SAS PROC GLIMMIX Explained

Introduction

Clinical-trial data frequently violate the assumptions of ordinary regression models.

Patients may contribute multiple observations over time. Patients may be clustered within centers. Subjects may contribute correlated binary outcomes. Counts may be overdispersed. Ordinal outcomes may have a natural ordering but cannot reasonably be analyzed with ordinary linear regression.

These situations require models that can account for both the distribution of the response and the dependence among observations.

PROC GLIMMIX is SAS's general-purpose procedure for fitting generalized linear mixed models and performing inference for them.

What Is a Generalized Linear Mixed Model?

A generalized linear mixed model, or GLMM, has three important components:

1. A probability distribution for the response.
2. A link function connecting the expected response to the predictors.
3. Fixed and random effects describing systematic and subject-specific variation.

Conceptually, the model can be written as:

where:

• \(g(\cdot)\) is the link function
• \(\mu_{ij}=E(Y_{ij}\mid b_i)\)
• \(\boldsymbol{\beta}\) contains fixed-effect parameters
• \(\mathbf{b}_i\) contains random effects for subject or cluster \(i\)
• \(\mathbf{x}_{ij}\) and \(\mathbf{z}_{ij}\) are design vectors

The random effects introduce dependence among observations from the same subject or cluster.

Why Not Just Use PROC LOGISTIC?

PROC LOGISTIC is an excellent procedure for ordinary logistic regression.

However, ordinary logistic regression assumes that the observations are independent, conditional on the fixed effects in the model.

Suppose every patient has four visits:

These observations are generally not independent because they come from the same patient.

A patient's treatment response at Week 8 is likely related to that same patient's response at Week 12.

PROC GLIMMIX allows the analysis to model this within-subject dependence.

PROC GLIMMIX Can Also Fit Ordinary Linear Mixed Models

An important conceptual point is that GLMMs include ordinary linear mixed models as a special case.

If the response is normally distributed and the identity link is used, the model becomes a conventional linear mixed model.

Thus, PROC GLIMMIX overlaps with PROC MIXED.

This does not mean that the two procedures should be treated as completely interchangeable. Their syntax and default approaches can differ, especially when modeling residual covariance structures.

The Basic PROC GLIMMIX Structure

A minimal PROC GLIMMIX program often looks like this:

proc glimmix data=analysis;
    class treatment sex;
    model response =
        treatment
        sex
        age
        / dist=binary
          link=logit;
run;

The major statements are:

The CLASS Statement

Categorical predictors generally belong in the CLASS statement.

proc glimmix data=analysis;
    class usubjid treatment visit sex;
    ...
run;

For a clinical-trial dataset, variables such as:

• Treatment arm
• Visit
• Sex
• Region
• Study center
• Subject identifier

may need to be treated as classification variables depending on their role in the model.

The MODEL Statement

The MODEL statement defines the response and fixed effects.

model response =
    treatment
    visit
    treatment*visit
    baseline
    / dist=binary
      link=logit;

Here:

• response is the dependent variable.
• treatment is a fixed treatment effect.
• visit is a fixed visit effect.
• treatment*visit tests whether treatment effects differ by visit.
• baseline is a continuous adjustment covariate.
• dist=binary specifies a binary response.
• link=logit specifies the logit link.

Distributions

One of the defining features of PROC GLIMMIX is the ability to select an appropriate response distribution.

The correct distribution should be determined from the scientific nature of the response and the statistical analysis plan rather than selected solely because a particular model converges.

Link Functions

The link function connects the expected response to the linear predictor.

Common examples include:

Binary Outcomes

Binary outcomes are one of the most common applications of PROC GLIMMIX in clinical research.

Examples include:

• Responder versus non-responder
• Occurrence versus non-occurrence of an adverse event
• Remission versus no remission
• Successful versus unsuccessful treatment outcome

A basic logistic GLMM is:

proc glimmix data=analysis;
    class treatment usubjid;
    
    model response(event='1') =
        treatment
        / dist=binary
          link=logit;
        
    random intercept / subject=usubjid;
run;

The random intercept allows each subject to have their own underlying propensity for the outcome.

Interpreting the Logistic Model

Suppose the treatment coefficient is:

Because the model uses a logit link, exponentiating the coefficient gives an odds ratio:

The interpretation is that the modeled odds of the outcome are approximately twice as high for the treatment comparison represented by that coefficient, holding other model terms constant.

