…the results reported by Clark do not reflect a universal law of social mobility. Quite to the contrary, other studies based on group-average data, even surnames data, frequently produce intergenerational coefficient estimates much smaller than Clark’s.
A second testable prediction of Clark’s hypothesis…is that instrumental variables (IV) estimation of the regression of son’s log earnings on father’s log earnings should yield a coefficient estimate in the 0.7-0.8 range if father’s long earnings are instrumented with grandfather’s log earnings. When Lindahl et al, estimated that regression with their data from Malmo, Sweden, the IV coefficient estimate was 0.15, considerably higher than their ordinary least squares (OLS) estimate of 0.303. They obtained a remarkably similar comparison of IV and OLS estimates when they used years of education instead of log earnings as the status measure. The pattern of IV estimates exceeding OLS estimates is consistent with Clark’s general story about measurement error in particular indicators as proxies for social status. It is equally consistent with all the alternative stories listed in section II for why grandparental status may not be “excludable” from a multigenerational regression. What the results are not consistent with is a universal law of social mobility in which the intergenerational coefficient is always 0.7 or more…
A third testable prediction…is that using an omnibus index that combines multiple indicators of social status should make the intergenerational coefficient estimate “much closer to that of the underlying latent variable.” [But]…The resulting estimate was not “much closer” to the 0.7-0.8 range.
In sum, when Clark’s hypothesis is subjected to empirical tests, it does not fare so well.
Here is an ungated version.