Multiplicative Effects In Poisson Regression

Why Poisson Effects Are Multiplicative

A Poisson regression with a log link describes how predictors change a conditional expected count. Its coefficients are additive on the logarithmic scale, but their effects on the expected count are multiplicative. This distinction matters when interpreting a fitted model: a coefficient does not usually represent a fixed number of extra events. It represents a proportional change for a specified change in the predictor. These interpretations are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Definition: For observations $i=1,\ldots,n$, suppose the conditional count $Y_i$, given predictors $x_{i1},\ldots,x_{ip}$, follows a Poisson distribution with mean $\mu_i>0$. A Poisson log-link regression specifies $$ Y_i\mid x_{i1},\ldots,x_{ip}\sim\operatorname{Poisson}(\mu_i), \qquad \log(\mu_i)=\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}. $$ Here $\beta_j$ is the change in log mean for a one-unit increase in $x_j$, with the other predictors held fixed. The model assumes conditional independence of observations and, given the predictors, a Poisson mean and variance both equal to $\mu_i$.

The multiplicative interpretation follows by comparing two predicted means. Let a predictor $x_j$ increase by $\Delta$, while all other predictors remain unchanged. The new linear predictor increases by $\beta_j\Delta$. Exponentiating converts the difference in log means into a ratio of means:

\frac{\mu_{\mathrm{new}}}{\mu_{\mathrm{old}}} = \exp(\beta_j\Delta).

For a one-unit increase, the multiplier is $\exp(\beta_j)$. A multiplier greater than 1 means the expected count increases; a multiplier between 0 and 1 means it decreases. The corresponding percentage change is $100\{\exp(\beta_j\Delta)-1\}\%$. The percentage can be negative for a decrease. These statements concern conditional means, not a guarantee that a particular observed count will change by that amount.

Why Exponentiating Gives a Ratio

The key justification is the structure of the log-link model. For simplicity, hold all predictors except $x_j$ fixed, and call their contribution $c$. Then the old and new means are $\exp(c+\beta_jx_j)$ and $\exp(c+\beta_j(x_j+\Delta))$. Their ratio is:

Write the two log means.
The old log mean is $c+\beta_jx_j$; the new log mean is $c+\beta_j(x_j+\Delta)$.

Exponentiate each log mean.
The corresponding means are $\exp(c+\beta_jx_j)$ and $\exp(c+\beta_jx_j+\beta_j\Delta)$.

Divide to compare.
The common factor $\exp(c+\beta_jx_j)$ cancels, leaving $\mu_{\mathrm{new}}/\mu_{\mathrm{old}}=\exp(\beta_j\Delta)$.

The cancellation explains why the multiplier does not depend on the starting value of $x_j$ in a model without interactions or nonlinear terms. The expected counts themselves still depend on the other predictors and on the starting value. If the modeled quantity is a rate, the same ratio compares rates when the exposure conditions are held equal; the interpretation of an individual coefficient still depends on the model specification and predictor units.

Interpreting the Size and Direction of an Effect

A positive coefficient gives a multiplier above 1 for a positive predictor change, while a negative coefficient gives a multiplier below 1. For example, if $\beta_j=-0.10$, a one-unit increase multiplies the mean by $\exp(-0.10)$, not by $-0.10$. The percentage change is found by subtracting 1 from the multiplier. This subtraction is essential: reporting the multiplier itself as a percentage change confuses a ratio with the change in that ratio.

For a change of $\Delta$ units, use $\exp(\beta_j\Delta)$, not just $\exp(\beta_j)$. This also means a two-unit effect is generally not twice the one-unit percentage change. The two-unit multiplier is the square of the one-unit multiplier, so the effects compound.

Worked Example: A Claim Count Predictor and Its Confidence Interval

In a hypothetical model for the number of claims during a fixed policy period, suppose a continuous risk score has estimated coefficient $\hat\beta=0.20$ and standard error $0.08$. Assume the score is measured in one-point units and the other predictors are held fixed. A policy at score zero has fitted mean count 4.00. Find the multiplier and expected count for a three-point increase, and find approximate 95% confidence intervals for one- and two-point multipliers.