Odds Ratio Estimates

PROC GLIMMIX can directly request odds-ratio estimates.

proc glimmix data=analysis;
    class treatment;
    
    model response(event='1') =
        treatment
        / dist=binary
          link=logit
          oddsratio;
run;

Odds ratios can also be obtained through LS-means and differences.

proc glimmix data=analysis;
    class treatment;
    
    model response(event='1') =
        treatment
        / dist=binary
          link=logit;
    
    lsmeans treatment /
        diff
        oddsratio
        cl;
run;

The current SAS documentation supports odds-ratio estimation through the MODEL statement and through LS-means-based comparisons.

Random Intercepts

The simplest random-effects model gives each subject their own random intercept.

random intercept / subject=usubjid;

Conceptually:

where:

Subjects with positive random intercepts have a higher underlying propensity for the outcome, while subjects with negative random intercepts have a lower propensity.

Why the Random Intercept Matters

Consider repeated binary response data:

The three observations from Patient 001 are not equivalent to three independent patients.

The random intercept accounts for patient-level heterogeneity and induces correlation among repeated observations.

Random Slopes

A random slope allows patients to have different trajectories over time.

random intercept visit / subject=usubjid;

Conceptually:

Now both the starting level and the time trend can vary across subjects.

Random-Effects Covariance

For a random intercept and random slope, PROC GLIMMIX estimates the variance of each random effect and their covariance unless the covariance structure is constrained.

The random-effects covariance matrix can be written as:

The covariance determines whether subjects with higher intercepts tend to have higher or lower slopes.

Repeated Measures and Residual Correlation

Repeated measurements can also be correlated through the residual covariance structure.

For example, measurements closer together in time may be more strongly correlated than measurements far apart.

An autoregressive structure might therefore be considered:

random _residual_ /
    subject=usubjid
    type=ar(1);

The `_RESIDUAL_` specification is important because PROC GLIMMIX handles residual covariance through the RANDOM statement rather than using the traditional PROC MIXED REPEATED syntax. SAS documents this distinction explicitly.

Common Covariance Structures

Compound Symmetry

Compound symmetry assumes approximately equal variances and a common correlation among repeated measurements.

This can be reasonable when the correlation does not depend strongly on the time separation.

AR(1)

An AR(1) structure assumes that observations become less correlated as the time gap increases.

For example:

• Adjacent visits: relatively high correlation
• Two visits apart: lower correlation
• Three visits apart: still lower correlation

A typical specification is:

random _residual_ /
    subject=usubjid
    type=ar(1);

Unstructured Covariance

An unstructured covariance matrix makes minimal structural assumptions.

For four visits:

This flexibility comes at a cost: many parameters must be estimated.

Binary Longitudinal Example

Suppose a clinical trial measures whether each patient has achieved a clinically meaningful response at Weeks 4, 8, 12, and 16.

The data may look like:

USUBJID   TRT01P   AVISIT   RESPONSE
001       Drug A   Week 4   0
001       Drug A   Week 8   1
001       Drug A   Week 12  1
001       Drug A   Week 16  1

002       Placebo  Week 4   0
002       Placebo  Week 8   0
002       Placebo  Week 12  0
002       Placebo  Week 16  1

A basic model might be:

proc glimmix data=analysis;
    class usubjid trt01p avisit;

    model response(event='1') =
        trt01p
        avisit
        trt01p*avisit
        / dist=binary
          link=logit;

    random intercept /
        subject=usubjid;

    lsmeans trt01p*avisit /
        ilink
        diff
        cl;
run;

What Does ILINK Do?

Model estimates are naturally expressed on the linear-predictor scale.

For a logistic model, that scale is the logit scale.

The inverse-logit transformation converts the estimate to a probability:

Thus, the ILINK option can make LS-means easier to interpret because the results are presented on the response scale rather than the logit scale.

LS-Means

Least-squares means are among the most useful features of PROC GLIMMIX for clinical-trial reporting.

For example:

lsmeans trt01p /
    ilink
    cl;

This requests model-adjusted estimates on the response scale.

For treatment-by-visit interactions:

lsmeans trt01p*avisit /
    ilink
    diff
    cl;

This can produce estimated probabilities for each treatment and visit combination and comparisons between them.

Estimated Probabilities

Suppose the model produces:

The model-estimated response probabilities are therefore approximately 62% and 38%.

The difference is:

or 24 percentage points.

Depending on the requested comparison and model, PROC GLIMMIX can also provide an odds ratio for the same treatment contrast.

