Plutzer, E. and McBurnett, M. (1991). "Family Life and American Politics: The 'Marriage Gap' Reconsidered," Public Opinion Quarterly 55(1): 113-127
This article had its genesis in two articles which examined the same data for the same question and reached mutually exclusive results. Both source articles, the Weisberg discussed here, and another article from the same year by Kingston and Finkel, used NES data from 1984 to test the existence of a "marriage gap" - i.e., the hypothesis that married people vote more Republican than unmarried people. Kingston and Finkel found it, while Weisberg failed. The authors here argue that Kingston and Finkel's more nuanced approach to "unmarried" brought out the effect, and test that hypothesis. They note that the prior authors used OLS regression, while they use logistic regression. This raises the question of whether Kingston and Finkel's and Weisberg's results should be thrown out for using the wrong kind of regression. Ultimately, these authors, using a nested logistic regression model, and testing multiple elections, find that the discrepancy in earlier results is an artifact of the particular elections analyzed by those authors. They then engage in some more statistical skullduggery and conclude that the effect is possible, but that it exists only for some set of thus far non-specified conditions.
My take: The only elections where significant family effects were found were 1972 and 1984. I'm left wondering if there's an artifact of the landslide quality of those elections. In other words, as Weisberg notes, since unmarried people constitute a lot less of the electorate than married people, both because they are fewer in number and because they turn out at lower rates, could the family effects be driven in those two elections not by any actual voting gap in the different groups, but rather in the fact that in both elections, the Democratic candidate was pummeled? Could McGovern and Mondale have lost because of low turnout among "unmarrieds"? We don't know - the authors don't test this question.