Poisson Regression for Count Data in Clinical Trials

\[ \lambda_i>0 \]

\[ X= \begin{cases} 0,&\text{placebo}\\ 1,&\text{active treatment} \end{cases} \]

\[ \log(\lambda_i) = \beta_0+\beta_1X_i \]

\[ \lambda_{\text{P}} = e^{\beta_0} \]

\[ \lambda_{\text{A}} = e^{\beta_0+\beta_1} \]

\[ \frac{\lambda_{\text{A}}}{\lambda_{\text{P}}} = e^{\beta_1} \]

\[ \boxed{\text{Rate Ratio}=e^{\beta_1}} \]

\[ 100\times(e^\beta-1)\% \]

\[ \beta=\log(0.70) \]

\[ e^\beta=0.70 \]

\[ \log(\mu_i) = \log(T_i) + \beta_0+ \beta_1X_{i1} +\cdots+ \beta_kX_{ik} \]

\[ \log(T_i) \] is the offset.

Equivalently, the event rate \(\lambda_i\) satisfies:

and:

This is one of the most important concepts in clinical Poisson regression.

Why the Offset Has a Fixed Coefficient

An offset differs from an ordinary predictor.

For an ordinary covariate:

the coefficient \(\beta\) is estimated from the data.

For the offset:

the coefficient is fixed at one.

This encodes the assumption that expected event counts are proportional to exposure time.

Clinical Example: Asthma Exacerbations

Suppose a randomized clinical trial compares an active treatment with placebo for prevention of asthma exacerbations.

The primary count outcome is the number of exacerbations experienced during follow-up.

Because patients may have different observation times, the analysis will model the exacerbation rate.

Suppose the analysis population is summarized as follows:

Treatment	Patients	Total Exacerbations	Total Person-Years
Placebo	200	100	90
Active	200	70	95

The crude event rates are:

and:

The crude rate ratio is approximately:

Thus, before covariate adjustment, the active treatment has approximately one-third fewer exacerbations per person-year.

Patient-Level Model

The corresponding Poisson regression model could be:

The treatment effect is:

If the fitted coefficient were:

then:

The estimated exacerbation rate would therefore be approximately 33% lower in the active-treatment group, conditional on the covariates in the model.

Confidence Intervals for the IRR

Suppose the estimated treatment coefficient is:

with standard error:

A Wald 95% confidence interval on the log scale is:

Therefore:

giving approximately:

Exponentiating both limits gives the confidence interval for the IRR:

The estimated IRR would therefore be approximately:

An IRR below one indicates a lower event rate in the active-treatment group.

Hypothesis Testing

For a treatment effect, the usual null hypothesis is:

Because:

this is equivalent to:

The alternative is:

or equivalently:

A two-sided 95% confidence interval that excludes one corresponds to rejection of the two-sided null hypothesis at the 5% level under the corresponding large-sample Wald framework.

Rate Ratio Versus Risk Ratio

These quantities should not be confused.

Measure	Typical Model	Question
Risk ratio	Binary regression	How does probability of an event differ?
Odds ratio	Logistic regression	How do event odds differ?
Hazard ratio	Survival model	How does instantaneous event hazard differ?
Incidence rate ratio	Poisson regression	How does event rate differ?

The IRR compares expected event rates, typically events per unit of person-time.

Poisson Regression With a Binary Treatment Variable

The simplest clinical-trial model contains only treatment:

where:

• \(X_i=0\) for placebo
• \(X_i=1\) for active treatment
• \(T_i\) is the patient's exposure time

The treatment IRR is:

This model provides an adjusted estimate only if additional covariates are included. With treatment as the only predictor, the result is essentially a model-based comparison of the treatment-specific event rates.

Adding Baseline Covariates

Clinical trials may include prognostic baseline variables to improve precision or address prespecified adjustment requirements.

For example:

The treatment IRR remains:

but it now represents the treatment association conditional on the covariates included in the model.

Categorical Predictors

Poisson regression can include categorical predictors in exactly the same general way as other generalized linear models.

Suppose disease severity has three categories:

• Mild
• Moderate
• Severe

If mild disease is the reference category, the model might contain indicators for moderate and severe disease.

The corresponding IRRs are:

and:

Continuous Predictors

A continuous predictor can also be included.

Suppose age is measured in years:

Then:

is the multiplicative change in event rate associated with a one-year increase in age.

