# Interpreting the "Population Attributable Fraction" or What Percent of Cancer Deaths are Due to Smoking?

/For the video version, click here.

This week in 150 seconds, smoking kills. OK I admit the study we're discussing is not rocking the world in terms of novel science, but it brings up an interesting epidemiologic metric that I rarely get to talk about, so here we are. So, smoking kills, yes. But how *much* does smoking kill? In other words, if we could get everyone in the country to quit smoking, how many lives would we save? To answer that question, we are going to talk about the "Population Attributable Fraction".

The PAF is a little epidemiologic metric that quantifies just how much of a certain disease can be blamed on a certain risk factor.

Let me try to give you some intuition.

Say we have 1000 people with cancer, and we want to know how many of these cases can be blamed on smoking.

Well, first, we get rid of the people who never smoked, since it's pretty hard to blame their cancer on an exposure they never had.

But what about the people who did smoke? Well, we can postulate that some of them had cancer caused by smoking, but some would have gotten cancer no matter what.

The latter should be the same proportion as the non-smoking population who develops cancer. So we can take that amount out as well.

What's left over is the population attributable fraction. For lung cancer, most estimates place that number between 75 and 90%.

Appearing in JAMA internal medicine, researchers used, for the first time, state level data to figure out the PAF for a variety of cancers based on smoking.

What they found is not surprising. Depending on the state, fully 15 to 34% of these cancers were attributable to smoking.

Kentucky was the top of the list here, with Utah at the bottom. The trend basically mirrors smoking prevalence in the states.

At the beginning, I said we were interested in figuring out how many lives we could save if we got everyone to quit smoking. But in full disclosure I now have to point out that the PAF is really a best-case scenario. Remember, all you need to calculate the PAF is data on smoking rates in people with cancer, and cancer rates in non-smokers. All of these data are still subject to the same confounding and biases and non-causality as a typical observational study – a reason I don't love the word "attributable" in the metric "population attributable fraction". Still, where there's smoke, there's often fire, and it's clear that significantly curbing smoking may go a very long way to advance public health.