Using the UK MES for electronics, retail and restaurants, NPR delivers plausible estimates. We find NPR attributes much more of cross-firm TFP dispersion to *persistent* productivity (64% vs 43% in electronics, 77% vs 49% in retail, 52% vs 33% in restaurants). (8/9)

And the data already exist. The UK Management and Expectations Survey, the US MOPS and Survey of Business Uncertainty, the Bank of England Decision Maker Panel, the China Employer–Employee Survey… We catalogue where the method is feasible today, and where small tweaks would make it so. (7/9)

Monte Carlo simulations (following ACF's design) show NPR is robust to optimisation errors in inputs. We also offer "EZ-NPR", building on Evdokimov and Zeleneev (2025), which withstands moderate measurement error in the elicited expectations. (6/9)

Our idea: instead of inferring productivity from what firms *do*, use firms’ *stated beliefs*. If productivity is Markov, a firm's expectations about its own future output and inputs are informative about its current productivity. Expectations become the proxy! (4/9)

The workhorse "proxy variable" methods (Olley–Pakes, Levinsohn–Petrin, Ackerberg–Caves–Frazer) invert a firm's investment or materials demand to back out productivity. That works only if input choices are strictly monotonic in productivity, which can be demanding in some settings. (3/9)

Production functions are behind a lot of things we care about: technical change, productivity dispersion, markups, policy evaluation. But estimating them is hard. Productivity is usually unobserved and correlated with input choices, so OLS is inconsistent. (2/9)

Agnes Norris Keiller, @johnvanreenen.bsky.social and I have recently issued a revision for our on "Production Function Estimation Using Subjective Expectations Data”! Here comes a short thread on what we do in the paper. @ucleconomics.bsky.social @poid-lse.bsky.social (1/9)

We do much more in the article! For example, we look at further measures of association and discuss our estimation protocol in detail – importantly, both skills and intergenerational regressions are estimated jointly. (6/7)

Interestingly, we also find that mother-child correlations are much higher than father-child! (This relationship remains even after robustness checks for alternative explanations such as different sample sizes.) (5/7)

This allows us to measure skills at the same age for parents and children. While correlations are lower than those measured contemporaneously and between other economic variables in related studies, they are still sizeable and significant! (4/7)

Here, we focus on “internalising” (e.g., focussing drive aothers) skills using the 1970 UK British Cohort Study, which follows all born in the UK on a week in Apr/1970 linking their info w their mums as well as (eventual) kids! (3/7)

‪The literature has registered association across several meaningful economic variables and outcomes across generations. Socio-emotional skills are also important! (2/7)

This is a (very) short thread on a recent paper with the great Orazio Attanasio and Alessandro Toppeta on “Intergenerational Mobility in Socio-Emotional Skills” (forthcoming at ‪@jpube.bsky.social)! (1/7) @ucleconomics.bsky.social @uclpolicylab.bsky.social‬ @sofi.su.se ‪@clscohorts.bsky.social

To employ conformal inference, we construct a conformity score function which accounts for the set-valued nature of the outcomes of interest. The procedure accommodates (irreducible) prediction uncertainty, modelling uncertainty due to partial identification and sampling uncertainty. (4/5)

To do so we first characterise the shortest prediction interval (or set, more broadly!) in such cases. We then employ “conformal inference” to construct prediction sets with particular finite sample guarantees under censoring while maintaining consistency as the sample size grows. (3/5)

Interval data is pervasive. Surveys usually employ “bracketing” to avoid item nonresponse. Censoring is also present in many contexts. Here, we offer a prediction protocol for outcomes that are interval-valued! (2/5)

I have a new paper on “Prediction Sets and Conformal Inference with Censored Outcomes” with the great Weiguang Liu and Elie Tamer. Prediction Sets and Conformal… what? I know, I know! Here comes a (very) short thread on what we do in the paper. (1/5)