2022 年的期刊文章 · 第 11 页

共 1481 篇,按发表时间排,新的在前,每页 40 篇。这一页是 2022-07-27 到 2022-07-20。

按期刊、按主题看 →

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 页

SMMR Survival Analysis · 发表于 2022-07-27

Ensemble methods for survival function estimation with time-varying covariates

Weichi Yao, Halina Frydman, Denis Larocque, Jeffrey S Simonoff

摘要

Survival data with time-varying covariates are common in practice. If relevant, they can improve on the estimation of a survival function. However, the traditional survival forests—conditional inference forest, relative risk forest and random survival forest—have accommodated only time-invariant covariates. We generalize the conditional inference and relative risk forests to allow time-varying covariates. We also propose a general framework for estimation of a survival function in the presence of time-varying covariates. We compare their performance with that of the Cox model and transformation forest, adapted here to accommodate time-varying covariates, through a comprehensive simulation study in which the Kaplan-Meier estimate serves as a benchmark, and performance is compared using the integrated [Formula: see text] difference between the true and estimated survival functions. In general, the performance of the two proposed forests substantially improves over the Kaplan-Meier estimate. Taking into account all other factors, under the proportional hazard setting, the best method is always one of the two proposed forests, while under the non-proportional hazard setting, it is the adapted transformation forest. [Formula: see text]-fold cross-validation is used as an effective tool to choose between the methods in practice.

Biometrics Other · 发表于 2022-07-27

Center-Augmented ℓ2-Type Regularization for Subgroup Learning

Ye He, Ling Zhou, Yingcun Xia, Huazhen Lin

摘要

The existing methods for subgroup analysis can be roughly divided into two categories: finite mixture models (FMM) and regularization methods with an ℓ1-type penalty. In this paper, by introducing the group centers and ℓ2-type penalty in the loss function, we propose a novel center-augmented regularization (CAR) method; this method can be regarded as a unification of the regularization method and FMM and hence exhibits higher efficiency and robustness and simpler computations than the existing methods. In particular, its computational complexity is reduced from the of the conventional pairwise-penalty method to only , where n is the sample size and K is the number of subgroups. The asymptotic normality of CAR is established, and the convergence of the algorithm is proven. CAR is applied to a dataset from a multicenter clinical trial, Buprenorphine in the Treatment of Opiate Dependence; a larger R2 is produced and three additional significant variables are identified compared to those of the existing methods.

SMMR Survival Analysis · 发表于 2022-07-27

Analysis of hospital readmissions with competing risks

Wenbo Wu, Kevin He, Xu Shi, Douglas E Schaubel, John D Kalbfleisch

摘要

The 30-day hospital readmission rate has been used in provider profiling for evaluating inter-provider care coordination, medical cost effectiveness, and patient quality of life. Current profiling analyzes use logistic regression to model 30-day readmission as a binary outcome, but one disadvantage of this approach is that this outcome is strongly affected by competing risks (e.g., death). Thus, one, perhaps unintended, consequence is that if two facilities have the same rates of readmission, the one with the higher rate of competing risks will have the lower 30-day readmission rate. We propose a discrete time competing risk model wherein the cause-specific readmission hazard is used to assess provider-level effects. This approach takes account of the timing of events and focuses on the readmission rates which are of primary interest. The quality measure, then is a standardized readmission ratio, akin to a standardized mortality ratio. This measure is not systematically affected by the rate of competing risks. To facilitate the estimation and inference of a large number of provider effects, we develop an efficient Blockwise Inversion Newton algorithm, and a stabilized robust score test that overcomes the conservative nature of the classical robust score test. An application to dialysis patients demonstrates improved profiling, model fitting, and outlier detection over existing methods.

Biometrics Spatial & Time Series Genetics & Omics Latent Variables & Mixtures · 发表于 2022-07-27

A Bayesian Multivariate Mixture Model for High Throughput Spatial Transcriptomics

Carter Allen, Yuzhou Chang, Brian Neelon, Won Chang, Hang J. Kim, Zihai Li, Qin Ma, Dongjun Chung

摘要

High throughput spatial transcriptomics (HST) is a rapidly emerging class of experimental technologies that allow for profiling gene expression in tissue samples at or near single-cell resolution while retaining the spatial location of each sequencing unit within the tissue sample. Through analyzing HST data, we seek to identify sub-populations of cells within a tissue sample that may inform biological phenomena. Existing computational methods either ignore the spatial heterogeneity in gene expression profiles, fail to account for important statistical features such as skewness, or are heuristic-based network clustering methods that lack the inferential benefits of statistical modeling. To address this gap, we develop SPRUCE: a Bayesian spatial multivariate finite mixture model based on multivariate skew-normal distributions, which is capable of identifying distinct cellular sub-populations in HST data. We further implement a novel combination of Pólya–Gamma data augmentation and spatial random effects to infer spatially correlated mixture component membership probabilities without relying on approximate inference techniques. Via a simulation study, we demonstrate the detrimental inferential effects of ignoring skewness or spatial correlation in HST data. Using publicly available human brain HST data, SPRUCE outperforms existing methods in recovering expertly annotated brain layers. Finally, our application of SPRUCE to human breast cancer HST data indicates that SPRUCE can distinguish distinct cell populations within the tumor microenvironment. An R package spruce for fitting the proposed models is available through The Comprehensive R Archive Network.

JRSSB Machine Learning 讨论 / 评论 · 发表于 2022-07-26

Yanbo Tang's Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Yanbo Tang

摘要

The method proposed by Vansteelandt and Dukes is an interesting combination of non-parametric and machine learning techniques with traditional ideas in statistics. We note in this discussion that the authors' argument is easily extended to obtain a quantitative rate in terms of the quality of the nonparametric conditional mean estimate, and comment on how this may impact a practitioner's choice of the machine learning model used to estimate the conditional mean. However the exact rate of consistency of the chosen estimator for E(Y | A, L) and the other conditional means matters if one wishes to quantify the speed at which the error term decays. For example, consider if all of the conditional means in Theorem 2 have a rate of consistency of Op {n − 1/2−ϵ}, for some ϵ > 0, then following the thread of calculation available in the appendix, we arrive at n1/2(⁠β-β̂⁠) = Zn + Op(⁠n-ϵ⁠). While the error in the approximation tends to 0, it may be doing so at a very slow rate which implies a large amount of data is needed to produce a confidence interval with the correct coverage. While if the model were a correctly specified GLM, using the traditional Wald statistic on the regression parameter β would have resulted in the classical error rate of Op (n − 1/2). This potential loss in the accuracy of the distributional approximation is not surprising as it reflects the difficulty of non-parametric estimation problems in general. This suggests that the method chosen to estimate the conditional mean needs to be flexible enough to produce a good estimator, while not so complex that the rate of consistency is too slow to be of practical use for a given dataset of size n. Thus some additional consideration is needed on the practitioner's part when selecting the machine learning or non-parametric method used to estimate the conditional mean, and in particular some knowledge of the true conditional mean function may be required, for example its smoothness in terms of higher-order derivatives. *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Thomas S. Richardson’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Thomas S. Richardson

摘要

I congratulate the authors on an interesting and thought provoking contribution to the literature on modelling associations. A particular strength of the approach advocated by the authors is that it may facilitate the estimation of main effects on the relative risk scale, by taking g = log. In contrast, a GLM specification such as (4) with log link presents difficulties owing to the variation dependence between β and ω(·). This dependence, also called non-congeniality, often leads to computational difficulties in practice (Lumley et al., 2006). While a partially linear GLM using the logit link function does not suffer from the problem of variation dependence it still faces the issue that, owing to the lack of collapsibility of the odds ratio, interpretation of a logistic ‘main effect’ can be challenging. In particular, even when a specification such as (4) with g = logit is correct, the main effect β will never be closer to zero than the log odds ratio for the marginal association. More generally, when (4) does not hold, the marginal (log) odds ratio may not even lie within the convex hull of the conditional (log) odds ratios as these vary over strata defined by L. In such a setting, any weighted average of the conditional associations will not reflect the marginal association. Perhaps for this reason, as noted by Lumley et al. (2006), applied researchers often express a preference for relative risks over (log) odds ratios. *(abstract via OpenAlex)*

JRSSB Causal Inference Survival Analysis Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Seconder of the Vote of thanks to Vansteelandt and Dukes and Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’

Vanessa Didelez

摘要

In my view, one of the most important contributions of the field of causal inference has been to place the target of inference, the desired estimand, at the centre of the analysis. The estimand is chosen in view of the research question, and typically reflects what decision problem we need to solve or what our ideal (target) trial would be. Crucially, the (causal) estimand is not automatically a parameter that happens to parametrise a chosen model. The role of models is mostly as a mere tool. For instance, with a survival outcome we may employ hazard regressions but still obtain effects in terms of survival probabilities. This is in some contrast to traditional ‘statistical modelling’, where the model often appears to be the starting point, somehow suggested by the data (e.g. a logistic regression for a binary outcome), and the natural coefficients are reported (e.g. log-odds-ratios). However, it has become abundantly clear that, as far as prediction is concerned, default regression models are regularly outperformed by machine learning methods such as random forests and similar flexible methods in competitions. The former are then typically defended with the argument of being more interpretable and their parameters being useful summaries of (conditional) dependence. The thoughtful proposal of Vansteelandt and Dukes (V&D) addresses many of these issues with an important lesson right at the start: models, even when only used as tools, may implicitly affect the meaning of our estimands and the desired summaries may be invalidated under misspecification. V&D therefore place a (particular) estimand at the core, aiming at a simple summary of a (high-dimensional) conditional dependence such that the estimand remains sensible regardless of whether a specific model holds. The influence function-based estimation method then uses flexible machine learning for the nuisance functions in a way that ensures valid inference. While I agree with many of V&D’s points, I am (very slightly) concerned that they might distract from asking scientifically relevant questions, which would conflict with the authors’ intention. Their proposal restricts our choice of estimand. For example, when the exposure A is continuous, V&D make the entirely valid point that the potential outcome Ya (and thus an estimand as E(Ya-Ya′)⁠) typically represents an unrealistic intervention of setting A to exactly a even for people whose ‘natural’ value of A would be very different from a. But, such an estimand can be scientifically meaningful, for example when A is the dosage of a drug; in contrast, it is less meaningful when A is BMI or income, for example. However, the proposed alternative estimand does not solve this problem—only when we actually formulate a scientifically relevant question will we (possibly) find scientifically relevant answers. Giving less weight to problematic covariate regions does not achieve this—and it does not absolve us from trying harder to elicit what a scientifically relevant estimand might be instead, for instance for the effect of BMI on an outcome. Moreover, it is a useful feature of approaches like IPW or propensity score matching that they alert us to problems, for example when we carry out diagnostics and find that there is a lack of overlap and then go on to characterise regions of sufficient overlap; or even when we just have extreme weights and confidence intervals get very wide, this rightly indicates that no useful statement can be made about our estimand because there is too little information in the data. I worry that such aspects are perhaps lost with V&D’s approach, or it would be interesting to know if it could be supplemented by something analogous. In motivating the proposed estimands, V&D further refer to our ‘familiarity’ with main effects and interactions in GLMs. However, familiarity is not per-se a relevant criterion: it does, again, not ensure scientific relevance. There are plenty of examples, for instance the problematic causal interpretation of hazard ratios despite many medical statisticians being extremely familiar with them. Besides, what exactly is it that we think we are familiar with? When regression models are used to describe conditional dependencies, the correct interpretation of regression coefficients as just a measure of conditional dependence, and not as an effect, is rare. A clear distinction between covariates included to adjust for confounding and those included as potential ‘effect’ modifiers is also rare, typically no rationale for the inclusion of particular covariates is given at all—what then motivates the choice of L in the basic quantity E(Y|A, L)? In summary, I very much welcome V&D’s proposal as a huge improvement on traditional statistical regression modelling; but with regard to causal analyses there is still room for further research into combining the truly impressive results on assumption-lean valid inference with the quest for more scientifically meaningful estimands. No doubt, V&D will lead the way. It is with greatest pleasure that I second the vote of thanks for this most stimulating and important paper. The vote of thanks was passed by acclamation. *(abstract via OpenAlex)*

