We develop a general framework for the identification of counterfactual parameters in a class of nonlinear semiparametric panel models with fixed effects and time effects. Our method applies to models for discrete outcomes (e.g., two-way fixed effects binary choice) or continuous outcomes (e.g., censored regression), with discrete or continuous regressors. Our results do not require parametric assumptions on the error terms or time-homogeneity on the outcome equation. Our main results focus on static models, with a set of results applying to models without any exogeneity conditions. We show that the survival distribution of counterfactual outcomes is identified (point or partial) in this class of models. This parameter is a building block for most partial and marginal effects of interest in applied practice that are based on the average structural function as defined by Blundell and Powell (2003, 2004). To the best of our knowledge, ours are the first results on average partial and marginal effects for binary choice and ordered choice models with two-way fixed effects and non-logistic errors.

This paper considers the transitory-permanent model for the earnings process, and allows for time-varying individual-specific unobserved heterogeneity in each shock. The cross-sectional heterogeneity in each shock is drawn from an unknown distribution at each time period. Sufficient conditions for the nonparametric identification of the cross-sectional density functions of the heterogeneity are provided, under different assumptions on the time series behavior of the transitory shock. The method proposed is then applied to earnings data to document a high degree of cross-sectional heterogeneity in each shock.

We provide new results showing identification of a large class of fixed-T panel models, where the response variable is an unknown, weakly monotone, time-varying transformation of a latent linear index of fixed effects, regressors, and an error term drawn from an unknown stationary distribution. Our results identify the transformation, the coefficient on regressors, and features of the distribution of the fixed effects. We then develop a full-commitment intertemporal collective household model, where the implied quantity demand equations are time-varying functions of a linear index. The fixed effects in this index equal logged resource shares, defined as the fractions of household expenditure enjoyed by each household member. Using Bangladeshi data, we show that women’s resource shares decline with household budgets and that half of the variation in women’s resource shares is due to unobserved household-level heterogeneity.

This paper develops nonparametric identification and estimation results for a single-spell hazard model, where the unobserved heterogeneity is specified as a Lévy subordinator. The identification approach solves a nonlinear Volterra integral equation of the first kind with an unknown kernel function defined on a non-compact support. Both the kernel of the integral operator, which models the distribution of the unobserved heterogeneity, and the functions that enter it nonlinearly are identified given regularity conditions and minimal variation in the observed covariates. The paper proposes a shape-constrained nonparametric two-step sieve minimum distance estimator. The second step estimates the kernel of the integral operator, exploiting a monotonicity property. Rates of convergence are derived and Monte Carlo experiments show the finite sample performance of the estimator.

Abadie et al. (2010) derive bounds on the bias of the Synthetic Control estimator under a perfect balance assumption on both observed covariates and pretreatment outcomes. In the absence of a perfect balance on covariates, we show that it is still possible to derive such bounds, but at the expense of relying on stronger assumptions on the effects of observed and unobserved covariates and of generating looser bounds. We also show that a perfect balance on pre-treatment outcomes does not generally imply an approximate balance for all covariates, even when they are all relevant. Our results have important implications for the implementation of the method.

This paper considers a dynamic panel model where a latent state variable follows a unit root process with nonparametric heteroskedasticity. We develop constructive nonparametric identification and estimation of the skedastic function. Applying this method to the Panel Survey of Income Dynamics (PSID) in the framework of earnings dynamics, we found that workers with lower pre-recession permanent earnings had higher earnings risk during the three most recent recessions.

This paper considers the difference‐in‐differences (DID) method when the data come from repeated cross‐sections and the treatment status is observed either before or after the implementation of a program. We propose a new method that point‐identifies the average treatment effect on the treated (ATT) via a DID method when there is at least one proxy variable for the latent treatment. Key assumptions are the stationarity of the propensity score conditional on the proxy and an exclusion restriction that the proxy must satisfy with respect to the change in average outcomes over time conditional on the true treatment status. We propose a generalized method of moments estimator for the ATT and we show that the associated overidentification test can be used to test our key assumptions. The method is used to evaluate JUNTOS, a Peruvian conditional cash transfer program. We find that the program significantly increased the demand for health inputs among children and women of reproductive age.

We consider estimation and inference of the effects of a policy in the absence of a control group. We obtain unbiased estimators of individual (heterogeneous) treatment effects and a consistent and asymptotically normal estimator of the average treatment effects, based on forecasting counterfactuals using a short time series of pre-treatment data. We show that the focus should be on forecast unbiasedness rather than accuracy. Correct specification of the forecasting model is not necessary to obtain unbiased estimates of individual treatment effects. Instead, simple basis function (e.g., polynomial time trends) regressions deliver unbiasedness under a broad class of data-generating processes for the individual counterfactuals. Basing the forecasts on a model can introduce misspecification bias and does not necessarily improve performance even under correct specification. Consistency and asymptotic normality of our Forecasted Average Treatment effects (FAT) estimator are attained under an additional assumption that rules out common and unforecastable shocks occurring between the treatment date and the date at which the effect is calculated.

We consider a class of fixed-T nonlinear panel models with time-varying link function, individual-specific unobserved heterogeneity, and endogenous regressors. We establish sufficient conditions for the identification of certain time-varying average partial effects. The latter follows from the identification of the regression coefficients, the link function, and the distribution of counterfactual outcomes. We propose estimators for all these objects and study their asymptotic properties. We demonstrate the relevance of our results by estimating the effects of teaching practices on student attainment as measured by test scores on standardized tests in mathematics and science. We use data from the Trends in International Mathematics and Science Study, and show that both traditional and modern teaching practices have positive effects of similar magnitudes on the performance of U.S. students on math and science tests.

In nonlinear panel models with fixed effects and fixed-T, the incidental parameter problem poses identification difficulties for structural parameters and partial effects. Existing solutions are model-specific, likelihood-based, impose time homogeneity, or restrict the distribution of unobserved heterogeneity. We provide new identification results for the structural function and for partial effects in a large class of Fixed Effects Linear Transformation (FELT) models with unknown, time-varying, weakly monotone transformation functions. Our results accommodate continuous and discrete outcomes and covariates, require only two time periods, and impose no parametric distributional assumptions. First, we provide a systematic solution to the incidental parameter problem in FELT. Second, we identify the distribution of counterfactual outcomes and a menu of time-varying partial effects without any assumptions on the distribution of unobserved heterogeneity. Third, we obtain new results for nonlinear difference-in-differences that accomodate both discrete and censored outcomes, and for FELT with random coefficients. Finally, we propose rank- and likelihood-based estimators that achieve square root-n rate of convergence.