Method and point estimate. For a three-point increase, $\Delta=3$, so the estimated multiplier is $\exp(3(0.20))=\exp(0.60)=1.8221$, rounded to four decimal places. The fitted mean is therefore $4.00(1.8221188)=7.2885$ claims, rounded to four decimal places. The model predicts a conditional mean, so a noninteger fitted value is appropriate.

Coefficient interval. Using the approximate Wald interval $\hat\beta\pm1.96\,SE(\hat\beta)$, the coefficient limits are $0.20\pm1.96(0.08)=0.20\pm0.1568$, or $0.0432$ to $0.3568$.

One-point multiplier interval. Exponentiating both limits gives $\exp(0.0432)=1.0441$ and $\exp(0.3568)=1.4288$, to four decimal places. Thus the approximate interval for the one-point multiplier is 1.0441 to 1.4288. In percentage-change terms, this is an increase of about 4.41% to 42.88%.

Two-point multiplier interval. Multiply each coefficient limit by 2 before exponentiating: $\exp(2(0.0432))=\exp(0.0864)=1.0902$, and $\exp(2(0.3568))=\exp(0.7136)=2.0413$, to four decimal places. The two-point interval is therefore 1.0902 to 2.0413, corresponding to an increase of about 9.02% to 104.13%.

Check and common mistake. The three-point fitted mean is greater than 4.00, consistent with a positive coefficient. A common error is to multiply the coefficient by 3 but use the one-point confidence limits unchanged. For a three-point multiplier interval, each coefficient endpoint must also be multiplied by 3 before exponentiation.

Changing the Predictor's Units

A coefficient is tied to the unit in which its predictor is measured. If $x$ is recorded in months and $z=x/12$ is the same quantity recorded in years, the model term $\beta_{\mathrm{month}}x$ can be rewritten as $(12\beta_{\mathrm{month}})z$. Thus the coefficient per year is 12 times the coefficient per month. Its standard error and confidence limits must be multiplied by 12 as well. Predictions do not change when the coefficient and predictor are converted consistently.

More generally, if one unit of a new scale equals $k$ units of the original scale, then a one-unit increase on the new scale has multiplier $\exp(k\beta)$. Always identify the scale before interpreting a reported coefficient.

Worked Example: Converting a Monthly Effect to an Annual Effect

A hypothetical model predicts the number of service requests during a fixed period. A machine's age is recorded in months, with estimated coefficient $\hat\beta_{\mathrm{month}}=0.025$ and standard error $0.010$. Its fitted mean at 24 months is 1.80. Find the multiplier and fitted mean for an increase from 24 to 36 months, and give an approximate 95% interval for the annual multiplier.

Identify the unit change. The difference is $36-24=12$ months, or one year. The estimated log-mean change is $12(0.025)=0.300$, so the multiplier is $\exp(0.300)=1.3499$, rounded to four decimal places. This is an increase of $100(1.3498588-1)=34.99\%$, rounded to two decimal places.

Calculate the new fitted mean. Multiply the fitted mean at 24 months by the annual multiplier: $1.80(1.3498588)=2.4297$ requests, rounded to four decimal places.

Construct the interval. On the annual scale, the coefficient estimate is $12(0.025)=0.300$ and its standard error is $12(0.010)=0.120$. The approximate 95% coefficient interval is $0.300\pm1.96(0.120)=0.300\pm0.2352$, or $0.0648$ to $0.5352$. Exponentiating gives $\exp(0.0648)=1.0669$ and $\exp(0.5352)=1.7078$, to four decimal places. The annual multiplier interval is 1.0669 to 1.7078.

Check and common mistake. The monthly estimate corresponds to $\exp(0.025)=1.0253$ per month, but the requested comparison spans 12 months. Using 1.0253 as the annual multiplier would use the wrong unit. The annual multiplier, 1.3499, is also the twelfth power of the monthly multiplier, apart from rounding.

Confidence Intervals on the Multiplier Scale

For a coefficient estimate $\hat\beta$ with standard error $SE(\hat\beta)$, an approximate two-sided 95% Wald interval is $\hat\beta\pm1.96SE(\hat\beta)$. This relies on the usual large-sample approximation that the estimator is approximately normal. Because the exponential function is increasing, exponentiating the lower and upper coefficient limits gives the corresponding interval for $\exp(\beta)$. For a $\Delta$-unit change, exponentiate $\Delta$ times each coefficient limit.