JRSSB Clinical Trials Causal Inference 讨论 / 评论 · 发表于 2022-07-26

Rachael V. Phillips and Mark J. van der Laan’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Rachael V. Phillips, Mark J. van der Laan

摘要

We commend the authors on this novel non-parametric extension of main and interaction coefficients in a generalised linear model (GLM), such as E(Y|A,L)=βA+g(L)⁠, and their efficient estimation. There is a wealth of non-parametric extensions one could pursue, including weighted squared error projections of a conditional treatment effect E(Y|A=a,L)-E(Y|A=0,L) and conditional interaction effect E(Y|A=(a1+1,a2+1),L)-E(Y|A=(a1,a2+1),L)-E(Y|A=(a1+1,a2),L)+E(Y|A=(a1,a2),L) onto a simple parametric models βa and β0+β1a1+β2a2⁠, respectively (Chambaz et al. (2012)). Such projection estimands are common in the causal inference literature. The authors propose a different estimand criterion that requires the efficient influence curve (EIC) to avoid (conditional) density estimation and inverse weighting. Inverse weighting can certainly cause instability but machine learning techniques have been well-adapted for conditional density estimation (for instance, Muñoz and van der Laan (2011), Rytgaard et al. (2021) and van der Laan (2010)). The above least squares projection for the main effect has an EIC that inverse weights by P(A=0|L)⁠, thereby achieving the first but not the second goal. The proposed main effect estimand E((A-E(A|L))(E(Y|A,L)-E(Y|L)))/EσA|L2 (5) satisfies this criteria for the EIC; however, it is harder to interpret outside the GLM. For example, if (E(Y|A,L)-E(Y|L)) is not linear in (A-E(A|L))⁠, then the numerator will generally average both negative and positive contributions E(Y|A,L)-E(Y|L)⁠, even for problems in which E(Y|A,L)-E(Y|A=0,L)≥0 everywhere. Consider A∼U(0,1) independent of L and E(Y|A,L)=ϵ-1AI(A≤ϵ)+I(A>ϵ)+g(L)⁠. Here, the numerator becomes E(A-1/2)(ϵ-1AI(A≤ϵ)+I(A>ϵ)) and approximates E(A-1/2)=0 as ϵ→0⁠, whereas the unweighted projection of E(Y|A,L)-E(Y|A=0,L) on βa⁠, would result in β=3/2-ϵ2/2⁠, correctly demonstrating a strong treatment effect. In addition to impacting the interpretation, these cancellations can hurt the power for testing a null hypothesis relative to testing with the parameter in Chambaz et al. (2012), even though the latter’s EIC might have larger variance. The interpretation of the interaction estimand (e.g., 10) for continuous A1⁠, A2 has an additional complication in the sense that it does not involve an average of L-specific interactions. Both of the proposed estimands and their EICs depend on the conditional distribution of A⁠, given L⁠, making the interpretations non-robust to irrelevant deviations in the study. Users should carefully consider knowledge regarding the model for P(A|L) and compute the EIC accordingly, potentially achieving efficiency gains relative to the non-parametric model considered. Finally, the authors proposed a one-step estimator of their estimands based on the EIC. We believe that practitioners that are used to maximum likelihood estimation (MLE) would find it more natural to use a plug-in estimator, and thereby targeted MLE (TMLE). It would be straightforward to develop a TMLE, as the EICs imply how to target initial estimators of E(A|L) and E(Y|A,L) for the main terms estimand, and their analogs for the interaction estimand. *(abstract via OpenAlex)*

JRSSB Causal Inference Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Proposer of the Vote of Thanks and Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Rhian M. Daniel

摘要

Traditional statistical modelling starts from a family F of observed data laws indexed by unknown parameters of interest β. The goal is to make inference about β under the assumption that F contains the true law. By labelling β ‘of interest’, it is implied that F can be expressed such that β naturally encompasses the main scientific goal, which is not always the case. Furthermore (e.g. ch.1 Cox & Hinkley, 1979) if F does not contain the truth then the inferential theory loses relevance and the interpretation of β is obscure. Such concerns sensibly lead to model checking procedures, which themselves raise further concerns, as VD describe. The causal inference and targeted learning schools (Hernán & Robins, 2020; van der Laan & Rose, 2011) start instead from an estimand, chosen to reflect the scientific question, without reference to any statistical model. Subsequent estimation and inference are tailored to this estimand, sometimes using a parametric model F⁠, but not to define the estimand. The targeted learning framework advocates replacing F with machine learning algorithms, using the estimand’s influence function and accompanying theory to derive estimators with well-understood asymptotic behaviour. Although the hygiene of the latter approach is eminently attractive, its implementation requires statistical expertise. In principle, each bespoke estimand demands that all subsequent steps be derived afresh, with no guarantees that the resulting estimator has good properties (e.g. when the estimand is too ambitious given the available data). Practical applications of targeted learning thus tend to focus on simple estimands (e.g. the marginal effect of a binary exposure) where off-the-shelf implementations are readily available. This leaves users in a quandary when their scientific question is more complex, for example when the exposure is continuous, as in the settings considered by VD. VD start, as in the traditional approach, from a generalised linear model F indexed by β. This has the advantage of restricting attention to quantities that are plausibly reliably estimable from the data. For any estimator β^⁠, consistent under F⁠, their philosophy is to consider its probability limit β* under A⁠, the set of all possible data laws. The honest estimand β* is only considered acceptable if it corresponds under A⁠, to a weighted average of parameters βl⁠, where each βl has the interpretation of β restricted to levels l of some covariates L, not of primary interest. VD set high standards for making inference about β*⁠, namely consistent estimation and parametric convergence rates under ‘lean’ regularity assumptions, honest inference after algorithm/variable selection, and no density estimation for continuous variables. They argue convincingly that such demands are necessary for the data to speak for themselves about β*⁠, and describe a general procedure that meets these standards in the case of any parameter motivated by a GLM. That VD propose essentially non-parametric estimation may seem alarming in the light of the curse of dimensionality (Stone, 1985). Indeed, the traditional approach based on one simple model is often justified on the grounds of a bias–variance trade-off: we assume a simple (‘wrong but useful’) model since it buys precision in modest-sized data sets. The simulation studies presented by VD illustrate that this intuition is faulty: their assumption-lean estimators are also relatively precise, but then at what cost? Our intuition was developed in the context of the traditional approach in which F plays the two roles described by VD: (i) estimand definition, and (ii) representing the set of possible data laws. Traditionally, choosing a more complex model F leads simultaneously to a less parsimonious estimand and a larger set of possible data laws. VD, on the other hand, propose a parsimonious estimand, coupled with only very lean restrictions on the set of data laws: parsimony in the first sense but not the second. Figure 1 gives a simple illustration of how parsimony in both senses increases efficiency, but with parsimony of type (i) having a greater impact than type (ii). This graph shows the increase in relative standard error for the estimator of two different types of estimands after subdividing a covariate L into progressively more strata. The 3000 simulated datasets each with sample size 1000 are from a hypothetical observational study with a continuous confounder L ∼ N(0, 1), a binary exposure A with Pr(A = 1|L) = expit(L) and a binary outcome Y with Pr(Y=1|A,L)=expit(-2+0.2AL2)⁠. Each dataset is divided into an increasing number s of approximately equally-populated strata based on the observed quantiles of L. We first plot the empirical standard deviation of the stratum-specific estimator of the average causal effect in each stratum separately, when splitting into s = 1, …, 30 strata relative to 1 (i.e. no stratification). Since the SE varies by stratum, the plot in fact takes the average of the SE over the s strata. We then plot the relative empirical standard deviation of the estimator of the average causal effect (marginalised over the strata) when the data analysis model is stratified into s strata relative to 1. Since the true model for Y given A and L has the same form regardless of the value of L, the models with and without stratification are all correctly specified. This allows us to explore, on the one hand, the impact of needless flexibility in F in the sense described in (ii) in the text (the slowly increasing lower line) compared with the additional impact of decreasing parsimony in the estimand of interest, i.e. sense (i) in the text (the more steeply increasing upper line). Since consistent estimation is guaranteed under very lean assumptions, and thus bias essentially avoided, the sacrifice made by VD’s parsimony (in the first sense) with which they buy precision is, I believe, not bias but bluntness. A more nuanced (less blunt) understanding of, say, a continuous exposure’s effect on an outcome, could be gained by choosing a less parsimonious summary, for example one that separately summarises the effect in more sub-groups, but at the cost of increased variance. VD start from the viewpoint that the two approaches in Section 1 are unsatisfactory. The traditional well-trodden path offers a comfortable ride but often to an unknown and uninteresting destination with a dishonest account of how we got there. On the other hand, the targeted learning path, in aiming admirably for the summit of a yet-to-be-conquered mountain, is often too perilous to navigate with our modest equipment and abilities. VD offer a third way, which feels on the surface much like the first, but leads to a well-defined destination that is both practically reachable and at least somewhere in the foothills of scientific interest. Beneath the surface lies much of the sophisticated technology from the targeted learning journey, but as passengers we need not necessarily know how to operate it, thanks to their general-purpose solution. I conclude by congratulating Vansteelandt and Dukes on their innovative yet pragmatic proposal presented in a wonderfully didactic fashion that provokes us to rethink fundamental aspects of statistical modelling. I enthusiastically propose the vote of thanks. *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Peng Ding’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Peng Ding

