The turnover problems saying that relative wage advantage declines? Their chart on a log scale shows that the wages are still really high compared to non-CS jobs. A more likely interpretation is that junior programmers are for whatever reason highly paid in their sample. Those error bars are suspiciously tight, so there's a good chance their sample is biased.
The selection problem data? First, they're missing a plot that shows that the same correlation (higher AFQT scoring individuals are slightly more likely to leave the field) doesn't hold for other majors/fields as well. Second, is the AFQT psychometrically valid when compared across different age groups? Third, is their sample valid here at all? It wouldn't take much bias to make these correlations disappear. The super bright principle engineer is probably not going to take the time to sit an AFQT and be part of this.
Then we have weird assertions without citation like "Some workers, endowed with superior ability, learn faster than others, picking up skills at a quicker pace. Those workers will tend to sort into high-skilled, fast-changing professions initially, maximizing their early career earnings. Less impressive workers will sort into low-skilled, slower-changing professions."
A quick glance through the original paper the data is from does not fill me with confidence. The idea that maybe conventions for job postings in different fields might be different and thus mess up their data doesn't seem to have occurred to them to start with. The NLSY data they work from for AFQT scores can't be used naively. People drop out steadily after the initial tracking period.
So: bad academic research, bad interpretation of it by a layman.
I really wish that I had something comprehensive to suggest. There are some things that I think everyone who deals with statistics should read, such as Freedman's 'Statistical Models and Shoe Leather'[^1] and Tukey's 'Exploratory Data Analysis'[^2]. Pearl's 'Causal Inference in Statistics' is the best introduction I know of to issues of causality as we understand them today. For actual inference, the basis is one player game theory (aka, decision theory). I learned it from Kiefer's 'Introduction to Statistical Inference,' which sets out the theory very nicely in the first few chapters. That's a starting point at least. There are some interesting courses[^3] that try to teach statistics via resampling that seem pedagogically very valuable. Resampling does build intuition nicely and using it gets people over their squeamishness around using randomized procedures.
The turnover problems saying that relative wage advantage declines? Their chart on a log scale shows that the wages are still really high compared to non-CS jobs. A more likely interpretation is that junior programmers are for whatever reason highly paid in their sample. Those error bars are suspiciously tight, so there's a good chance their sample is biased.
The selection problem data? First, they're missing a plot that shows that the same correlation (higher AFQT scoring individuals are slightly more likely to leave the field) doesn't hold for other majors/fields as well. Second, is the AFQT psychometrically valid when compared across different age groups? Third, is their sample valid here at all? It wouldn't take much bias to make these correlations disappear. The super bright principle engineer is probably not going to take the time to sit an AFQT and be part of this.
Then we have weird assertions without citation like "Some workers, endowed with superior ability, learn faster than others, picking up skills at a quicker pace. Those workers will tend to sort into high-skilled, fast-changing professions initially, maximizing their early career earnings. Less impressive workers will sort into low-skilled, slower-changing professions."
A quick glance through the original paper the data is from does not fill me with confidence. The idea that maybe conventions for job postings in different fields might be different and thus mess up their data doesn't seem to have occurred to them to start with. The NLSY data they work from for AFQT scores can't be used naively. People drop out steadily after the initial tracking period.
So: bad academic research, bad interpretation of it by a layman.