For example, if:

then each additional year of age is associated with a 3% higher modeled event rate, assuming the linear effect on the log-rate scale is appropriate.

Scaling Continuous Predictors

A one-unit increase is not always clinically meaningful.

Suppose baseline biomarker concentration is measured in units where a one-unit difference is trivial. Instead, the predictor can be scaled per 10 units.

Then:

represents the IRR associated with a 10-unit increase rather than a one-unit increase.

Interactions

Poisson regression can also evaluate whether a treatment effect differs across subgroups.

For example:

The interaction coefficient tests whether the treatment effect differs by sex on the log-rate scale.

The Most Important Assumption: Equidispersion

The standard Poisson model assumes:

In real clinical data, the variance frequently exceeds the mean.

For example, suppose:

but:

The observed variance is much larger than the mean.

This is evidence of possible overdispersion.

Why Does Overdispersion Matter?

If the mean model is appropriate but the variance is larger than assumed by the Poisson distribution, the standard Poisson model can underestimate standard errors.

That can lead to:

• Confidence intervals that are too narrow
• P-values that are too small
• Overly optimistic evidence against the null hypothesis

Sources of Overdispersion

Clinical count data can be overdispersed for many reasons.

• Unobserved patient heterogeneity
• Patients with unusually high event rates
• Clustering of events within patients
• Different underlying disease severity
• Excess zeros combined with a small number of high counts
• Dependence among recurrent events
• Incorrect model specification
• Unmodeled treatment-by-covariate interactions

Checking for Overdispersion

One simple diagnostic compares the residual deviance with its degrees of freedom.

Values substantially above one can indicate overdispersion.

Another rough diagnostic compares the Pearson chi-square statistic with its degrees of freedom:

A value close to one is consistent with the Poisson variance assumption, while a materially larger value suggests overdispersion.

Quasi-Poisson Regression

One way to account for overdispersion is a quasi-Poisson model.

The mean model remains:

but the variance is allowed to be:

where:

represents overdispersion.

The estimated regression coefficients retain the same basic interpretation, but the variance estimation is adjusted.

Negative Binomial Regression

Another widely used approach is negative binomial regression.

A common parameterization has:

and:

where:

allows the variance to exceed the mean.

The mean structure can still use the log link:

and the exponentiated treatment coefficient remains an incidence rate ratio.

Poisson Versus Negative Binomial

Feature	Poisson	Negative Binomial
Count outcome	Yes	Yes
Log link	Typically	Typically
Mean	\(\mu\)	\(\mu\)
Variance	\(\mu\)	\(\mu+\alpha\mu^2\)
Overdispersion	Not explicitly modeled	Yes
Exposure offset	Yes	Yes
Effect measure	IRR	IRR

Excess Zeros

Clinical count data can contain many patients with zero events.

For example, in a trial of a rare disease flare:

• Most patients may have zero flares.
• A small number may have one or two.
• A few patients may have many recurrent flares.

If the number of zeros is substantially greater than expected under a standard Poisson or negative binomial model, a zero-inflated or hurdle model may be considered.

Zero-Inflated Models

A zero-inflated model combines two processes.

One process determines whether an observation is in a structural-zero group.

The second process generates counts, potentially including additional zeros.

Conceptually:

Zero-inflated models should not be selected simply because the dataset contains many zeros.

There should be a scientifically defensible reason for assuming that two distinct processes generate the observations.

Hurdle Models

A hurdle model also separates zero from positive counts.

The first component models:

while the second component models the positive counts.

Hurdle models can therefore be useful when the scientific question naturally separates:

• Whether a patient experiences any events
• How many events occur among patients who experience at least one

Repeated Events and Patient-Level Dependence

A major issue in clinical trials is that recurrent events within the same patient are generally not independent.

For example, a patient who experiences one hospitalization may be at greater risk of experiencing another hospitalization.

Therefore, simply treating all events as independent Poisson observations can understate uncertainty.

The appropriate model depends on the estimand and the scientific question.

Population-Level Rate Versus Recurrent-Event Modeling

There is an important distinction between:

• Comparing overall event rates between treatment groups
• Modeling within-patient recurrent-event dependence

A simple Poisson rate model can be useful for the first question, especially when robust variance estimation or other appropriate methods account for within-patient dependence.

