Showing posts with label Effect Size. Show all posts
Showing posts with label Effect Size. Show all posts

Monday, September 21, 2015

Dynamic scoring

This is Joseph.

I have noticed this issue in the data on economic growth for a while, as a general impression.  But Robert Waldman has a past pointing out that the current data suggests that it is quite possible that lowering tax rates is an effective way to decrease growth:
I haven’t run the numbers (I can’t) but I am confident that Jeb’s plan would cause lower growth. Partly this is because of post WWII data on top marginal tax rates (I admit on labor income) and growth in OECD countries. Atheoretic estimates, if taken literally, suggest that the growth is maximized at a top rate of over 50% (some estimates are 70%). In contrast, I know of essentially no evidence published in the peer reviewed literature that lower rates cause more rapid growth (the essentially is in regressions which consider convergence that is include initial per capita GDP as a regressor — it’s negative coefficient is overwhelmingly statistically significant for the sample.
 Now it is true, as one of the commenters points out, that there is a lot of unexplained variance in the data and so it possible that this estimate is incorrect.  But think carefully about the implications: the best available estimate not only has to be wrong but the sign needs to switch (and the magnitude still needs to be large as a very small positive effect might still be essentially zero).

So what if we decide that the data is utterly unreliable.  All we would have are theoretical models that we really cannot test (as the hypothesis was proposed under Reagan and we have since seen two switches in tax rates, both going in the direction of the data -- if that isn't enough data we are basically giving up on a data based answer for decades). 

So what happens if you justify tax cuts, assuming no effects on economic growth?  Or, worse, if you model deficits as decreasing growth but not tax cuts? 

I think you would end up with some very different policy conclusions. 

Postscript: Paul Krugman also has a very nice plot of this data, after both a tax cut and a tax hike. 

Tuesday, October 4, 2011

Effect Sizes

I think that this is very insightful:

They mistake small truths for large ones, and use the small truth to obfuscate the big one. So, the truth - that a few of the unemployed don’t want to work - is exaggerated and used to hide the bigger truth, that the vast majority of unemployment has other causes.


Mark and I have often discussed how effect size can be easily overlooked in modern debates. In epidemiology, for example, it can be the case that a drug has a serious side effect that is so rare that it basically cannot change the risk-benefit calculus. So, for example, statins can cause rhabdomyolysis (as an adverse drug side effect) despite have massive benefits on all-cause mortality (in secondary prevention of cardiovascular disease). But the rare side effect is often newsworthy and may discourage patients from seeking a beneficial therapy. Fortunately, we have randomized trials to sort out what the net impact of the benefits and risks of the drugs is like across a whole population.

I think lacking these experiments makes it easy to get focused on the details in macroeconomics. Policies that may increase utility across the whole population (e.g. immigration) may have costs to individual workers. Failing to properly specify the relative effect size of different interventions may lead to a focus on "second or third order effects". Or, even worse, to misjudging the net impact of a policy.

I think that might well be correct in the example above, as well. It is certain that there are people who would hire more if the minimum wage was to drop. But it is unclear that adjusting the minimum wage would have a major impact on the >9% unemployment rate we have in the United States. We may have to look elsewhere for solutions.

Now, implementing this advice is rough. Which is why I am pleased we have experiments over here in epidemiology.