Tractable Estimation of Nonlinear Panels with Interactive Fixed Effects

1. Problem Statement

The paper addresses the computational and theoretical challenges associated with estimating nonlinear panel models that include interactive fixed effects (unobserved heterogeneity following a linear factor structure).

• Context: While interactive fixed effects are standard in linear panel models (e.g., Bai, 2009; Pesaran, 2006), their application in nonlinear settings (e.g., Probit, Logit, Poisson) has been limited.
• The Bottleneck: The existing state-of-the-art estimator, proposed by Chen, Fernández-Val, and Weidner (2021), is the Fixed Effects (FE) Maximum Likelihood Estimator. However, this estimator requires solving a high-dimensional, non-convex optimization problem.
  • The objective function involves $d_X + R(N+T)$ parameters (where $N$ is units, $T$ is time, and $R$ is the number of factors).
  • The non-convexity arises from the rank constraint on the unobserved effects matrix ($\text{rank}(\Theta) \leq R$).
  • Standard optimization methods (like EM algorithms) lack global convergence guarantees in non-convex settings, and multi-start strategies become computationally intractable as $N$ and $T$ grow large.
• Goal: To develop a computationally efficient estimator that is asymptotically equivalent to the Chen, Fernández-Val, and Weidner (2021) FE estimator, thereby enabling feasible inference in large nonlinear panels.

2. Methodology: Two-Step Estimation Procedure

The authors propose a two-step estimation procedure that combines convex relaxation with local optimization.

Step 1: Nuclear Norm Regularized (NNR) Estimation (Convex Relaxation)

• Idea: Replace the non-convex rank constraint ($\text{rank}(\Theta) \leq R$) with a convex nuclear norm penalty ($\|\Theta\|_{nuc}$), which acts as the $\ell_1$-norm of the singular values, inducing low-rank solutions.
• Optimization: Solve the following convex problem:
  $$(\hat{\beta}_{nuc}, \hat{\Theta}_{nuc}) = \arg\min_{\beta, \Theta} \left\{ L_{NT}(\beta, \Theta) + \phi_{NT}\sqrt{NT}\|\Theta\|_{nuc} \right\}$$
  where $L_{NT}$ is the negative log-likelihood and $\phi_{NT}$ is a tuning parameter.
• Algorithm: The authors use a Proximal Gradient Descent method (Algorithm 1) to solve this efficiently.
• Initialization of Factors: The preliminary factor loadings ($\hat{\Lambda}_{nuc}$) and factors ($\hat{\Gamma}_{nuc}$) are extracted via Singular Value Decomposition (SVD) of $\hat{\Theta}_{nuc}$.

Step 2: Local Estimation (Global Solution Recovery)

• Idea: Use the NNR estimators from Step 1 as initial values for a standard gradient descent algorithm to solve the original non-convex problem (2.2).
• Mechanism: The authors prove that the original objective function is locally convex in a shrinking neighborhood around the true parameters. If the Step 1 estimator falls within this neighborhood (which they prove it does with high probability), the local optimization is guaranteed to converge to the global minimum (the true FE estimator).
• Algorithm: Standard Gradient Descent (Algorithm 2) initialized at $(\hat{\beta}_{nuc}, \hat{\Lambda}_{nuc}, \hat{\Gamma}_{nuc})$.

3. Key Theoretical Contributions

The paper makes significant theoretical advances to justify the two-step approach in nonlinear settings:

1. Asymptotic Equivalence: The authors prove that their two-step estimator is asymptotically equivalent to the global FE estimator defined by Chen, Fernández-Val, and Weidner (2021). This allows researchers to use the existing inferential theory (bias correction, standard errors, confidence intervals) from the FE framework with their new, faster estimator.
2. Local Convexity in a Shrinking Neighborhood:
   • Unlike linear models where the objective is often globally or broadly locally convex, the nonlinear case is only locally convex in a shrinking neighborhood of the true parameters as $N, T \to \infty$.
   • The paper establishes the specific rate at which this neighborhood shrinks ($\delta_{NT} \sim \log(NT) \min(N^{-3/8}, T^{-3/8})$).
   • Crucially, they show that the NNR estimator converges fast enough to enter this shrinking neighborhood with probability approaching 1.
3. Improved Convergence Rates for NNR:
   • The authors derive a new error bound for the NNR estimator in nonlinear models.
   • They improve upon previous results (e.g., Moon and Weidner, 2018) by showing that the convergence rate is sufficient to satisfy the strict requirements of the shrinking neighborhood in nonlinear settings.
   • They provide primitive conditions to verify the Restricted Strong Convexity (RSC) condition, extending existing literature to allow for predetermined covariates (e.g., lagged dependent variables), which is a common feature in dynamic panels.
4. Handling Normalization: The paper addresses the identification issue of factor models (rotational invariance) by proposing a specific normalization constraint that simplifies the Hessian analysis.

4. Empirical and Simulation Results

Monte Carlo Simulations

• Setup: The authors simulate Binary Response (Probit/Logit) and Poisson models with varying sample sizes ($N, T$) up to $(1000, 200)$.
• Findings:
  • Bias: The pooled estimator (ignoring factors) is heavily biased. The NNR estimator reduces bias but retains some regularization bias. The two-step estimator significantly reduces bias, converging rapidly to zero.
  • Bias Correction: When combined with analytical or Jackknife bias correction methods (from Chen et al., 2021), the two-step estimator achieves negligible bias even in small samples.
  • Factor Selection: The proposed data-driven method for selecting the number of factors ($\hat{R}$) performs with high accuracy.
  • Computation: The method is computationally feasible for large panels where the direct FE optimization fails or is too slow.

Empirical Application: Gravity Model of Trade

• Data: Bilateral trade flows between 157 countries (1986).
• Model: Poisson regression with additive fixed effects and interactive fixed effects.
• Results: The estimates from the proposed two-step estimator are nearly identical to those obtained by Chen, Fernández-Val, and Weidner (2021) using an EM algorithm.
  • Coefficients for distance, borders, legal systems, and currency unions are consistent with economic theory.
  • The log-likelihood values are virtually indistinguishable, confirming the asymptotic equivalence in finite samples.

5. Significance and Impact

• Bridging Theory and Practice: The paper removes the primary barrier (computational intractability) preventing the use of rigorous interactive fixed effects models in nonlinear empirical research.
• Scalability: It enables the analysis of "Big Data" panels (large $N$ and $T$) in fields like macroeconomics, finance, and international trade, where nonlinearities are prevalent.
• Software Implementation: The authors provide an optimized R package (NNRPanel) that implements the two-step algorithm, tuning parameter selection, and factor number determination, making the method immediately accessible to applied researchers.
• Theoretical Rigor: By establishing local convexity in shrinking neighborhoods and providing verifiable conditions for RSC with predetermined covariates, the paper advances the theoretical understanding of low-rank matrix estimation in high-dimensional econometrics.

Conclusion

Zeleneev and Zhang successfully demonstrate that a convex relaxation (NNR) followed by local optimization is a viable and theoretically sound strategy for estimating nonlinear panel models with interactive fixed effects. Their method offers the computational efficiency of convex optimization while retaining the statistical properties (consistency, asymptotic normality, and bias properties) of the computationally expensive global FE estimator.