We host an annual retreat in support of our research activities. We organize the event around the research codes that are under active development in our group to facilitate a productive exchange of ideas. While participants were initially limited to economists, we hope to expand the event and include the broader mathematics and computational science community going forward.
Date September 7th - 9th, 2020
Location IAME, University of Bonn
We will host regular conference talks on the first two days and then organize a social event for the third day to facilitate networking.
Kenneth L. Judd, Hoover Institution
Date September 13th, 2019
Location IAME, University of Bonn
Effects of taxes and safety net pensions on life-cycle labor supply, savings and human capital: The case of Australia
Fedor Iskhakov, Australian National University
The paper structurally estimates a life-cycle model of consumption/savings, labor supply and retirement, using data from the Australian HILDA panel. The model is used to evaluate effects of Australia's Age Pension system and income tax policy on labor supply, consumption and retirement. The model accounts for human capital, savings, uninsurable wage risk and credit constraints. The authors account for "bunching" of hours by assuming a discrete set of hours levels, and they investigate labor supply on both the intensive and extensive margins. The model allows to quantify the effects of anticipated and unanticipated tax and pension policy changes at different points of the life-cycle. The results imply that Australia's Age Pension system as currently design is poorly targeted. The simulations suggest that a doubling of taper rates, combined with 5.9% reduction of income tax rates, would be budget neutral and Pareto improving. (The current version of the paper, on which this abstract is based, can be accessed at the homepage of Fedor Iskhakov)
Boryana Ilieva, DIW Berlin
Do women have biased expectations about returns to experience and about the part-time penalty? Using SOEP data the paper provides empirical evidence about returns to experience and the part-time penalty. In a dynamic setting, the authors investigate how biased expectations (about the part-time penalty) may affect employment choices and how policies can increase female employment and working hours (under those biased beliefs). To elicit expectations about returns to experience in full-time/part-time employment in the short/long run, newly designed survey questions are incorporated in the SOEP-IS module. Empirical findings based on SOEP data reveal that expected returns to experience in part-time and full-time work do not differ. There is no expected part-time penalty in experience. Nevertheless, reduced form evidence points towards evidence of a part-time penalty in experience. The estimation of a structural model reveals the existence of biased expectations. A policy simulation shows that de-biasing reduces the part-time choice rates in the structural estimation.
Rafael Suchy, University of Bonn
The presentation compares different integration methods to solve the so-called EMax() encountered in Keane and Wolpin (1994). Since then, many discrete choice dynamic programming models required to solve multi-dimensional integral with discontinuous integrands. To evaluate whether different methods, e.g. quasi-Monte Carlo, are suitable to reduce the required points for a given accuracy, and hence increase computational efficiency, a simulation study implemented through respy is conducted. The simulation studies shows that switching to quasi-Monte Carlo methods (particularly Halton or Sobol) increases efficiency and has effects on the policy effect (although small).
Gregor Boehl, Goethe University Frankfurt
Occasionally binding constraints (OBCs) play a central role in macroeconomic modelling since major developed economies have hit the zero lower bound (ZLB) on interest rates. The paper proposes a solution method for rational expectations models with OBCs and a Bayesian filter/smoother that, both combined, can be used for quick and accurate Bayesian estimation of large-scale models. The quasi-analytic solution method calculates the endogenous duration at the constraint while avoiding matrix inversions and simulations at runtime for gains in computational speed. The TEnKF (transposed Ensemble Kalman Filter) is a hybrid of the particle filter and the Kalman filter that, requiring only a very small number of particles, can be used to approximate the likelihood of nonlinear models with high accuracy. The IPA (iterative path-adjusting transposed-ensemble RTS) smoother adds a smoother and an iterative procedure to filter and can be applied to estimate the state distributions while fully respecting the nonlinearity of the transition function. Further, a massively parallelized combination of different forms of heuristic global optimizers is proposed to avoid local maxima of the likelihood function. Techniques from astrophysics for efficient sampling from the posterior distribution are suggested. As an example, these methods are used to estimate the simple New Keynesian model. (The current version of the paper can be accessed on the homepage of Gregor Boehl (2019))
Jasmin Maag, University of Zurich
A primer that presents a method to efficiently estimate hidden Markov models with continuous latent variables using maximum likelihood estimation. The name Recursive Maximum Likelihood Integration (RLI) is derived from its design: to evaluate the (marginal) likelihood function the integral is decomposed over the unobserved state variables into a series of lower dimensional integrals. Those integrals are recursively approximated by numerical quadrature and interpolation. Advantages include: 1) computational complexity grows linearly in the number of periods, 2) numerical error accumulates sublinearly in the number of time periods integrated. The method is applied to the Rust bus engine replacement model of Rust (1987). The presentation is based on a publication by Gregor Reich (2018).
