2024 年的期刊文章 · 第 12 页

共 1515 篇,按发表时间排,新的在前,每页 40 篇。这一页是 2024-07-18 到 2024-07-01。

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

SiM Survey & Sampling Spatial & Time Series Bayesian Methods · 发表于 2024-07-18

Bayesian modeling of spatial ordinal data from health surveys

Miguel Ángel Beltrán‐Sánchez, Miguel‐Angel Martinez‐Beneito, Ana Corberán‐Vallet

摘要

Health surveys allow exploring health indicators that are of great value from a public health point of view and that cannot normally be studied from regular health registries. These indicators are usually coded as ordinal variables and may depend on covariates associated with individuals. In this article, we propose a Bayesian individual‐level model for small‐area estimation of survey‐based health indicators. A categorical likelihood is used at the first level of the model hierarchy to describe the ordinal data, and spatial dependence among small areas is taken into account by using a conditional autoregressive distribution. Post‐stratification of the results of the proposed individual‐level model allows extrapolating the results to any administrative areal division, even for small areas. We apply this methodology to describe the geographical distribution of a self‐perceived health indicator from the Health Survey of the Region of Valencia (Spain) for the year 2016.

SiM Longitudinal Data Latent Variables & Mixtures · 发表于 2024-07-18

A latent variable approach to jointly modeling longitudinal and cumulative event data using a weighted two‐stage method

Madeline R. Abbott, Inbal Nahum‐Shani, Cho Y. Lam, Lindsey N. Potter, David W. Wetter, Walter H. Dempsey

摘要

Ecological momentary assessment (EMA), a data collection method commonly employed in mHealth studies, allows for repeated real‐time sampling of individuals' psychological, behavioral, and contextual states. Due to the frequent measurements, data collected using EMA are useful for understanding both the temporal dynamics in individuals' states and how these states relate to adverse health events. Motivated by data from a smoking cessation study, we propose a joint model for analyzing longitudinal EMA data to determine whether certain latent psychological states are associated with repeated cigarette use. Our method consists of a longitudinal submodel—a dynamic factor model—that models changes in the time‐varying latent states and a cumulative risk submodel—a Poisson regression model—that connects the latent states with the total number of events. In the motivating data, both the predictors—the underlying psychological states—and the event outcome—the number of cigarettes smoked—are partially unobservable; we account for this incomplete information in our proposed model and estimation method. We take a two‐stage approach to estimation that leverages existing software and uses importance sampling‐based weights to reduce potential bias. We demonstrate that these weights are effective at reducing bias in the cumulative risk submodel parameters via simulation. We apply our method to a subset of data from a smoking cessation study to assess the association between psychological state and cigarette smoking. The analysis shows that above‐average intensities of negative mood are associated with increased cigarette use.

Biometrics Causal Inference · 发表于 2024-07-16

Multiply robust estimation of marginal structural models in observational studies subject to covariate-driven observations

Janie Coulombe, Shu Yang

摘要

Electronic health records and other sources of observational data are increasingly used for drawing causal inferences. The estimation of a causal effect using these data not meant for research purposes is subject to confounding and irregularly-spaced covariate-driven observation times affecting the inference. A doubly-weighted estimator accounting for these features has previously been proposed that relies on the correct specification of two nuisance models used for the weights. In this work, we propose a novel consistent multiply robust estimator and demonstrate analytically and in comprehensive simulation studies that it is more flexible and more efficient than the only alternative estimator proposed for the same setting. It is further applied to data from the Add Health study in the United States to estimate the causal effect of therapy counseling on alcohol consumption in American adolescents.

