Predictive algorithms are everywhere these days, as are concerns that they embed & perpetuate discrimination

In a new (short!) working paper with David Arnold & Will Dobbie, we develop + apply new quasi-experimental tools to address these concerns!



(🧵) dropbox.com/scl/fi/g0s83e3… Consider the pretrial setting, where judges increasingly use algorithmic risk scores meant to predict a defendant's potential for pretrial misconduct

The algorithmic inputs (e.g. past criminal convictions) may embed systemic biases (in e.g. past policing / judge decisions...)

Regression is a tool for making comparisons

If you don't know / can't easily explain what comparisons you're trying to make, then you don't understand the regression you're running This goes for IV too btw

Happy New Year! Kirill @Borusyak and I have a New (short) Paper on the infamous "negative weights" issue recently raised for TWFE and other popular OLS/IV specifications



Here's an (even shorter) summary thread dropbox.com/scl/fi/gfvv9bu…
We show that design-based specifications, which leverage assumptions on the assignment process of exogenous shocks, also have negative "ex-post" weights (i.e. ones that depend on the realized shocks)

However....

Hi! I'm back long enough to tell you about some awesome @Brown_Economics JMCs I'm lucky enough to write letters for this year

It's a great cohort overall, and you should check 'em all out here: . But here are the six I know the best (in alphabetical order)economics.brown.edu/job-market-can… First up is Tommaso Coen (), an econometrician studying robust welfare analysis in the presence of behavioral biases

His JMP shows how gains from "de-biasing" interventions can be informatively bounded w/ tools from the treatment effects literature. Neat! tommasocoen.com

Apropos of nothing, here’s a brief thread on a key point about “negative weights” in regression analysis Consider some outcome Y_i and some randomly assigned non-negative treatment X_i >= 0

We posit a causal model of Y_i = Y0_i + beta_i*X_i, where beta_i captures heterogeneous treatment effects across different units i

We regress Y_i on X_i. What do we get?

I'm teaching a new grad applied metrics course this spring; inspired by @paulgp, I've decided to post slides here

First, Ch. 1-3: a review of regression basics and discussions of design- & model-based ID

dropbox.com/s/8gx1oj69vz9n…

dropbox.com/s/z55anktz4anh…

dropbox.com/s/8jnlcjshek8v… Next up, in Chapters 4-5, an overview of recent findings on "negative weights" in regression and a brief interlude on clustered standard errors

dropbox.com/s/2po53rfdc7lu…

dropbox.com/s/t2fez008rk84…

Very excited about this new working paper with @aislinnbohren & @alexoimas, "Systemic Discrimination: Theory and Measurement"

dropbox.com/s/sp72pogz0lem…

We develop theoretical & empirical tools to model & measure the systemic drivers of discrimination in many settings

Summary 🧵: Econ has long studied direct discrimination - causal effects of race/gender/etc holding all else fixed - both in theory (eg taste/statistical disc) and empirics (eg audit studies)

Other fields take a systemic view: discrimination can arise *indirectly*, thru accumulated actions

Ok this is a fun one which may be useful for ~87% of you, if you've never heard of "the stacking trick"

It's relatively straightforward to get SEs on linear/nonlinear functions of the coefficients in a *single* OLS/2SLS regression. e.g. in @Stata you can use lincom or nlcom...

https://twitter.com/instrumenthull/status/1492589189442981888

But what if we want SEs on functions of coefs across *multiple* regressions, run on the same or different data?

E.g. what if we want to see if the coefficients in one OLS and one IV specification are significantly different?

Lots of potential answers here, but my favorite is...

Ok, so I come bearing good news for ~93% of you: esp. those bootstraping complex models (e.g. w/many FEs)

Instead of resampling, which can be seen as reweighting by a random integer W that may be zero, you can reweight by a random non-zero non-integer W

https://twitter.com/instrumenthull/status/1487204842988781570

This is the "Bayesian bootstrap" of Rubin (1981) jstor.org/stable/2240875…

Rubin derived it as "the Bayesian analogue of the bootstrap," but as with many Bayesian things it has attractive properties in a frequentist world too

Shao and Tu (1995) have a good treatment of both sides

Just posted a new working paper, with @metrics52, @prof_parag, and @c_r_walt

It’s called “Simple and Credible Value-Added Estimation Using Centralized School Assignment”

bit.ly/2WLeFmF

Here's a short summary thread 👇 Our starting point is two recent trends in US public schooling

1) The rise of value-added/report card quality metrics for individual schools

2) The spread of centralized mechanisms for assigning students to schools

We ask: how can data generated by (2) be used to improve (1)?

I just posted a new working paper, with David Arnold (@PrincetonEcon) and Will Dobbie (@Kennedy_School)

It's called “Measuring Racial Discrimination in Bail Decisions”

bit.ly/2QWz4Cu

Here's a short summary thread 👇 Racial disparities are pervasive in the criminal justice system. But do they reflect racial discrimination, or unobserved differences in criminal behavior?

We develop new quasi-experimental methods to answer this question in the context of bail decisions