With Alexandra Killewald, Demography, June 2017, [paper] [open access] [replication materials] [accessible summary]
Does getting or staying married cause men to earn more per hour? Economists have long argued that marriage should increase men's productivity through specialization, as they develop human capital in the workplace while their wives focus on home production. Sociologists have likewise argued that marriage might motivate men to work harder and thus earn more. Against each of these perspectives, Alexandra Killewald and I argue that marriage has no causal effect on men's hourly wages because a time-varying variable confounds the association: the transition to adulthood. Using panel data, we reproduce recent literature showing that hourly wages grow prior to marriage and decline prior to divorce - a nuance missed by fixed-effects models that focus on intercept shifts alone. We then adjudicate between theories of the sources of these trajectories by exploring their empirical predictions for different subgroups. If premarital wage growth is attributable to causal anticipation of marriage, then it should be smaller among men for whom marriage may be less planned: those whose marriage is followed quickly by a birth. However, we find similar premarital wage growth in both groups. Further, men who marry at older ages experience no unusual wage growth prior to marriage. We argue that individuals mature at heterogeneous times, and these heterogeneous latent maturation processes cause both marriage and wage growth.
How prevalent is housing eviction among urban American children? (with Louis Donnelly, forthcoming in Demography) [paper][accepted manuscript (no paywall)][replication code] Ethnographic research suggests that housing eviction has become a common experience among urban renters, particularly those with children. With Louis Donnelly, I use 5 waves of panel data from the Fragile Families and Child Wellbeing Study to estimate that 14.8 % of U.S. children born in 1998-2000 in large cities experienced an eviction by age 15. Estimates are higher for the most disadvantaged subgroups; more than 1 in 4 of those born into deep poverty (below 50 % of the poverty line) were evicted.
Methodologically, this paper uses several approaches to draw conclusions despite missing data, under several assumptions. Following a set identification approach, we provide a lower bound estimate that shows that eviction is surprisingly common even under assumptions that would tend to dramatically underestimate this quantity. We additionally provide a preferred estimate under more typical assumptions about missing data. Finally, we demonstrate the robustness of our central claims to a nonparametric random forest estimator that relaxes the parametric estimation assumptions of our preferred estimator. Together, these estimates suggest that eviction is alarmingly prevalent.
Winner of 2018 Charles F. Westoff Prize in Demography, Princeton University. Working paper presented at the 2018 Annual Meeting of the Population Association of America. [draft]
Under what assumptions are sibling and cousin correlations informative about long-run attainment processes? Recent literature suggests that grandparents play a previously unrecognized role in the status attainment process (Mare 2011), casting doubt on the simplicity of classic models of parent-to-child transmission such as those of Blau and Duncan (1967) and Becker and Tomes (1986). Those claiming the largest role for extended families build this evidence by examining cousins' outcomes, arguing that similarities must be attributable to the extended kin that cousins share. I show that this evidence is equally consistent with several alternative theories in which extended kin play no direct role. One mundane alternative --- measurement error in the outcome variable --- is sufficient to explain away the observed outcomes. Additionally, I develop a Bayesian estimation approach which reveals that statistical uncertainty about sibling and cousin correlations is substantial. This provides further reason that authors should be cautious when interpreting sibling and cousin correlations. Conclusions clarify that sibling and cousin correlations are informative only in light of an assumed theoretical model, and they point toward a return to the classic framework of socioeconomic transmission from parent to child.
I am a Ph.D. candidate in sociology and social policy at Princeton University. I study methodological issues in inequality.
I have shown that housing eviction is more common than previously thought, that sibling and cousin correlations are often misinterpreted with respect to multigenerational social mobility, and that some adolescent and family outcomes are difficult to predict even with the best machine learning techniques. I have also formalized the gap-closing estimand as a research goal that combines pre-existing categorical inequality with a causal treatment. The methodological theme across these papers is that precise statement of the estimand—the quantity at the core of the theoretical claim—is an essential starting point for subsequent methodological choices. The estimand links the project to a broader literature, enables estimation by machine learning, and gives meaning to the parameter estimate that results. My research agenda moving forward is to continue to find estimands we have missed in the study of inequality and developing statistical estimators for those quantities.
One concrete example of current work involves the role of treatment heterogeneity in inequality. Social scientists frequently collapse a treatment variable measured with many values into a small number of categories. Yet, if these collapsed categories lead to distinct potential outcomes and are unequally distributed across subgroups, this hides an important source of inequality: those who are already disadvantaged take the less-beneficial versions of treatment. When we collapse treatments to a few categories, we completely miss this tendency. Treatment heterogeneity produces particularly misleading conclusions about effect heterogeneity and mediation. By instead defining the research goal as involving a multi-valued treatment, I directly develop estimators with good properties in this setting.
In my spare time, I enjoy backpacking in national parks, finding all the trail runs in the Princeton area, and getting to the beach as often as I can. You can reach me at ilundberg [at] princeton.edu.
I completed general exams in three areas of specialization. Princeton's sociology department requires students to demonstrate depth of knowledge in three subfields of our choice before advancing to candidacy. These fields represent areas of focus in which I hope to publish in the future.
Frontiers of causal inference: Nonparametrics, dimensionality, and time
Driven by rising computational power and waves of new sources of "big data," statisticians and computer scientists have rapidly developed new methods for nonparametric statistics and causal inference. Nevertheless, social scientists have only just begun to import these methods to address substantive questions. Under the guidance of Brandon Stewart, I developed a reading list covering new developments in causal inference. I chose to focus on two subfields: nonparametric approaches designed for high-dimensional settings where the functional form of associations may be unknown, and methods for dynamic causal inference in which units are exposed to dynamic treatments that change over time. These two areas of research epitomize a core principle of this emerging field: the best science is achieved when we combine the expertise of social scientists at defining an estimand and arguing for identification assumptions with the abilities of machine learning models to relax estimation assumptions about which humans are less able to reason. In addition to being examined on the reading list, I developed slides for one subset of the readings that bring social science and machine learning together in a beautiful way: tree-based methods for causal inference. (reading list) (tutorial slides)
Inequality and American Families
I designed a syllabus under the guidance of Sara McLanahan for a hypothetical graduate-level course that I could teach sometime in the future, focusing on three areas of family sociology: partnership, parenthood, and labor supply. This syllabus summarizes my substantive area of specialization in family and inequality. (syllabus here)
I completed a general exam in Demography administered by German Rodriguez and Thomas J. Espenshade. The exam covered the basics of demographic methods (following the Preston, Hueveline, and Guillot and the contents of this course) as well as substantive findings from the field (following this reading list).
My teaching experience is in statistics courses aimed at Ph.D. students in sociology.
I precepted Soc 500 (Applied Social Statistics) in Fall 2016 and Soc 504 (Advanced Social Statistics) in Spring 2017, both under Brandon Stewart at Princeton University. In Spring 2018, I am again precepting Soc 504 under Matthew J. Salganik. In each course, I led a 2-hour section reviewing materials from the course every other week. (link to sample handout on generalized linear models) (links to sample slides on random variables, likelihood inference, binary outcome models, duration models, and missing data)
My committee includes Brandon Stewart (chair), Matthew Salganik, Dalton Conley, and Sara McLanahan. Any of them can serve as references.
My CV is available here.