SiM Missing Data & Measurement Error Longitudinal Data Nonparametric & Semiparametric · 发表于 2024-07-16

Extending the DeLong algorithm for comparing areas under correlated receiver operating characteristic curves with missing data

Lily Zou, Yun‐Hee Choi, Leonardo Guizzetti, Di Shu, Joshua Zou, Guangyong Zou

摘要

A nonparametric method proposed by DeLong et al in 1988 for comparing areas under correlated receiver operating characteristic curves is used widely in practice. However, the DeLong method as implemented in popular software quietly deletes individuals with any missing values, yielding potentially invalid and/or inefficient results. We simplify the DeLong algorithm using ranks and extend it to accommodate missing data by using a mixed model approach for multivariate data. Simulation results demonstrate the validity and efficiency of our procedure for data missing at random. We illustrate our proposed procedure in SAS, Stata, and R using the original DeLong data.

JRSSB Bayesian Methods Testing & Uncertainty Nonparametric & Semiparametric · 发表于 2024-07-16

Authors’ reply to the Discussion of ‘Safe testing’

Peter Grünwald, Rianne de Heide, Wouter Koolen

摘要

We thank all discussants, and especially the proposer and the seconder of the vote of thanks, for their insightful and encouraging contributions. We are particularly delighted by the diversity of philosophical positions found among the respondents (both nullistic and non-nullistic Bayesians, frequentists, likelihoodists)—apparently we struck a chord among all these groups, which was exactly what we had hoped for with this paper. We received so many comments that it is impossible for us to respond to each of them—freely paraphrasing (Diaconis & Freedman, 1986, page 86), we hope discussants whose remarks were not or not all singled out for reply will not feel insulted, on the theory that silence is consent, and no news is good news. We grouped major themes, brought up by several discussants, into Sections 2–5. This is followed by a final section with some more specific issues and questions. Our main result, Theorem 1, requires the alternative Q and the composite null to all have densities relative to the same underlying measure, and also imposes the finite KL conditionD(Q‖PW)<∞⁠: the KL divergence between Q and at least some subset of the convex hull of the null has to be finite. All extensions of Theorem 1 require analogous conditions. As noted by Vovk and Larsson, Ramdas, and Ruf, in nonparametric settings these conditions are quite strong. Relatedly, our basic growth optimality notion GRO can sometimes trivialize if the finite KL condition does not hold. Luckily though, over the course of the last year, both Theorem 1 and the GRO notion have been significantly extended. First, Lardy et al. (2023) replaced the KL divergence by the more general notion of description gain and extended GRO accordingly. Second, as Larsson, Ramdas and Ruf discuss, in Larsson et al. (2024) they managed to come up with further sweeping generalizations of the GRO e-variable, the reverse information projection and the result linking them, i.e. our Theorem 1. Their generalization of the GRO e-variable, which they call the numeraire e-variable, exists in complete generality, under no regularity conditions whatsoever. We are of course very happy that our result can be generalized in this way, thereby showing that the basic idea underlying our Theorem 1 is really the same idea as that employed in the construction of various nonparametric e-processes such as those in Waudby-Smith and Ramdas (2024) and those implicitly derived by e.g. Honda and Takemura (2010) and Agrawal et al. (2021). In our view, this essentially implies full generality of the GRO/numeraire e-variable for the case that the alternative Q is (a) a singleton which (b) does not assign positive probability to any event that is null under every distribution in the convex hull of the null. But what if (a) or (b) is violated? As to (a), as demonstrated in our Example 8, in the case of composite alternatives in combination with worst-case growth optimality (GROW and relative GROW), it can be necessary to move to a coarser filtration to obtain the optimal e-variable. Tian and Yu ask when such coarsening can lead to more (worst-case) growth or power. Our answer is that at present we only have a partial understanding of this matter, and it is an exciting area of ongoing research; see for example (Choe & Ramdas, 2024; Pérez-Ortiz et al., 2024). Violation of (b) often occurs if the outcome Y really represents an infinite sequence of data. Vovk considers, for example, the case that Y is an infinite binary sequence and the null hypothesis expresses that the sequence is exchangeable. In this case, (b) is necessarily violated (Ramdas et al., 2022), for any Q that is not in the null. Then the extended GRO/numeraire e-variable against any such Q will still exist, but it will have an expected logarithmic growth of ∞ (Larsson et al., 2024, Theorem 2.5) and this will not tell us anything about whether it is a useful choice in an e-process defined on finite subsequences of Y⁠. In this case, again, changing the filtration can come to our rescue: Vovk and collaborators have developed elegant methods for designing e-processes in a particular coarsened filtration. Interestingly, once one restricts e-processes to the chosen filtration, expected logarithmic growth again becomes a natural notion of optimality. Given the fact that different types of filtration-coarsenings (i.e. not just Vovk’s) have