A modified version of the inference function for margins and interval estimation for the bivariate Clayton copula SUR Tobit model: An simulation approach

TL;DR

This paper introduces a modified inference method for bivariate Clayton copula models within SUR Tobit frameworks, improving dependence estimation in censored data through simulation and bootstrap techniques.

Contribution

It proposes the MIFM method, combining data augmentation and a modified inference function for margins, to better estimate copula parameters in censored SUR Tobit models.

Findings

Simulation results show accurate copula parameter estimation.

Bootstrap confidence intervals have good coverage probabilities.

Empirical application demonstrates model effectiveness on real data.

Abstract

This paper extends the analysis of bivariate seemingly unrelated regression (SUR) Tobit model by modeling its nonlinear dependence structure through the Clayton copula. The ability in capturing/modeling the lower tail dependence of the SUR Tobit model where some data are censored (generally, at zero point) is an additionally useful feature of the Clayton copula. We propose a modified version of the inference function for margins (IFM) method (Joe and Xu, 1996), which we refer to as MIFM method, to obtain the estimates of the marginal parameters and a better (satisfactory) estimate of the copula association parameter. More specifically, we employ the data augmentation technique in the second stage of the IFM method to generate the censored observations (i.e. to obtain continuous marginal distributions, which ensures the uniqueness of the copula) and then estimate the dependence parameter. Resampling procedures (bootstrap methods) are also proposed for obtaining confidence intervals for the model parameters. A simulation study is performed in order to verify the behavior of the MIFM estimates (we focus on the copula parameter estimation) and the coverage probability of different confidence intervals in datasets with different percentages of censoring and degrees of dependence. The satisfactory results from the simulation (under certain conditions) and empirical study indicate the good performance of our proposed model and methods where they are applied to model the U.S. ready-to-eat breakfast cereals and fluid milk consumption data.