A large portion of the social science and policy-related literature uses a different approach to the problems of underdetermination or confounding. The basic logic of these methods can be seen in Abram Harris’s classic study of the role of bank failures in the history of African-American capitalist enterprise (1936). Harris asked whether the banks failed because they were owned and run by African-Americans, or for prosaic financial reasons. He reasoned that if the ratios of real estate to business loans and bank size (a known confounder) were the cause, and race was a causally irrelevant confounder, it should be possible to divide bank failures into two groups by race, and see if the relation between financial causes and bank failure held up within each group. If race was causally irrelevant, the relationship between the ratios and bank size and bank failure should continue to hold in both groups. If race were the cause and size and loan ratios were irrelevant, that relation should disappear in both groups.
This is the basic logic of statistical approaches to confounding. In this case there is background knowledge, knowledge about the kinds of financial variables that might be relevant to bank failure, and background knowledge about race. But little depends on the general validity of theoretical formulae. The statistical test of partialing or dividing the cases into groups determines which was the cause and which was the confounder.
Question: Which way does the causal relationship go in Table 2a? Table 2b?
Bank Failures in %
Note: These are not real data. Reality is never so simple