been successfully used in e-value design (Ramdas et al., 2023), we speculate that currently unknown yet useful filtration-coarsening are still to be discovered, and even that designing e-variables and processes by filtration change may somehow be unified with design by growth optimality and information projection. It is here that we turn to the insightful comments of Ruodu Wang, proposer of the vote of thanks. Professor Wang notes that the interpretation of the e-power (i.e. expected logarithmic growth) of individual e-variables is not clear. We agree that there is a sore point here: in our paper, we define and optimize it for individual e-variables, but it derives its meaning from the optional continuation setting, in which more and more batches of data may be added. Wang also provides some intriguing examples in which the e-power is minus infinity yet the e-variable will quite likely grow large, and cannot become zero, under the alternative. The reciprocal of this very e-variable can be used to show that the opposite can happen as well: infinite e-power yet a large probability of being very small in practice. This clearly shows that there is more to e-variable theory than just GRO and its variations! One option is to turn attention to concave functions f(E) of E different from the logarithm, i.e. those that are induced if we construct E-variables by Rényi-projection instead of reverse information projection—Lardy et al. (2023) and Larsson et al. (2024) recently showed that such projections give rise to E-variables as well, and it appears that the corresponding functions f also admit a monetary interpretation (Soklakov, 2020). Also the developments by Grünwald (2023) suggest that when loss functions may be set in a data-dependent manner, we may sometimes want to diverge from GRO. Nevertheless, we feel that the GRO criterion and the reverse information projection will remain central to the area, and we concur with Wang that the coming years ‘will bring better understanding of its advantages and limitations, after which it will remain the default notion of e-power, with properly acknowledged caveats’. Hennig writes that the GRO-criterion, namely optimizing EQ[logE]⁠, may not look so attractive to a gambler who is interested in a payout of E rather than logE⁠. As stated, we agree that GRO may not always be the criterion of choice, but this particular remark seems besides the point: in our testing-by-betting game, the pay-off is E, not logE⁠, and we clearly state in the paper why it is not at all a good idea to optimize for E directly if optional continuation is a possibility. We are similarly puzzled by the remark ‘I am not convinced that their [i.e. the author’s monetary] interpretation of e-values is clearer than the interpretation of p-values. Probabilities and consequently p-values have been given monetary interpretations as well in the literature’. We would like to turn this remark on its head and argue that the monetary interpretation of p-values on the contrary makes clear that they are problematic as a notion of evidence in testing. Such a monetary interpretation goes as follows: suppose a collection of lottery tickets {Tr:0<r<1} is for sale. Ticket Tr pays out 1/r euro if the event p≤r happens, where p is the observed p-value and 0 otherwise. Now, under the null hypothesis, the fair price of each ticket Tr is 1 euro. Now suppose you observe a p-value of p. This indicates that you have observed an event for which, if you had bought the ticket Tp in advance, you would now have gained 1/p euro. The problem is with the in advance: you didn’t know in advance that Tp is the ticket you should have bought (giving you the highest pay-off among all tickets). Hence, stating p as your evidence against the null, in betting terms, amounts to a form of cheating. We refer to Shafer (2021) who emphasizes the point and interprets p-to-e calibration as making p-values honest. Indeed, in his discussion of our paper, Shafer even writes regarding this issue ‘…the authors concede too much. Testing with e-values, they tell us, requires more data than testing with p-values. Was this concession extracted by referees? When we make p-values honest, e-values do not need more data to catch up’. We agree with this statement and hasten to say that, when we wrote ‘requires more data’, we referred to a test that outputs merely a binary accept/reject decision, without a continuous measure of evidence accompanying it. Greenland hopes we can explain by example how e-variables should be chosen in practice, and in particular asks us to consider the case of observing a k-degree-of-freedom χ2 statistic Y taking value y with p-value p. Greenland himself notes that two possible e-variables which suggest themselves for this null hypothesis perform terribly in terms of power, and challenges us to do better. We happily take up the challenge! We first note that in many practical scenarios in which such a test is done, Y will have a k-degree noncentral χ2-density pδ under the alternative, with some (usually unknown) noncentrality parameter δ. This suggests to take as our e-variable a likelihood ratio Eδ(y):=pδ(y)/p0(δ)(y)⁠, or a mixture EW(δ)(y):=∫Eδ(y)dW(δ)⁠; since the null is simple, all such Eδ and EW are e-variables. Since the family of noncentral χ2-distributions is strictly of Pólya (∞)-type (Karlin, 1956), it constitutes a monotone-likelihood ratio family and therefore, for each δ and W, the corresponding likelihood ratio is a monotone function of the p-value, facilitating comparison. Like Greenland