Interpreting and Converting Poisson Regression Coefficients

1. The Coefficient as a Log Rate Ratio

In the last lesson, we derived that for a one-unit change in a predictor $x$, the ratio of the new mean to the old mean is $e^{\beta }$.

This ratio of means is what we call the Rate Ratio (RR), sometimes also called the Incidence Rate Ratio (IRR).

If we take the natural log of both sides, we get:

This shows us something fundamental: the coefficient $\beta$ from the model output is the log of the rate ratio.

To build your intuition for this, the following short video provides a very clear mathematical walkthrough. It starts from the definition of a rate ratio and shows how the Poisson model equation naturally leads to the conclusion that the coefficient is the log rate ratio.

The key takeaway is this two-step process:

1. The model estimates the coefficient $\beta$, which is the log rate ratio.
2. We exponentiate it ($e^{\beta }$) to get the rate ratio, which is what we actually interpret.

The following video reinforces this concept from a slightly different angle, referring to the exponentiated coefficient as a "multiplicative factor" and the "Incidence Rate Ratio". Seeing these terms used will help you recognize them in different contexts.

Why "Rate" Ratio?

You might be wondering why we call it a "rate" ratio when we are modeling counts. The term "rate" implies a count of events occurring over some exposure period (e.g., time, space, or population size). For example:

• Number of accidents per month.
• Number of goals per game.
• Number of Covid cases per 100,000 people.

Even when the exposure is constant for all observations (e.g., everyone is observed for one year), we are implicitly modeling a rate (events per year). The interpretation as a ratio of rates is therefore more general and accurate.

2. Practical Interpretation with Examples

Now for the most practical part: how do you take a coefficient from your model output and turn it into a meaningful sentence? The interpretation changes slightly depending on the type of predictor variable.

The following article is an excellent, practical guide that we will use as a reference for the next section.

Let's break it down by predictor type, using the hospitalization example from the article. The model is: Log(Hospitalization Count) = β0 + β1 × Smoking.

Case 1: Binary Predictor

Suppose Smoking is coded as 0 for non-smokers and 1 for smokers. The model output gives ${\hat{\beta }}_1=0.23$.

1. Calculate the Rate Ratio (RR):
   $RR=e^{{\hat{\beta }}_1}=e^{0.23}\approx 1.26$

2. Interpret the RR:

   • As a multiplicative factor: "The rate of hospitalization for smokers is 1.26 times the rate for non-smokers."
   • As a percentage change: A rate ratio of 1.26 corresponds to a 26% increase. You calculate this as $(RR-1)\times 100\%$. So, "The rate of hospitalization is 26% higher for smokers compared to non-smokers."

If the coefficient were negative, say ${\hat{\beta }}_1=-0.23$, then $RR=e^{-0.23}\approx 0.79$.

• Interpretation: "The rate of hospitalization is 21% lower for smokers compared to non-smokers." (Calculated as $(0.79-1)\times 100\%$).

Case 2: Continuous Predictor

Suppose we modeled Age (in years) instead, and our coefficient is ${\hat{\beta }}_1=0.04$.

1. Calculate the Rate Ratio (RR):
   $RR=e^{{\hat{\beta }}_1}=e^{0.04}\approx 1.041$

2. Interpret the RR: The interpretation is now per unit increase in the predictor.

   • "For each one-year increase in age, the rate of hospitalization increases by a factor of 1.041."
   • "For each one-year increase in age, the rate of hospitalization increases by 4.1%."

This interpretation is analogous to how we interpret coefficients in standard linear regression, but the effect is multiplicative, not additive.

3. Confidence Intervals for Rate Ratios

A coefficient from a model is only a point estimate. To capture the uncertainty in our estimate, we must also report a confidence interval (CI). We cannot simply add and subtract a margin of error from the rate ratio itself because its sampling distribution is skewed.

Instead, the process is:

1. Calculate the 95% CI for the coefficient $\beta$ on the log-scale, which is symmetric. The formula is the familiar:
   $\hat{\beta }\pm 1.96\times SE(\hat{\beta })$
   Your software output will always provide the standard error, $SE(\hat{\beta })$.

2. Exponentiate the lower and upper bounds of this interval to get the 95% CI for the rate ratio.
   \text{95% CI for RR} = [e^{(\hat{\beta} - 1.96 \times SE)}, e^{(\hat{\beta} + 1.96 \times SE)}]

Example:
For the smoking example, ${\hat{\beta }}_1=0.23$ and the standard error was $SE=0.1$.

1. CI for ${\beta }_1$:
   $0.23\pm 1.96\times 0.1=[0.034,0.426]$

2. CI for the Rate Ratio:
   $[e^{0.034},e^{0.426}]=[1.03,1.53]$

Interpretation of the CI: "We are 95% confident that the true rate ratio for smoking lies between 1.03 and 1.53. This means smokers have a hospitalization rate that is anywhere from 1.03 to 1.53 times that of non-smokers."

A crucial point for hypothesis testing: If the 95% CI for a rate ratio does not contain the value 1, the association is statistically significant at the 0.05 level. In our example, since the interval [1.03, 1.53] is entirely above 1, we can conclude that smoking is associated with a significantly higher rate of hospitalization.

Conclusion

In this lesson, we moved from the abstract formulation of the Poisson model to the practical skill of interpreting its results. This is the key to communicating your findings to others.

Key Takeaways:

• A Poisson regression coefficient, $\beta$, represents the log rate ratio.
• To make it interpretable, we exponentiate it ($e^{\beta }$) to get the Rate Ratio (RR).
• The RR is a multiplicative factor. An RR of 1.25 means a 25% increase in the rate for a one-unit change in the predictor, while an RR of 0.75 means a 25% decrease.
• The specific wording of the interpretation depends on whether the predictor is binary or continuous.
• Confidence intervals for RRs are calculated by exponentiating the endpoints of the confidence interval for the original coefficient, $\beta$.
• If the CI for an RR does not include 1, the effect is statistically significant.

Preview of the Next Lesson:
The Poisson model rests on a strong assumption: that the mean and variance of the count data are equal. In reality, data is often more variable than the model predicts. In the next lesson, we will explore this common problem, known as overdispersion, and understand its consequences for our model's validity and how to detect it.