For more complex recurrent-event questions, alternatives may include:

• Negative binomial models
• Generalized estimating equations
• Mixed-effects count models
• Frailty models
• Andersen-Gill-type recurrent-event approaches
• Joint or multi-state models

Robust Variance Estimation

When the mean structure is appropriate but independence or variance assumptions are questionable, robust or sandwich variance estimation can provide a useful approach.

The model can retain a Poisson mean structure:

while the covariance matrix is estimated using a robust method.

This can be particularly relevant when multiple events occur within the same patient.

Worked Example: Hospitalizations

Suppose a randomized trial evaluates a treatment intended to reduce hospitalizations.

The analysis includes 300 patients.

Treatment	Patients	Hospitalizations	Person-Years
Placebo	150	120	110
Active	150	84	115

The crude rates are:

and:

The crude rate ratio is:

Thus, the active treatment has approximately 33% fewer hospitalizations per person-year in this crude comparison.

Poisson Model for the Hospitalization Example

The patient-level model is:

Suppose the fitted treatment coefficient is:

Then:

The treatment effect can be expressed as:

Therefore, the estimated hospitalization rate is approximately 33% lower under active treatment than placebo.

Predicted Rates

The model can also be used to estimate expected event rates.

For placebo:

For active treatment:

The ratio remains:

This makes the IRR especially convenient for communicating treatment effects.

Expected Counts for a Fixed Follow-Up Period

Suppose the modeled placebo rate is:

hospitalizations per person-year.

For a patient followed for two years, the expected count is:

If the treatment IRR is 0.70, the active-treatment rate is:

and the expected number over two years is:

R Implementation

In R, Poisson regression can be fitted using the generalized linear model function glm().

Suppose the dataset contains:

• events = number of events
• treatment = treatment indicator
• exposure = follow-up time
• age = baseline age
• severity = baseline severity

The basic model is:

fit <- glm(
  events ~ treatment + age + severity +
    offset(log(exposure)),
  family = poisson(link = "log"),
  data = trial
)

summary(fit)

Extract the Incidence Rate Ratio in R

The treatment coefficient is on the log-rate scale.

Exponentiating it gives the IRR.

beta_treatment <- coef(fit)["treatment"]

irr <- exp(beta_treatment)

irr

A convenient way to obtain exponentiated coefficients is:

exp(coef(fit))

The treatment coefficient should be identified explicitly rather than assuming a particular coefficient position.

95% Confidence Interval for the IRR in R

The confidence interval can be obtained on the model coefficient scale and then exponentiated.

ci_beta <- confint.default(
  fit,
  parm = "treatment"
)

ci_irr <- exp(ci_beta)

ci_irr

For a likelihood-based confidence interval:

ci_beta_profile <- confint(
  fit,
  parm = "treatment"
)

exp(ci_beta_profile)

The choice between Wald and profile-likelihood intervals should follow the analysis specification and software-validation strategy.

Checking Overdispersion in R

A simple diagnostic is:

dispersion <- sum(
  residuals(fit, type = "pearson")^2
) / df.residual(fit)

dispersion

A value substantially greater than one suggests possible overdispersion.

For example:

dispersion
# approximately 1.8

would indicate substantially more variability than the standard Poisson model expects.

Quasi-Poisson in R

A quasi-Poisson model can be fitted using:

fit_qp <- glm(
  events ~ treatment + age + severity +
    offset(log(exposure)),
  family = quasipoisson(link = "log"),
  data = trial
)

summary(fit_qp)

The treatment effect can again be exponentiated:

exp(coef(fit_qp)["treatment"])

Negative Binomial Regression in R

Negative binomial regression can be fitted using the MASS package.

library(MASS)

fit_nb <- glm.nb(
  events ~ treatment + age + severity +
    offset(log(exposure)),
  data = trial
)

summary(fit_nb)

The treatment IRR is:

exp(coef(fit_nb)["treatment"])

SAS Implementation

In SAS, Poisson regression can be fitted using PROC GENMOD.

proc genmod data=trial;
    class treatment severity;

    model events =
        treatment
        age
        severity
        / dist=poisson
          link=log
          offset=log_exposure
          type3;

    estimate "Active vs Placebo"
        treatment 1 -1
        / exp cl;
run;

Here log_exposure should contain the logarithm of the patient's exposure time.

