,100 occur within each treatment level
• Within each plot you measure y, the number of trees within
the plot that are infected by some disease
• The objective is to see whether the treatment has an effect
on the incidence of the disease
Example (cont.)
Modeling this scenario…
yij | pij ~ Binomial m, pij
exp i ij pij
pij or log i ij
1 exp i ij 1 p
ij
ij ~ N 0, 2
Example (cont.)
Consequences of model:
exp i
E yij m i where i E pij
1 exp i
Varyij m i 1 i 1 2 m 1 i 1 i
Notice that the inflation factor is a function of i , and
not constant (i.e. it changes for each treatment level)
Example (cont.)
How do we fit this model in SAS?
• Proc Catmod or Nlin – crude model without any
overdispersion
• Proc Logistic or Genmod – simple overdispersed
model
• Glimmix or Nlinmix macros – iterative GEE
approach
• Proc Nlmixed – exact (sort of) approach
Proc Genmod code
proc genmod data=fake;
class treat;
model y/m=treat / scale=p link=logit
dist=binomial type3;
title 'Random-effects Logistic Regression
using Proc Genmod';
output out=results pred=pred;
run;
Proc Genmod Output
Criteria For Assessing Goodness Of Fit
Criterion DF Value Value/DF
Deviance 297 646.7835 2.1777
Pearson Chi-Square 297 607.1128 2.0442
Log Likelihood -2119.8941
Analysis Of Parameter Estimates
Standard Wald 95% Chi-
Parameter DF Estimate Error Confidence Limits Square Pr > ChiSq
Intercept 1 0.1389 0.0523 0.0363 0.2415 7.04 0.0080
treat 1 1 -2.0194 0.0931 -2.2019 -1.8368 470.18 <.0001
treat 2 1 1.8726 0.0964 1.6838 2.0615 377.65 <.0001
treat 3 0 0.0000 0.0000 0.0000 0.0000 . .
Scale 0 1.4297 0.0000 1.4297 1.4297
NOTE: The scale parameter was estimated by the square root of Pearson's
Chi-Square/DOF.
LR Statistics For Type 3 Analysis
Source Num DF Den DF F Value Pr > F
treat 2 297 929.55 <.0001
Glimmix macro code
%inc 'h:\SASPROGS\Glimmix macro\glmm800.sas' /
nosource;
%glimmix(data=fake, procopt=%str(method=reml
covtest maxiter=100), maxit=100, out=results,
stmts=%str(
class treat ident;
model y/m = treat / ddfm=residual; random
ident;
parms (0.25) (1) / hold=2;
title 'Random-effects Logistic Regression
using the Glimmix macro';
),
error=binomial, link=logit);
run;
Glimmix macro output
Covariance Parameter Estimates
Standard Z
Cov Parm Estimate Error Value Pr Z
ident 0.2283 0.03780 6.04 <.0001
Residual 1.0000 0 . .
Fit Statistics
-2 Res Log Likelihood 639.5
Solution for Fixed Effects
Standard
Effect treat Estimate Error DF t Value Pr > |t|
Intercept 0.1423 0.06058 297 2.35 0.0195
treat 1 -2.0654 0.09475 297 -21.80 <.0001
treat 2 1.8989 0.09614 297 19.75 <.0001
treat 3 0 . . . .
Type 3 Tests of Fixed Effects
Effect Num DF Den DF F Value Pr > F
treat 2 297 723.42 <.0001
Proc Nlmixed code
proc nlmixed data=fake tech=trureg df=297;
bounds sigma2>0;
parms mu=1 t1=-2 t2=2 sigma2=0.25;
if treat=1 then eta=mu + t1 + e;
else if treat=2 then eta=mu + t2 + e;
else eta=mu + e;
prob=exp(eta)/(1+exp(eta));
model y ~ binomial(m, prob);
random e ~ normal(0, sigma2) subject=ident;
contrast 'treat' t1, t2;
predict prob out=results;
title 'Random-effects Logistic Regression
using Proc Nlmixed';
run;
Proc Nlmixed output
Fit Statistics
-2 Log Likelihood 1473.3
Parameter Estimates
Standard
Parameter Estimate Error DF t Value Pr > |t|
mu 0.1455 0.06199 297 2.35 0.0195
t1 -2.1197 0.09782 297 -21.67 <.0001
t2 1.9499 0.09887 297 19.72 <.0001
sigma2 0.2423 0.04118 297 5.88 <.0001
Contrasts
Num Den
Label DF DF F Value Pr > F
treat 2 297 694.22 <.0001
Results
True Parameter estimates
Parameter value
Nlin Genmod Glimmix Nlmixed Nlinmix
0.0 0.14 0.14 0.14 0.15 0.14
1 -2.0 -2.02 -2.02 -2.07 -2.12 -2.02
2 2.0 1.87 1.87 1.90 1.95 1.87
2 0.25 n/a 0.144* 0.228 0.242 0.301
F ? 3163.4 929.6 723.4 694.2 699.2
• treatment is significant
• estimates are similar
• Nlimixed works
Core syntax
• Proc Nlmixed statement options
• tech=
• optimization algorithm
• several available (e.g. trust region)
• default is dual quasi-Newton
• method=
• controls method to approximate integration of
likelihood over random effects
• default is adaptive Gauss-Hermite quadrature
Syntax (cont.)
