Donald Rubin

Life and career

Donald B. Rubin was born in 1943 in Washington, D.C., and arrived at Princeton as an undergraduate intending to study physics. He left with a degree in psychology, then went to Harvard, where he took a master's in computer science and, in 1970, a PhD in statistics under William Cochran — the great figure of survey sampling and observational-study design. That lineage matters: Cochran spent his career worrying about how to draw valid conclusions from data that were not produced by a randomized experiment, and Rubin took that worry and gave it a formal language.

His first professional years were spent not in a university but at the Educational Testing Service in Princeton, with periods at the Environmental Protection Agency and elsewhere. This is another detail that repays attention. ETS dealt in massive survey and testing data with pervasive nonresponse, and the EPA in observational environmental data where randomization was impossible. Rubin's three great contributions — potential outcomes, EM, and multiple imputation — are all, at bottom, responses to the same practical circumstance: the data you have are not the data you wanted, and pretending otherwise is not an option.

He held a professorship at the University of Chicago and then moved to Harvard in 1983, where he spent three decades, chaired the statistics department for many years, and trained an extraordinary number of the people who now run causal inference in statistics, economics, epidemiology, and political science. He later took a position at Tsinghua University in Beijing.

Rubin is also, it should be said, one half of a long and pointed methodological disagreement. His potential-outcomes framework and Judea Pearl's structural causal models are two formalizations of causality that are provably capable of expressing much of the same content, and their proponents have spent decades arguing about which is the better vehicle for thinking. Rubin's position emphasizes explicit assumptions about the *assignment mechanism* — how units came to receive the treatment they received — and treats causal inference as a missing-data problem. Pearl's emphasizes graphs and the identification calculus they support. Students should learn both and be suspicious of anyone who tells them the argument is settled.

He is among the most cited statisticians alive, has been elected to the National Academy of Sciences, and has received the field's major honors.

Key contributions

**Potential outcomes and the Rubin Causal Model.** For each unit i and each treatment level, posit a potential outcome: Y_i(1) is what unit i's response *would be* under treatment, Y_i(0) under control. The individual causal effect is Y_i(1) − Y_i(0). The fundamental problem of causal inference is that you observe exactly one of them — the other is missing, not merely unmeasured. Causal inference therefore *is* a missing-data problem, which is the unifying idea of Rubin's entire career. Neyman had used potential outcomes for randomized experiments in 1923; Rubin's 1974 paper and its successors extended the framework to observational studies and made it the standard.

Three pieces of formal apparatus follow. **SUTVA** (the stable unit treatment value assumption): no interference between units, and no hidden versions of the treatment. It sounds like bookkeeping and is in fact where most real analyses go wrong — vaccines, social programs, and network interventions all violate the no-interference clause. **Ignorability** (unconfoundedness): treatment assignment is independent of the potential outcomes given observed covariates, which is what licenses adjustment. And the **assignment mechanism** itself as the central object: Rubin's insistent methodological advice is to design the observational study — decide who is comparable to whom, using only covariates and treatment status — *before* looking at outcomes, so that the analysis cannot be tuned to produce a preferred result.

**The propensity score.** With Paul Rosenbaum in 1983, Rubin introduced e(x) = P(T = 1 | X = x), and proved that it is a *balancing score*: conditional on e(X), treatment is independent of X. Under ignorability, therefore, adjusting for the scalar e(X) suffices — a startling dimension reduction, turning adjustment for a high-dimensional covariate vector into adjustment for a single number, implementable by matching, stratification, or weighting. The Rosenbaum–Rubin sensitivity analysis, which asks how strong an unmeasured confounder would have to be to overturn a conclusion, comes from the same collaboration.

**The EM algorithm.** With Arthur Dempster and Nan Laird, the 1977 paper that unified a scattered set of iterative procedures under one framework. Given incomplete data, alternate between the E-step — compute the expected complete-data log-likelihood given current parameters and the observed data — and the M-step — maximize it. The key theoretical guarantee is monotone ascent: the observed-data likelihood never decreases. EM is the standard fitting method for mixture models, hidden Markov models, factor analysis, and any latent-structure model where the complete-data likelihood is tractable and the marginal one is not.

**Multiple imputation and the missing-data taxonomy.** Rubin's 1976 *Biometrika* paper "Inference and Missing Data" defined MCAR, MAR, and nonignorable missingness, and stated precisely the conditions under which the response mechanism can be ignored in likelihood and Bayesian inference. Multiple imputation, developed through the late 1970s and 1980s and consolidated in his 1987 book, is the operational answer: impute the missing values m times from a predictive distribution, analyze each completed dataset with standard methods, and combine using **Rubin's rules** — the point estimate is the average, and the variance is within-imputation variance plus (1 + 1/m) times the between-imputation variance. That second term is the whole point: it is the uncertainty due to not having observed the data, made explicit rather than hidden. Single imputation understates uncertainty; multiple imputation admits it. Census bureaus and epidemiological cohorts worldwide run on this.

**Bayesian computation and model checking.** Rubin contributed the sampling/importance resampling (SIR) algorithm, the Bayesian bootstrap, and — with Andrew Gelman — the potential scale reduction factor R̂, the standard convergence diagnostic for MCMC based on comparing between-chain and within-chain variance. His work on posterior predictive checking gave Bayesians a principled way to ask whether a model could have generated data like the data observed. He is a co-author of *Bayesian Data Analysis* and, with Roderick Little, of *Statistical Analysis with Missing Data*.

**Instrumental variables and principal stratification.** With Angrist and Imbens, the potential-outcomes reconciliation of instrumental variables with causal inference, identifying the local average treatment effect for compliers under exclusion and monotonicity. Principal stratification generalizes this to post-treatment variables — the correct way to handle noncompliance, truncation by death, and surrogate outcomes.

In battle

Rubin's computed profile is a top-tier one: mean 49.1 across 101 problems, **twenty-two dominant scores at 80 or above**, thirty-two at 70 or above. He and Pearl are the roster's two causal superpowers, and they are dangerous in overlapping but distinguishable regions.

His single best score is **P060 — Missing, not at random** (99), which the judge calls his biography — the problem's year is the year of "Inference and Missing Data," and its closing requirement, that the estimate admit the uncertainty of what was never observed, is Rubin's rules almost verbatim. **P184 — Fill in the hidden variables** (97) is EM. **P009 — When you can't randomize** (97) is the observational-study problem the propensity score was built for. **P117 — Randomize the villages, not the people** (97) is cluster randomization, where SUTVA and design-based reasoning are exactly the framework needed. **P121 — When treatment leaks across friends** (95) is interference — the *violation* of his own assumption, which he named and which he is therefore the natural person to reason about. **P059 — The confounder you can't measure** (96) is Rosenbaum–Rubin sensitivity analysis. **P215 — The therapy the trial reversed** (93) is noncompliance and principal stratification. **P212 — Sample from the impossible posterior** (93) draws on his Bayesian computation work, including R̂.

His categories run: fairness 71.7, causality 69.1 across fourteen problems, experimental design 65.3 across fifteen, small-sample 55.8 across thirteen. That experimental-design number is the one that distinguishes him from Pearl, whose strength is identification given a graph; Rubin's is design, assignment mechanisms, and what to do when the data collection itself was imperfect.

The losses are clean. **P079 — The language for the job** (7) is 1959 programming-language design, three decades before his career and several layers of the stack away. **P030 — Will it ever halt?** (8) is the halting problem, and computability sits at 13. **P195 — Let the machine find the proof** (8) is automated theorem proving and **P054 — The grandmaster's intuition** (8) is chess expertise modeling — symbolic AI problems with no assignment mechanism and no missing data. **P242 — The memory that completes the pattern** (10) is associative memory in the Hopfield sense, a dynamical-systems model of recall rather than a statistical estimator. His systems average of 7.5 and search average of 11.5 are near the roster floor.

One result in the profile is worth flagging to students as a live question rather than a settled fact: **P250 — The cluster that iterates** scores only 8, despite EM being the canonical iterative clustering procedure. The likely reading is that the problem is posed as an algorithmic clustering question — Lloyd's algorithm, initialization, convergence in a computational sense — rather than as maximum likelihood in a latent-class model, and Rubin's framing is relentlessly inferential rather than algorithmic. It is a good problem to battle him on and see what the judge actually says.