摘要

I want to congratulate the authors on a thought-provoking paper on statistical inference with misspecified models. It deepens our understanding of the role of models in causal inference. I offer two comments. Comparison of the conditional and marginal estimands based on 105 Monte Carlo samples. The parameters of the distributions of L and Y are all drawn independently from Uniform(0, 1) The authors provide inspiring derivations of the novel average effect and interaction measures under possibly misspecified models. I am wondering how general their strategy is beyond these two measures. *(abstract via OpenAlex)*

JRSSB Clinical Trials Causal Inference Testing & Uncertainty 讨论 / 评论 · 发表于 2022-07-26

Pallavi Basuʼs Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Pallavi Basu

摘要

The authors propose an ingenious way to improve on providing inference for the main and interaction effects for generalised linear model parameters. With the use of the efficient influence function, they can extend the allowable bias introduced by machine learning estimators up to op(n −1/4). Noting that cross-fitting has been suggested to counter the Donsker condition, it may be a useful alternative to contrast with machine learning estimates, averages over sample splits or even averages over several models. A minimiser of the Kullback–Leibler divergence style model selection procedure of the ‘best’ model within the class of considered models has the advantage of having a model available at disposal. This may be useful in explaining to applied collaborators or providing out-of-sample estimates or prediction intervals. The philosophy of addressing model misspecification used in this work loses that connection. It could be not very clear, a priori, how to choose between the main effect versus an interaction effect estimand. Suppose there is an interest in the main effect. However, it is expected for the linear model to have interaction terms; how would one interpret the main effect estimand in the interaction model? Although a little far-fetched, practitioners are often interested in testing the interaction effect of a group of covariates, say, the interactions of a policy and socioeconomic factors on health outcome; a model then aids in simpler joint statistical tests. Analogously, simultaneous intervals or confidence sets will be useful here. Furthermore, in the spirit of the recommended ‘hierarchical principle’, how would we interpret estimated effects when the confidence interval (CI) for the interaction estimand does not include the null and the CI for the main effect includes zero or no effect? Further, to connect the dots well, it will be great to understand the causal identifying conditions, considering that inference rather than prediction is the goal, an edge that will greatly help declutter the uses of black-box methodologies. Thereafter, connecting to unmeasured confounding either via Rosenbaum’s sensitivity or VanderWeele's E-value will be beneficial. However, the inferential nature of the causal estimand, either testing or providing valid confidence intervals, should remain. *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Oliver Hines and Karla Diaz-Ordazʼs Contribution to the Discussion of ‘Assumption-Lean Inference For Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Oliver Hines, Karla Diaz-Ordaz

摘要

We congratulate the authors on their work, which contributes to a growing movement in Statistics, moving away from (semi)parametric-based regression model assumptions, in favour of understanding working models as non-parametric projections of the true distribution. Here, we highlight this aspect of the work, and draw connections with other recent developments. The parametric modeller assumes that for some particular β, ω(.), the remainder term is exactly zero. Inference is usually carried out under this assumption, resulting in the dishonest inference described in (Vansteelandt & Dukes, 2020). The projectionist, however, instead aims to report β such that the remainder term is, in some sense, small. The projectionist argues that when the remainder is small, β captures the main effect of A on Y. Indeed when the parametric modeller is correct, both the reported main effects coincide. We believe that model projection estimands represent an exciting frontier in statistical research. *(abstract via OpenAlex)*

JRSSB Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Niwen Zhou and Xu Guo’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Niwen Zhou, Xu Guo

摘要

We thank Professors Vansteelandt and Dukes for their innovative and stimulating paper. They introduce a new estimand which reduces to the exposure effect parameter when the (semi)parametric model holds, is generic for continuous or discrete exposure, and still captures conditional association when model assumption fails. We make some additional understanding of the new estimand and introduce a related new estimand. Here we consider the expectation of the ratio. Here β· and ω· are unknown functions. Different from the model (4) in Vansteelandt and Dukes (2021), we now allow A–L interaction, and extend to the exposure effect heterogeneity setting, with respect to L. While under model (1), γ=Eβ(L)⁠, which measures the average exposure effect. When there is no exposure effect heterogeneity, i.e. β(L)≡β⁠, γ reduces to β as the estimand in Vansteelandt and Dukes (2021). Even when the model fails, the estimand γ is still meaningful to capture conditional association. Clearly, the new estimand in Vansteelandt and Dukes (2021) can be viewed as an extension of weighted average treatment effect; while our new estimand is an extension of the classical average treatment effect. Lastly, current literature only focus on the conditional expectation of the exposures, which is sensitive to outliers. Conditional quantile can provide a more robust and complete view of the association between response and exposure. Thus it would be very interesting to extend the insightful idea in Vansteelandt and Dukes (2021) to quantile setting. *(abstract via OpenAlex)*

JRSSB Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Michael Lavine and James Hodges’ Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Michael Lavine, James Hodges

摘要

The authors advocate a style and rhetoric of statistical analysis that begins with specifying an estimand and ends with an estimate and interval. They say, ‘The starting point… is to come up with an estimand that is meaningful when the above [generalized partially linear] model [(4)] does not hold, but reduces to [the familiar regression coefficient] β when the model holds; this… allows for nonparametric inference to be developed for that estimand’. However, in our opinion, such an analysis is not always appropriate. For example, consider the second of Anscombe's (1973) data sets, in which Y is a quadratic function of X. The authors’ Y is Anscombe's Y; the authors’ A is Anscombe's X; and the authors’ L has just one value. The authors’ estimand appears to be given by (5), where g is the identity function. For a given data set, including Anscombe's, (5) reduces to one number which is to be interpreted similarly to a slope. In our opinion, that is a poor way to summarize Anscombe's data. The problem arises because the authors insist on choosing an estimand before seeing the data, thereby precluding the possibility of summarizing the data with a statement like ‘E[Y|X] appears to be a roughly unimodal function of X with a maximum around X = 11’. Contrary to the authors’ claim that ‘the resulting analysis can be pre-specified, which is essential if one aims for an honest data analysis that reflects all uncertainties’, we think it is dishonest, or at least a serious mistake, to summarize Anscombe's data with one slope-like number. Instead, we prefer a style and rhetoric of sensitivity, which begins with looking at the data to see what sorts of models provide useful descriptions; then uses familiar, interpretable quantities and a simple analysis; and then considers how substantive conclusions change with deviations from the assumptions of the simple analysis, elaborated only as needed. The product is a leading result and a summary of the effects of variations. Barr et al. (2012) is but one of many possible examples of this style and rhetoric, the regression diagnostics literature being another. From our experience, a rhetoric of sensitivity provides a more intelligible and better-fitted answer to our scientific collaborators’ questions and concerns, at least for the range of problems we have worked, as well as being more transparent. *(abstract via OpenAlex)*

JRSSB Econometrics & Economics Causal Inference Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Mats J Stensrud and Aaron L. Sarvet’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Mats J. Stensrud, Aaron L. Sarvet

摘要

We congratulate Vansteelandt and Dukes (V & D) with their innovative and interesting article. Here we further explore the interpretation of V & D’s main effects estimand. An algorithm for causal inference. There has been tremendous progress in causal inference by approaching causal queries in the following way (Hernan & Robins, 2020; Richardson & Robins, 2013; Robins, 1986): Choose a causal target: an estimand that corresponds to a scientific question of interest (usually a causal parameter). Specify a set of (reasonable) assumptions to define a model, and deduce a functional that equals the causal target at every law in the model. If the functional is fully comprised of observed data parameters, then we have full identification. If the functional is partially comprised of observed data parameters, then we have partial identification. Implement an estimator of the observed components of the identification functional. An analysis that results in partial identification, that is, it goes through 2.b, might be considered the classical ‘assumption-lean’ approach. The assumptions and the causal target are fixed. The investigators leverage their (limited) assumptions to derive bounds, that is ‘worst-cases’, for the value of the causal target. V & D do not follow this algorithm. Their algorithm fixes a functional of observed data—what they call an ‘estimand’. Then they consider an assumption (a parametric modelling assumption) that may or may not hold, and they derive what we might call a lower bound on the interpretation of their ‘estimand’. For example, V & D's Expression (5) corresponds to an average treatment effect under usual identification assumptions and a strong parametric assumption. In the absence of the strong parametric assumption, it corresponds to a variance weighted treatment effect (Robins et al., 2008). But, those who make decisions based on data cannot avoid discrete commitments about assumptions in their analyses. V & D’s algorithm does provide novel information (in the form of a worst-case interpretation) when the investigator is not willing to commit to the (parametric) model assumption. However, in this case decision-makers must be aware of the implications of the worst-case interpretation and how it differs from a causal target that more naturally corresponds to their scientific question of interest. Failure to do so could lead to wrong conclusions, as illustrated in the following example. Thus VD(p, q) > 0 when |q − 0.5| |p − 0.5|. This example illustrates a more general point. V & D’s main effects estimand does not necessarily correspond to an ATE, even when the ATE is 0 and even when there is no unmeasured confounding. In more complex settings, with higher dimensional treatments and non-linear link functions, these issues can still exist. *(abstract via OpenAlex)*

JRSSB Clinical Trials Bayesian Methods 讨论 / 评论 · 发表于 2022-07-26

Kuldeep Kumar’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Kuldeep Kumar

摘要

According to Leo Breiman (2001), there are two broad cultures for analysing and modelling to reach conclusions from the data. The first one is data modelling culture where the value of the parameters are estimated from the data and then the model is used for information and/or prediction. The second one is algorithmic modelling culture, where the approach is to find a function f(x) using an algorithm that operates on x to predict y. In the first case, statistical tools like OLS are used whereas machine learning tools perform much better in the second case. I should congratulate the authors for developing a hybrid model using machine learning and traditional statistical techniques like OLS. However, I do have few queries on the real-life applications of the proposed model. Most of the simulation results for main effects, effect modification and misspecified link function are valid for sample size n = 500 or more. Also, the data analysis example presented in section 7 has a sample size of 2500. How is the empirical bias affected when the sample size is small? Bayesian models incorporate prior information, which can be specified quantitatively in the form of a distribution. Do the authors think a Bayesian parameter generalised linear model can perform better if some prior information is available? Also, I am not sure how this model tackles the problem of multicollinearity? Finally, in the context of high dimensional variable selection, what will happen if the covariates are not distributed normally? *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Jiwei Zhao’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Jiwei Zhao

