Research
Research
Embedded Files
Job Market Paper
A Score-Driven Filter for Causal Regression Models with Time-Varying Parameters and Endogenous Regressors [pdf] [Working paper]
(with Francisco Blasques)
Abstract: This paper proposes a score-driven model for filtering time-varying causal parameters using instrumental variables. In the presence of suitable instruments, we show that we can uncover dynamic causal relations between variables, even in the presence of regressor endogeneity which may arise because of simultaneity, omitted variables, or measurement errors. Due to the observation-driven nature of score models, the filtering method is simple and practical to implement. We establish the asymptotic properties of the maximum likelihood estimator and show that the instrumental-variable score-driven filter converges to the unique unknown causal path of the true parameter. We further analyze the finite sample properties of the filtered causal parameter in a comprehensive Monte Carlo exercise. Finally, we reveal the empirical relevance of this method in an application to aggregate consumption in macroeconomic data and we provide a time-varying estimate of price elasticity of demand for a dataset on recorded market prices.
Objective: estimate path of causal time-varying parameter, where regressors (x) are endogenous
Solution: a causal score-driven filter using valid instruments (z)
Properties: MLE is consistent, filter converges to true unobserved causal path
Application: Time-varying excess sensitivity of consumption to personal income.
US data: real per capita disposable income (Y), real consumption expenditures on non-durable goods and services (C)
Results:
2SLS, OLS not signficant, but IV-score is signficant in some periods
Overall, endogeneity does not seem to be a concern (Hausman test on 2SLS, OLS p-val 0.55 and Score/IV-score largely overlap), but in financial crisis estimate IV-score twice as large as score. In line with break-down Ricardian equivalence in crisis periods: consumption depends strongly on current income
Figure: static OLS and 2SLS estimators vs. the time-varying estimates with and without instruments (IV-score, score) incl. 90% bootstrap bounds
Working Papers
Mitigating Estimation Risk: A Data-Driven Fusion of Experimental and Observational Data [Working paper]
(with F. Blasques, S.J. Koopman, P. Gorgi)
Abstract: The identification of causal effects of marketing campaigns (advertisements, discounts, promotions, loyalty programs) require the collection of experimental data. Such data sets frequently suffer from limited sample sizes due to constraints (time, budget) which can result in imprecise estimators and inconclusive outcomes. At the same time, companies passively accumulate observational data which oftentimes cannot be used to measure causal effects of marketing campaigns due to endogeneity issues. In this paper we show how estimation uncertainty of causal effects can be reduced by combining the two data sources by employing a self-regulatory weighting scheme that adapts to the underlying bias and variance. We also introduce an instrument-free exogeneity test designed to assess whether the observational data is significantly endogenous and experimentation is necessary. To demonstrate the effectiveness of our approach, we implement the combined estimator for a real-life data set in which returning customers were awarded with a discount. We demonstrate how the indecisive result of the experimental data alone can be improved by our weighted estimator and arrive to the conclusion that the loyalty discount has a notably negative effect on net sales.
Research in Progress
Extending Generalised Synthetic Control: Total Effect Estimation in the Presence of Mediators
(with Francisco Blasques)
Abstract: Generalised synthetic control (Xu, 2017) offers an alternative to traditional synthetic control methods, by directly modelling the time varying heterogeneity using an interactive fixed effects panel data model, which allows for the incorporation of time varying covariates. The counterfactual outcome is then constructed by simply taking the fitted values, using the treated unit's own covariates and the estimated parameters, factors and loadings. However, when a treated unit's covariates are also affected by treatment (mediators), then this counterfactual does not form a reliable basis for what would have truly happened in absence of the intervention. With GSC one would only estimate the direct effect, and not the total effect that includes the change in regressors. In this paper we propose an extension of the GSC estimator, that allows for the estimation of the total effect of an intervention.
References:
Xu, Y. (2017). Generalized synthetic control method: Causal inference with interactive fixed effects models. Political Analysis, 25 (1), 57–76.
Conferences and Seminars
2024 VU seminar, Econometrics department, Amsterdam
2024 Netherlands Econometrics Study Group (NESG), Maastricht
2023 International Association for Applied Econometrics (IAAE) in Oslo
2023 International Conference on Econometrics and Statistics (EcoSta) in Tokyo, Japan
2023 Netherlands Econometrics Study Group (NESG), Rotterdam
2021 VU seminar, Econometrics department, Amsterdam
Page updated
Google Sites
Report abuse