Odds Ratio Versus Risk Difference

These are different estimands.

A strong statistical report should state clearly which quantity is being reported.

Count Data

PROC GLIMMIX is also useful for repeated or clustered count outcomes.

Examples include:

• Number of exacerbations
• Number of hospitalizations
• Number of adverse events
• Number of rescue-medication uses
• Number of disease episodes

A Poisson model may begin with:

proc glimmix data=analysis;
    class usubjid treatment;

    model count =
        treatment
        age
        / dist=poisson
          link=log;

    random intercept /
        subject=usubjid;
run;

Interpreting a Poisson Coefficient

Suppose:

Exponentiating:

The corresponding rate or mean ratio is approximately 0.80, depending on the model and exposure specification.

This can be described as an estimated 20% lower expected event rate relative to the reference group, holding other model terms constant.

Poisson Versus Negative Binomial

A major issue with count data is overdispersion.

The Poisson distribution imposes:

Real clinical data often violate this assumption.

If the variance substantially exceeds the mean, a negative binomial model may be more appropriate.

proc glimmix data=analysis;
    class treatment usubjid;

    model count =
        treatment
        / dist=negbin
          link=log;

    random intercept /
        subject=usubjid;
run;

Offset Variables

Count-rate analyses may require an exposure or observation-time offset.

For example, if patients have different amounts of follow-up, modeling raw event counts can be misleading.

A log exposure offset can be incorporated into the linear predictor:

The offset adjusts the expected count for differing exposure times.

Ordinal Outcomes

Clinical trials frequently use ordered outcomes.

Examples include:

• Mild / moderate / severe
• Grade categories
• Clinical severity scales
• Ordered patient-reported outcome categories

An ordinal model respects the ordering of the categories rather than treating them as unrelated nominal categories.

A cumulative-logit formulation can be expressed as:

where each category boundary has its own intercept or threshold.

Nominal Versus Ordinal Outcomes

For example, tumor subtype categories may be nominal, while toxicity grades have a natural severity ordering.

Random Center Effects

Clinical trials may involve many investigative sites.

If the scientific objective is to account for site-to-site heterogeneity, center can potentially be represented as a random effect.

proc glimmix data=analysis;
    class treatment site;

    model response(event='1') =
        treatment
        / dist=binary
          link=logit;

    random intercept /
        subject=site;
run;

This model treats sites as sampled from a broader population of possible sites.

Nested Random Effects

Hierarchical data may have several levels.

For example:

Random effects can represent multiple levels of this hierarchy when supported by the design and analysis plan.

Crossed Versus Nested Effects

An effect is nested when its levels occur within another effect.

For example, patients are nested within study centers:

Crossed effects occur when every level of one factor can potentially occur with multiple levels of another factor.

Understanding whether effects are nested or crossed is essential when specifying random effects.

Treatment-by-Visit Interactions

Longitudinal clinical-trial analyses often require a treatment-by-visit interaction.

model response(event='1') =
    treatment
    visit
    treatment*visit
    / dist=binary
      link=logit;

The interaction asks whether the treatment difference changes across visits.

Without the interaction, the model imposes a common treatment effect across all visits on the model scale.

Contrasts

The CONTRAST statement lets you test specific linear combinations of fixed effects.

contrast "Drug A vs Placebo at Week 12"
    treatment 1 -1
    treatment*visit 1 -1
    / chisq;

The exact coefficients depend on the model parameterization and reference levels.

In production clinical programming, contrasts should be explicitly derived from the model design rather than copied blindly from another analysis.

ESTIMATE Statements

The ESTIMATE statement can obtain a specific linear combination of parameters.

estimate "Treatment effect"
    treatment 1 -1
    / exp cl;

For logistic models, EXP transforms a log-odds contrast into an odds ratio.

Model-Based Versus Empirical Covariance

PROC GLIMMIX supports different approaches to estimating the covariance of fixed-effect estimates.

The model-based approach assumes the specified covariance model is correct.

Empirical or sandwich-style approaches can provide robustness against some forms of covariance misspecification.

However, robust methods do not automatically solve every problem involving small samples, sparse data, or a misspecified mean model.

Estimation Methods

GLMMs involve integration over random effects.

For nonlinear mixed models, these integrals generally do not have simple closed-form solutions.

PROC GLIMMIX therefore provides several computational approaches, including linearization methods and numerical integration approaches such as quadrature or Laplace-type methods.

Linearization

A common approach approximates the nonlinear model locally so that mixed-model methods can be applied.