摘要

First I congratulate Vansteelandt and Dukes for a fascinating and stimulating paper, which I believe will have great impact on the practical data analysis as well as on the application of semiparametric techniques. In applications, any fitted model could be a misspecified model, for example Ê(A|L) converges to E*(A|L)⁠, where E* denotes the expectation that is evaluated under a working model. The similar notations are also used for Cov and Var throughout. Nonetheless, I think the framework the authors propose can be naturally extended to the case where E(A | L) is misspecified. This comment is strongly relevant to the locally efficient estimator in semiparametric models. *(abstract via OpenAlex)*

JRSSB Causal Inference Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Ilya Shpitser’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Ilya Shpitser

摘要

A hallmark of principled causal inference is being careful and explicit about assumptions underlying data analysis. In their timely paper, the authors, Professor Stijn Vansteelandt and Oliver Dukes, adopt this view to provide a general roadmap for cautious statistical inference about target parameters. The authors distinguish between assumptions made for substantive reasons and those made for convenience. An important feature of their approach is that an estimand with respect to an unrestricted model (hereafter, the ‘non-parametric estimand’) must be specified prior to imposing any model restrictions. The authors illustrate their roadmap by defining and estimating main effect and interaction parameters in regression models. These estimands are defined without reference to any model restrictions, but at the same time specialise to known special cases if restrictions hold, for example if the partially linear regression model is true. The proposal recapitulates developments in mediation analysis, where direct and indirect effect parameters were initially defined as simple functions of regression coefficients in linear causal models, with the definition eventually generalising to a fully non-parametric one based on nested potential outcomes (Pearl, 2001; Robins & Greenland, 1992) or separable effects based on treatment components (Robins & Richardson, 2010). In causal inference, an estimand in an unrestricted model is often implied by the combination of the causal model, substantive considerations, and identification theory yielding an identified functional for the target causal parameter. The difficulty with the authors’ proposal is that the considerations one should appeal to when defining the non-parametric estimand are unclear. The authors propose a number of desiderata such an estimand must satisfy, but (a) these are perhaps arguable, and (b) given a particular specialised parameter defined in a parametric or semi-parametric model, for instance a regression coefficient, it seems there will be many non-parametric estimands that reduce to the specialised parameter under restricted models, and otherwise satisfy the authors’ desiderata. Different estimands will yield substantively different conclusions, and it is not clear how to choose among them. It would not be surprising if pragmatic considerations (e.g. ease of inference, availability of software) guided such a choice in practice. Indeed, it appears that the authors themselves were guided by pragmatic considerations when choosing among multiple possibilities for the non-parametric interaction parameter. This would seem to be contrary to the spirit of the authors’ proposal. As a causal inference researcher myself, I applaud the call to clarity about assumptions and estimands made by the authors. However, absent clear substantive guidance or an analogue of identification theory in causal inference, making the crucial choice of one non-parametric estimand out of many appears to be a ‘black art’. *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Ian Hunt's Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Ian Hunt

摘要

This paper offers a generic and principled addition to conventional statistical modelling. It is a real advance in applicable methodology. But there is no need in the paper for the following claims: that statisticians routinely use ‘dishonest’ modelling assumptions, that statisticians should aim for ‘purely evidence-based’ inferences and that modelling assumptions are ‘almost always a pure mathematical convenience’. These sorts of claims are redundant or misleading for three reasons. First, text-book statistical inferences look like deductions. But nearly everyone wants to make inductive inferences. Statisticians assist in the leap from deduction to induction via iterations of a two-step process: step one, specify different models and estimate their parameters (adding and dropping interaction terms, using different error assumptions and so on); step two, assess and compare the models (including assumption checking). This inductive process is necessarily vague, but it embodies the key statistical tasks defined by Fisher (1959, pp. 6–8). Properly disclosed assumptions that cannot be proved, or are openly false, play an honest and essential role in inductive processes—we have to take risks to go beyond the data and what we already know. Adding assumption-lean or ‘less risky’ methods to inductive processes may prove useful but this would not be ‘more honest’. Second, if by ‘evidence’ the authors really mean data then I would argue that aiming for ‘purely evidence-based’ inferences is not ambitious enough. Seeking inductive or causal inferences is to aim higher, but ‘no causes in, no causes out’ (Cartwright, 1994) and, I argue, ‘no assumptions in, no inductions out’. In other words, substantive and risky modelling assumptions help turn data into the sort of evidence that supports causal and inductive claims. Third, I acknowledge the convenience-factor of conventional modelling assumptions; proverbial examples include simplistic error structures and the absence of model mis-specification. But with finite samples, the methodology promoted in this paper has its own hazards tucked away. For example, estimators may depend on the particular machine learning method used (kitchen-sink approaches, like SuperLearner, are no panacea); and sample-splitting or cross-fitting can cause biases (such as a tendency to over-penalize more complex models that require larger samples for adequate parameter estimation) as well as noise. Assumption-lean methods are not necessarily superior to assumption-rich methods, but they are surely complimentary. In future applied work, I look forward to adding some of these generic estimand-centric methods to my usual, honestly-reported, model comparisons and tentative inductive leaps. *(abstract via OpenAlex)*

JRSSB Survival Analysis 讨论 / 评论 · 发表于 2022-07-26

Heather Battey’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Heather Battey

摘要

For situations in which there is uncertainty over the underlying probabilistic model, there are at least three broad approaches. One is to seek reliable inference for interest parameters or perhaps, as the authors advocate, for quantities retaining at least a degree of interpretability under misspecification. Another is to acknowledge more explicitly the model uncertainty. A third approach, loosely connected to the first, is to encapsulate uncertainty over the model in a possibly large number of nuisance parameters, to be eliminated in the analysis by suitable conditioning arguments or other problem-specific manoeuvres (e.g. Bartlett, 1937). A helpful example is the use of partial likelihood to evade the baseline hazard function (an infinite-dimensional nuisance parameter) of a proportional hazards model. The appropriateness of each of the three approaches depends largely on context. I will constrain my discussion to the first two. If the interpretation of an interest parameter is stable over models, it appears that first-order reliable inference via maximum likelihood estimation is possible in spite of considerable misspecification in the nuisance part of the model only when the interest parameter is orthogonal (in the sense of Jeffreys, 1948, pp. 158–184) to the notional nuisance parameters, whose interpretation then has to be in terms of Kullback–Leibler projection. This would be a necessary condition rather than a sufficient one. Note that the true model is also implicit in the definition of parameter orthogonality. It is, as far as I am aware, an open problem to characterise the class of models whose interest and notional nuisance parameters are orthogonal under arbitrary model misspecification, perhaps after interest-respecting reparameterisation. The second-order properties are always affected, sometimes severely, which is problematic beyond point estimation. On a historical point, the limit in probability of the maximum likelihood estimator under model misspecification and its connection to the Kullback–Leibler divergence was derived by Cox (1961, 1962), who also noted the failure of Bartlett's second identity and gave a generalisation of the result (Cox, 1961, equations (28)–(43)) which later became known as the sandwich formula. A more rigorous discussion of regularity conditions was given by Huber (1967). Similar results were obtained independently by White (1982a,1982b). It could be argued, contrary to the paper under discussion, that when the effects of interest are represented by parameters whose interpretations differ according to the model used, the appropriate approach is to acknowledge the model uncertainty rather than seek inference on a quantity whose interpretation is stable but perhaps only tangentially relevant when the assumed model is false. The role of sufficiency in assessment of model adequacy, implicit in R. A. Fisher's work is perhaps best approached via Barndorff-Nielsen and Cox (1994, p.29). When the ideas can be operationalised, there are no difficulties associated with double use of the data for model assessment and parametric inference. The conclusion may be that some, all or none of the a priori plausible representations are compatible with the data. If multiple models with different interpretations are not significantly contradicted, it is sometimes appropriate to report as many as feasible, a point emphasised repeatedly by D. R. Cox (e.g. Cox, 1968, 1995; Cox & Snell, 1974, p. 55; 1989, p. 193). See also Davison (1995). This underpins the development of confidence sets of models (Cox & Battey, 2017). *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Eric J Tchetgen Tchetgen’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Eric J. Tchetgen Tchetgen

摘要

Finally, it is worth noting that in the case of identity link, Robins et al. (2008), Li et al. (2011) and Mukherjee et al. (2017) proposed minimax estimators of functional (5) of V&D, which are root-n consistent, asymptotically normal and semiparametric efficient under rate conditions for estimating nuisance functions that are significantly weaker than those of V&D in Theorem 2 (nuisance estimation rate conditions for the theorem are therefore sufficient but not necessary at least in the identity link case); their estimator also attains the minimax lower bound of Robins et al. (2009) in non-root-n regimes. More recently, Liu et al. (2021) obtained optimal adaptive minimax estimation results (both in the sense of upper and lower bounds) which simultaneously described both root-n and non-root-n regimes of estimation of a general class of non-parametric functionals which includes (5) in the identity link case. It remains unknown and much of interest whether similar results can be obtained for functional (5) in non-linear link cases. *(abstract via OpenAlex)*

JRSSB Causal Inference 讨论 / 评论 · 发表于 2022-07-26

Elizabeth L Ogburn, Junhui Cai, Arun K Kuchibhotla, Richard A Berk and Andreas Buja’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Elizabeth L. Ogburn, Junhui Cai, Arun K. Kuchibhotla, Richard A. Berk, Andreas Buja