If $\Delta$ is negative, multiplication by $\Delta$ reverses the order of the coefficient endpoints. The transformed interval must still be reported from its smaller endpoint to its larger endpoint. Exponentiating preserves endpoint order after this reversal. These are approximate confidence intervals, not exact finite-sample guarantees.

Worked Example: A Negative Effect Over Several Units

In a hypothetical model of the number of maintenance incidents per fixed operating period, an equipment-age coefficient is $\hat\beta=-0.12$, with standard error $0.05$. Find the multiplier and percentage change for a two-unit increase. The fitted mean at the starting age is 3.20. Also find an approximate 95% interval for the two-unit multiplier.

Point estimate. For $\Delta=2$, the estimated log-mean change is $2(-0.12)=-0.24$. The multiplier is $\exp(-0.24)=0.7866$, to four decimal places. The percentage change is $100(0.7866279-1)=-21.34\%$, to two decimal places. The fitted mean at the higher age is $3.20(0.7866279)=2.5172$ incidents, to four decimal places.

Interval. The coefficient interval is $-0.12\pm1.96(0.05)=-0.12\pm0.098$, or $-0.218$ to $-0.022$. Multiplying both limits by 2 gives $-0.436$ to $-0.044$. Exponentiating yields $\exp(-0.436)=0.6466$ and $\exp(-0.044)=0.9570$, to four decimal places. The two-unit multiplier interval is 0.6466 to 0.9570, corresponding to decreases of about 35.34% to 4.30%.

Check and common mistake. Both multiplier limits are below 1, consistent with the coefficient interval being entirely negative. A common error is to describe 0.7866 as a decrease of 78.66%. It is the proportion of the starting mean that remains; the decrease is $1-0.7866=0.2134$, or 21.34%.

Scope and Common Interpretation Traps

The multiplier interpretation assumes the stated log-link model and a specified predictor change, with other modeled predictors held fixed. It applies to the conditional mean. It does not say that each observation changes by a fixed percentage, nor does it imply that the model is correctly specified. If the Poisson variance assumption is unsuitable, inference based on the model's standard errors may be unreliable even though exponentiating a fitted log-link coefficient still gives the model's mean ratio.

Do not interpret $\beta$ as a count difference. The count-scale effect depends on the starting mean; $\exp(\beta\Delta)$ is the ratio.
Do not omit the predictor units. A coefficient per month is not a coefficient per year.
Do not forget the change size. A five-unit comparison uses $\exp(5\beta)$.
Do not call a multiplier a percentage change. Subtract 1, then multiply by 100%.
Do not leave an interval on the wrong scale. To obtain a multiplier interval, transform the coefficient limits, not just the point estimate.

Key takeaway: In a Poisson log-link model, a $\Delta$-unit predictor increase multiplies the conditional mean by $\exp(\beta\Delta)$. Convert to a percentage change by subtracting 1, and exponentiate the appropriately scaled coefficient confidence limits to describe uncertainty about that multiplier.

Check Your Understanding

For each problem, state the multiplier or interval on the requested scale. Round multipliers to four decimal places and percentage changes to two decimal places.

A coefficient is $0.08$ per one-unit increase. Find the one-unit multiplier and percentage change. (Answer: multiplier $1.0833$; increase $8.33\%$)
A predictor has coefficient $0.03$ per week. Find the multiplier for a five-week increase. (Answer: $\exp(0.15)=1.1618$, an increase of $16.18\%$)
A coefficient is $-0.20$. Find the multiplier and percentage change for a three-unit increase. (Answer: $\exp(-0.60)=0.5488$, a decrease of $45.12\%$)
A coefficient estimate is $0.12$ with standard error $0.04$. Find the approximate 95% interval for the one-unit multiplier. (Answer: coefficient limits $0.0416$ to $0.1984$; multiplier limits $1.0425$ to $1.2194$)
A coefficient is $0.04$ per 1,000 dollars of a predictor. Find the multiplier and percentage change for an increase of 2,500 dollars. (Answer: $\exp(0.04(2.5))=1.1052$, an increase of $10.52\%$)

What You'll Learn