This is computationally attractive and often useful for large longitudinal datasets.

However, approximations can be less accurate when the response distribution is highly nonnormal or the random-effect variance is large.

Quadrature

Numerical quadrature approximates the integration over random effects directly.

Adaptive quadrature can provide more accurate likelihood-based calculations in some low-dimensional random-effects problems.

The computational burden increases as the dimensionality of the random-effects distribution increases.

When Models Become Difficult to Estimate

GLMMs can be substantially harder to estimate than ordinary regression models.

Common causes include:

• Very sparse binary outcomes
• Complete or quasi-complete separation
• Too many random-effect parameters
• Very few clusters
• Unstructured covariance with many visits
• Near-zero random-effect variance
• Highly correlated random effects
• Extremely unbalanced treatment groups
• Rare outcome categories
• Overly complex interactions

Convergence Warnings

A model that produces output is not automatically a valid model.

Always examine the SAS log and convergence information.

Warning signs include:

• Failure to converge
• Nonpositive definite covariance matrices
• Parameters at boundaries
• Very large standard errors
• Extreme estimates
• Near-zero variance components
• Optimization warnings

Complete Separation

Suppose every responder is in the treatment group and every non-responder is in placebo.

A logistic model may attempt to send the treatment coefficient toward infinity.

Conceptually:

This can cause estimation problems.

Possible solutions depend on the analysis objective and may involve:

• Reconsidering the model specification
• Reducing unnecessary parameters
• Using an appropriate alternative model
• Considering exact or penalized methods when justified
• Following the prespecified statistical methodology

Boundary Estimates

Suppose the random-intercept variance is estimated as essentially zero.

This suggests that the data provide little evidence for subject-level random intercept heterogeneity under the specified model.

It does not automatically mean the random effect should be deleted without considering the scientific model, inferential target, and prespecified analysis.

Model Selection

PROC GLIMMIX can support model comparisons, but statistical model selection should be driven by the analysis objective rather than by automated searching through every possible covariance structure.

Candidate models may differ in:

• Fixed effects
• Random effects
• Covariance structure
• Distribution
• Link function

Information criteria such as AIC can be useful descriptive tools for comparing candidate models under appropriate circumstances.

AIC and BIC

AIC balances model fit and complexity:

where \(L\) is the likelihood and \(k\) is the number of estimated parameters.

BIC imposes a stronger complexity penalty:

Lower values indicate a better fit-complexity tradeoff under the criterion.

Missing Data

Longitudinal GLMMs can use incomplete repeated-measures data under the assumptions associated with the likelihood-based analysis.

For example, one patient might have:

Week 4    0
Week 8    1
Week 12   .
Week 16   1

The missing Week 12 measurement does not necessarily require deleting the entire patient from the analysis.

However, the missing-data assumptions remain critical.

Missing at Random

Many likelihood-based longitudinal analyses rely on a Missing At Random, or MAR, framework.

Informally, after conditioning on observed information in the model, the probability of missingness does not depend on the unobserved value itself.

MAR is an assumption about the missing-data mechanism, not a guarantee provided by PROC GLIMMIX.

Clinical-Trial Example: Binary Endpoint

Suppose the primary endpoint is clinical response at multiple visits.

A possible model specification is:

proc glimmix data=adae;
    class usubjid trt01p avisit;

    model response(event='1') =
        trt01p
        avisit
        trt01p*avisit
        baseline
        / dist=binary
          link=logit;

    random intercept /
        subject=usubjid;

    lsmeans trt01p*avisit /
        ilink
        diff
        oddsratio
        cl;
run;

This model includes:

• Treatment
• Visit
• Treatment-by-visit interaction
• Baseline adjustment
• Patient-level random intercept

Clinical-Trial Example: Repeated Count Endpoint

Suppose patients contribute repeated counts of disease-related events.

proc glimmix data=analysis;
    class usubjid trt01p avisit;

    model events =
        trt01p
        avisit
        trt01p*avisit
        / dist=negbin
          link=log;

    random intercept /
        subject=usubjid;

    lsmeans trt01p*avisit /
        ilink
        cl;
run;

If follow-up differs substantially between patients, the model may also need an appropriate exposure-time offset.

Clinical-Trial Example: Ordinal Outcome

Suppose a patient-reported outcome is categorized as:

1. None
2. Mild
3. Moderate
4. Severe

The ordering contains information.

An ordinal model can account for that ordering rather than treating the four categories as unrelated nominal outcomes.