摘要

Not all conditional associations between outcomes and exposures are of interest. Those that are tend to be directional: up or down. The simplest way to assess directionality is to fit a confounder- adjusted linear exposure term, as the authors propose. We agree with this approach as some of us have argued that linear slopes are meaningful and interpretable even if the directional association is not linear (Buja et al., 2019, section 10). The authors, and Whitney et al. (2019), remind us that severely misspecified adjustment can result in distortions of linear exposure slopes. In their examples, the A−L distributions have U-shaped nonlinearities and, as a result, naive linear adjustment produces a biased estimate of the true slope. Thorough data analysis could unearth such exposure-confounder structure if present in real data. A greater worry for practitioners is missing an essential confounder that biases or reverses the direction of association. The authors’ inferential framework does not require L to control for all A−Y confounding, but meaningful use of the estimand likely does—and therefore practitioners should select L with care and interpret estimates in conjunction with sensitivity analyses. The authors’ project of assumption lean inference rests on the assumption that nuisance parameters can be estimated nonparametrically at rate n1/4. It is surprising to us that this property is widely assumed to hold for machine learning methods. The authors are in good company with this assumption, but, for example, the random forests included in the authors’ analyses can have large bias if a tuning parameter is chosen badly (Olson, 2018), and as far as we know cross-validation has not been shown to reliably choose good tuning parameters. Even if n1/4 rates are achieved asymptotically, slower rates of convergence may require large samples before asymptotic approximations are useful. This points to the importance of methods to test or help ensure that the required rates are achieved (Liu et al., 2020; Robins et al., 2008; van der Laan et al., 2021), or to perform valid inference under slower rates (Cattaneo and Jansson, 2018; Kuchibhotla et al., 2021). We re-ran the authors’ code and applied HulC, a new method for the construction of assumption—lean confidence intervals (Kuchibhotla et al., 2021).1 We found that the point estimates are indeed sensitive to choice of tuning parameters. Although HulC intervals are wider, they are valid even if approximate normality does not hold, as would be the case if the nuisance estimators converge slower than n − 1/4, as long as the estimator satisfies a weaker median unbiasedness property (Kuchibhotla et al., 2021). *(abstract via OpenAlex)*

JRSSB Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Christian Hennig's contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Christian Hennig

摘要

There is a tendency in statistics to talk about model assumptions in a misleading way. Most of us probably agree with George Box's ‘all models are wrong but some are useful’, yet there is much communication that implies that for applying methods ‘assuming’ certain models, these models have to be true. If this were so, no model-based method could ever be used! Generally model assumptions do not have to be fulfilled. A model assumption just means that certain theoretical results regarding a statistical procedure hold assuming the model. A procedure may well deliver useful results if its model assumptions do not hold. This can be addressed by investigating what happens if other models hold. The advantage of ‘assumption-lean’ methods is that they come with theory that applies under a wider range of models, so we know more, but none of this wider range of models will ultimately be ‘correct’ either, and the theory does not necessarily guarantee a good behaviour in practice. The authors shift the focus on estimands rather than models, although their theory and even their very definition still assumes a model, be it ‘big’ and non-parametric. Tukey (1997) rather put a focus on procedures, to be challenged by studying their behaviour under different models, without assuming any of them to be true. An important issue not treated by Tukey is addressed by discussing estimands, namely whether what we estimate is what we really want to know. This requires arguments other than evidence from the data. It may not be enough to choose an estimand so that its estimator fulfils nice asymptotic theory. The expected value E is central for the authors' estimands. Concentrating on one-dimensional location, standard robustness theory (Huber & Ronchetti, 2009) teaches that E is a non-robust functional. Looking at a model such as Pϵ,μ1,σ12,μ2,σ22=(1-ϵ)N(μ1,σ12)+ϵN(μ2,σ22) with very small ϵ⁠, theoretically all is fine with estimating E, but N(μ2,σ22) may model irregular outliers, and the interest maybe in estimating µ 1. But EPϵ,μ1,σ12,μ2,σ22 can be arbitrarily far away from µ1. The influence function of E is unbounded (as is (13) in the paper, due to the key role of E in the estimand), which is a problem from the robustness point of view. Paraphrasing Buja et al. (2019), having a general nonparametric theory for an estimand does not necessarily imply that the estimand behaves well over the range of models covered by the theory. Whether an estimand is well chosen requires more than being ‘assumption-lean’. *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Chaohua Dong, Jiti Gao and Oliver Linton’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Chaohua Dong, Jiti Gao, Oliver Linton

摘要

The title of this paper is ironically self-fulfilling, since there are almost no meaningful assumptions made throughout! The starting point is that we have some plausible semi-parametric model, which is a special case of a more general non-parametric model, but we wish to allow for misspecification and in particular define an estimand that is meaningful in the non-parametric model and that specialises in the plausible model to a slope coefficient. However, since the estimator that is proposed is not the semi-parametric efficient estimator of that slope coefficient under the semi-parametric model, we wonder what is the role of the model at all? The theory side of it seems to assume in Theorem 2, for example that E(Y | A, L) is consistently estimated in L2 under the full unrestricted -parametric setting. But if that is possible, then why bother with the model? The authors talk casually about machine learning methods being used to estimate E(Y | A, L), but if that is a silver bullet, then who needs the model? The model embodies some structure around A, L but the discussion is focussed away from the dimensionality of L, which is a big reason why one might want a structured model such as additivity (Linton & Nielsen, 1995) or the partial linear model (Robinson, 1988). Perhaps it would help if a full model was written down for the effect of high-dimensional L. Perhaps the point is that the parameter of interest is only defined in terms of low-dimensional conditional expectations, but this does not appear to be the case in the sense that high-dimensional smoothing is employed in (a) of p.14, which is then projected down by conditional expectation onto L, but if A is binary, then this has not reduced dimensionality at all, the dimensionality issue sits in L and what structure is assumed about its effect on Y. The authors say that the usual choice of smoothing parameter is not tuned to the estimand. This point has been made in Linton (1995) who derived an optimal bandwidth for estimation of slope parameters and Wald statistics in the partially linear model based on local polynomial estimators; the optimal rates are indeed different from those that minimise the mean squared error of the non-parametric regression involved. *(abstract via OpenAlex)*

JRSSB Machine Learning Networks & Graphs 讨论 / 评论 · 发表于 2022-07-26

Blair Bilodeau's Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Blair Bilodeau

摘要

I commend Vansteelandt and Dukes for (a) advocating that statistics move beyond well-specified assumptions and (b) providing a practical estimator to address this for generalised linear models (GLMs). Furthermore, their derivation of theoretical guarantees for this estimator is highly attractive. However, the theorem assumptions exclude many standard machine learning methods, which may violate the spirit of ‘assumption-lean’ inference guarantees. Theorems 2 and 4 assume that the three (non-parametric) conditional mean estimates have expected square loss converging at rate strictly faster than n − 1/2. Theorem 7 of Rakhlin et al. (2017) can be extended to show that if the empirical L2 entropy of the model class depends on the scale ε ∈ (0, 1) at rate ε − p for some p > 0, then the minimax expected square loss is lower bounded by n − 2/(p+2) even in the well-specified setting. That is, in order for the theory of Vansteelandt and Dukes to apply without additional assumptions, the three conditional mean models must each be strictly Donsker (p < 2), and this cannot be readily side-stepped using sample splitting as Vansteelandt and Dukes do elsewhere. The strict Donsker assumption is satisfied by models with the number of ‘parameters’ (e.g. weights of a neural network) growing strictly slower than n. However, a fixed number of parameters is exactly a parametric assumption, and in practice the number of parameters in machine learning models is often larger than n (corresponding to the interpolation regime, see Belkin et al., 2019). In the interpolation regime, such models are better understood from a non-parametric perspective; unfortunately, this means that they suffer from an exponential dependence on the dimension of the inputs. For linear models, the dimension-free entropy growth rate is ε − 2 (Mendelson & Schechtman, 2004; Zhang, 2002), and consequently, the present theory does not apply. For integer α and γ ∈ (0, 1], entropy for the class of (α, γ)-Hölder smooth functions on Rd grows at rate ε − d/(α+γ) (Wainwright, 2019, Example 5.11), requiring univariate inputs to apply the present theory in the Lipschitz (α = 0, γ = 1) setting. These examples include neural networks with linear and Lipschitz activations, as well as certain variants of random forests (Mourtada et al., 2020). In the notation of Vansteelandt and Dukes, the input dimension corresponds to the dimension of A and L jointly, and in practice L can be quite high dimensional. The authors clearly acknowledge that their convergence requirements may not be satisfied by ‘very flexible machine learning methods’. However, these requirements exclude many methods of interest, including standard neural network and random forest architectures. Ultimately, it remains an interesting open question whether Vansteelandt and Dukes' estimator enjoys convergence guarantees (even with appropriately slow rates) when used with such canonical nonparametric methods. *(abstract via OpenAlex)*

JRSSB Causal Inference Clinical Trials Machine Learning · 发表于 2022-07-26

Authors' reply to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Stijn Vansteelandt, Oliver Dukes