For example:

data trial;
    set trial;
    log_exposure = log(exposure);
run;

Negative Binomial Regression in SAS

The same general procedure can fit a negative binomial model.

proc genmod data=trial;
    class treatment severity;

    model events =
        treatment
        age
        severity
        / dist=negbin
          link=log
          offset=log_exposure
          type3;

    estimate "Active vs Placebo"
        treatment 1 -1
        / exp cl;
run;

The exponentiated treatment estimate is interpreted as an incidence rate ratio.

Choosing the Reference Treatment

The treatment coding determines the interpretation of the coefficient.

Suppose placebo is the reference category.

Then:

An IRR below one favors active treatment when lower event rates are clinically desirable.

If the reference category is reversed, the estimated coefficient changes sign and the IRR becomes its reciprocal.

Interpreting an IRR of 0.65

Suppose the analysis reports:

Then the active-treatment event rate is estimated to be 65% of the placebo event rate.

The relative reduction is:

or:

Thus, a useful clinical interpretation is: the estimated event rate was approximately 35% lower with active treatment than with placebo.

Interpreting an IRR Greater Than One

Suppose:

Then the event rate is 1.40 times the reference rate.

The percentage increase is:

Therefore, the event rate is approximately 40% higher.

Model-Based Rate Predictions

Poisson regression can also provide predicted rates for specific patient profiles.

Suppose:

and:

For a placebo patient:

For an otherwise identical active-treatment patient:

The ratio is:

which is exactly the treatment IRR:

Model Diagnostics

Poisson regression should not be treated as a black box.

Useful diagnostic questions include:

• Is the count outcome defined correctly?
• Is exposure time correctly measured?
• Is the offset appropriate?
• Is the mean-variance relationship plausible?
• Is there evidence of overdispersion?
• Are there influential observations?
• Are there extreme counts?
• Are there excess zeros?
• Are events clustered within patients?
• Are important covariates or interactions omitted?

Residual Diagnostics

Residuals can help identify observations that do not fit the model well.

Common residual types include:

• Pearson residuals
• Deviance residuals
• Working residuals

For example, in R:

plot(
  fitted(fit),
  residuals(fit, type = "pearson"),
  xlab = "Fitted values",
  ylab = "Pearson residuals"
)

abline(h = 0, lty = 2)

The goal is not simply to obtain a visually perfect plot. Residual patterns should be interpreted in relation to the clinical data and the assumed mean structure.

Influential Observations

A small number of patients with unusually large event counts can have a substantial influence on a Poisson model.

For example, suppose most patients have zero to two events, but one patient has 25 events.

That patient may be clinically legitimate, but the analyst should determine whether:

• The observation is correct
• The patient truly belongs in the analysis population
• The event definition was applied consistently
• The high count reflects genuine clinical heterogeneity
• The model adequately accommodates such patients

Exposure Time Must Be Defined Carefully

The offset is only as good as the exposure definition.

Potential exposure measures include:

• Time on treatment
• Time at risk
• Time under observation
• Person-years
• Person-months

The appropriate denominator depends on the estimand.

For example, if the question is treatment-emergent adverse-event rate during actual treatment exposure, exposure time might be based on time on treatment.

If the question concerns events during a fixed follow-up period, a different exposure definition may be appropriate.

Person-Time Units

The units of exposure affect the numerical rate but not the treatment IRR, provided the same unit is used consistently.

For example, suppose the rate is measured per year:

The same rate expressed per month is:

The numerical intercept changes with the time unit, but the treatment IRR does not change when the exposure unit is consistently rescaled.

Poisson Regression and Adverse Events

Adverse events are frequently represented as counts.

However, analysts must first determine what is actually being modeled.

Potential endpoints include:

• Number of patients with at least one adverse event
• Total number of adverse events
• Number of treatment-emergent adverse events per patient
• Rate of adverse events per patient-year
• Number of recurrent events of a specific preferred term

These are different estimands and may require different statistical methods.

Poisson Regression and Hospitalizations

Hospitalization counts are a classic use case because patients may experience multiple admissions and may have different follow-up times.

A rate model might estimate:

and compare treatment groups using an IRR.

For example:

would indicate a 28% lower modeled hospitalization rate for active treatment relative to placebo.

Poisson Regression and Disease Flares

Suppose a clinical trial measures the number of disease flares during follow-up.