Gregor Reich, University of Zurich
Gregor Reich (2018) estimates the models of Rust (1987) with serially correlated EV1 errors. So far, states were either observed or unobserved. But what changes if the states are occasionally observed? It turns out that the method of Recursive Maximum Likelihood Integration (RLI) is still applicable but with respect to the model, there is need to take care for the one-sided laws of motion. For example, assuming that buses are new when entering the data set but only the mileage state at engine replacement is observable, then only the cost trade-off can be consistently estimated, but not the law of motion of mileage. The problem: a comparison of two estimates does not reveal much in this setting. Consequently, the distribution of the estimator (parametric bootstrapping) at various observation frequencies is compared. The main finding shows that a similar distribution does imply a similar point estimate.
Maximilian Blesch, University of Bonn
The talk revisits Rust (1987) and evaluates the sensibility of the decision rule performance to the estimation quality of the transition probabilities. The authors explicitly address, whether the agent accounts for uncertainty in his estimates and if accounting for uncertainty does lead to a better decision strategy? In a first step it is shown, that in the presence of ambiguity in the estimates decision making under risk only decreases the performance of the decision rule. Using methods from distributionally robust optimization the authors account for ambiguity and construct decision rules that explicitly take potential model misspecification into account.
Philipp Mueller, University of Zurich
The paper proposes a novel method to estimate structural parameter efficiently and systematically, even in models with one poorly identified parameter and in the presence of multiple equilibria. Inference is based on the full (profile) likelihood function, instead of point estimates. The approach includes (1) to formulate the structural estimation as a parameterized version of Su and Judd (2012), and (2) to solve the problem efficiently by homotopy parameter continuation. The presentation of the concept focuses on dynamic discrete choice models where the discount parameter is generally poorly identified. Using the original data set and model from Rust (1987) they can reject the instance that the discounting parameter is unidentified, and estimate it to be statistically significant larger than one.
Robert Erbe, University of Zurich
To answer the question whether stock returns are predictable one could simply run a regression on returns, using a forecasting variable that is very persistent and slow-moving. In the literature the forecasting variable is often modeled as AR(1) process and confined to be stationary. Using this approach, finite sample bias can inflate the coefficient on the stock returns. The authors use a vector autoregression model by Stambaugh (1986)(1999), Mankiw and Shapiro (1986) in order to contrast estimation by likelihood to a regression approach. Using the “full” likelihood assures stationarity of the VAR, whereas least squares does not (although assumed). The paper presents MLE and inference based on constrained optimization. Power gains from stationarity are realized. On the other hand finite-sample bias leads to power loss. The authors conclude that likelihood estimation provides a natural way to preserve stationarity for both, estimation and inference. They demonstrate that the computation of Likelihood is computationally not prohibitive. Finally, they show that their test has too much bite and propose a bias correction that might resolve this problem.
We concluded our first retreat with a bike trip from Bonn to Cologne along the river Rhine.
Participants in front of Cologne Cathedral
Philipp Eisenhauer and Fedor Iskhakov