摘要

We thank all discussants for their interesting and thoughtful comments on our paper. In this rejoinder, we will focus on common themes amongst the commentaries and will close with a discussion of some open issues. Didelez, Shpitser, and Stensrud and Sarvet consider the framework described in our paper as being to some extent at odds with the philosophy of causal inference. There, one translates a scientific question into a causal estimand; ideally this definition should also be model-free. Didelez fears that in spite of the conveniences of our framework, it may lead researchers to bypass the first step of formulating a meaningful question. Stensrud and Sarvet argue that even in simple settings, the parameter that our ‘algorithm’ for generating a target parameter outputs may deviate from the natural causal quantity of interest. The hygienic causal inference approach is clearly ideal in the simple point-treatment example of Stensrud and Sarvet, and has led to enormous progress in statistical research. In particular, inference for the average treatment effect is generally preferable to our proposal in such settings, provided that ‘treatment’ or ‘no treatment’ are feasible options for all. However, by being somewhat divorced from the specific complexities of the considered data, those hygienic principles can rarely be strictly adhered to; this is especially so as more complex analyses are needed, a point made excellently clear by Daniel. Analyses that were intended to be hygienic, then turn somewhat into a black art, as Shpitser would call it. This is commonly seen in popular marginal structural model analyses for the effect of a time-varying treatment. Here, adjustment for baseline covariates is common, but hardly ever motivated by the strive for a scientifically relevant estimand. It is merely a statistical attempt to trade bias for variance by avoiding the need for inverse weighting to eliminate confounding induced by baseline covariates, which also motivated our work. The lack of guiding principles regarding the choice of baseline covariate set, which Didelez rightly judges to be ad hoc in our proposal, is as much ad hoc in this causal modelling context. Moreover, for similar reasons as explained in Section 2 of our paper, misspecification of the baseline covariate effects in the marginal structural model turns the intended treatment effect estimand into a generally poorly understood functional of the observed data law, which may no longer even summarise that effect. What remains may well be a pale reflection of the hygienic analysis that was intended. We agree with Didelez that problems due to highly variable inverse probability weights elucidate that no useful statements can be made about the considered estimand due to it being too ambitious for the data at hand. However, the sad truth is that these problems often end up being hidden by heuristic truncation of inverse probability weights, which has become the default in software packages. In contrast, our proposal, which could analogously be developed for marginal structural models, does not suffer these problems to the same extent; it is not hiding them, as Didelez seems to worry. This is achieved by targeting an estimand that is not too ambitious for the data at hand. While the weighted average of baseline-covariate-specific treatment effects that we target may appear less appealing, at least it is much better understood than the above marginal structural model estimand when the model is misspecified or inverse probability weights are being truncated. We agree with Didelez that causal analyses should ideally be handled on a case-by-case basis. The difficulties experienced in the above example may in particular be remedied by targeting the effect of specific dynamic interventions designed to be feasible for all. However, translating a question into a causal estimand is often a highly subtle exercise. In practice, even in the causal inference literature, researchers are therefore commonly drawn to estimands that have been well studied, even when their relevance for the scientific question is dubious (e.g. what if the same BMI, or the same level of pack-years of smoking applied uniformly in the population)? It is therefore not uncommon to see exposures being dichotomised/categorised for mathematical convenience, leading to estimands that are deceptively simple, but still remote from the real world (e.g. what if all people in the study population were obese?). The difficulty of constructing an estimand is further compounded by the fact that some study participants may well be ineligible for the considered interventions, or that the considered exposure cannot be well linked to a specific intervention. The latter is for instance the case in studies on the effect of obesity. Such studies generally have a causal pursuit, but often of a more exploratory nature; attempting to infer the effect of specific interventions on body weight may then go well beyond what the data allow to infer as well as beyond the researchers' initial aim. Our focus on weighted averages of stratum-specific conditional association measures is therefore purposefully less ambitious. The considered weights downweigh individuals for whom few exposure values are plausible. Such individuals would also be less likely recruited if an experiment were conducted. We therefore find Stensrud and Sarvet's example misleading in that there is nothing wrong in finding a treatment effect different from zero when there is treatment effect heterogeneity. In such case, any scalar summary is deficient. Whether the average treatment effect (ATE) is the most relevant target is then context-dependent. It is tempting to believe that conclusions should be drawn for the entire study population and that the marginal causal effect is most relevant (see also Ding), but in practice—even in clinical trials—we often work with convenience samples. Without careful restriction of the study population (Hernán & Robins, 2016), the ATE may well end up focussing on a less relevant population than the retargeted variance-weighted population on which our estimands focus. As such, the considered estimands also address and overcome the formidable task of how to restrict the study population (e.g. consider how difficult it would be to identify individuals in whom obesity could be a plausible exposure status). Given the difficulty of choosing a proper estimand, we believe that the ideal causal analysis is often not within reach of the many data analysts who have no expert on causal inference within their research network. To connect with applied practice, it is therefore important to provide general purpose strategies that move well beyond the simple binary point-treatment example of Stensrud and Sarvet, while being sufficiently safe to use without necessitating ‘black art’ remedial measures, such as weight truncation. We believe that our framework offers this. It is partly driven by practical considerations, which Shpitser understood to be against the spirit of our proposal, but whose relevance on the contrary motivated this work. It is a pity that the commentaries did not attempt to demonstrate how an assumption-lean ‘algorithm for causal inference’, as endorsed by Stensrud and Sarvet, would function in real applied settings that involve more complex queries (e.g. with continuous exposures). The above criteria, along with the criteria in the paper, leave surprisingly few choices of weights; in fact, we found the construction of an interaction estimand which satisfied all of these criteria to be a non-trivial task. In our proposal, there may however be many ways to define β(L) when A is not dichotomous. In this paper, we have chosen to work with linear projections as this is visually attractive and drastically simplifies the resulting inference. In future work, we will also consider defining β(L) as the solution to the population maximum likelihood score equation restricted to the stratum L. Regarding the first criterion in the paper, Didelez questioned the relevance of choosing an estimand which reduces to a regression coefficient when the model restriction (4) holds. We do not entirely agree. First, there is an abundance of causal queries aimed at developing etiological insight without the immediate ambition of doing a specific intervention. In such settings, it is much easier to reason about the causal data-generating mechanism, than about what estimands might be relevant for the data at hand. This partly explains the popularity of causal diagrams, which enable more intuitive reasoning than that based on counterfactuals. It also explains the popularity of regression-based methods, which continue to dominate applied practice. By connecting to regression models, we believe that we may often connect better to researchers' a priori understanding of the causal data-generating mechanism, while merely inferring specific features of it. Though the postulated model could be misspecified, our estimands retain close connections to (and sometimes equal) average derivative effects (Hines et al., 2021), which—by virtue of focussing on the effect of a small change in in everyone's observed exposure—tend to be quite ‘safe’ for general use. An alternative would be to focus on shift interventions that express the effect of increasing the exposure uniformly by, say, 1 unit in the population. The greater appeal of the resulting effects is somewhat deceptive, however, as shift interventions are rarely planned in practice. This then raises the question what would happen when increasing the exposure with 0.5 units, 2 units, … Flexibly answering these questions calls for some form of modelling, which our framework (when adapted to shift interventions) provides. Second, a key strength of our proposal is that it enables the investigator to work on the scale of choice. In response to Ding's concern, one may therefore choose to model risk differences or relative risks even for a dichotomous outcome. We agree that the interpretation of model coefficients (causal or otherwise) in more general non-linear models is not always obvious. Nevertheless, if the generalised partially linear model (4) holds, then a given choice of β enables one to work out how specific means or risks E(Y|A = 0, L = l) in the unexposed would translate into the corresponding means or risks E(Y|A = a, L = l) at other exposure levels a. When the model restriction (4) fails, the resulting point estimate may still be useful in terms of giving a rough impression of the strength of association. For the second criterion, Phillips and van der Laan questioned the importance of choosing estimands for which non-parametric inference does not require estimation of a conditional density. They state that machine learning methods are ‘well-adapted’ for conditional density estimation. Although some proposals have certainly been made, including those by the authors, we are concerned that such estimators may still suffer from unstable performance in finite samples. In fact, unstable performance is often expected with a continuous exposure, even if its conditional density is a priori known, as a result of influential weights for individuals in the tail of the density. Our intention was to develop procedures that could be used safely by non-experts. For the third criterion, Buja et al. argue that averaging slopes over L-specific strata is ‘incorrect’. We disagree. Summarising the different slopes obtained for the L-specific strata in terms of a weighted average is perfectly well aligned with the standard notion of summarising in statistics. Overall, we agree that other criteria may be worth considering; part of the intention was to stimulate discussion on how to choose an estimand. for β‾ ≡ E{w(L)β(L)}. If one is willing to settle for a scalar summary, then it is still questionable whether the average treatment effect best corresponds to the scientific question of interest—at least if the goal is generalisability. Stensrud and Sarvet consider how the overlap-weighted effect changes by changing the conditional distribution of the exposure (given L), but fix the distribution of L. This is reasonable, given the emphasis in causal inference on a well-defined study population. Nevertheless, varying P(L = l) in their example could also change the direction of the average treatment effect. In cases where the treatment-assignment policy is similar between populations but the covariate distribution changes, as is also reasonable, it is possible that the overlap-weighted treatment effect is better transportable than the average treatment effect. Phillips and van der Laan also highlight that the numerator of the main effect estimand (5) averages positive and negative contributions E(Y|A, L) − E(Y|L), such that the estimand may equal zero in the presence of a strong individual-level treatment effect, and tests of the null hypothesis can suffer from low power relative to tests of other ‘projection-type’ parameters. The example they provide is interesting, but dependent on a lucky choice of reference value (0 in Phillips and van der Laan, and x0 in Chambaz et al. (2012)). An unlucky choice may likewise make their estimand zero in the presence of a strong individual-level treatment effect. More generally, the connection of our results to the optimality results in Crump et al. that better power can be expected when the model is In of the that the model is we will non-parametric modelling in the As a Phillips and van der Laan also the of the estimand on the conditional distribution of the However, shift which have been developed in part by those & van der Laan, also this estimands consider interventions that the conditional distribution of the we effects over a retargeted population in terms of the conditional distribution of the that even when the model restriction (4) holds, the estimand (5) will not to a marginal causal effect. It is that the latter parameter is most relevant for Our intention was not to into the about which is most but we do believe that marginal and conditional causal effects have and that are important to For model restriction the conditional effect may better to different populations it is to in the distribution of L. A of for conditional treatment effects is that they on modelling in a The of is given that these effects are and therefore the construction of non-parametric is not generally our proposal conditional treatment non-parametric inference and that our estimands could also be as specific projections for example the conditional association between exposure and that by the The of is highly relevant and has some in the on non-parametric inference. In this literature, or no is being to the of the resulting estimand. This is especially as is commonly the entire data-generating model is the In that case, misspecification in of the model may all model as we in Section 2 of the paper. This is we have chosen to merely the conditional association between exposure and a analysis for considered The proposal by and a for more general estimands along these It will be of to how specific on the terms in their translate into estimands with specific to and we would to that our considered estimand for treatment effect by a weighted average of conditional treatment effects them, we have chosen not to work with averages as these do not to continuous exposures and inference for such effects inverse model and and question whether it is useful to target an estimand for which the interpretation is but which may be misleading about the association of interest. and an example of when the association between and A is our estimand merely the linear association between and A and so may poorly summarise the and and therefore a which the results from A key of our proposal is that it the need for such analysis with to models for the between and the covariates L. However, we are the that a linear association between and exposure A the scale of a may sometimes a It is for that reason that the discussion of our paper how the proposal may be to estimate that association the estimation of in of their when it to inference and Our focus on remains of even more so as linear are often for example they sometimes express how much the average would change if observed exposure were (Hines et al., analyses are but the truth is that researchers will often to results for a In the example of and one may use the data to a in a regression of on but a the coefficients in the model or the model which for the in the step is then for relevant et al., Although of may be in settings, as by as as we are they cannot be to for the many model that in data We the connection that and to the which may parameter estimation. et al. and that the that parameter estimators at a than in our 2 and may out many machine learning are the same in the maximum likelihood estimation and machine learning et al., van der Laan & these may be in (see et al. for and for and and et al. for we that these results may not how machine learning methods are in practice. further in statistical learning may our understanding of the of different in for applied researchers in the of our it may be to the of the often used in this to In the of we some practical a should be used of a Although the in the paper often on a this was for convenience (given the of different to in the data analysis we were the in van Laan and van der et al. and van der Laan et al. that the performance of the should be as as that of the and we a of including simple methods and regression greater (e.g. or but may have a of and better finite performance when the one may even to to these methods, which then standard analyses by the we are by performance in but that remains In settings, it may be that even the with the best still than We therefore agree with et al. that an ideal analysis should inference based on with for example tests of whether bias standard et al., or use alternative that are et al., This is an of and further will no the proposal made particular are methods that are and can be applied by non-experts. that when the effects of are by whose to the model the approach is to the model than inference on a quantity whose interpretation is but relevant when the model is We that the considered parameter is First, it is a weighted average of stratum-specific association focus on to poorly understood especially in a (see the for and how would be handled in our We have chosen to one scalar parameter at a so that a on one parameter to a choice of does not the projections on other parameters. For when the in the main effect β of a analysis is from when an interaction between A and some covariate is Moreover, if the in the β then then the obtained in the we would the function of β and work with it. This of that for instance our for do not the main effect of A to be This with If are then can still be developed based on the distribution of the for the different considered We to allow for misspecification of the but that this is not The reason is that the of the function would require of the population of the machine learning of the score of the observed data Such would be difficult to as they on the features of the considered machine learning an that we have aimed to We find alternative of the relative risk model attractive to the standard However, in our proposal one is to choose any for the E(Y|A, L) and that may may the on the parameter may a model for E(Y|A, L) and still target a relative for This is we are concerned that the resulting and interpretation for the developed for the partially linear model in et al. are to of model restriction This is especially so if the odds model is for example variable For that it may be preferable to non-parametric inference for the probability of as for the odds when model restriction (4) fails, the resulting odds by no longer to be if = = 0, L) or = = 0, L) is This with our that the of estimators obtained the generalised partially linear model may of the This the for example by et al. where the regression may be to at a for example or if the score can be at a The of methods remains useful in our as we better how to parameter estimators in this that target estimators of the parameter of with low et al., & & We agree with and that our data analysis the of the This was so on purpose we to the to generalised linear the to data models is important and is being out along the same principles that we More generally, we agree with and and that to more general estimands (e.g. remains for the important of influential that the of our estimators to such values is of the We are to may the analysis Our is that model are often and it is generally even when this is to how the resulting inference may have been In that we would the as misleading even if all were and the were they would not the truth at the We agree that by expert cannot be in a real data especially as causal are Such are not not based on the data being and were therefore not in the of our paper. we have our methods they do require relative to the of the However, even if methods were we believe that our proposal may still standard by inference when the model holds. the paper and many of the on the of adjustment for In many data analyses (causal or otherwise) are to some form of & for example data, In a modelling framework, a we can this bias in terms of how an estimand is and how inference is When the statistical model is estimators target estimands that on the estimands may be with but may be difficult to and between In this work, we have chosen to target estimands that on the exposure mechanism, so that they to However, in choosing estimands more generally, should we as those that are to even if they too much on should we that are less The is not in data to more complex for example if there is in L. Here, non-parametric inference a could be complex and a may be more plausible et al., such an an estimand that on the conditional variance of the exposure given covariates in the the of this will often be If one that target should be of a as much of the causal inference has then there is much for estimands and practical in a choice. *(abstract via OpenAlex)*

