Unions and Inequality Over the Twentieth Century: New Evidence from Survey Data∗ Henry S. Farber, Daniel Herbst, Ilyana Kuziemko, Suresh Naidu October 9, 2020 Abstract U.S. income inequality has varied inversely with union density over the past hundred years. But moving beyond this aggregate relationship has proven difficult, in part be- cause of limited micro-data on union membership prior to 1973. We develop a new source of micro-data on union membership dating back to 1936, survey data primarily from Gallup (N ≈ 980,000), to examine the long-run relationship between unions and inequality. We document dramatic changes in the demographics of union members: when density was at its mid-century peak, union households were much less educated and more non-white than other households, whereas pre-World-War-II and today they are more similar to non-union households on these dimensions. However, despite large changes in composition and density since 1936, the household union premium holds relatively steady between ten and twenty log points. We then use our data to examine the effect of unions on income inequality. Using distributional decompositions, time- series regressions, state-year regressions, as well as a new instrumental-variable strategy based on the 1935 legalization of unions and the World-War-II era War Labor Board, we find consistent evidence that unions reduce inequality, explaining a significant share of the dramatic fall in inequality between the mid-1930s and late 1940s. ∗We thank our research assistants Obaid Haque, Chitra Marti, Brendan Moore, Tamsin Kan- tor, Amy Wickett, and Jon Zytnick and especially Fabiola Alba, Divyansh Devnani, Elisa Jacome, Elena Marchetti-Bowick, Amitis Oskoui, Paola Gabriela Villa Paro, Ahna Pearson, Shreya Tan- don, and Maryam Rostoum. We have benefited from comments by seminar participants at Berke- ley, Columbia, Georgetown, Harvard, INSEAD, SOLE, the NBER Development of the American Economy, Income Distribution and Macroeconomics, and Labor Studies meetings, McGill Uni- versity, Princeton, Rutgers, Sciences Po, UMass Amherst, UC Davis, Universitat Pompeu Fabra, Stanford, and Vanderbilt. We are indebted to Devin Caughey and Eric Schickler for answering questions on the early Gallup data. We thank John Bakija, Gillian Brunet, Bill Collins, An- gus Deaton, Arindrajit Dube, Barry Eidlin, Nicole Fortin, John Grigsby, Ethan Kaplan, Thomas Lemieux, Gregory Niemesh, John Schmitt, Stefanie Stantcheva, Bill Spriggs, and Gabriel Zucman for data and comments. All remaining errors are our own. Farber: Princeton University and NBER, [email protected]. Herbst: University of Arizona, [email protected]. Kuziemko: Prince- ton University and NBER, [email protected]. Naidu: Columbia University and NBER, [email protected]. 1 Introduction Understanding the determinants of the U-shaped pattern of U.S. income inequality over the twentieth century has become a central goal among economists over the past few decades. While most economists agree that both redistributive institutions such as unions and taxation as well as market forces such as technology and trade have roles to play in explaining this pattern, there remains widespread disagreement over the relative importance of the two. While there is a substantial literature in labor economics and sociology that argues for a causal relationship from labor unions to lowered labor market inequality (Card, 2001; DiNardo, Fortin, and Lemieux, 1996; Western and Rosenfeld, 2011), another view holds that more fundamental drivers, namely technological developments that increase the demand for educated labor faster than increases in educational attainment, better explain the time-series variation in inequality (Acemoglu and Autor, 2011; Goldin and Katz, 2008; Goldin, Katz, and Autor, 2020). In the aggregate, there is a well-documented inverse relationship between income in- equality and union membership in the US (see Figure 1). But moving beyond this aggregate relationship has proven difficult. While aggregate measures of union density date back to the early twentieth century, it is not until the Current Population Survey (CPS) introduces a question about union membership in 1973 that labor economists have had a consistent source of microdata that includes union status. Put differently, it is not until unions are in steady decline that they can be studied with U.S. microdata. By contrast, the U.S. Census has tracked Americans' education and wages consistently since 1940, allowing historical analysis of models emphasizing supply and demand of skill as determining levels of inequality (see, e.g., the seminal work by Goldin and Katz, 2008). In this paper we bring a new source of household-level data to the study of unions and inequality. While the Census Bureau