Overview

This package implements non-parametric, unbiased estimators of the distribution of teacher effects described in Rose, Schellenberg, and Shem-Tov (2022). These unbiased estimators are \(U\text{-}\mathrm{statistics}\), which provide minimum-variance unbiased estimators of population variances and covariances of latent parameters. The approach overcomes several issues experienced by Empirical Bayes (EB) techniques when estimating the distribution of teachers’ ‘value-added,’ but can applied in any setting where the researcher seeks to estimate the distribution of agent-specific effects.

Brief introduction

Rose, Schellenberg, and Shem-Tov (2022) employ this approach in the context of estimating teachers’ effects on students across multiple outcomes (e.g., test scores, suspensions, and future crime). Throughout the package and its documentation, we use the same ‘teacher’ vocabulary, but the estimators in the ustat_var package would also apply in other settings.

The package assumes the researcher observes for each teacher \(j = 1, 2, ..., J\) and outcome \(k = 1, 2, ..., K\)

\[y^k_j = (y^k_{j1}, ..., y^k_{jT_j})\]

where \(y^k_{jt} = a_j^k + e_{jt}^k\). The parameter \(a_j^k\) represents teacher j’s effect on outcome \(k\). Different outcomes could refer to separate measures (e.g., math test scores and reading test scores) or separate sub-populations (e.g., male and female students). The term \(e_{jt}^k\) represents estimation error. The key asumptions are that:

  • \(\operatorname{E}[e_{jt}^k | a_j^k] = 0\) for all \(j,k,t\)

  • \(\operatorname{E}[e_{jt}^ke_{jt'}^{l}] = 0\) for \(t \neq t\) and all \(j,k,l\).

The package produces estimates of \(\operatorname{Var}(a_j^k)\) and \(Cov(a_j^k,a_j^l)\), as well as estimates of their sampling variance. There are options to equally weight each of these variance/covariance parameters as well as to apply user-given weights. The package can also accomodate heavily unbalanced data, where \(T_j\) differs across teachers and/or across outcomes within teacher.

  • The user manual summarises the formulae for the estimators (used in the core ustat_var package functions) for transparency.

  • Rose, Schellenberg, and Shem-Tov (2022) contains the full discussion of the empirical setup and the required assumptions to interpret the estimates causally.

Authors

Installation

To install, execute

python3 -m pip install ustat_var

Contents

  • Usage offers a brief explanation of how to use the core functions.

  • References offers a reference guide for each of the package’s functions, including helper functions.

Suggested Citation

If you use this package, please cite the original paper:

@workingpaper{rss2022effects,
  title        = {The Effects of Teacher Quality on Adult Criminal Justice Contact},
  author       = {Rose, Evan K. and Schellenberg, Jonathan T. and Shem-Tov, Yotam},
  institution  = {National Bureau of Economic Research},
  number       = {30274},
  year         = {2022},
  month        = jul,
  doi          = {10.3386/w30274},
  url          = {https://www.nber.org/papers/w30274},
  type         = {NBER Working Paper}
}