JRSSB Clinical Trials Nonparametric & Semiparametric · 发表于 2022-07-26

Assumption-lean Inference for Generalised Linear Model Parameters

Stijn Vansteelandt, Oliver Dukes

摘要

Inference for the parameters indexing generalised linear models is routinely based on the assumption that the model is correct and a priori specified. This is unsatisfactory because the chosen model is usually the result of a data-adaptive model selection process, which may induce excess uncertainty that is not usually acknowledged. Moreover, the assumptions encoded in the chosen model rarely represent some a priori known, ground truth, making standard inferences prone to bias, but also failing to give a pure reflection of the information that is contained in the data. Inspired by developments on assumption-free inference for so-called projection parameters, we here propose novel nonparametric definitions of main effect estimands and effect modification estimands. These reduce to standard main effect and effect modification parameters in generalised linear models when these models are correctly specified, but have the advantage that they continue to capture respectively the (conditional) association between two variables, or the degree to which two variables interact in their association with outcome, even when these models are misspecified. We achieve an assumption-lean inference for these estimands on the basis of their efficient influence function under the nonparametric model while invoking flexible data-adaptive (e.g. machine learning) procedures.

JRSSB Causal Inference Longitudinal Data Clinical Trials 讨论 / 评论 · 发表于 2022-07-26

Anna Choi and Weng Kee Wong’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Anna Choi, Weng Kee Wong

摘要

We congratulate the authors for major advances in the methodology of estimands, in particular assumption-lean reference for the estimands in possibly misspecified generalised linear models. Whereas the simulation studies in Section 6 indeed support the ‘assumption-lean’ claim and the authors’ software in Practice of Epidemiology187 (pp. 1079–1084, 2018) facilitates such inference, the empirical example in Section 7 seems problematic to illustrate applications of the methodology. The example focuses on ‘data obtained from birth certificates from 2500 children born in King County, Washington in 2001’ in the First Steps program that was ‘set up in 1989’ in King County. These are longitudinal data, and the program is still ongoing. The methodology for longitudinal studies developed by Laird and Ware (1982) and applied to the Six Cities Study (Ryan, 2015, section 3) using generalised linear mixed models seems more appropriate for the re-analysis of the longitudinal data from the First Steps program of King County. Although Laird and Ware's methodology does not involve causal inference, Laird mentions how such inference can be added to the analysis in Ryan (2015, section 9 on statistical genetics). Chen et al. (2018, sections 3.6.3, 4.5, 5.4, 6.3, 6.4, 6.5, 7.4, 7.5) describe how causal inference/conclusions can be derived via multi-criteria statistical decision theory, exposure-adjusted incidence rates and multiple testing with familywise error or false discovery rate control for clinical trials data, control for confounding using inverse probability weighting, propensity scores, instrumental variables and research designs for unmeasured confounders, structural causal models and symbolic causal calculus for post-marketing safety data, empirical Bayes and Bayesian approaches to signal detection from adverse event databases. In addition, Sections 2.1.3 and 2.1.4 of that book describe recent advances in statistical learning methods which involve nonlinear basis functions such as neural networks and classification/regression trees and in validation of the prediction/regression model, in contrast to the linear (or partial linear) basis functions used in the present paper and complementary to the sample splitting asymptotic theory in its Theorem 2. *(abstract via OpenAlex)*

JRSSB Other 讨论 / 评论 · 发表于 2022-07-26

Andreas Buja, Richard A. Berk, Arun K. Kuchibhotla, Linda Zhao and Ed George’s Contribution to the Discussion of ‘Assumption-Lean Inference for Generalised Linear Model Parameters’ by Vansteelandt and Dukes

Andreas Buja, Richard A. Berk, Arun K. Kuchibhotla, Linda Zhao, Ed George

摘要

We thank the authors for their stimulating article. At the highest level, it agrees with our thinking as expressed in the second of two articles in Statistical Science (Buja et al., 2019). The guiding principle is to focus on quantities of interest that are meaningful statistical functionals, beyond the statistical models that motivate them, even when those models are fully misspecified. In some form, this focus has existed for decades in the semi-parametric literature. We suggest below a reinterpretation of the authors’ approach and a modification of their third principle. 1.Interpretation: The authors’ formulation can be reinterpreted as follows: The authors’ proposal (5) is the solution to the minimization problem (1), which shows that it is a projection functional based on a non-standard loss function. When the model is 1st order correct, gE[Y|A,L]=βA+ωL⁠, it holds that βw is the same for all weight functions w(L): βw ≡ β. When the model is 1st order misspecified, there exist weight functions w1(L) and w2(L) such that βw1≠βw2⁠, which is equivalent to ‘effect heterogeneity’, i.e., β | L is not the same (a.s.) for all confounder strata L. This lends itself to a ‘well-specification’ diagnostic for regression functionals, as elaborated in Buja et al. (2019). It specializes here to a diagnostic for effect heterogeneity. *(abstract via OpenAlex)*

SMMR Clinical Trials · 发表于 2022-07-25

The benefits of covariate adjustment for adaptive multi-arm designs

Kim May Lee, David S. Robertson, Thomas Jaki, Richard Emsley

摘要

Covariate adjustment via a regression approach is known to increase the precision of statistical inference when fixed trial designs are employed in randomized controlled studies. When an adaptive multi-arm design is employed with the ability to select treatments, it is unclear how covariate adjustment affects various aspects of the study. Consider the design framework that relies on pre-specified treatment selection rule(s) and a combination test approach for hypothesis testing. It is our primary goal to evaluate the impact of covariate adjustment on adaptive multi-arm designs with treatment selection. Our secondary goal is to show how the Uniformly Minimum Variance Conditionally Unbiased Estimator can be extended to account for covariate adjustment analytically. We find that adjustment with different sets of covariates can lead to different treatment selection outcomes and hence probabilities of rejecting hypotheses. Nevertheless, we do not see any negative impact on the control of the familywise error rate when covariates are included in the analysis model. When adjusting for covariates that are moderately or highly correlated with the outcome, we see various benefits to the analysis of the design. Conversely, there is negligible impact when including covariates that are uncorrelated with the outcome. Overall, pre-specification of covariate adjustment is recommended for the analysis of adaptive multi-arm design with treatment selection. Having the statistical analysis plan in place prior to the interim and final analyses is crucial, especially when a non-collapsible measure of treatment effect is considered in the trial.

Biometrika Econometrics & Economics · 发表于 2022-07-25

Sample-constrained partial identification with application to selection bias

Matthew J Tudball, Rachael A Hughes, Kate Tilling, Jack Bowden, Qingyuan Zhao

摘要

Summary: Many partial identification problems can be characterized by the optimal value of a function over a set where both the function and set need to be estimated by empirical data. Despite some progress for convex problems, statistical inference in this general setting remains to be developed. To address this, we derive an asymptotically valid confidence interval for the optimal value through an appropriate relaxation of the estimated set. We then apply this general result to the problem of selection bias in population-based cohort studies. We show that existing sensitivity analyses, which are often conservative and difficult to implement, can be formulated in our framework and made significantly more informative via auxiliary information on the population. We conduct a simulation study to evaluate the finite sample performance of our inference procedure, and conclude with a substantive motivating example on the causal effect of education on income in the highly selected UK Biobank cohort. We demonstrate that our method can produce informative bounds using plausible population-level auxiliary constraints. We implement this method in the $\texttt{R}$ package $\texttt{selectioninterval}$.

