Qingyuan Zhao’s research while affiliated with University of Cambridge and other places

It is crucial to be able to explain black-box prediction models to use them effectively and safely in practice. Most existing tools for model explanations are associational rather than causal, and we use two paradoxical examples to show that such explanations are generally inadequate. Motivated by the concept of genetic heritability in twin studies, we propose a new notion called counterfactual explainability for black-box prediction models. Counterfactual explainability has three key advantages: (1) it leverages counterfactual outcomes and extends methods for global sensitivity analysis (such as functional analysis of variance and Sobol's indices) to a causal setting; (2) it is defined not only for the totality of a set of input factors but also for their interactions (indeed, it is a probability measure on a whole ``explanation algebra''); (3) it also applies to dependent input factors whose causal relationship can be modeled by a directed acyclic graph, thus incorporating causal mechanisms into the explanation.

The increasing availability and scale of biobanks and “omic” datasets bring new horizons for understanding biological mechanisms. PathGPS is an exploratory data analysis tool to discover genetic architectures using Genome Wide Association Studies (GWAS) summary data. PathGPS is based on a linear structural equation model where traits are regulated by both genetic and environmental pathways. PathGPS decouples the genetic and environmental components by contrasting the GWAS associations of “signal” genes with those of “noise” genes. From the estimated genetic component, PathGPS then extracts genetic pathways via principal component and factor analysis, leveraging the low-rank and sparse properties. In addition, we provide a bootstrap aggregating (“bagging”) algorithm to improve stability under data perturbation and hyperparameter tuning. When applied to a metabolomics dataset and the UK Biobank, PathGPS confirms several known gene–trait clusters and suggests multiple new hypotheses for future investigations.

Understanding the causal pathogenic mechanisms of diseases is crucial in clinical research. When randomized controlled experiments are not available, Mendelian Randomization (MR) offers an alternative, leveraging genetic mutations as a natural "experiment" to mitigate environmental confoundings. However, most MR analyses treat the risk factors as static variables, potentially oversimplifying dynamic risk factor effects. The framework of life-course MR has been introduced to address this issue. However, current methods face challenges especially when the age-specific GWAS datasets have limited cohort sizes and there are substantial correlations between time points for a single trait. This study proposes a novel approach, estimating a unified system of structural equations for a sequence of temporally ordered heritable traits, requiring only GWAS summary statistics. The method facilitates statistical inference on direct, indirect, and path-wise causal effects and demonstrates superior efficiency and reliability, particularly with noisy GWAS data. By incorporating a spike-and-slab prior for genetic effects, the approach can address extreme polygenicity and weak instrument bias. Through this methodology, we uncovered a protective effect of BMI on breast cancer during a confined period of childhood development. We also analyzed how BMI, systolic blood pressure (SBP), and low-density cholesterol levels influence stroke risk across childhood and adulthood, and identified the intriguing relationships between these risk factors.

Consider a simultaneous hypothesis testing problem where each hypothesis is associated with a test statistic. Suppose it is difficult to obtain the null distribution of the test statistics, but some null hypotheses--referred to as the internal negative controls--are known to be true. When it is reasonable to assume that the test statistics associated with the negative controls are exchangeable with those associated with the unknown true null hypotheses, we propose to use a statistic's Rank Among Negative Controls (RANC) as a p-value for the corresponding hypothesis. We provide two theoretical prospectives on this proposal. First, we view the empirical distribution of the negative control statistics as an estimate of the null distribution. We use this to show that, when the test statistics are exchangeable, the RANC p-values are individually valid and have a positive regression dependence on the subset of true nulls. Second, we study the empirical processes of the test statistics indexed by the rejection threshold. We use this to show that the Benjamini-Hochberg procedure applied to the RANC p-values may still control the false discovery rate when the test statistics are not exchangeable. The practical performance of our method is illustrated using numerical simulations and a real proteomic dataset.

Mendelian randomization (MR) is a term that applies to the use of genetic variation to address causal questions about how modifiable exposures influence different outcomes. The principles of MR are based on Mendel’s laws of inheritance and instrumental variable estimation methods, which enable the inference of causal effects in the presence of unobserved confounding. In this Primer, we outline the principles of MR, the instrumental variable conditions underlying MR estimation and some of the methods used for estimation. We go on to discuss how the assumptions underlying an MR study can be assessed and describe methods of estimation that are robust to certain violations of these assumptions. We give examples of a range of studies in which MR has been applied, the limitations of current methods of analysis and the outlook for MR in the future. The differences between the assumptions required for MR analysis and other forms of epidemiological studies means that MR can be used as part of a triangulation across multiple sources of evidence for causal inference. Mendelian randomization is a technique for using genetic variation to examine the causal effect of a modifiable exposure on an outcome such as disease status. This Primer by Sanderson et al. explains the concepts of and the conditions required for Mendelian randomization analysis, describes key examples of its application and looks towards applying the technique to growing genomic datasets.

This article reviews some early investigations and research studies in the first weeks of the coronavirus disease 2019 (COVID‐19) pandemic from a statistician's perspective. These investigations were based on very small datasets but were momentous in the initial global reactions to the pandemic. The article discusses the initial evidence of high infectiousness of COVID‐19 and why that conclusion was not reached faster than in reality. Further reanalyses of some published COVID‐19 studies show that the epidemic growth was dramatically underestimated by compartmental models, and the lack of fit could have been clearly identified by simple data visualization. Finally, some lessons for statisticians are discussed.

Regression adjustment is broadly applied in randomized trials under the premise that it usually improves the precision of a treatment effect estimator. However, previous work has shown that this is not always true. To further understand this phenomenon, we develop a unified comparison of the asymptotic variance of a class of linear regression-adjusted estimators. Our analysis is based on the classical theory for linear regression with heteroscedastic errors and thus does not assume that the postulated linear model is correct. For a completely randomized binary treatment, we provide sufficient conditions under which some regression-adjusted estimators are guaranteed to be more asymptotically efficient than others. We explore other settings such as general treatment assignment mechanisms and generalized linear models, and find that the variance dominance phenomenon no longer occurs.