we henceforth focus on the case with 1 degree of freedom. As we show in Section 3.1, Proposition 3 of our paper, if we want to test δ=0 (i.e. the null hypothesis of central χ2 is correct) vs. δ≥δ* for some fixed minimal relevant effect size δ* then the results of our paper can be directly applied: by the monotone likelihood ratio property, the GROW criterion prescribes to use the e-value Eδ*⁠. In Figure 1 we plot Eδ(y) as function of δ for fixed data y=3.84⁠, the observation corresponding to p-value p=0.05⁠. Now, in practice often no minimally relevant δ* is given. In this case, we may still follow the prescription of our paper and equip δ with a prior to optimize for the REGROW (relative growth) criterion of Section 4.1-4.3. Exact computation of the REGROW-optimal prior is difficult; here we merely give an indication of what happens by taking an intuitively noninformative prior which puts heavy mass near 0 and ∞⁠: the prior with such a we for y=3.84⁠, i.e. and we as as corresponding to a very but an to two the of which the if we would like to do better. The REGROW criterion has been for settings in which data in and will perform especially well in combination with optional as in Section of our paper. In the e-values are still we may want to with e-values the REGROW criterion is not the of to obtain power, when is then with monotone likelihood ratio one may use the as an e-variable by a point prior on δ which on is to among all the when if it value being for a specific it its under optional continuation for any as In our case, this prior puts all mass on and we obtain if i.e. if better than what is with the noninformative This is the that can be for e-variables of the likelihood ratio We to the use of as to obtain e-variables more in Eδ(y) as function of δ for fixed y corresponding to p=0.05⁠. seconder of the vote of thanks, with has recently & to design e-processes on with and show that, up to first this to growth to our optimal growth asks if there are practical or to growth optimality to growth a was by Tian and the answer on the it out that growth optimality can be in some example, when testing of a it can be that and GRO have the same expected logarithmic growth e-power, in up to terms, yet if is not than the both in and in e-power at can be This is by results from and especially the discussion by et al., and also by our in ongoing we make a of both methods for family which in the same We hasten to that we really like idea to use as an alternative to e.g. of or the of in the occurs of GRO by and seems to the use of also notes that methods (i.e. do not they can be to of When is somehow this will lead to with than our which are to for optional under all under the given filtration. course we concur with this the of is one of just and we very look to see intriguing that can be to the the we would like to that a set and then methods with optional a but not with optional continuation from one to the which was our main in this e-values this one of us recently a rather different for testing with in to p-value they can be used to obtain a of of loss that are only after the data has been We that this is by of We agree with and and further we call nullistic that the central of our paper is relevant for the and we and interested to have a look at Pérez-Ortiz et al. Section the prior is a choice in but it has been for not being defined in This problem if we consider rather than further that it to how to construct the least prior on the null hypothesis on a general especially from a and whether it into a point Our reply is that the of GRO that is to is the REGROW criterion of Section if the null is simple, it amounts to a of prior for the alternative. Nevertheless, we concede that it is not clear to us how to construct it in general in a In to it will not lead to point mass In fact one may of our as who are to use REGROW on alternative and corresponding on the null, and point optimizing to they test the same null hypothesis on different they may their e-values even these are on different underlying a and that, if the alternative is the family with a the null is a family to between the corresponding GRO e-variable would use a prior in the that all mass on the It is not clear whether this would be the in not exactly the settings et al., that the mass would be out over many or even form a in the null the prior over the alternative had a then would be on a In fact not just e-variables for is a of ongoing with results in Grünwald et al. a collection of e-variables with some asks if we still an e-variable (i.e. its under the null should not be than to we would use the e-variable where is by the e.g. the value which on the given data Y⁠. We to do since we and as one of the attractive of our the one may take the idea as a and turn into a e-variable. necessarily of this would be i.e. this would lead to a of the GROW criterion in which, in information terms, is replaced by a is by and We are currently this idea seems an in where GROW optimality the use of point In these the point of the point prior a minimal relevant effect size from the This is often happens in to show the at the same the result of a null hypothesis test null corresponding to and a The should (a) be i.e. one should have the that its to 0 as size the same (b) one the test and the to be with each i.e. the test should 0 of the observation implies that for GROW optimal one cannot have (a) and (b) at the same We agree that this this is an and we not this to be the case when we developed the GROW asks whether some of GROW