Baseline Covariates

Baseline covariates can be included when specified by the statistical analysis plan.

model response(event='1') =
    treatment
    baseline
    age
    sex
    / dist=binary
      link=logit;

Continuous baseline variables should generally remain continuous unless there is a prespecified scientific reason to categorize them.

Continuous Covariates

Suppose age is included as a continuous covariate.

The coefficient for age describes the change in the linear predictor for a one-unit increase in age, holding the other model terms constant.

Interactions With Continuous Covariates

A treatment-by-age interaction can test whether the treatment association changes with age:

model response(event='1') =
    treatment
    age
    treatment*age
    / dist=binary
      link=logit;

Because this is a nonlinear model, interpretation should generally use estimated probabilities or contrasts at clinically meaningful ages rather than relying exclusively on individual regression coefficients.

Centering Covariates

Continuous covariates can sometimes be centered to improve interpretability.

For example:

data analysis2;
    set analysis;

    age_c = age - 60;
run;

Now the model intercept corresponds to an age of 60 rather than age zero.

Modeling Time

Time can be modeled categorically or continuously.

For clinical-trial visits such as baseline, Week 4, Week 8, Week 12, and Week 24, categorical visit is often attractive when no simple functional form is scientifically justified.

Random Effects Versus Repeated Covariance

These concepts are related but not identical.

Random effects model latent subject-level variation.

Residual covariance models dependence remaining after accounting for the fixed and random effects.

A model may contain both.

proc glimmix data=analysis;
    class usubjid treatment visit;

    model response(event='1') =
        treatment
        visit
        treatment*visit
        / dist=binary
          link=logit;

    random intercept /
        subject=usubjid;

    random _residual_ /
        subject=usubjid
        type=ar(1);
run;

Whether both structures are appropriate depends on the response distribution, design, estimation method, and analysis plan.

PROC MIXED Versus PROC GLIMMIX

This is one of the most common sources of confusion for SAS programmers.

SAS describes PROC MIXED as a special case within the broader GLIMMIX framework when the response is normal and the identity link is used.

Converting a PROC MIXED Model

Consider a PROC MIXED model:

proc mixed data=analysis;
    class usubjid visit;

    model y = treatment visit treatment*visit;

    repeated visit /
        subject=usubjid
        type=ar(1);
run;

A related GLIMMIX formulation for a normal response can use:

proc glimmix data=analysis;
    class usubjid visit;

    model y =
        treatment
        visit
        treatment*visit
        / dist=normal
          link=identity;

    random _residual_ /
        subject=usubjid
        type=ar(1);
run;

The two procedures can therefore express closely related models, but the programmer should validate the parameterization, estimation method, covariance specification, and resulting inference rather than assuming that a textual conversion is automatically equivalent.

Estimated Marginal Means

For a treatment comparison, LS-means can be requested with:

lsmeans treatment /
    diff
    cl;

For a binary outcome:

lsmeans treatment /
    ilink
    diff
    oddsratio
    cl;

The ILINK option is particularly useful because it returns estimates on the response scale.

Multiple Comparisons

When there are more than two treatment groups, multiple comparisons can be requested.

lsmeans treatment /
    diff=all
    adjust=tukey
    cl;

Other multiplicity adjustments may be appropriate depending on the analysis objective.

OUTPUT Statements

PROC GLIMMIX can create an output dataset for model-based quantities.

For example:

output out=pred
    pred(ilink)=pred_prob
    lower(ilink)=lower_prob
    upper(ilink)=upper_prob;

This can be useful for creating:

• Predicted probability plots
• Model-based longitudinal figures
• Diagnostic displays
• Custom clinical-trial tables

Predicted Probabilities

Suppose the output dataset contains predicted response probabilities across visits.

These can be plotted to show model-estimated treatment trajectories.

proc sgplot data=pred;
    series x=avisit y=pred_prob / group=trt01p;
    band x=avisit
         lower=lower_prob
         upper=upper_prob
         / group=trt01p
           transparency=0.65;

    yaxis label="Estimated Probability of Response";
    xaxis label="Visit";
run;

This provides a model-based longitudinal visualization rather than a plot of raw patient observations.

Raw Proportions Versus Model-Based Estimates

These should not be confused.

A clinical report should clearly distinguish observed summaries from model-derived estimates.

Model Diagnostics

Model diagnostics should address both statistical fit and computational stability.

