StATS: Modeling a declining count variable (created 2008-06-14).

I've been working on an interesting project that requires Poisson regression. A company sends out a mailing and gets a certain number of telephone calls back on each of the days following. The number of phone calls is typically (but not always) highest on the first day afterwards and declines rapidly on successive days. I wanted to develop a simple Poisson regression model for this data.

Here's an example of this data

1 9
2 4
3 2
4 3
5 0
6 0
7 1

You can fit a simple Poisson regression model in R using the following code:

ct <- c(9,4,2,3,0,0,1)
tm <- 0:6
pmod <- glm(ct~tm,family=poisson)

Using 0 as the starting point rather than 1 makes the interpretation of this model simpler. When you list the contents of  pmod, you get the following

Call: glm(formula = ct ~ tm, family = poisson)

Coefficients:
(Intercept)           tm
     2.1063      -0.5505

Degrees of Freedom: 6 Total (i.e. Null); 5 Residual
Null Deviance:      22.06
Residual Deviance: 5.289        AIC: 24.21

How would you interpret these coefficients?  The Poisson regression model uses a log link function. This means that the average value for the data is assumed to be linear on a log scale.

In a linear regression model, the intercept is the estimated average value of the dependent variable when the independent variable is equal to zero. In a Poisson regression model, the antilog of the intercept represents the estimated average count when the independent variable is zero. Since exp(2.1) = 8.2, you would say that the estimated average count at time zero is 8.2.

You might be tempted to make the intercept equal to 9. While this is not an outrageous thing to do, it is not quite as efficient. The Poisson regression model has to fit all of the counts well, not just the count at 0.

In a linear regression model, the slope represents the estimated average change in the dependent variable when the independent variable increases by one unit. In a Poisson regression model, the antilog of the slope represents the estimated average change in the count when the independent variable increases by one unit. An important difference, though, is that the change is a multiplicative change. Since exp(-0.55) = 0.58, you would say that the estimated average count declines by a factor of 0.58 (a 42% decline) when time increases by one day.

A zero slope in a Poisson regression model corresponds to a multiplicative change of 1 which, of course, is no change. A positive slope in a Poisson regression model corresponds to a multiplicative change greater than 1. You could call this an exponential rise. A negative slope in a Poisson regression model corresponds to a multiplicative change less than 1. This is an exponential decline.

The predict function in R produces estimated average values on a log scale for each time point.

> round(predict(pmod),4)
     1      2      3      4       5       6       7
2.1063 1.5558 1.0053 0.4548 -0.0957 -0.6462 -1.1967

I've rounded these data slightly from the actual output to save space. Note that R places an index of 1 through 7 above each predicted value. This is a bit confusing, perhaps, because the first value corresponds to time=0 and the last value corresponds to time=6.

If you take the antilog of these values, you get the estimated average count at each time point.

> round(exp(predict(pmod)),4)
     1      2      3      4      5      6      7
8.2177 4.7388 2.7327 1.5758 0.9087 0.5240 0.3022

To demonstrate that this is an exponential decline, note that

8.2*0.58^0 = 8.2
8.2*0.58^1 = 4.8
8.2*0.58^2 = 2.8
8.2*0.58^3 = 1.6
8.2*0.58^4 = 0.93
8.2*0.58^5 = 0.54
8.2*0.58^6 = 0.31

There is a bit of rounding here, but the series matches the values shown using the antilog of the predicted values.

You might wonder if you are likely to see any more values beyond time=6. The estimated average counts for times 7, 8, etc. are

8.2*0.58^7 = 0.18
8.2*0.58^8 = 0.11
 .
 .
 .

You can sum the infinite series from time=0 to time=infinity if you like. Notice that the terms in the above expression can be represented by a classic geometric series

a
ar
ar2
ar3
 .
 .
 .

For r<1, the sum of this infinite series is

a/(1-r)

For a=8.2 and r=0.58, this equals 19.5. Since we've already seen 19 events, we would expect to see half an event on average for the remainder the time. You can use the dpois function to estimate probabilities for a Poisson random variable with a mean of 0.5.

> round(dpois(0:4,0.5),2)
[1] 0.61 0.30 0.08 0.01 0.00

You would estimate that it is more than likely (61% probability) that you will not see any more events. There is still a reasonable chance (30%) that you will see one more event. It is unlikely (9% probability) that you will see 2 or more events.

This page was written by Steve Simon while working at Children's Mercy Hospital. Although I do not hold the copyright for this material, I am reproducing it here as a service, as it is no longer available on the Children's Mercy Hospital website. Need more information? I have a page with general help resources. You can also browse for pages similar to this one at Category: Poisson regression.