What’s this about?
Poisson regression is used when the dependent variable is a count from a Poisson process.
Outcomes can be left-censored if they are not observed when they are below a certain level and can be right-censored if are not observed when they are above another level.
New command cpoisson fits Poisson regression models on count data and allows the counts to be left-censored, right-censored, or both. The censoring can be at constant values, or it can differ across observations.
An example of a right-censored count outcome is the number of cars in a family, where data might be top-coded at 3 or more.
An example of a left-censored count outcome is the number of cookie boxes sold by Girl Scouts if the first outcome value recorded is 10 or fewer boxes.
Left- and right-censoring combined is also known as interval-censoring.
Distinguish between censored and truncated. With censored outcomes, it is the outcomes that are not observed even though the observation is in our data; we observe the other values for the person. In truncated data, it is the observation that is entirely missing from our data. Stata has an estimator for truncated Poisson data, see [R] tpoisson.
Let’s see it work
Below we study the number of car accidents a person has during a year. The number recorded is 0, 1, 2, or 3, and 3 means 3 or more accidents. The number is right-censored.
We will model the determinants of accidents as the number of previous accidents, whether the driver is a parent, and the number of traffic tickets the driver received during the previous year.
We type
cpoisson accidents i.past i.parent i.ntickets, ul(3) irr
initial: log likelihood = -3352.1349
rescale: log likelihood = -3352.1349
Iteration 0: log likelihood = -3352.1349
Iteration 1: log likelihood = -3348.7553
Iteration 2: log likelihood = -3348.737
Iteration 3: log likelihood = -3348.737
Censored Poisson regression Number of obs = 3,000
LR chi2(8) = 312.86
Log likelihood = -3348.737 Prob > chi2 = 0.0000
| accidents | IRR Std. Err. z P>|z| [95% Conf. Interval] | |
| 1.past | 1.325026 .1026562 3.63 0.000 1.138355 1.542308 | |
| 1.parent | .644582 .0288632 -9.81 0.000 .5904226 .7037095 | |
| ntickets | ||
| 1 | 1.028125 .0511614 0.56 0.577 .9325849 1.133452 | |
| 2 | 1.094165 .0783527 1.26 0.209 .950886 1.259032 | |
| 3 | 3.015248 .241031 13.81 0.000 2.577984 3.526679 | |
| 4 | 2.615793 .4061223 6.19 0.000 1.929513 3.546166 | |
| 5 | 4.317464 1.580035 4.00 0.000 2.107268 8.845809 | |
| 6 | 2.339149 1.655636 1.20 0.230 .5842281 9.365548 | |
| _cons | .8550119 .026662 -5.02 0.000 .8043201 .9088984 | |
| 0 left-censored observations 2,827 uncensored observations 173 right-censored observations | ||
We interpret the model coefficients (or incidence-rate ratios) as if the censoring had not occurred. That is to say, as though we had seen all of the data, uncensored.
We find that past accidents predict more future accidents, that being a parent predicts fewer future accidents, and that the number of tickets generally predicts more future accidents, although having just 1 or 2 tickets has little significance.
Because of the censoring, we do not know which of the people coded as having 3 accidents really had exactly 3 accidents, or which had more.
We can, however, now make predictions of the expected uncensored number of accidents and the probabilities of any specified number of accidents, including values greater than 3.
We wonder, what are the chances anyone had more than 3 accidents in our data? Our data were officially top-coded, but were they practically top-coded? We can obtain each driver’s probability of having four or more accidents by typing
predict fourplus, pr(4,.)
We now have the probability that each driver in our sample had four or more accidents. To get the expected number of drivers who had 4 or more accidents, we simply sum these probabilities
total fourplus Total estimation Number of obs = 3,000
| Total Std. Err. [95% Conf. Interval] | ||
| fourplus | 52.32738 2.496397 47.43256 57.2222 | |
We expect 52.3 drivers in our data had more than 3 accidents, and top-coding almost certainly affected our data.
Almost certainly? We have a standard error above, but the standard error and confidence interval do not account for the probabilities having themselves been estimated. If we use margins to perform the computation, it will produce the correct standard error and confidence interval
margins , expression(3000*predict(pr(4,.))) Predictive margins Number of obs = 3,000 Model VCE : OIM Expression : 3000*predict(pr(4,.))
| Delta-method | ||
| Margin Std. Err. z P>|z| [95% Conf. Interval] | ||
| _cons | 52.32738 4.377615 11.95 0.000 43.74741 60.90735 | |
margins wants to report a mean, so we had to trick it into giving us a total by multiplying the probabilities by our sample size of 3000.
With such a small standard error and a lower bound of 43.7 on our confidence interval, we can definitively say, or at least as definitively as any statistician can say, that top-coding affected our data.