that the optimal a prior with full which this The answer the REGROW criterion we on in the paper has this with it to that agree with But the of REGROW is that it does not us to make an prior of the effect size under the alternative. We are currently the of and (2010) which, it may give us the of both that the e-variable are not in the of e.g. and e-variables on et al., and that e-variables call for a we agree with this we do not the is as as of may but we may not The is that the GRO e-variable relative to null hypothesis a e-variable relative to any null hypothesis that is a subset of the null hypothesis In practice we often with one null for which we the GRO e-variable. We can then the same e-variable when we consider of and the data us to on this e-variable, it will also us to any only the once we the GRO criterion also to design e-variables for two who both the same alternative, but null by and with may but a who first and may is to to the likelihood and that e-values the likelihood in of the that any is with such We that the of this just like on quite specific and of a to us, and may be a In any case, we are by to such as and the likelihood and we to these in and what they call nullistic which of testing We would like to that we are in fact very to with or e-values as a function of the as they by the in the by one of But to in such one first has to the testing in as we do in this if one does not to use it for testing. Shafer that the value of our will become even clearer once no their to into measure and we feel that optional as Shafer recently continuation is really given a than and as in fact the example Shafer in his discussion clearly We have to admit that, such a would be it may also be quite to in by who have been to since their Wang, proposer of the vote of thanks, notes that theory of is for this [i.e. to and to this theory in the theory of e-values has successfully and to the of and provides this We and we hope and that this of between and will turn out to become even more in the coming In the same extensions of the theory to more general agree this is a and have about whether can make the more As they may still over as the can be to in practice. example, in will require and may as such the of the We the main point of our paper is optional not we want to several with the same null, where the to an may be by The fact that you can often see Section of our do a is a but not directly relevant to the main in which we We do agree that in may not be such a good it suggests in every and this is of course impossible to This point was also by we should have given the paper a different We agree with all these discussants that every should be used with example, if may issue by we may want to against it by any given e-variable E which is still an e-variable. In betting terms, this to a of in can then in fact be from the data by e.g. of by it with a Relatedly, more about the of agree that this is a as do Ramdas et al. that Eδ is an e-variable for data Y relative to null hypothesis of effect size δ. suggests to use as an of effect size and asks how it under optional This is an intriguing and have not at it but we agree it further and Greenland note that the e-value is used with a different meaning in different of We agree that this is quite had no idea at the the was but by now use of the e-value had become so that it seems impossible to we do not see an and ask whether e-values can be for One option is to to nonparametric e-values Ramdas et al., for an which the of to some though, one really to use some of a quite specific that may still be the a in such is to equip the likelihood with a or to use a defined on a set of rather than probability with the likelihood replaced by the loss of the on the observed that one of us has for in the It would be to see if our results can somehow be to this setting, since the of and e-variables quite some both being to Grünwald & 2020). in a first in this has recently been by and the by the with a of the and that it on We feel this is an exciting and call for a corresponding notion of growth optimality to better just like GRO provides better growth and than in the and note that under likelihood ratio the developments of et al. suggests that it may be to the by making the to on an e-value with a of to the probability of being than thereby e-value and methods even We In one of us has been by the very same for a and the paper by et al. was a major of for the present paper. partial answer is given by Grünwald Section rather than the it is that under likelihood a e-value can be which if a e-value would it is also clear that such an e-variable cannot be into an and provides only for optional continuation and not for optional Tian and Yu ask several which are in example, they ask about practical for our can be found in e.g. et al. (2024) and et al. with is possible et al., and by now there is a of on nonparametric e-processes (Ramdas et al., Waudby-Smith & Ramdas, 2024). Their construction can often be of as a or generalization of the GRO criterion of our paper, as by Larsson et al. also ask about of an is given by Ramdas et al. as to whether is the or the in our It to power. also asks whether there are there any examples where or is more we can what are the to The answer is that into as as we with & whether is or not on the We may be in in which they both also asks how the e-value with reply is that, in to e-values, the p-value has not been to continuation of more than two reply would really require an for the very *(abstract via OpenAlex)*

