Donglin Zeng

• Education

  B.S.(major in pure mathematics), Department of Mathematics, University of Science and Technology of China, 1993
  M.S. (major in partial differential equations), Department of Mathematics, University of Science and Technology of China, 1995
  Ph.D., Department of Statistics, University of Michigan, 2001

• Teaching Activities

  Course I. Semiparametric Models in Health Science, Fall 2002
  Course II. Advanced Probability and Statistical Inference (I), Fall 2004
  Course III. Advanced Probability and Statistical Inference (I), Fall 2005
  Course IV. Advanced Probability and Statistical Inference (I), Fall 2006
  Course V. Advanced Probability and Statistical Inference (I), Fall 2007
  Workshop. Empirical Processes and Statistical Applications, Jul-Oct 2008
  Course VI. Advanced Probability and Statistical Inference (I), Fall 2009
  Course VII. Statistical Learning and High-Dimensional Data, Spring 2011
  Course VIII. Advanced Probability and Statistical Inference (I), Fall 2011
  Course VIV. Advanced Survey Sampling Methods, Spring 2013
  Course X. Advanced Probability and Statistical Inference (I), Fall 2015
  Course XI. Advanced Probability and Statistical Inference (I), Fall 2016
  Course XII. Statistical Learning and Personalized Medicine, Spring 2018
  Course XIII. Advanced Probability and Statistical Inference (I), Fall 2019

• Research Interest

  My current interest is method development on machine learning, personalized medicine, semiparametric inference and high dimensional inference, with particular applications to EHR data, clinical trials, survival data and genetics in biomedical studies. Other applied areas I involve include medical diagnostics in breast imaging and AIDS.
  Some recent work can be found in my curriculum vitae.
  Some algorithms and codes related to my research on semiparametric transformation models are also available. Another page displays simulations for the meta-analysis.
  My personal research interest, probably future interest, also includes linear/nonlinear functional analysis, which I respectfully view as the most pretty pearl in abstract sciences.

• Selected Publications

  1. D. Zeng, Estimating Marginal Survival Function by Adjusting for Dependent Censoring Using Many Covariates, Annals of Statistics, 2004.
  2. D. Zeng, Likelihood Approach for Marginal Proportional Hazards Regression in the Presence of Dependent Censoring, Annals of Statistics, 2005.
  3. D. Zeng and J. Cai, Asymptotic Results for Maximum Likelihood Estimators in Joint Analysis of Repeated Measurements and Survival Time, Annals of Statistics, 2005.
  4. D. Zeng, D.Y. Lin and G. Yin, Maximum Likelihood Estimation in the Proportional Odds Model with Random Effects, Journal of the American Statistical Association, 2005.
  5. D. Zeng, G. Yin and J. Ibrahim, Inference for a Class of Transformed Hazards Model, Journal of the American Statistical Association, 2005.
  6. D.Y. Lin and D. Zeng, Likelihood-Based Inference on Haplotype Effects in Genetic Association (with discussions), Journal of the American Statistical Association, 2006.
  7. D. Zeng, G. Yin and J. Ibrahim, Semiparametric Transformation Models for Survival Data with a Cure Fraction, Journal of the American Statistical Association, 2006.
  8. D. Zeng and D.Y. Lin, Maximum likelihood estimation in semiparametric transformation models for counting processes , Biometrika, 2006 and its technical report.
  9. D. Zeng and D.Y. Lin, Semiparametric Transformation Models With Random Effects for Recurrent Events, Journal of the American Statistical Association, 2007.
  10. D. Zeng and D.Y. Lin, Maximum Likelihood Estimation in Semiparametric Models with Censored Data (with discussion)., Journal of the Royal Statistical Society B, 2007.
  11. D. Zeng and D.Y. Lin, Efficient Estimation in the Accelerated Failure Time Model, Journal of the American Statistical Association, 2007.
  12. Q. Chen, D. Zeng and J. Ibrahim, Sieve Maximum Likelihood Estimation for Regression Models with Covariates Missing at Random, Journal of the American Statistical Assoication, 2007.
  13. B. Johnson, D. Y. Lin, and D. Zeng, Penalized Estimating Functions and Variable Selection in Semiparametric Regression Models, Journal of the American Statistical Assoication, 2008.
  14. G. Yin, H. Li, and D. Zeng, Partially Linear Additive Hazards Regression with Varying Coefficients, Journal of the American Statistical Association, 2008.
  15. D. Zeng, Q. Chen and J. Ibrahim, Gamma-Frailty Transformation Models for Multivariate Survival Times, Biometrika, 2009.
  16. D. Zeng and D.Y. Lin, A Generalized Asymptotic Theory for Maximum Likelihood Estimation in Semiparametric Regression Models with Censored Data, Statistica Sinica, 2009.
  17. D. Zeng and J. Cai Semiparametric Additive Rate Model for Recurrent Events with Informative Terminal Event, Biometrika, 2010.
  18. D.Y. Lin and D. Zeng On The Relative Efficiency of Using Summary Statistics versus Individual Level Data in Meta-Analysis , Biometrika, 2010.
  19. Y. Wang, T. Chen and D. ZengSupport Vector Hazards Machine: A Counting Process Framework for Learning Risk Scores for Censored Outcomes, Journal of Machine Learning Research, 2016.
  20. Y. Wang, P. Wu, Y. Liu, C., C. Weng and D. Zeng Learning Optimal Individualized Treatment Rules form Electronic Health Record Data, IEEE International Conference on Healthcare Informatics: ICHI 2016 proceedings, 2016.