Important checks include:

• Convergence
• Covariance parameter estimates
• Residual or Pearson-type diagnostics where appropriate
• Influential observations or clusters
• Outlying counts
• Rare categories
• Overdispersion
• Model-predicted versus observed patterns

Overdispersion

For Poisson models, overdispersion is particularly important.

If:

the Poisson assumption may be inadequate.

Possible explanations include:

• Unmodeled heterogeneity
• Clustering
• Excess zeros
• Incorrect mean structure
• Genuine negative-binomial variation

The appropriate response depends on the source of the extra variability.

Zero-Inflated Data

Some count datasets contain substantially more zeros than expected under a standard Poisson or negative-binomial distribution.

For example, an adverse-event count may contain many patients with zero events.

Before choosing a zero-inflated model, investigate why the zeros occur.

A large number of zeros does not automatically prove that a zero-inflated model is appropriate.

Modeling Adverse Events

PROC GLIMMIX can be useful when the adverse-event endpoint is structured as a repeated binary or count outcome.

For example, whether an adverse event is present at each assessment:

proc glimmix data=adae;
    class usubjid trt01p avisit;

    model ae_present(event='1') =
        trt01p
        avisit
        trt01p*avisit
        / dist=binary
          link=logit;

    random intercept /
        subject=usubjid;
run;

However, standard clinical-trial adverse-event summaries often use subject-level incidence rather than repeated-visit binary modeling. The analysis method should therefore follow the endpoint definition.

Subject-Level Versus Visit-Level Endpoints

This distinction is critical.

A subject-level endpoint might ask:

A visit-level endpoint instead contains multiple observations per patient.

The appropriate GLIMMIX model depends on which estimand is being analyzed.

Clinical Trial Programming Workflow

A Complete Binary GLMM Example

The following example combines many of the concepts discussed above.

proc glimmix data=analysis
              method=laplace;

    class usubjid
          trt01p
          avisit;

    model response(event='1') =
          trt01p
          avisit
          trt01p*avisit
          baseline
          / dist=binary
            link=logit
            solution
            oddsratio;

    random intercept /
          subject=usubjid;

    lsmeans trt01p*avisit /
          ilink
          diff
          oddsratio
          cl;

    ods output
          ParameterEstimates = parm
          LSMeans            = lsmeans
          Differences         = diffs;

run;

The exact estimation method, model structure, and requested output tables should be consistent with the statistical analysis plan.

Reading the Fixed-Effects Table

A typical parameter table might contain:

For a logistic model, these coefficients are on the log-odds scale.

They should generally not be interpreted as direct percentage-point changes.

Reading the Odds Ratio Table

If a treatment contrast has:

then:

An odds ratio of approximately 1.68 indicates higher modeled odds in the numerator treatment relative to the reference comparison.

Reading the LS-Means Table

Suppose the ILINK-transformed results are:

These values are much easier to communicate clinically than the corresponding logit-scale estimates.

Reference Coding Matters

SAS parameter estimates depend on the coding and ordering of CLASS variables.

Therefore, the meaning of:

treatment  1  -1

depends on which treatment level is represented by each coefficient.

Always inspect the CLASS-level information and verify the reference category.

ORDER= Options and Reference Levels

For reproducible clinical programming, explicitly controlling category ordering can be useful.

class treatment(ref='Placebo') / param=ref;

This makes the intended reference treatment explicit.

Parameterization

PROC GLIMMIX supports different CLASS-variable parameterizations.

Reference-cell coding is common in clinical-trial work:

class treatment(ref='Placebo') / param=ref;

The choice affects the interpretation of individual parameter estimates but does not change the fitted model when equivalent parameterizations are used correctly.

Why LS-Means Are Often Better Than Raw Coefficients

In a model containing interactions, interpreting a single coefficient can be misleading.

Suppose the model contains:

treatment
visit
treatment*visit

The treatment coefficient may represent the treatment difference at the reference visit rather than at every visit.

LS-means and explicit contrasts allow the programmer to estimate the treatment difference at each clinically relevant visit.

Interaction Example

Suppose treatment A and placebo have:

The treatment effect is clearly not constant on the probability scale.

A treatment-by-visit interaction can represent this changing pattern.

Model-Based Probability Plot

A useful figure can show these model-based probabilities over time with confidence intervals for each treatment group.

When to Use G-Side Random Effects

In mixed-model terminology, random effects are sometimes described as G-side effects.

Examples include:

• Random subject intercept
• Random subject slope
• Random site intercept
• Random cluster effect

These effects represent variation among subjects or clusters.

