TL;DR
This paper revisits a comparison between two methods for estimating structural dynamic models: MPEC and NFXP. The authors demonstrate that a previous finding of MPEC being significantly faster than NFXP was due to an inefficient NFXP implementation. When a more efficient NFXP variant is used, both methods show comparable speed and performance.
What Problem Does This Paper Address?
Estimating structural dynamic models, which are crucial for understanding economic behavior and policy impacts, often involves solving complex mathematical problems. Two prominent approaches are Mathematical Programming with Equilibrium Constraints (MPEC) and Nested Fixed Point (NFXP). A previous study by Su and Judd (SJ) concluded that MPEC was substantially faster than NFXP for estimating dynamic discrete choice models. This paper addresses whether this speed advantage holds when a more computationally efficient version of NFXP is employed, as the original NFXP implementation used by SJ relied on a method known to be inefficient for certain problems.
Methods
The core of this paper is a re-evaluation of the computational performance of MPEC and NFXP. The authors replicate the Monte Carlo study conducted by Su and Judd (2012) but replace the NFXP implementation used by SJ (referred to as NFXP-SA, which uses successive approximations) with a more advanced version called NFXP-NK. NFXP-NK, originally proposed by Rust (1987), combines successive approximations with Newton-Kantorovich iterations to solve the inner fixed-point problem more efficiently, especially when discount factors are close to one. The authors compare the CPU times and numerical performance of MPEC (using the AMPL implementation) against this improved NFXP-NK.
Datasets
The study uses a simulated dataset based on the bus engine replacement model from Rust (1987). This model is a standard benchmark in the field of structural estimation of dynamic discrete choice models. The authors replicated the sample size and parameter values used in the original Su and Judd study.
Key Insights and Ideas
- The devil is in the details of the algorithm: The significant speed difference reported by Su and Judd was not an inherent property of NFXP versus MPEC, but rather a consequence of using an inefficient successive approximation method for the fixed-point problem within NFXP.
- NFXP-NK is a game-changer: The Newton-Kantorovich variant of NFXP dramatically improves computational efficiency, making it competitive with MPEC. This highlights the importance of choosing appropriate numerical methods for solving fixed-point problems.
- MPEC and NFXP-NK are comparable: When the efficient NFXP-NK is used, MPEC and NFXP exhibit similar speed and numerical performance in the tested model. This suggests that the choice between MPEC and NFXP-NK might depend more on user preference, ease of implementation, or specific model characteristics rather than a clear computational advantage of one over the other.
- Robustness matters: The NFXP-NK method is found to be more reliable than MPEC-AMPL in their experiments, suggesting that the improved NFXP variant offers greater stability.
Key Results
The paper's main quantitative finding is that when using the NFXP-NK algorithm, the CPU times for estimating the bus engine replacement model are comparable to those of MPEC-AMPL. Both methods take approximately 0.07 to 0.08 seconds per estimation. This contrasts sharply with Su and Judd's finding that MPEC was orders of magnitude faster than their NFXP-SA. For instance, at a discount factor of $\beta = 0.995$, NFXP-SA required an average of nearly 750,000 successive approximation iterations, while NFXP-NK required only about 145 iterations (including both successive approximation and NK iterations).
Limitations and Future Work
The authors acknowledge that their study focuses on a specific model (the bus engine replacement model) and a particular implementation of MPEC (MPEC-AMPL). While they suggest MPEC might still be a preferred method for its ease of use, especially with intuitive software like AMPL, they don't extensively explore its performance on a wider range of models or compare it to other MPEC implementations. Future work could involve testing these algorithms on more complex and diverse structural models to further validate their findings and explore the trade-offs between MPEC and NFXP-NK in different contexts.
Summary updated