
Initial sequence estimators are commonly used to estimate Monte Carlo standard errors for reversible Markov chains used in Markov chain Monte Carlo simulations. Despite their widespread use and subsequent methodological extensions, they are known only to be asymptotically conservative, with consistency still an open question. We address this gap by investigating the statistical properties of the initial positive sequence estimator. Leveraging a nonnegativity property of sums of adjacent autocovariances from reversible Markov chains, the initial positive sequence estimator is obtained by summing sample autocovariances until a sum of adjacent sample autocovariances becomes negative. We characterize the convergence behavior of this random truncation point. We also propose an alternative initial sequence type estimator based on a modified truncation rule. For both estimators, we establish consistency and derive bounds on rates of convergence, showing that, under geometric ergodicity, both converge to the asymptotic variance no slower than the near-parametric rate of (log n)/n1/2. Finally, through empirical studies using both simulated and real-world data, we validate our theoretical findings and compare the empirical performance of the two initial sequence-type estimators with the standard overlapping batch mean estimator.
We consider GLS and OLS estimation in a linear regression model with serially correlated errors, and we provide the following contributions. First, we clarify when OLS is consistent or not. Second, we give sufficient conditions such that GLS is valid without the assumption of exogenous regressors (uncorrelated with past innovations). Third, we devise a feasible GLS procedure valid whether or not the regressors are exogenous, and which achieves a MSE close to that of the correctly specified infeasible GLS. Fourth, we derive the asymptotic distribution of this feasible GLS estimate: with exogenous regressors, the limit distribution is the same as that of the infeasible GLS estimator. However, with non-exogenous regressors, an additional term is present in the asymptotic variance. We provide a variance estimator, free of tuning parameters, which delivers confidence intervals with asymptotically correct coverage in both cases. We also illustrate how GLS can be more robust than OLS when the regressors are exogenous, even when GLS is based on an incorrect correction. The main assumptions are: a) the regressors are pre-determined (uncorrelated with future innovations); b) the errors are stationary thereby admitting a Wold representation in terms of some unpredictable innovations; c) the moving-average representation of the errors is invertible and implies a short-memory process. We also briefly address issues related to heteroskedastic errors.
Treatment effect heterogeneity with respect to covariates is common in instrumental variable (IV) analyses. An intuitive approach, which we call the interacted two-stage least squares (2sls), is to regress the outcome on the treatment, covariates, and treatment-covariate interactions, using the IV, covariates, and IV-covariate interactions as instruments. We clarify the causal interpretation of the interacted 2sls under the local average treatment effect (LATE) framework when the IV is valid conditional on the covariates. Our main findings are twofold. First, we show that the coefficients on the treatment-covariate interactions from the interacted 2sls identify treatment effect heterogeneity with respect to covariates among compliers for any outcome-generating process if and only if the product of the IV propensity score and the covariates is linear in the covariates, a requirement we call the linear IV-covariate interactions condition. Second, when the covariate vector has dimension K and includes a constant term, we show that the linear IV-covariate interactions condition requires the IV propensity score to take at most K distinct values. The condition holds when (a) the covariates take exactly K distinct values, or (b) the IV is randomly assigned independently of the covariates, but is restrictive beyond these cases. These results underscore the difficulty of interpreting regression coefficients from specifications with treatment-covariate interactions when the covariates are not saturated and the IV is not unconditionally randomized, absent correct specification of the outcome model. We then connect these results to the literature on LATE estimation, and show that under the linear IV-covariate interactions condition, the interacted 2sls with covariates demeaned by their complier means recovers the LATE as a single coefficient.
This paper develops the long-span asymptotic theory for the maximum likelihood (ML) estimator of parameters in the discretely sampled fractional Ornstein-Uhlenbeck (fOU) process. It is shown that the convergence rate is a continuous function of the Hurst parameter (i.e., H), and the asymptotic distribution is always Gaussian, facilitating statistical inference. The asymptotic theory is different from that of existing estimators in the literature. Simulation results indicate that the ML method offers more accurate parameter estimates compared to the existing estimators and the derived asymptotic distribution can closely approximate the finite-sample distribution. In the empirical analysis, we fit the fOU process to daily logarithmic RV and daily logarithmic trading volume series for ten exchange-traded funds. The ML estimate of H always lies substantially below 1/2, supporting the roughness hypothesis documented in the literature. Moreover, the fOU process with the ML estimate and the conditional mean predictor outperformed two sets of competitors — the fOU process with existing estimates and predictive formula, and established models, including HAR, log-HAR, ARFIMA, Cauchy processes and Brownian semistationary processes with the power law kernel — in out-of-sample forecasting.
Particle Markov Chain Monte Carlo (PMCMC) is a powerful approach to Bayesian computation for general state space models. Our article enables PMCMC to handle a larger number of observations and parameters compared to previous approaches by generating the parameters that are highly correlated with the states in a pseudo-marginal step(s) with the states 'integrated out'; the rest of the parameters are generated conditional on the states. We make the pseudo-marginal step much more efficient than previous approaches by positively correlating the numerator and denominator in the Metropolis-Hastings acceptance probability in a novel way by expressing the target density of the PMCMC in terms of the basic uniform or normal random numbers used in the sequential Monte Carlo algorithm, instead of the standard way which expresses the target density in terms of the state particles. We also show that the new sampler combines and generalizes two separate particle MCMC approaches: particle Gibbs and the correlated pseudo-marginal Metropolis-Hastings. We investigate the performance of this hybrid sampler empirically by applying it to univariate and multivariate stochastic volatility models having both a large number of parameters and a large number of latent states and show that it is much more efficient than competing PMCMC methods. An online supplement to the article and computer code are available online.