Biometrics Genetics & Omics · 发表于 2024-07-15

PathGPS: discover shared genetic architecture using GWAS summary data

Zijun Gao, Qingyuan Zhao, Trevor Hastie

摘要

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

SiM Networks & Graphs High-Dimensional · 发表于 2024-07-15

Dynamic undirected graphical models for time‐varying clinical symptom and neuroimaging networks

Erin I. McDonnell, Shanghong Xie, Karen Marder, Fanyu Cui, Yuanjia Wang

摘要

In this work, we propose methods to examine how the complex interrelationships between clinical symptoms and, separately, brain imaging biomarkers change over time leading up to the diagnosis of a disease in subjects with a known genetic near‐certainty of disease. We propose a time‐dependent undirected graphical model that ensures temporal and structural smoothness across time‐specific networks to examine the trajectories of interactions between markers aligned at the time of disease onset. Specifically, we anchor subjects relative to the time of disease diagnosis (anchoring time) as in a revival process, and we estimate networks at each time point of interest relative to the anchoring time. To use all available data, we apply kernel weights to borrow information across observations that are close to the time of interest. Adaptive lasso weights are introduced to encourage temporal smoothness in edge strength, while a novel elastic fused‐ penalty removes spurious edges and encourages temporal smoothness in network structure. Our approach can handle practical complications such as unbalanced visit times. We conduct simulation studies to compare our approach with existing methods. We then apply our method to data from PREDICT‐HD, a large prospective observational study of pre‐manifest Huntington's disease (HD) patients, to identify symptom and imaging network changes that precede clinical diagnosis of HD.

Biometrics Causal Inference · 发表于 2024-07-12

Optimal refinement of strata to balance covariates

Katherine Brumberg, Dylan S Small, Paul R Rosenbaum

摘要

What is the best way to split one stratum into two to maximally reduce the within-stratum imbalance in many covariates? We formulate this as an integer program and approximate the solution by randomized rounding of a linear program. A linear program may assign a fraction of a person to each refined stratum. Randomized rounding views fractional people as probabilities, assigning intact people to strata using biased coins. Randomized rounding is a well-studied theoretical technique for approximating the optimal solution of certain insoluble integer programs. When the number of people in a stratum is large relative to the number of covariates, we prove the following new results: (i) randomized rounding to split a stratum does very little randomizing, so it closely resembles the linear programming relaxation without splitting intact people; (ii) the linear relaxation and the randomly rounded solution place lower and upper bounds on the unattainable integer programming solution; and because of (i), these bounds are often close, thereby ratifying the usable randomly rounded solution. We illustrate using an observational study that balanced many covariates by forming matched pairs composed of 2016 patients selected from 5735 using a propensity score. Instead, we form 5 propensity score strata and refine them into 10 strata, obtaining excellent covariate balance while retaining all patients. An R package optrefine at CRAN implements the method. Supplementary materials are available online.