did not ask about union membership until the 1973 CPS, public opinion polls regularly asked about household union membership, together with extensive questions on demographics, socio-economic status and political views. We harmonize these surveys, primarily Gallup public opinion polls, going back to 1936. Our new dataset draws from over 500 surveys over the period from 1936-1986 and has over 980,000 observations, each providing union status at the household level. We combine these data with more familiar microdata sources (e.g., the CPS) to extend the analysis into the present day. These new data sources allow us to revisit the role of unions in shaping the income dis- tribution and to contribute to the long-running \institutions versus market forces" debate on the causes of inequality, particularly their role during the mid-century \Great Compres- 1 sion".1 The competitive model focusing on the supply and demand for skilled workers offers hypotheses on the joint movement of relative wages and relative quantities and can be used to assess the economic forces at work. Given the increase in relative college wages since the 1960s, authors in this tradition (with a long pedigree stretching back to Douglas (1930), Tinbergen (1970), and Freeman (1976)) have focused on changes in demand resulting from technology (Katz and Murphy, 1992; Autor, 2014; Card and Lemieux, 2001; Katz and Au- tor, 1999; Autor, Katz, and Kearney, 2008)) interacting with the rate of schooling increases. Adaptations of the relative skill model to account for recent patterns in wage inequality include Beaudry, Green, and Sand (2016), Acemoglu and Autor (2011), Autor, Levy, and Murnane (2003), and Deming (2017). On the institutions side, the literature includes Bound and Johnson (1992), DiNardo, Fortin, and Lemieux (1996) and Lee (1999), with recent liter- ature incorporating firms as important determinants of inequality (Song et al., 2015; Autor et al., 2020; Card, Heining, and Kline, 2013). A third strand of literature has attempted to \horse race" these two forces in ecological regressions across countries (Blau and Kahn, 1996; Jaumotte and Osorio Buitron, 2020), albeit with limited identifying variation. Bringing new micro data to the study of unions allows us to present several new results suggesting unions played a significant role in reducing income inequality at mid-century, when unions were at their peak and inequality at its lowest. These results fall into two broad sets. Our first set of results replicates many of the stylized facts about unions established with CPS data and extends them back to earlier decades. We begin by showing that patterns of selection into unions has varied substantially over time: the education of union members relative to non-union members has followed a marked U-shaped pattern, mirroring the pat- tern of inequality itself and the sharp inverse of union density. That is, at mid-century, when density was the highest, unions were drawing in the least educated workers. Today, as in the 1930s, unions are smaller and union and non-union households look similar in terms of education. A similar pattern emerges for minorities: unions were relatively less white at mid-century than either before or after, even conditioning on education. A key stylized fact about CPS-era unions is that members enjoy a wage premium, but did this advantage exist as union density was growing in the 1930s and 1940s and at their peak in the 1950s and 1960s? We show that the income advantage accruing to union households relative to non-union households with the same demographics and skill proxies holds quite steady (between ten and twenty log points) over our eighty-year period, despite the huge 1Collins and Niemesh (2019) is another recent paper emphasizing the role of unions in the Great Compression. They use the industry measures of union density constructed by Troy (1965) and form proxies of union density using 1940 IPUMS industry allocations within state economic areas. We build on this by providing direct measures of household union membership at the annual level over this period. 2 swings in union density and composition. As unobserved selection is typically a challenge in interpreting cross-sectional union premium regressions, we use a panel survey from 1956{ 1960, and we can show that the cross-sectional and respondent fixed-effects estimates are very close in magnitude. The household union premium is larger for the less-educated households, decreasing by four log points for every additional year of education, and effects for non-white households are also remarkably stable over our entire sample period. We show that unions not only reduce the differentials paid to observable traits such as education and race, but shrink the differentials