Road Deaths Down 12%? Counts, Rates and Random Noise

"Road deaths fell 12% this year." It is the kind of headline that affects everyone who drives, walks or cycles, and the kind that invites a policy victory lap. But accident counts are small, random numbers, and a drop of a few dozen may or may not mean anything.

Counts jump around by themselves

Suppose a region recorded 250 road deaths last year and 220 this year, a fall of 12%. Counts of rare events follow roughly a Poisson distribution, whose standard deviation is the square root of the mean. With a typical level of 250, the standard deviation is about 16, so a normal range of ±1.96 standard deviations is roughly 219 to 281.

The Poisson distribution calculator shows that, if the underlying level were really 250, a count of 220 or fewer would occur about 3% of the time. That is unusual but not astonishing, and it ignores that last year's 250 was itself a noisy estimate.

Comparing two years fairly

A cleaner test uses both counts. If nothing changed, each of the 470 deaths is equally likely to fall in either year, so 220 of 470 can be examined with the binomial distribution calculator. The two-sided p-value is about 0.18, and the chi-square goodness-of-fit calculator gives about 0.17. Neither is convincing evidence of a real improvement.

Rates, not just counts

Counts also ignore exposure. With 5 million residents, 250 deaths is 5.0 per 100,000 and 220 is 4.4. Distance driven, weather and population change all matter, so safety analysts prefer rates per vehicle-kilometer. The difference of two proportions calculator can compare rates when you have the populations.

Regression to the mean

Road crash frequency at a given place varies from year to year around a fairly stable mean, so a bad year tends to be followed by a better one (TRB). Road-safety research uses Poisson models for exactly this reason (SWOV).

This matters for policy claims. If a measure is introduced after an unusually bad year, part of the later improvement would have happened anyway. Studies have found a substantial regression-to-the-mean effect in crash data (Hasselt University). Longer histories and comparison areas protect against being fooled.

Seasons, small places and standard errors

Monthly counts show seasonal swings, so comparing the same months across years avoids mistaking a season for a trend. When counts are very small, as in a single town or junction, even a 50% change may sit inside random variation, which is why local safety reviews pool several years.

A useful guide is the ratio of the change to its standard error. For counts of 250 and 220, the difference of 30 has a standard error of about 22, since the variances add. A change of 1.4 standard errors is not compelling, while one of 3 or more usually is. Good reporting states the counts, the period, the exposure and the uncertainty, not just a percentage.

It also helps to ask what else changed. Weather, fuel prices, enforcement and reporting rules can all move the numbers.

None of this means the improvement is imaginary. It means one year is a weak basis for celebration or alarm. Several years of data, with rates and exposure, tell a clearer story.

A steady downward trend over many years is far harder to explain by luck.