Research

Digitalization, Factor-Biased Productivity, and Market Power

with Subal C. Kumbhakar

Download Paper (PDF) Research Statement

Abstract

Digitalization can improve labor efficiency, raise overall input efficiency, or increase firm markups. To disentangle these three channels, we develop a sequential production and markup estimator. Technology is first identified from quantity-side moments without imposing a specific model of competition. Markups are then recovered from revenue shares. We apply this framework to Spanish manufacturing firms from 2000 to 2019. Our central finding is that accounting for market power significantly changes the estimated composition of digital returns. On a common sample under perfect competition, a one-point increase in digitalization raises overall input efficiency by 6.5% and labor efficiency by 4.8%. Under imperfect competition, these responses shift to 4.3% and 5.5%, while markups rise by 0.7%. Paired bootstrap tests confirm a significant drop in the overall efficiency response and a 3.0 percentage-point shift toward labor efficiency. Two patterns further shape these gains. First, labor-efficiency gains are significantly larger for small firms, even though large firms adopt more digital technology. Second, among distinct digital technologies, physical automation is the sole component that significantly increases markups. Ignoring market power and technology differences thus distorts the sources of digital productivity growth.

Causal Forests versus Penalized Splines for Heterogeneous Treatment Effects

with Ivan Korolev and Nency Dhameja

Economics Letters, 268 (2026), 113205.

Journal Article Working Paper (PDF) Replication Code

Abstract

This paper compares causal forests with varying-coefficient penalized spline estimators for heterogeneous treatment effects. We conduct simulations with randomized treatment assignment that vary heterogeneity, predictor dimension, and discrete nuisance structure. Performance is measured by out-of-sample mean squared error. Forests perform best in Wager–Athey designs with sharp treatment-effect transitions. Penalized splines tend to dominate when heterogeneity is smooth, when the baseline mean and treatment effect differ in structure, and in mixed designs with continuous, binary, and group-level covariates. As group cardinality rises, forest performance deteriorates, while mgcv random-effect smooths handle group structure effectively and remain easy to implement.

The Scale-Technology Paradox: Technological Progress and Productivity Decline in Nigerian Banking

with Subal C. Kumbhakar

Economic Modelling (2026), 107859.

Journal Article Working Paper (PDF) Replication Code

Abstract

Nigeria’s banking sector has undergone substantial consolidation and technological modernization, yet bank productivity has continued to decline. This study examines this scale–technology paradox using annual data for 17 commercial banks over 2000–2023. We estimate a multi-output input distance function using a control-function approach to distinguish productivity change from technical change and examine the roles of scale economies and input costs. Bank productivity declined by 3.5% per year on average (3.1% at the median), with larger and internationally licensed banks experiencing steeper declines. Technical change was cost-reducing, while returns to scale averaged 1.46, indicating substantial unexploited scale economies. Lending was the main source of cost pressure. The findings show that technological progress and scale economies did not translate into productivity growth because scale economies remained underexploited and lending costs persisted. Strengthening credit information, collateral enforcement, bad-loan resolution, and lending efficiency is therefore central to translating technological progress and scale economies into sustained productivity growth.

Mediation Analysis in Production Functions with Unobserved Productivity

with Subal C. Kumbhakar

Request Paper

Summary

Digitalization can change output through additional labor hours or through greater productive efficiency. More productive firms may both adopt technology and adjust employment, making these channels difficult to distinguish. We develop a framework that controls for unobserved productivity in both the production and labor-hours equations. Simulations show that omitting this control from the labor-hours equation can substantially overstate the indirect effect. An application to Spanish manufacturing distinguishes producing more through additional labor from producing more efficiently.

Regularization in Translog Production Models

Solo-authored

Can better output prediction also improve the economic conclusions drawn from production models? Flexible production models can predict output accurately while producing unstable input elasticities because their squared and interaction terms are highly correlated. I am developing a comparison of Lasso, Ridge, Elastic Net, principal component regression, and partial least squares that evaluates the recovery of input elasticities, returns to scale, and inferred markups. Planned simulations will vary input correlations, sample size, and production relationships, and assess economically implausible estimates such as negative marginal products. Regularization alone does not resolve endogenous input choices.

Artificial Intelligence and Spatial Factor-Biased Productivity

Solo-authored

My next solo project asks who benefits from AI beyond the adopting firm. It extends the adoption framework to examine how AI changes the efficiency of particular inputs and how knowledge flows, worker mobility, and supply-chain relationships transmit gains to other firms. I plan to compare exposure through geographic proximity and economic linkages and assess whether the framework distinguishes spillovers from common shocks. The empirical design will require information on the AI applications firms use and the production activities they affect.