We're all running estimations all the time. If we've a modicum of truth-seeking and are trying to figure out how the world works, we're all estimating implicit bivariate and multivariate relationships between things all the time. A lot of them are going to be wrong if we're not doing it properly in Stata, but we get a basic working model of how the world works.
Whether you're running that regression implicitly or explicitly, things go wrong if you're omitting variables. Fremling and Lott called this
Bias Towards Zero in Aggregate Perceptions.
Suppose that youth unemployment rates are a function of the business cycle, the youth minimum wage, and regulations on ease of dismissing workers that don't pan out - and interaction effects among those.
And suppose further that everyone who considers all of those variables will get the relationships right, on average. There's variance around a true mean. So among the set of people running the implicit regressions, you've got a reasonable picture of the world.
But if half the population doesn't know that they need to think about the youth minimum wage, then that's omitted variable bias. It'll screw up the coefficients on the other variables on which the variation from the omitted variables will load, and it will bias towards zero the aggregate perception of the true relationship between youth unemployment and the youth minimum wage.
I don't want to argue about youth minimum wages and youth unemployment - the relationship could be as weak as you like. But if there's a relationship and you leave it out, you get the problem.
There's been a ton of
press coverage around poor outcomes in maternity care in New Zealand. The explanations that have shown up thus far are around the number of beds in maternity care centres and overall funding.
There's another potential variable that should be considered. I don't know how big the effect of that variable is, but I know that aggregate perceptions of it are going to be biased towards zero because nobody's talking about it.
In the 1990s, maternity care shifted from being GP-led to being midwife-led. Lots of GPs left the market because the payment wasn't high enough to keep them interested, but was high enough to interest midwives with less training.
One paper at the NZAEs in 2008 presented some evidence that this affected neonatal deaths. The paper's a decade old now and likely needs strengthening, but suggests there's something there in need of investigating.
When we had our kids, we knew that some midwives had only weak training. So we made sure to book in real early to get a midwife with proper nursing training as well - and we went for a shared-care arrangement with an obstetrician at the same time. The people who most need the best trained midwives will not know that they need to move early to get one. And then we get some rather poor outcomes.
Maybe the coefficient on the variable "We changed from GP-led to midwife-led maternity care" is small, maybe it's big. But aggregate perceptions of the coefficient are going to be biased towards zero. It would be nice if it were part of the mix when folks are thinking about maternity outcomes.