• Selected Personal Views

  
  
• Nowadays, everyone would like to call him/herself a data scientist. But nobody knows what is exactly data science? I create the following contrasts that hopefully can help a bit in thinking.
  Data Engineering vs Data Science
  Precision Science vs Population Science
  Exploratory Science vs Mechanism Data Science
  Empirical Learning vs Generalizable Learning
  Interpretable Machine Learning vs Justifiable Machine Learning
  
• Data analytics driven by mathematics or by computation? This conflict has become even more obvious with advance of computing technology and urgence of anlayzing big data. Traditionally, mathematicians (more precisely, mathematical statisticians) tend to come up with flexible and robust models for analyzing data such as nonparametric or semiparametric models. They provide smart ways (well, need a lot of theory) to produce results in optimal ways. However, there is a growing trend in statistical fields (publications) that people like to use complex and parametric models (generative models, hiearchical neural networks) for data analysis, without worrying much about model misspecification or theory. This is certainly useful with need to analyze big data and takes advantage of "free" computing power. Optimality of the results is less concern as compared to computability. Hence, an interesting question is how to balance data analytics driven by math and by computation. My view is that the latter is useful for model discovery and exploration while the former is useful for inference and decision making, kind of aligning with two levels in my data science pyramid. Another prediction is that similar to evoluation of math and phycics which is known to take sin and cos function shapes, peak time of mathematical data sciene is down time of computational data science and vice versa. We will see.
  
• I often think about where we, as statisticians, are in the era of Big Data. As more and more data emerge these days, researchers from all different fields including us are certainly excited to seek more data evidence to improve and test their own models and methods. However, Big Data does not necessarily mean Big Evidence and it is usually not; moreover, Cheap Data are often Poor Data. Therefore, it is important for us to ensure scientific quality of data, representativeness of data, and generalizability of findings, which is exactly the whole driving force of Statistics in terms of handling random errors, sampling and probabilistic inference. By saying that, I think that we should maintain this important value and unique identity of our researches, instead of being torn away by case-specific methods purely driven by individual data and specific tasks. Evidence-based decision making should always be combined with theory-guided decision making! At the same time, we need to learn from different disciplines, especially computer science, to effectively handle data storage, processing and analysis, combine global analytic objectives and local analytic objectives, and eventually efficiently translate our work into practice.
  
  I create this pyramid for data science (I am sure that workforce in data science has the same pyramid shape--hope salary too -:)).

  Data Science Pyramid

  !!! As statisticians, we should be proud where we are in the pyramid --- it is us who can directly interact with and influence decision makers and clinical users--HOORAY!!!
  
  
  
• High-dimensional variable selection has been the most eye-catching topic in journal publications. However, only till recently do we start to look into post-estimation inference concerning non-regularity of inference. I think that this should be the most important goal from now on, given plenty of regularization methods in this area--maybe every single statistical paper on high-dimensional estimation should have a big component on obtaining correct inference (not point-wise but should be locally regular). For prediction purpose, certain oracle selection may be a good choice; however, concerning inference, non-oracle selection should be used, which exactly reflects the fact of Uncertainty Principle (trade-off between precision and stability). I notice there have been interesting development from Zhang, van der Geer and Belloni's groups using orthogonalization ideas. However, one of my general thoughts is whether we can cast this problem into semiparametric inference problems (for semiparametric models, nuisance parameters are usually infinite-dimensional so certainly high-dimensional), treating the vector of coefficients in some l_2 space (infinite dimensional vector) so the sparsity being some compact set of this space. There have been well-known results in semiparametric framework to construct efficient estimation for parameter of interest and obtain valid inference even if the nuisance parameter is estimated far slower than parametric-rate. I am sure that those results can be useful. Furthermore, we can not only obtain correct inference for one single coefficient but also some "smooth" functional of the whole vector of coefficients (this is known in semiparametric inference). This may be a good direction towards to high-dimensional inference with local regularity, and certainly, it will revive interest of learning semiparametric inference from our students.
  
  
• I wrote a short blog regarding causal inference and I know the same concept has been discussed by other before. I am not trying to change any existing causality framework but just think it may be useful to accelerate drug development, given the difficulty of conducting a perfect randomization study but with the increasing knowledge of patient's information in the Big Data era.
  
  

• Some stories to share

  
  
  Story 1: A while ago, someone told someone who told me this--Any science that tries to name itself "Science" is actually not real science (he reminded me about Mathematics, Physics, Chemistry, Biology)--I don't think they meant to be offensive but it is worth our thinking.
  
  Story 2: This is my own story. The other day, I went to a museum with my data science friend and saw this display:
  
  
  Our immediate response was "Wow! Artificial Intelligence Data Science!" Akwardly enough, it turned out to be a display for AIDS (Acquired Immune Deficiency Syndrome). Later after I came back, I found it very intriguing to ask myself the following question:
  Is A-I-D-S (Artificial Intelligence Data Science) really an "aid-Science" (aid-S) by making human more intelligent, life more efficient and society more fair, or,
  can it become Acquired Immune Deficiency Syndrome (AIDS) for science by dysfunctioning human evolution and impairing value systems?
  Another thing worth deep thought and learning, right?