Research Papers

We combined a special class of deep neural networks with directed acyclic graphs to improve our ability to draw inferences about entire systems of causal relationships.

Social science theories often postulate causal relationships among a set of variables or events, but when it comes to empirically evaluating them, researchers typically resort to methods that impose unrealistic assumptions on the data, leading to incorrect inferences. Our approach can model these relationships without invoking such restrictive assumptions.

Our approach allows researchers to more accurately learn causal relationships from data. It therefore has a very wide scope for application, ranging from studies of inequality and social mobility to health, education, and psychology.

We evaluated the hypothesis that neighborhood effects on academic achievement are mediated or interact with school quality using data from the Early Childhood Longitudinal Study.

We aimed to explore the mechanisms through which neighborhood poverty affects academic achievement, focusing on the role of school quality. By examining diverse measures of school effectiveness, resources, and climate, we sought to provide insights into whether and how these factors mediate or interact with neighborhood effects.

Our findings indicate that living in a disadvantaged neighborhood reduces academic achievement, but find little evidence that this effect is mediated by measures of school quality. These results challenge prevailing hypotheses and suggest that neighborhood effects operate independently of school quality factors. Addressing concentrated poverty directly may therefore be crucial in reducing educational inequalities.

We developed and tested a model explaining how growing up in a poor neighborhood impedes cognitive development, focusing on early life exposure to neurotoxic air pollution. We analyzed data from a national sample of American infants, examining their exposure to over 50 pollutants known or suspected to harm the central nervous system, using causal inference methods and supervised machine learning.

We aimed to uncover the mechanisms by which neighborhood poverty affects cognitive development. By identifying specific factors, such as neurotoxic air pollution, that mediate these effects, we sought to provide insights that could inform interventions and policies aimed at mitigating the adverse impacts of growing up in impoverished neighborhoods.

Our study shows that living in a high-poverty neighborhood increases exposure to harmful air pollutants during infancy, reducing cognitive abilities at age 4 by about one-tenth of a standard deviation. Understanding these mechanisms is crucial for developing targeted interventions to improve cognitive outcomes for children in poor neighborhoods and addressing environmental health disparities.

We followed 1,266 children in Chicago from birth through school entry, tracking their risk of lead exposure and neighborhood socioeconomic composition to estimate the joint effects on receptive vocabulary ability.

We aimed to address the gap in research on the impact of disadvantaged neighborhoods on cognitive development during early childhood, as most studies focus on school-age children or adolescents.

Our findings show that sustained exposure to disadvantaged neighborhoods reduces vocabulary skills in early childhood through lead contamination. Understanding this causal pathway is crucial for developing interventions to improve cognitive outcomes for children in impoverished neighborhoods and addressing environmental health disparities, promoting better long-term educational and developmental prospects.

We developed a new method for estimating randomized intervention analogues to the natural direct effect (rNDE) and the natural indirect effect (rNIE) in the presence of treatment-induced confounders.

We aimed to address the challenge of estimating mediation effects when treatment-induced confounders complicate the mediator-outcome relationship. Existing methods require cumbersome weighting procedures, making them difficult to use. Our method provides a more practical and accessible approach for researchers by utilizing simpler regression-based techniques.

Our approach allows researchers to estimate mediation effects more efficiently and accurately, even when treatment-induced confounders are present. By simplifying the estimation process, this method makes causal mediation analysis more accessible and robust.

We introduced “residual balancing” as an alternative method for constructing weights in marginal structural models (MSMs) to estimate time-varying treatment effects, aiming to address issues inherent in inverse probability weighting (IPW).

We sought to overcome limitations of IPW, such as inefficiency, sensitivity to model misspecification, and difficulty with continuous treatments.

Residual balancing improves the efficiency and robustness of estimating marginal treatment effects in the presence of time-varying confounders. By offering a more reliable and user-friendly alternative to IPW, it enhances causal inference accuracy in observational studies.

We evaluated hypotheses about education’s impact on beliefs about racial inequality using cross-sectional, sibling, and panel models. Analyzing data from the General Social Survey (GSS), the 1994 Study of American Families, and the 2006–2010 GSS panels, we assessed whether education promotes structural explanations for racial disparities or if observed effects are superficial and influenced by unobserved factors.

We aimed to clarify conflicting perspectives on education’s influence on racial attitudes: whether it genuinely fosters more enlightened views or if associations are superficial and due to unobserved biases. By employing different models and datasets, we sought to disentangle these complexities and provide a nuanced understanding of education’s role in shaping beliefs about racial inequality.

Our findings suggest that while cross-sectional analyses indicate education promotes structural explanations for racial inequality, sibling and fixed-effects models suggest these effects may be superficial and inflated by unobserved factors. This challenges simplistic interpretations and underscores the need for careful methodological considerations in studying the relationship between education and racial attitudes.

We employed a novel rank-based index of stratification to measure how occupational classes occupy distinct, non-overlapping, hierarchical layers in the US income distribution from 1980 to 2016.

We aimed to address the gap in empirical studies that focus on inequality rather than stratification. By developing a rank-based stratification index, we sought to accurately capture class stratification and provide insights into the structural dynamics of the US income distribution over time, enhancing the understanding of class structure and social stratification.

Our findings show a substantial increase in class stratification in the US from 1980 to 2016, driven by factors like income returns to education, deunionization, and deindustrialization. These results highlight growing economic divides based on occupational class, informing theories of class structure and guiding policies to reduce economic stratification and promote social mobility.

We explored an alternative regression-based method, regression-with-residuals (RWR), for estimating marginal effects of time-varying treatments on end-of-study continuous outcomes. This method was compared with inverse-probability-of-treatment-weighted (IPTW) estimation of a marginal structural model.

We aimed to address challenges in estimating marginal treatment effects with observational data, particularly when dealing with time-varying confounders affected by prior treatments.

Our findings suggest that RWR estimation of a constrained structural nested mean model may outperform IPTW estimation in certain scenarios for estimating marginal treatment effects of time-varying treatments on continuous outcomes.

We introduced a new method, regression with residuals, to estimate moderated causal effects in longitudinal studies where both treatment and effect moderators vary over time.

We aimed to address the limitations of traditional regression models in longitudinal settings, where time-varying moderators of future treatment can be influenced by prior treatment.

Our method provides a more accurate approach to understanding how individual characteristics modify the effects of treatments over time. This is crucial for social scientists seeking to identify the nuanced ways in which interventions impact different subgroups.