SiM Bayesian Methods · 发表于 2022-07-25

Capturing the pool dilution effect in group testing regression: A Bayesian approach

Stella Self, Christopher McMahan, Stefani Mokalled

摘要

Group (pooled) testing is becoming a popular strategy for screening large populations for infectious diseases. This popularity is owed to the cost savings that can be realized through implementing group testing methods. These methods involve physically combining biomaterial (eg, saliva, blood, urine) collected on individuals into pooled specimens which are tested for an infection of interest. Through testing these pooled specimens, group testing methods reduce the cost of diagnosing all individuals under study by reducing the number of tests performed. Even though group testing offers substantial cost reductions, some practitioners are hesitant to adopt group testing methods due to the so‐calleddilution effect. The dilution effect describes the phenomenon in which biomaterial from negative individuals dilute the contributions from positive individuals to such a degree that a pool is incorrectly classified. Ignoring the dilution effect can reduce classification accuracy and lead to bias in parameter estimates and inaccurate inference. To circumvent these issues, we propose a Bayesian regression methodology which directly acknowledges the dilution effect while accommodating data that arises from any group testing protocol. As a part of our estimation strategy, we are able to identify pool specific optimal classification thresholds which are aimed at maximizing the classification accuracy of the group testing protocol being implemented. These two features working in concert effectively alleviate the primary concerns raised by practitioners regarding group testing. The performance of our methodology is illustrated via an extensive simulation study and by being applied to Hepatitis B data collected on Irish prisoners.

SiM Survival Analysis Bayesian Methods · 发表于 2022-07-24

Bayesian criterion‐based assessments of recurrent event models with applications to commercial truck driver behavior studies

Yiming Zhang, Ming‐Hui Chen, Feng Guo

摘要

Summary: Multitype recurrent events are commonly observed in transportation studies, since commercial truck drivers may encounter different types of safety critical events (SCEs) and take different lengths of on‐duty breaks in a driving shift. Bayesian nonhomogeneous Poisson process models are a flexible approach to jointly model the intensity functions of the multitype recurrent events. For evaluating and comparing these models, the deviance information criterion (DIC) and the logarithm of the pseudo‐marginal likelihood (LPML) are studied and Monte Carlo methods are developed for computing these model assessment measures. We also propose a set of new concordance indices (C‐indices) to evaluate various discrimination abilities of a Bayesian multitype recurrent event model. Specifically, the within‐event C‐index quantifies adequacy of a given model in fitting the recurrent event data for each type, the between‐event C‐index provides an assessment of the model fit between two types of recurrent events, and the overall C‐index measures the model's discrimination ability among multiple types of recurrent events simultaneously. Moreover, we jointly model the incidence of SCEs and on‐duty breaks with driving behaviors using a Bayesian Poisson process model with time‐varying coefficients and time‐dependent covariates. An in‐depth analysis of a real dataset from the commercial truck driver naturalistic driving study is carried out to demonstrate the usefulness and applicability of the proposed methodology.

SiM Spatial & Time Series Bayesian Methods Genetics & Omics · 发表于 2022-07-24

A Bayesian modified Ising model for identifying spatially variable genes from spatial transcriptomics data

Xi Jiang, Guanghua Xiao, Qiwei Li

摘要

A recent technology breakthrough in spatial molecular profiling (SMP) has enabled the comprehensive molecular characterizations of single cells while preserving spatial information. It provides new opportunities to delineate how cells from different origins form tissues with distinctive structures and functions. One immediate question in SMP data analysis is to identify genes whose expressions exhibit spatially correlated patterns, called spatially variable (SV) genes. Most current methods to identify SV genes are built upon the geostatistical model with Gaussian process to capture the spatial patterns. However, the Gaussian process models rely on ad hoc kernels that could limit the models' ability to identify complex spatial patterns. In order to overcome this challenge and capture more types of spatial patterns, we introduce a Bayesian approach to identify SV genes via a modified Ising model. The key idea is to use the energy interaction parameter of the Ising model to characterize spatial expression patterns. We use auxiliary variable Markov chain Monte Carlo algorithms to sample from the posterior distribution with an intractable normalizing constant in the model. Simulation studies using both simulated and synthetic data showed that the energy‐based modeling approach led to higher accuracy in detecting SV genes than those kernel‐based methods. When applied to two real spatial transcriptomics (ST) datasets, the proposed method discovered novel spatial patterns that shed light on the biological mechanisms. In summary, the proposed method presents a new perspective for analyzing ST data.

JASA Testing & Uncertainty · 发表于 2022-07-22

Hypothesis Tests for Structured Rank Correlation Matrices

Samuel Perreault, Johanna G. Nešlehová, Thierry Duchesne

摘要

Joint modeling of a large number of variables often requires dimension reduction strategies that lead to structural assumptions of the underlying correlation matrix, such as equal pair-wise correlations within subsets of variables. The underlying correlation matrix is thus of interest for both model specification and model validation. In this article, we develop tests of the hypothesis that the entries of the Kendall rank correlation matrix are linear combinations of a smaller number of parameters. The asymptotic behavior of the proposed test statistics is investigated both when the dimension is fixed and when it grows with the sample size. We pay special attention to the restricted hypothesis of partial exchangeability, which contains full exchangeability as a special case. We show that under partial exchangeability, the test statistics and their large-sample distributions simplify, which leads to computational advantages and better performance of the tests. We propose various scalable numerical strategies for implementation of the proposed procedures, investigate their behavior through simulations and power calculations under local alternatives, and demonstrate their use on a real dataset of mean sea levels at various geographical locations. Supplementary materials for this article are available online. *(abstract via OpenAlex)*

Biometrics Latent Variables & Mixtures Testing & Uncertainty · 发表于 2022-07-21

Translocation Detection from Hi-C Data via Scan Statistics

Anthony Cheng, Disheng Mao, Yuping Zhang, Joseph Glaz, Zhengqing Ouyang

摘要

Recent Hi-C technology enables more comprehensive chromosomal conformation research, including the detection of structural variations, especially translocations. In this paper, we formulate the interchromosomal translocation detection as a problem of scan clustering in a spatial point process. We then develop TranScan, a new translocation detection method through scan statistics with the control of false discovery. The simulation shows that TranScan is more powerful than an existing sophisticated scan clustering method, especially under strong signal situations. Evaluation of TranScan against current translocation detection methods on realistic breakpoint simulations generated from real data suggests better discriminative power under the receiver-operating characteristic curve. Power analysis also highlights TranScan's consistent outperformance when sequencing depth and heterozygosity rate is varied. Comparatively, Type I error rate is lowest when evaluated using a karyotypically normal cell line. Both the simulation and real data analysis indicate that TranScan has great potentials in interchromosomal translocation detection using Hi-C data.

Biometrics Nonparametric & Semiparametric Missing Data & Measurement Error · 发表于 2022-07-21

Quantile Regression for Nonignorable Missing Data with Its Application of Analyzing Electronic Medical Records

Aiai Yu, Yujie Zhong, Xingdong Feng, Ying Wei

摘要

Over the past decade, there has been growing enthusiasm for using electronic medical records (EMRs) for biomedical research. Quantile regression estimates distributional associations, providing unique insights into the intricacies and heterogeneity of the EMR data. However, the widespread nonignorable missing observations in EMR often obscure the true associations and challenge its potential for robust biomedical discoveries. We propose a novel method to estimate the covariate effects in the presence of nonignorable missing responses under quantile regression. This method imposes no parametric specifications on response distributions, which subtly uses implicit distributions induced by the corresponding quantile regression models. We show that the proposed estimator is consistent and asymptotically normal. We also provide an efficient algorithm to obtain the proposed estimate and a randomly weighted bootstrap approach for statistical inferences. Numerical studies, including an empirical analysis of real-world EMR data, are used to assess the proposed method's finite-sample performance compared to existing literature.

Biometrika Testing & Uncertainty · 发表于 2022-07-21

Localized conformal prediction: a generalized inference framework for conformal prediction

Leying Guan

摘要

Summary: We propose a new inference framework called localized conformal prediction. It generalizes the framework of conformal prediction by offering a single-test-sample adaptive construction that emphasizes a local region around this test sample, and can be combined with different conformal scores. The proposed framework enjoys an assumption-free finite sample marginal coverage guarantee, and it also offers additional local coverage guarantees under suitable assumptions. We demonstrate how to change from conformal prediction to localized conformal prediction using several conformal scores, and we illustrate a potential gain via numerical examples.

JRSSB Spatial & Time Series Missing Data & Measurement Error · 发表于 2022-07-20

The Debiased Spatial Whittle Likelihood

Arthur P. Guillaumin, Adam M. Sykulski, Sofia C. Olhede, Frederik J. Simons

摘要

We provide a computationally and statistically efficient method for estimating the parameters of a stochastic covariance model observed on a regular spatial grid in any number of dimensions. Our proposed method, which we call the Debiased Spatial Whittle likelihood, makes important corrections to the well-known Whittle likelihood to account for large sources of bias caused by boundary effects and aliasing. We generalize the approach to flexibly allow for significant volumes of missing data including those with lower-dimensional substructure, and for irregular sampling boundaries. We build a theoretical framework under relatively weak assumptions which ensures consistency and asymptotic normality in numerous practical settings including missing data and non-Gaussian processes. We also extend our consistency results to multivariate processes. We provide detailed implementation guidelines which ensure the estimation procedure can be conducted in O(nlogn) operations, where n is the number of points of the encapsulating rectangular grid, thus keeping the computational scalability of Fourier and Whittle-based methods for large data sets. We validate our procedure over a range of simulated and realworld settings, and compare with state-of-the-art alternatives, demonstrating the enduring practical appeal of Fourier-based methods, provided they are corrected by the procedures developed in this paper.

SiM Clinical Trials Diagnostics & Prediction Survival Analysis · 发表于 2022-07-20

Sample size methods for evaluation of predictive biomarkers

Kevin K. Dobbin, Lisa M. McShane

摘要

Treatment selection biomarkers are those that can be useful in guiding choice of therapy. Just as new therapies require evaluation in appropriately designed clinical trials to determine their benefit, therapy selection biomarkers require evaluation in appropriately designed studies. These studies may be prospective clinical trials or retrospective studies based on specimens stored from a completed clinical trial. Ideally, patient treatment assignments should be randomized, and consideration should be given to an appropriate sample size—either for prospective planning of a new study or access to a sufficient number of stored specimens. Here, we develop a novel sample size method for estimation of a confidence interval of specified average width, for an intuitively appealing previously proposed parameter that reflects the expected benefit of using biomarker‐guided therapy relative to a standard‐of‐care therapy. The estimation approach combines Monte Carlo and regression to result in a procedure that performs well over a range of scenarios. Although derived under a specific Cox proportional hazards regression model, robustness to model violations is demonstrated by evaluation under accelerated failure time and cure models. The sample size method produces adequate or conservative sample size estimates under a range of scenarios. Computer code in R and C++, and applications for Mac and Windows are made available for implementation of the sample size estimation procedure. The method is applied to a real data setting and results discussed.

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 页