When to Use R-Side Covariance

R-side covariance describes dependence among residual observations after accounting for the modeled fixed and random effects.

In PROC GLIMMIX this can be represented through a residual random effect.

random _residual_ /
    subject=usubjid
    type=cs;

The distinction between G-side and R-side modeling is important when designing a repeated-measures model.

Random Intercept Versus AR(1)

These structures answer different questions.

Both may be scientifically reasonable in some models, but they should not be added indiscriminately.

Small Number of Clusters

Random-effects inference can become unreliable when the number of independent clusters is very small.

For example, a model with only six study centers may not support the same inference about center-level variability as a model with hundreds of centers.

The number of subjects is not necessarily the same as the number of independent clusters.

Degrees of Freedom

Mixed-model procedures may use approximations for denominator degrees of freedom.

The choice can affect standard errors, confidence intervals, and p-values.

For clinical-trial analyses, the selected inferential method should follow the SAP and the methodology appropriate for the model.

Confidence Intervals

Confidence intervals should generally accompany key effect estimates.

For an odds ratio:

where \(L\) and \(U\) are the confidence limits on the log-odds-ratio scale.

Interpreting an Odds Ratio Confidence Interval

Suppose:

The estimated odds are 1.75 times the reference odds, and the confidence interval excludes 1.

Whether this constitutes statistical significance depends on the prespecified inferential framework.

Interpreting a Rate Ratio

For a log-linked count model:

If:

the modeled event rate is approximately 28% lower for the comparison represented by the ratio.

Interpreting an Identity-Link Estimate

For a normal response with an identity link:

a treatment coefficient can be interpreted directly in the response units.

For example, an estimate of −4.2 could mean a 4.2-unit lower adjusted mean, depending on the treatment contrast.

Choosing the Distribution

A useful decision framework is:

Common Mistakes

1. Treating repeated observations as independent. Ignoring within-subject dependence can produce inappropriate inference.
2. Using the wrong distribution. A convenient model is not necessarily an appropriate model.
3. Confusing odds ratios with probability differences. These are different effect measures.
4. Ignoring the link function. Regression coefficients live on the model's linear-predictor scale.
5. Overfitting random effects. Complex random structures require sufficient information.
6. Adding unstructured covariance automatically. UN covariance can require many parameters.
7. Ignoring convergence warnings. A numerical result is not automatically a valid statistical result.
8. Interpreting interaction coefficients in isolation. Use appropriate LS-means and contrasts.
9. Forgetting the reference category. Treatment contrasts depend on coding.
10. Confusing observed proportions with model-estimated probabilities. They answer different questions.
11. Choosing a model solely by AIC. Scientific interpretation and the SAP remain central.
12. Reporting only p-values. Effect estimates and confidence intervals are generally more informative.

Production Clinical Programming Checklist

Validation Strategy

For regulatory clinical-trial programming, a GLIMMIX analysis should be validated at multiple levels.

Dataset Validation

• Confirm the analysis population.
• Verify treatment assignments.
• Verify visit derivations.
• Verify baseline values.
• Check duplicates.
• Check missingness.

Model Validation

• Verify the distribution.
• Verify the link.
• Verify CLASS variables.
• Verify reference levels.
• Verify fixed effects.
• Verify random effects.
• Verify covariance structure.
• Verify estimation method.

Output Validation

• Verify parameter estimates.
• Verify LS-means.
• Verify contrasts.
• Verify confidence intervals.
• Verify odds ratios or rate ratios.
• Verify p-values.

Independent Recalculation

One useful validation strategy is to independently reproduce selected results.

For example, if the primary result is a treatment odds ratio, the independent program should verify:

and independently confirm the confidence interval and treatment coding.

Traceability

Every reported model-based result should be traceable through:

When PROC GLIMMIX Is a Good Choice

PROC GLIMMIX is particularly useful when all three of the following are true:

• The response is not adequately represented by an ordinary normal linear model.
• The observations have a hierarchical or repeated-measures structure.
• A mixed-effects framework is scientifically appropriate.

Examples include repeated binary outcomes, clustered count outcomes, ordinal longitudinal endpoints, and other generalized responses with subject-level heterogeneity.

When PROC GLIMMIX May Not Be the Best Choice

PROC GLIMMIX is powerful, but it is not automatically the correct procedure for every nonnormal or repeated-measures problem.

Other methods may be preferable depending on the estimand and study design.

