Online debiased lasso estimation and inference for heterogenous updating regressions

In the era of big data, online updating problems have attracted extensive attention. In practice, the covariates set of the models may change according to the conditions of data streams. In this paper, we propose a two-stage online debiased lasso estimation and inference method for high-dimensional heterogenous linear regression models with new variables added midway. At the first stage, the homogenization strategy is conducted to represent the heterogenous models by defining the pseudo covariates and responses. At the second stage, we conduct the online debiased lasso estimation procedure to obtain the final estimator. Theoretically, the asymptotic normality of the heterogenous online debiased lasso estimator (HODL) is established. The finite-sample performance of the proposed estimators is studied through simulation studies and a real data example.