Career
Chen started his academic career as an assistant professor at the University of Waterloo in 1991. He was promoted to the rank of associate professor in 1996 and became a professor in 2001. He was then recruited as a professor at the University of British Columbia in 2007 and was appointed to a Tier I Canada Research Chair in the application of finite mixture models in statistical genetics, a position he held until 2020.
Chen has also held appointments in China. He was the director of the Big Data Institute at Yunnan University from 2016 to 2021 and held a part-time post at Yunnan University through the Thousand Talents Plan of China.
He was the president of the International Chinese Statistical Association (ICSA) from 2005 to 2006 and the president of the Survey Methods Section of the Statistical Society of Canada in 2007. His editorial service includes a term as editor-in-chief of the Canadian Journal of Statistics (2010–2012) and as a series editor of the ICSA Book Series in Statistics published by Springer (2013–2019).
Research
Chen's research spans finite mixture models, empirical likelihood, variable selection, statistical genetics, survey sampling, the design of experiments and asymptotic theory. His doctoral work, supervised by C. F. Jeff Wu, concerned minimum-aberration fractional factorial designs, a criterion for selecting efficient experimental designs.
One theme of Chen's research is the theory of finite mixture models, whose irregular parameter space causes standard likelihood asymptotics to break down. In a 1995 paper in the Annals of Statistics he established the best possible rate of convergence for estimating the mixing distribution, showing that although a √n-consistent rate is attainable when the number of components is known, the optimal rate falls to n−1/4 when it is unknown, a rate achieved by suitable minimum-distance estimators under a strong-identifiability condition. Later authors have treated the finding as a foundational reference point in the convergence theory of mixture models, describing it as a surprising departure from regular parametric behavior. Subsequent work by other researchers refined the precise rate. Heinrich and Kahn (2018), also in the Annals of Statistics, identified a gap in Chen's upper-bound argument. It did not account for distinct mixture components converging toward a common location, and showed that the optimal local minimax rate is instead n−1/(4(m−m0)+2) for a target with m0 components estimated within the class of m-component mixtures, which is slower than n−1/4 once the model permits more components than the truth and degrades further as that gap widens.
Chen's 1993 paper with Jing Qin showed how empirical likelihood can incorporate auxiliary information when estimating finite-population quantities. Later, he extended his technique to handle the case of constrained problems, and with A. M. Variyath and Bovas Abraham, came up with the idea of adjusted empirical likelihood. In this technique, an artificial data point is included into the data set such that the convex hull of the augmented data includes the origin, thus creating confidence regions that have better coverage probability. With Yukun Liu, Chen later showed that, by tuning the level of adjustment to the Bartlett correction factor, it attains high-order precision.
In the area of variable selection, Chen and Zehua Chen introduced the extended Bayesian information criterion (EBIC). EBIC adds a penalty on the size of the model space to the ordinary BIC so that sparse models are favoured in case the number of candidate covariates grows with the sample size. The criterion is consistent for linear regression with large model spaces, and later work extended the idea to outside the linear full-likelihood case such as composite likelihood model selection, where the linear-model EBIC is a special case.
Chen also worked on the treatment of item non-response in surveys through imputation. He and Jun Shao wrote a series of papers studying nearest-neighbour imputation, in which a missing value is replaced by the observed value of the most similar responding unit. Chen and Shao later showed that the resulting estimator of the population mean is asymptotically unbiased without requiring a linear relationship between the imputed variable and the auxiliary information, and that the empirical distribution function obtained from nearest-neighbour imputation is asymptotically unbiased, so that quantile estimators are also unbiased. They then derived design-consistent variance estimators for means and quantiles when the imputed values cannot be identified in the data set, and showed that the ordinary adjusted jackknife overestimates the variance under nearest-neighbour imputation, and proposed modified jackknife estimators that are asymptotically valid.