Examples include:

• GEE methods when the marginal population-average effect is the primary target.
• PROC MIXED for straightforward Gaussian mixed models.
• PROC GENMOD for generalized linear models without random effects.
• Survival-analysis procedures for time-to-event endpoints.
• Specialized ordinal or multinomial approaches when the outcome structure requires them.

GLMM Versus GEE

This distinction is especially important for longitudinal clinical-trial analysis.

Neither framework is universally superior.

The correct choice depends on the scientific question and estimand.

Subject-Specific Versus Population-Average Effects

Suppose a logistic model estimates a treatment effect.

A GLMM treatment coefficient is naturally interpreted conditionally on the random effects.

A marginal model instead targets the population-average effect.

Because logistic models are nonlinear, these two quantities generally differ.

Practical Example: Why This Matters

Suppose the conditional odds ratio from a GLMM is:

A marginal population-average odds ratio need not also equal 2.4.

Therefore, analysts should not casually substitute one modeling framework for another simply because both can analyze binary repeated measurements.

A Minimal GLIMMIX Template

For day-to-day programming, the following template is a useful starting point:

proc glimmix data=analysis;

    class
        usubjid
        treatment
        visit;

    model response =
        treatment
        visit
        treatment*visit
        baseline
        / dist=binary
          link=logit
          solution;

    random intercept /
        subject=usubjid;

    lsmeans treatment*visit /
        ilink
        diff
        cl;

run;

The analyst should then customize the distribution, link, covariance structure, random effects, estimation method, contrasts, and multiplicity handling according to the prespecified analysis.

Practical Interpretation Framework

When reviewing PROC GLIMMIX output, use the following sequence.

Example Reporting Table

The exact estimand and model-based interpretation should be stated in the table footnote.

Example Footnote

Debugging PROC GLIMMIX

When a model does not behave as expected, work systematically.

Step 1: Fit the Simplest Model

proc glimmix data=analysis;
    class treatment;

    model response(event='1') =
        treatment
        / dist=binary
          link=logit;
run;

Step 2: Add the Subject Structure

random intercept / subject=usubjid;

Step 3: Add Visit

visit

Step 4: Add the Interaction

treatment*visit

Step 5: Evaluate Covariance Complexity

Only after the basic model is stable should additional covariance complexity be considered.

Why Incremental Modeling Helps

If a complex model fails to converge immediately, it can be difficult to know whether the problem is caused by:

• The distribution
• The fixed-effects structure
• The random-effects structure
• The covariance structure
• Data sparsity
• Category separation

Building the model incrementally makes the source of the problem much easier to identify.

What to Check in the SAS Log

Do not rely exclusively on the printed output tables.

Inspect the SAS log for:

• Warnings
• Notes about convergence
• Iteration information
• Invalid likelihood calculations
• Boundary estimates
• Singular covariance warnings
• Missing or invalid observations

Clinical-Trial Programming Best Practices

• Use analysis datasets rather than raw source domains whenever appropriate.
• Make the reference treatment explicit.
• Use meaningful variable names.
• Document the model specification.
• Capture ODS output datasets.
• Validate critical contrasts independently.
• Review the SAS log as part of QC.
• Keep the production model aligned with the SAP.
• Document deviations from the prespecified analysis.
• Do not rely on numerical convergence alone.

The Most Important Concept

PROC GLIMMIX is best understood as a framework for answering a specific question:

Summary

PROC GLIMMIX is one of the most versatile mixed-model procedures in SAS.

Its primary strength is the ability to combine:

• Generalized response distributions
• Nonlinear link functions
• Fixed effects
• Random effects
• Repeated-measures structures
• Flexible covariance structures
• Model-based treatment comparisons

For clinical programmers, the most important skills are not memorizing syntax. They are understanding what each component of the model means.

A robust PROC GLIMMIX analysis should therefore proceed in this order:

References

SAS Institute Inc. The GLIMMIX Procedure. SAS/STAT User's Guide. SAS Help Center.

SAS Institute Inc. Comparing the GLIMMIX and MIXED Procedures. SAS/STAT User's Guide. SAS Help Center.

McCulloch, C.E., Searle, S.R., and Neuhaus, J.M. (2008). Generalized, Linear, and Mixed Models. Wiley.

Stroup, W.W. (2012). Generalized Linear Mixed Models: Modern Concepts, Methods and Applications. CRC Press.

Diggle, P., Heagerty, P., Liang, K.-Y., and Zeger, S. (2002). Analysis of Longitudinal Data. Oxford University Press.