On white flight (from the comments)
Are whites fleeing from Asian-heavy California public schools? One recent paper suggested maybe so, but abc raises some doubts:
I don’t want to dismiss the paper out of hand, as I have seen time and again the challenges communities face both in and outside of the school setting in accommodating demographic change.
However, I don’t think the headline result in this paper is particularly credible. First, there isn’t a well-articulated research question to guide the choice of regression. Second, the authors implicitly rely on the “an instrument is always better” fallacy rather than explaining why their instrument yields more reliable estimates than naive OLS for the (unstated) question of interest. Taken together, the paper is undergrad-thesis level material elevated only by a click bait topic and result. If we want to make bold claims about White animosity towards Asians (a claim that also constructive of such animosity and counter-animosity from Asians towards Whites) we should demand substantive evidence. This paper does not present such evidence.
Some key takeaways:
(1) The authors note that a mechanical housing market replacement would suggest a one-for-one effect, but say that their -1.47 effect is above that threshold. However, if we check the confidence interval using a conservative 1.96 critical value and the estimated standard error of the coefficient estimate, we have -1.47 + 1.96*0.268 = -0.96 so that we are not statistically significantly different from -1 by this measure.
(2) The naive OLS estimate in high-SES regions is -0.6, well below the fixed enrollment effect of -1. The authors speculate that OLS may be biased downward because the error term include unmeasured district quality changes that draw in both Asians and Whites. (Note such a correlation only operates if enrollment is not capped, so inconsistent with that model.) The authors don’t document any of these omitted variable issues, however, and just assert that their instrument will be better.
(3) Authors do not substantively engage issues with their IV. First, the IV doesn’t account for changes in composition of immigrants over time (increasing wealth and education of Asian arrivals relative to earlier waves) nor does it account for movement of second-generation Asian families. If there is no omitted variable bias but the instrumented entry is lower than the actual entry, then mechanically the coefficient will have to be higher to offset this effect and restore least-squares minimization.
(4) The instrumented Asian inflows coefficient could pick up effects from Asian-agglomeration effects. A one unit increase in Asian enrollment from pure fixed-pattern immigration flows made lead to shifts of previously settled Asians or shift the direction of subsequent immigration. For example, a settled Korean in Riverside who sees large increases in Korean population in Orange County may see OC as being more attractive than before and move into the area. This induced shift may be only partially captured by the first-stage prediction, leaving the 2nd stage coefficient of interest to increase in magnitude.
(5) Various sensitivities lead to surprising results. First, the instrument behaves poorly in some subsamples, e.g. the bottom-half of the SES scale. Why should we believe an instrument in one data subset when it plainly fails in the complement? Second, the instrument is insignificant in the Bottom Tercile of the above-median SES group (appendix table 2). Third, the IV estimate is only -0.841 in the top tercile of the above-median SES group, again below the key -1 threshold if enrollment caps are binding. Taken together, are we to think that we can identify white flight using this instrument only for the 66.6th to 83.3th percentile bucket?
(6) There’s just a big background trend issue that one has to worry about here. The theory of white flight begs the question of “flight to where?” However if we just look at Appendix Figure 2 during this time period there is a big drop in total White enrollment (and a small decline in Black enrollment) while Asian and Hispanic enrollment see big increases. To what extent are we just finding that aging out of whites in high-SES regions is being replaced disproportionately by Asians?
(7) A couple other wrinkles: how are mixed-race students handled? how would demographic shifts in total enrollment by district affect the 1-to-1 threshold? If child population is shrinking over time (e.g. because families are leaving CA, children per family is declining) then normal churn would predict more than 1-to-1 replacement of new-cohort race versus previous-cohort race.
So perhaps the right answer is “no”?