How many children are abducted each year? Did you know anyone who died in Vietnam?
Wikipedia explains that big data is about correlations, and that small data is either about the causes of effects, or is an inference from big data. None of that captures what I mean by small data.
Most people in my circles instead think small data deals with inferences about populations made from the sparse data from within those populations. For Bayesians, this means making best use of an intuitive informative prior distribution for a model. For wise non-Bayesians, it can mean bullshit detection.
In the early 90’s I taught a course on probabilistic risk analysis in aviation. In class we were discussing how to deal with estimating equipment failure rates where few previous failures were known when Todd, a friend who was attending the class, asked how many kids were abducted each year. I didn’t know. Nor did anyone else. But we all understood where Todd was going with the question.
Todd produced a newspaper clipping citing an evangelist – Billy Graham as I recall – who claimed that 50,000 children a year were abducted in the US. Todd asked if we thought that yielded a a reasonable prior distribution.
Seeing this as a sort of Fermi problem, the class kicked it around a bit. How many kids’ pictures are on milk cartons right now, someone asked (Milk Carton Kids – remember, this was pre-internet). We remembered seeing the same few pictures of missing kids on milk cartons for months. None of us knew of anyone in our social circles who had a child abducted. How does that affect your assessment of Billy Graham’s claim?
What other groups of people have 50,000 members I asked. Americans dead in Vietnam, someone said. True, about 50,000 American service men died in Vietnam (including 9000 accidents and 400 suicides, incidentally). Those deaths spanned 20 years. I asked the class if anyone had known someone, at least indirectly, who died in Vietnam (remember, this was the early 90s and most of us had once owned draft cards). Almost every hand went up. Assuming that dead soldiers and our class were roughly randomly selected implied each of our social spheres had about 4000 members (200 million Americans in 1970, divided by 50,000 deaths). That seemed reasonable, given that news of Vietnam deaths propagated through friends-of-friends channels.
Now given that most of us had been one or two degrees’ separation from someone who died in Vietnam, could Graham’s claim possibly be true? No, we reasoned, especially since news of abductions should travel through social circles as freely as Vietnam deaths. And those Vietnam deaths had spanned decades. Graham was claiming 50,000 abductions per year.
Automobile deaths, someone added. Those are certainly randomly distributed across income, class and ethnicity. Yes, and, oddly, they occur at a rate of about 50,000 per year in the US. Anyone know someone who died in a car accident? Every single person in the class did. Yet none of us had been close to an abduction. Abductions would have to be very skewed against aerospace engineers for our car death and abduction experience to be so vastly different given their supposedly equal occurrence rates in the larger population. But the Copernican position that we resided nowhere special in the landscapes of either abductions or automobile deaths had to be mostly valid, given the diversity of age, ethnicity and geography in the class (we spanned 30 years in age, with students from Washington, California and Missouri).
One way to check the veracity of Graham’s claim would have been to do a bunch of research. That would have been library slow and would have likely still required extrapolation and assumptions about distributions and the representativeness of whatever data we could dig up. Instead we drew a sound inference from very small data, our own sampling of world events.
We were able to make good judgments about the rate of abduction, which we were now confident was very, very much lower than one per thousand (50,000 abductions per year divide by 50 million kids). Our good judgments stemmed from our having rich priors (prior distributions) because we had sampled a lot of life and a lot of people. We had rich data about deaths from car wrecks and Vietnam, and about how many kids were not abducted in each of our admittedly small circles. Big data gets the headlines, causing many of us to forget just how good small data can be.