• Model statement
• specify the conditional distribution of the data given
the random effects
• e.g. y | u ~ N Xβ Zu, R
• Valid distributions:
• normal(m, v)
• binary(p)
• binomial(n, p)
• gamma(a ,b)
• negbin(n, p)
• Poisson(m)
• general(log likelihood)
Syntax (cont.)
Fan-shaped error model:
yi ~ N a bxi , xi
2
proc nlmixed;
parms a=0.3 b=0.5 sigma2=0.5;
var = sigma2*x;
pred = a + b*x;
model y ~ normal(pred, var);
run;
Syntax (cont.)
• Binomial:
model y ~ binomial(m,prob);
• General:
combin = gamma(m+1)/(gamma(y+1)*gamma(m-y+1));
loglike = y*log(prob)+(m-y)*log(1-prob)+
log(combin);
model y ~ general(loglike);
Syntax (cont.)
• Random statement
• defines the random effects and their distribution
• e.g. u ~ N 0, G
• The input data set must be clustered according to the
SUBJECT= variable.
• Estimate and contrast statements also available
Summary
Pros
• Syntax fairly straightforward
• Common distributions (conditional on the random
effects) are built-in – via the model statement
• Likelihood can be user-specified if distribution is non-
standard
• More exact than glimmix or nlinmix and runs faster than
both of them
Summary (cont.)
Cons
• Random effects must come from a (multivariate) normal
distribution
• All random effects must share the same subject (i.e.
cannot have multi-level mixed models)
• Random effects cannot be nested or crossed
• DF for Wald-tests or contrasts generally require manual
intervention
A Smidgeon of Theory…
Linear mixed-effects model
y Xβ Zu ε
ε ~ N 0, R and u ~ N 0, G
Ey | u Xβ Zu
y | u ~ N Xβ Zu, R
y ~ N Xβ, ZG Z R
Theory (cont.)
Generalized linear mixed-effect model
Ey | u g Xβ Zu
g 1 E y | u Xβ Zu
u ~ N 0, G
y | u ~ fY y | u
The marginal distribution of y cannot usually be
simplified due to nonlinear function g
Theory (cont.)
Nonlinear mixed-effect model
Ey | u hX, β, Z, u
• Link h-1 (yielding linear combo) does not exist
u ~ fU u
y | u ~ fY y | u
• The marginal distribution of y is unavailable in closed form
Conclusion
• Flexibility of proc nlmixed makes it a good choice for
many non-standard applications (e.g. non-linear
models), even those without random effects.
References
• Huet, S., Bouvier, A., Poursat, M.-A., and E. Jolivet. 2004. Statistical
Tools for Nonlinear Regression. A Practical Guide with S-PLUS and R
Examples, Second Edition. Springer-Verlag. New York.
• Littell, R.C., Milliken, G.A., Stroup, W.W., and R.D. Wolfinger. 1996.
SAS System for Mixed Models. Cary, NC. SAS Institute Inc.
• McCullogh, C.E. and S.R. Searle. 2001. Generalized, Linear, and
Mixed Models. John Wiley & Sons. New York
• Pinheiro, J.C. and D.M. Bates. 2000. Mixed-effects Models in S and
S-PLUS. Springer-Verlag. New York.
• SAS Institute Inc. 2004. SAS OnlineDoc® 9.1.3. Cary, NC: SAS
Institute Inc.