A count model may be more informative than a binary endpoint because:

• One flare and five flares are clinically different.
• Patients can experience repeated flares.
• Follow-up may vary.

A Poisson or negative binomial model can therefore directly target the rate of recurrent flares.

When Poisson Regression Is Not Appropriate

Poisson regression should not be used simply because the outcome happens to be an integer.

Other methods may be more appropriate when:

• The endpoint is binary.
• The endpoint is continuous.
• Time to first event is the primary estimand.
• There is strong overdispersion requiring a different variance model.
• There are structural zeros requiring a two-part model.
• Within-patient recurrent-event dependence is central to the analysis.
• The count process has a fundamentally different distribution.

Poisson Regression Versus Logistic Regression

Feature	Logistic Regression	Poisson Regression
Outcome	Binary	Count
Examples	Responder / nonresponder	Number of exacerbations
Link	Logit	Log
Typical effect measure	Odds ratio	Incidence rate ratio
Exposure offset	No	Often appropriate

Poisson Regression Versus Negative Binomial Regression

The most important practical distinction is the variance structure.

Poisson:

Negative binomial:

when the dispersion parameter is positive.

Both models can use exposure offsets and produce IRRs.

Poisson Regression Versus Time-to-Event Analysis

Poisson regression and survival analysis can both involve event rates, but they answer different questions.

A time-to-event analysis often focuses on:

A Poisson rate analysis may instead focus on:

If recurrent events are clinically important, reducing the outcome to time to first event can discard information.

Conversely, if only the first event is scientifically relevant, a recurrent count analysis may answer the wrong question.

A Practical Model Selection Framework

Reporting the Statistical Model

A clinical-trial report should make the count model reproducible.

A useful description should include:

• Definition of the count endpoint
• Analysis population
• Exposure-time definition
• Offset specification
• Treatment coding
• Covariates
• Link function
• Distributional assumption
• Method for handling overdispersion
• Method for handling within-patient dependence
• Effect measure
• Confidence interval method
• Hypothesis-testing procedure

Example Statistical Analysis Plan Language

A statistical analysis plan might specify that the number of recurrent events will be analyzed using a log-linear count model with treatment as the primary predictor and the logarithm of patient-level exposure time included as an offset.

The treatment effect would be summarized using an incidence rate ratio with a 95% confidence interval.

If overdispersion is identified or prespecified, a negative binomial model may be used instead of a standard Poisson model.

The exact method should be prespecified before unblinding treatment-group comparisons when required by the trial design.

Common Mistakes

1. Ignoring exposure time. If patients have substantially different follow-up times, analyzing raw event counts can produce misleading comparisons.
2. Using follow-up time as an ordinary predictor instead of an offset. When the scientific model assumes proportional exposure, the logarithm of exposure belongs in the model with its coefficient fixed at one.
3. Assuming the Poisson distribution automatically fits count data. Count-valued data can be overdispersed, zero-inflated, or otherwise incompatible with a simple Poisson model.
4. Reporting the coefficient instead of the IRR. The regression coefficient is on the log-rate scale. Clinical interpretation usually requires exponentiation.
5. Calling an IRR a risk ratio. These are different effect measures.
6. Ignoring recurrent-event dependence. Multiple events from the same patient are generally not independent.
7. Removing patients with high counts without justification. Extreme observations should be investigated, not automatically discarded.
8. Selecting a negative binomial model solely because it produces a more favorable P-value. Model selection should be driven by the endpoint, data-generating mechanism, prespecified assumptions, and diagnostic evidence.
9. Interpreting an IRR as an absolute treatment effect. An IRR is relative. Absolute rates should also be reported when clinically useful.

Absolute Rates and Relative Effects Should Be Reported Together

An IRR alone does not tell the reader how frequent the events actually were.

Suppose:

This could represent:

events per person-year, or:

events per person-year.

The relative reduction is the same, but the clinical implications are very different.

From IRR to Absolute Rate Difference

Suppose the placebo rate is:

and:

Then the active-treatment rate is:

The absolute rate difference is:

Thus the treatment is associated with 0.36 fewer events per person-year in this example.

Prediction at Different Exposure Times

Suppose a treatment-specific rate is:

For a six-month period:

For two years:

The expected count changes with exposure time even though the underlying event rate is unchanged.

What the Offset Does Mathematically

Starting with:

subtracting the offset gives:

Since: