Shoki Okubo

Shoki Okubo

I am an Assistant Professor of Sociology at Toyo University, Tokyo. I work on causal inference under model uncertainty — which covariates a study should adjust for, how robustness should be reported honestly, and what data-driven variable selection does to statistical inference — and on what measurement itself does to the survey data we analyse.

Three identities from current work
Where a multiverse's disagreement comes from
\[ \operatorname{Var}_{G,K}(b_K\mid D)=\underbrace{\textstyle\sum_{g} w_g\,v_g}_{\text{within graphs}}\;+\;\underbrace{\textstyle\sum_{g} w_g\,(\mu_g-\bar\mu)^2}_{\text{between graphs}} \]
Restrict a multiverse to the specifications each candidate causal graph licenses. Draw a graph \(g\) with weight \(w_g\) and then one of its specifications \(K\), and let \(b_K\) be that specification's estimate on the observed data \(D\); \(\mu_g\) and \(v_g\) are the mean and variance of the estimates within graph \(g\). The between-graph share \(\rho=\text{between}/\text{total}\) says how much of the spread is disagreement between the candidate graphs rather than among the specifications each licenses. It is a conditional summary of one data set that depends on the candidate set, its weights and a common estimand, and it carries no repeated-sampling uncertainty of its own. Preprint: arXiv:2609.16618.
When an outcome predictor lowers precision for the effect on the treated
\[ V_{\mathrm{att}}(S\cup W)-V_{\mathrm{att}}(S)=\frac{1}{p^{2}}\,E\!\left[e_S\left\{\frac{1-2e_S}{1-e_S}\,V_0(S)-2\,C_{10}(S)\right\}\right] \]
The display is the change in the efficiency bound of the effect on the treated when a covariate \(W\) with \(W \perp\!\!\!\perp A \mid S\) is added to the adjustment set \(S\). For the average treatment effect such a covariate never raises the bound. For the effect on the treated, if the propensity score \(e_S\) is at most one half and the two arms' predictions are not positively correlated, \(C_{10}(S)\le 0\), the bound cannot fall, and it rises strictly when \(W\) predicts the untreated outcome where \(e_S<1/2\). Whether \(W\) helps therefore depends on the distribution, and no rule based on the graph alone selects the optimal adjustment set for this estimand. Preprint: arXiv:2609.11222.
Panel conditioning from a cohort's own first wave
\[ \begin{aligned} h^{S} &= E[Y_t\mid S_{e:t}{=}1,\ \text{cont}]\;-\;E[Y^{*}_t\mid \text{fresh}] \\ &\quad -\;\underbrace{\big\{E[Y_e\mid S_{e:t}{=}1,\ \text{cont}]-E[Y_e\mid \text{cont}]\big\}}_{\text{attrition selection, measured before any conditioning}} \end{aligned} \]
A continuing cohort's entry-wave answers are unconditioned for survivors and attriters alike, so they measure how selected the survivors were at entry; a refreshment sample supplies the unconditioned mean at the comparison wave. If the survivors' selection differential is the same at the comparison wave as at entry, the combination identifies the survivors' mean conditioning effect \(h^{S}\). It is the observable functional of Das, Toepoel and van Soest (2011), here read as a survivor effect for general outcome means. The design needs no follow-up of the fresh cohort. In the paper's model of attrition it remains unbiased when selection on persistent traits changes over time, but it carries the same bias as survival matching under stationary state-dependent attrition. Preprint: arXiv:2610.01654.

Preprints; full statements, assumptions, and proofs are in the papers.

My methodological work sits between social-science methodology and statistics. Current projects develop causally disciplined multiverse analysis — running robustness analyses only within specifications that a stated causal model licenses, and decomposing the licensed multiverse’s dispersion into within-graph and between-graph components — together with valid post-selection inference for outcome-adaptive covariate selection and estimand-specific theory of optimal adjustment sets. A second strand asks what repeated interviewing changes in survey answers, what can and cannot be identified about it from a panel's own design, and what an analyst should believe when it cannot be identified.

Substantively I use Japan's long-running panel surveys and administrative data to study inequality over the life course: care work and long-term care policy, including households that combine childcare with elder care; why Japan's gender gaps in pay, careers, and workplace authority have proved so immobile; what staying put does in a society where most adults never move, using residential histories linked to small-area geographic data; how employment careers and family formation are jointly produced; a programme of natural-experiment studies of Japanese social policy built on administrative data that covers every municipality; and the long-run persistence of historical institutions, studied with historical sources I digitize and release myself.

I am a member of the Japanese Life Course Panel Surveys (JLPS) team at the Institute of Social Science, The University of Tokyo, where I was on the faculty from 2018 to 2025. I hold a PhD in Human Sciences from Osaka University (2017).

ORCID 0009-0007-6925-5174 · Research interests: causal inference · social-science methodology · survey methodology and measurement error · panel data · social stratification and the life course · gender inequality · residential mobility and neighbourhood effects · social policy and natural experiments · long-run persistence and historical data construction.

Selected work

All publications →

News