Insurance Fraud! Headlines As Persuasion Plays
26th January 2010
People always have short attention spans and limited capacity for processing, so the better you are at Headlines, the more persuasive you can be. Headlines can make it look like we’re on the Central Route with complex, scientific, multisyllable words, but when you look behind them, you find they are Cues for Peripheral Route processors. Consider this health story.
Here’s the title of the research study:
Health Insurance and Mortality in Adults
Sounds like we’ve got three things: Insurance, death, and adults.
Here’s the “headline” sentence from the research abstract:
Uninsurance Is Associated with Mortality.
Sounds like we’ve got two things: Uninsurance and death.
You don’t need to read the paper, do you? People without insurance are more likely to die than people with insurance. Heck, that is so commonsensical that you don’t need to do research. Sure, it’s more complicated than the Headline and if you want to read a report with a lot of unfamiliar words and worse still with a lot of numbers and those funny Greek symbols, you can Read More About It. But, you know the Headline:
No Insurance? You Die!
Okay, let’s check that Headline.
The report actually provides the information. Using data collected as part of NHANES (Google it up; your tax dollars at work), the research team followed nearly 9,000 adults over a 3 year period. Over that 3 year time, 6,655 insured people had a death rate of 3.0% while the 2,350 uninsured people had a death rate of 3.3%, a 0.3% difference. Now, since these data come from a sampling procedure that involves randomization (Amen, brother, you don’t see that often enough in epidemiological health research), it means that the sampling procedure itself can cause numbers to vary against each other. So, we need to test that 0.3 percentage point difference to see if it is inside or outside the range expected from sampling error.
The simplest test here is called the z-test for the difference between proportions and it compares 3.3% and 3.0% and adjusts for the differences in sample size (2,350 versus 6,655). If you do that test the results are z = 1.15, p < .25, h = .02.
The key number is that “p < .25.” The Industrial Standard is typically a value of .05 meaning that if your test is less than .05 (like .04 or .003 or .0001), then your results are different from sampling error and likely to be reliable, trustworthy, real. Results greater than .05 (like .06 or .10 or .25!!!) are not different from sampling error and therefore unlikely to be reliable, trustworthy, real.
Thus, this research report finds that the test comparing death rates of 3.3% versus 3.0% fails to cross the Industrial Standard for Truth, Justice, and the American Way and therefore should be rejected as False, UnJust, and UnAmerican!
Yet it wasn’t rejected. It was accepted by the reviewers and editors of the American Journal of Public Health.
The Headline is No Insurance? You Die! And, that is not true, not even close. And, even if we call it a “gimme,” it still is a piddling effect size, that h = .02. In Windowpane terms, an h of .2 (not .02, but .2) is Small, 45/55 effect. The effect here is almost 10 times smaller than Small!
What gives here? How does a scientific, peer review, respectable and self-respecting journal publish this? Well, you get out your statistics guitar and start playing in different keys.
What the researchers do is “adjust” the data. In geekspeak, the problem they’ve got is a big error term and a small explained term. It means their ignorance is wildly larger than their knowledge. The trick, therefore, is to add variables to the test that reduce the error term making it appear that the small difference of 0.3% is actually a Small difference.
Gee, what factors besides lack of insurance kill people?
Age. Smoking. Drinking. Exercise. Self reported health.
Let me illustrate. Let’s look at two big ones, smoking and age. Compare smokers versus people who never smoked. The death rates are 4.6% and 1.7%, respectively. The comparison test shows that z = 10.003, p < .000001, and h = .17, a small Windowpane (45/55).
Compare people aged 17-24 (death rate 0.7%) versus people 55-64 (death rate 10.7%). The comparison test shows that z = 17.37, p < .00001, and h = .50, a moderate Windowpane effect (35/65).
These are much more important factors in risk of mortality especially compared to insurance status. Now, the researchers run all of these factors in front of insurance status, which reduces the error/ignorance term enormously, making that 0.3% difference look bigger and bigger and bigger until it finally crosses the magic number of p < .05 and baby, you’ve got a Headline!
And, of course, it is important to note that the authors are affiliated with Harvard University and their Medical School. A little Crimson Veritas helps as another Cue here, obviating the need to read past the Headline.
Now, it is true that if you take their data set and run all the adjustments that you will find Lack of Insurance Is Associated with Mortality. They did not make that up.
They just made up the Headline so you didn’t have to read the report and discover just how persuasive Harvard Medical School researchers are.