TL;DR
Standard rank-rank regressions (RRR) are a favorite tool for economists to measure social mobility, but they become mathematically "messy" and hard to interpret as soon as you add control variables. This paper introduces Conditional Rank-Rank Regression (CRRR), a new framework that uses conditional ranks to provide a clean, interpretable measure of average within-group association that stays strictly within the bounds of a correlation coefficient.
What Problem Does This Paper Address?
In fields like labor economics, researchers often regress a child’s income rank on their parent’s income rank to measure intergenerational mobility. However, we often want to "control" for covariates—like age, education, or region—to see how much of that mobility is happening within specific groups versus between them.
The problem is that the two standard ways of doing this are flawed:
- RRRX (Adding covariates as regressors): This destroys the "rank" nature of the variables. The resulting slope is no longer a Spearman correlation and can even produce nonsensical results, like coefficients greater than 1 or less than -1.
- RRR.res (Using residuals): This approach only works under very restrictive assumptions (like location-shift models) that rarely hold in real-world data.
Essentially, when you add controls to a standard rank regression, you lose the "ceteris paribus" (all else being equal) interpretation that economists crave.
Methods
The authors propose Conditional Rank-Rank Regression (CRRR). Instead of using a person's rank in the entire population, they use their rank conditional on their characteristics ($X$).
The Secret Sauce: Distribution Regression
To estimate these conditional ranks, the authors use Distribution Regression (DR). Unlike standard regressions that just look at the mean, DR estimates the entire conditional distribution function. This allows for a much more flexible model where covariates can affect the shape, spread, and tails of the income distribution differently.
Theoretical Rigor
Because the ranks are estimated (making them "generated regressors"), standard linear regression inference doesn't work. The authors develop a theoretical framework using the functional delta method and Hadamard differentiability to prove that their estimator is consistent and follows a normal distribution in large samples. They use an exchangeable bootstrap method to provide reliable standard errors and confidence intervals.
Datasets
The paper evaluates the method using:
- Simulated Data: A bivariate normal design to prove the estimator converges at the expected rate ($\sqrt{n}$).
- WiSiER Dataset: A rich administrative dataset from Switzerland (1986–2016) covering 11 cantons. It links tax records, social insurance, and surveys for fathers and their children, specifically comparing them both at age 35.
Key Insights and Ideas
- Order of Operations Matters: Standard methods rank first and then "partial out" the covariates. CRRR reverses this: it partials out the covariates through the ranking process. This ensures the variables remain true ranks (uniformly distributed between 0 and 1).
- The "Aha" Interpretation: The CRRR slope has a beautiful property: it is exactly equal to the average conditional Spearman correlation. This provides a natural summary measure of within-group association.
- Variance Decomposition: CRRR allows researchers to decompose overall persistence into "within-group" (unexplained by $X$) and "between-group" (explained by $X$) components. This is a powerful tool for understanding the drivers of inequality.
Key Results
Using the Swiss data, the authors found:
- The Gender Gap: Intergenerational income persistence is significantly stronger for sons than for daughters.
- Persistence Drivers: Within-group persistence (the CRRR result) explains 62% of the overall income persistence for sons and 52% for daughters.
- Subgroup Findings: Persistence is higher in families with highly educated fathers and in smaller families. Interestingly, they found no significant difference in mobility for children of immigrants in Switzerland, contrasting with some US-based studies.
- Robustness: The results held up regardless of whether they used logistic or Gaussian link functions, proving the method is stable.
Limitations and Future Work
- The Tail Problem: Estimating distributions at the very extreme ends (the 1% or the bottom 1%) is difficult because data is sparse. While the authors propose a tail-restriction model to handle this, it remains a sensitive area for any distribution-based method.
- High Dimensionality: The current framework is designed for a moderate number of covariates. The authors mention that an extension using Double/Debiased Machine Learning (DML) is currently underway to handle "Big Data" scenarios with hundreds or thousands of control variables.
- Discrete Variables: The current theory focuses on continuous variables. While the authors suggest it can be extended to discrete data, they leave the formal proof for future research.
Summary updated