Does weighting industries differently matter? In finite samples ordinary and weighted least squares yield different parameter estimates. On balance, the answer appears to be “maybe.” Instances where weighting does not change the qualitative results include Krueger’s work on the 1990s (and the extensions thereof in Table 1). BC offer the most insight on this issue with their very thorough reporting of both unweighted and weighted results using both employment and output as weights. For the 1960s and 1980s results look very similar across the weighted and unweighted specifications. But for manufacturing industries during the 1970s weighting seems to matter much more. Their unweighted results suggest a mandated fall in inequality but their weighted results–particularly with output weights–suggest a mandated rise in inequality. This difference seems particularly important when compared with Leamer’s conclusion that the 1970s was the “Stolper-Samuelson decade” with a large rise in mandated inequality. Leamer’s reports this finding only for weighted specifications, not unweighted ones. Employment and value-added weights give similar results, but that “unweighted regressions are entirely different” (fn. 5, p. 29).
Robustness to Different Decades
Different decades appear to have different product-price trends. Assuming for the moment that the evidence presented in the previous studies is correct, on balance it appears that the U.S. prices of less-skilled-labor-intensive products relative to more-skilled-labor-intensive products held relatively constant during the 1960s and 1980s but declined during the 1970s and 1990s. Thus it is important not to generalize about “the” pattern in U.S. product prices without specifying the particular time period. For example, Krueger concludes that his evidence of declining relative prices during the 1990s “is consistent with Sachs and Shatz (1994), but inconsistent with Lawrence and Slaughter (1993)” (p. 19). Because SS and LS analyze the 1980s, not the 1990s, it is not clear that there should be any necessary relationship among results across different decades.
Importantly, the results for each given decade appear to be fairly robust to the endpoints chosen. For example, LS, Leamer, and BC all find no strong trend in manufacturing relative prices during the 1980s even though each study dates the decade slightly differently.