ST313 course page
Ethics for data science: the statistical choices hidden inside technical steps, and who they affect.
Each week has one lecture and one class (80 minutes each, lecture first). The notes are the main reference; the slides are what is shown in the lecture; the class sheet is what you work through in class, and its .qmd link downloads the source to run the R code. Bring a laptop with R and RStudio to the class.
Week 1: Selection and the replication crisis
Why a literature of honest tests can still be wrong on average: power, the significance filter, and what a published estimate is conditioned on. The standing question, with two cases.
Reading: ms4ds, chapter 3: Ethical data science, sections 3.1 to 3.6 · Wikipedia: Replication crisis, skim
Class: class sheet (.qmd) · class sheet with solutions
Week 2: Mindless statistics and sophisticated bigotry
Cargo cult science, cargo-cult statistics and the null ritual, and Feynman’s first principle: you must not fool yourself. Then numbers that are computed correctly and read as answers to other questions: a correlation across fifty states with the opposite sign among the people who live in them, what \(R^2\) measures and what averaging does to it, decisions about a person made from the average of their group, what a small difference in spread does far out in a tail, and what the largest of \(n\) items shows you of a group.
Required reading: Stark and Saltelli, Cargo-cult statistics and scientific crisis (in full) · Hyde and Mertz, Gender, culture, and mathematics performance (first paragraph of “Do Gender Differences Exist Among the Mathematically Talented?”, Figure 1, Table 2; free full text) · Freedman, Ecological Inference and the Ecological Fallacy (sections 1 to 3) · one of two psychology studies, as announced (Connellan and colleagues 2000, or Alexander and colleagues 2009), skimmed for its data and analysis
Optional: Rathje, Van Bavel and van der Linden, Out-group animosity drives engagement on social media · Gigerenzer and Marewski, Surrogate Science
Class: class sheet (.qmd) · class sheet with solutions
Week 3: Ethics and causation
Four frameworks of normative ethics for judging whether an action is right, utilitarianism and its difficulties, the applied principles written after the Tuskegee study, and the five Cs, a practical checklist for data work. A consequentialist judges an action by what it brings about, which is a causal question: how much does a job-search course help people find work, when the people who attend it are more likely to find work anyway? Causal graphs, conditioning and intervening, the average treatment effect and what an average hides, confounding and how to adjust for it, and the choice of graph that any adjustment rests on.
Required reading: Barocas, Hardt and Narayanan, Fairness and Machine Learning, chapter 5: Causality (“The limitations of observation” through “Confounding”, pages 105 to 116) · ms4ds, chapter 3: Ethical data science (sections 3.2 and 3.3, a re-read) · skim the Wikipedia articles Consequentialism, Deontology, Virtue ethics, Pragmatic ethics and Belmont Report
Optional: Cunningham, Causal Inference: The Mixtape, first edition (section 1.2 and sections 3.1 to 3.1.4) · Tukey, Conclusions vs Decisions · Bross, The Role of the Statistician: Scientist or Shoe Clerk
Class: class sheet (.qmd) · class sheet with solutions
Week 4: What’s the score?
A score measures one thing and is taken to stand for another. A hospital’s risk score predicted next year’s medical cost and was used to decide who was offered extra care; at the same cost, Black patients were sicker, so the score flagged fewer of them. A model of a measure with an offset for one group, who it selects and how ill they are, and a noisy measure that flattens a slope. Reinforcement learning: an agent given a reward, and the good we care about. What happens once a score is rewarded: Goodhart’s law as an intervention, strategic classification, performative prediction, and a platform that ranked by views.
Required reading: Obermeyer, Powers, Vogeli and Mullainathan, Dissecting racial bias in an algorithm used to manage the health of populations (Science, 2019; the abstract and the sections “Health disparities conditional on risk score,” “Mechanism of bias” and “Problem formulation,” pages 448 to 451) · Barocas, Hardt and Narayanan, Fairness and Machine Learning, chapter 1: Introduction (“The trouble with measurement” through “Feedback and feedback loops”, pages 7 to 14) · skim the Wikipedia article Goodhart’s law
Optional: Amodei and colleagues, Concrete Problems in AI Safety (sections 2 and 4) · Meyerson, YouTube now: why we focus on watch time (YouTube Official Blog, 2012) · Miller, Milli and Hardt, Strategic Classification is Causal Modeling in Disguise (section 1) · Perdomo, Zrnic, Mendler-Dünner and Hardt, Performative Prediction (section 1) · Jacobs and Wallach, Measurement and Fairness · Manheim and Garrabrant, Categorizing Variants of Goodhart’s Law
Bring to the class: one score or algorithm used to make decisions about people, with what it measures and what it stands in for.
Class: class sheet (.qmd)
Week 5: Who counts as a case
Every dataset has a rule for who gets into it: who was born small, who was hired, who answered the survey, who was released on bail. When that rule depends on two things at once, looking only at the people it let in creates an association between those two things that is not there in the population. Colliders on a causal graph; a model that leaves out a sensitive attribute and becomes both less accurate and more unequal between groups; data missing completely at random, at random, and not at random; what dropping the incomplete rows does to an estimate, and what filling them in assumes.
Required reading: van Buuren, Flexible Imputation of Missing Data (second edition, 2018; sections 1.1 to 1.3) · re-read Barocas, Hardt and Narayanan, Fairness and Machine Learning, chapter 5: Causality (the subsection “Colliders”, pages 112 and 113, and the list of what controlling for each kind of variable does, page 115)
Optional: Elwert and Winship, Endogenous Selection Bias: The Problem of Conditioning on a Collider Variable (Annual Review of Sociology, 2014; section 4, and the examples of truncation, nonresponse and attrition in section 5) · Lakkaraju, Kleinberg, Leskovec, Ludwig and Mullainathan, The Selective Labels Problem (KDD 2017; section 1) · van Buuren, section 2.7: When not to use multiple imputation · Hernández-Díaz, Schisterman and Hernán, The birth weight “paradox” uncovered? (American Journal of Epidemiology, 2006)
Bring to the class: one dataset or study where some people who should be in it are not, with who is missing and in which direction their absence pushes the result.