Biometrics Diagnostics & Prediction Nonparametric & Semiparametric · 发表于 2024-07-12

Nonparametric receiver operating characteristic curve analysis with an imperfect gold standard

Jiarui Sun, Chao Tang, Wuxiang Xie, Xiao-Hua Zhou

摘要

This article addresses the challenge of estimating receiver operating characteristic (ROC) curves and the areas under these curves (AUC) in the context of an imperfect gold standard, a common issue in diagnostic accuracy studies. We delve into the nonparametric identification and estimation of ROC curves and AUCs when the reference standard for disease status is prone to error. Our approach hinges on the known or estimable accuracy of this imperfect reference standard and the conditional independent assumption, under which we demonstrate the identifiability of ROC curves and propose a nonparametric estimation method. In cases where the accuracy of the imperfect reference standard remains unknown, we establish that while ROC curves are unidentifiable, the sign of the difference between two AUCs is identifiable. This insight leads us to develop a hypothesis-testing method for assessing the relative superiority of AUCs. Compared to the existing methods, the proposed methods are nonparametric so that they do not rely on the parametric model assumptions. In addition, they are applicable to both the ROC/AUC analysis of continuous biomarkers and the AUC analysis of ordinal biomarkers. Our theoretical results and simulation studies validate the proposed methods, which we further illustrate through application in two real-world diagnostic studies.

Biometrics Survival Analysis · 发表于 2024-07-12

Absolute risk from double nested case-control designs: cause-specific proportional hazards models with and without augmented estimating equations

Minjung Lee, Mitchell H Gail

摘要

We estimate relative hazards and absolute risks (or cumulative incidence or crude risk) under cause-specific proportional hazards models for competing risks from double nested case-control (DNCC) data. In the DNCC design, controls are time-matched not only to cases from the cause of primary interest, but also to cases from competing risks (the phase-two sample). Complete covariate data are available in the phase-two sample, but other cohort members only have information on survival outcomes and some covariates. Design-weighted estimators use inverse sampling probabilities computed from Samuelsen-type calculations for DNCC. To take advantage of additional information available on all cohort members, we augment the estimating equations with a term that is unbiased for zero but improves the efficiency of estimates from the cause-specific proportional hazards model. We establish the asymptotic properties of the proposed estimators, including the estimator of absolute risk, and derive consistent variance estimators. We show that augmented design-weighted estimators are more efficient than design-weighted estimators. Through simulations, we show that the proposed asymptotic methods yield nominal operating characteristics in practical sample sizes. We illustrate the methods using prostate cancer mortality data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial Study of the National Cancer Institute.

SiM Spatial & Time Series · 发表于 2024-07-11

Prefiltered component‐based greedy (PreCoG) scan method

Joshua P. French, Mohammad Meysami, Ettie M. Lipner

摘要

The spatial distribution of disease cases can provide important insights into disease spread and its potential risk factors. Identifying disease clusters correctly can help us discover new risk factors and inform interventions to control and prevent the spread of disease as quickly as possible. In this study, we propose a novel scan method, the Prefiltered Component‐based Greedy (PreCoG) scan method, which efficiently and accurately detects irregularly shaped clusters using a prefiltered component‐based algorithm. The PreCoG scan method's flexibility allows it to perform well in detecting both regularly and irregularly‐shaped clusters. Additionally, it is fast to apply while providing high power, sensitivity, and positive predictive value for the detected clusters compared to other scan methods. To confirm the effectiveness of the PreCoG method, we compare its performance to many other scan methods. Additionally, we have implemented this method in the smerc R package to make it publicly available to other researchers. Our proposed PreCoG scan method presents a unique and innovative process for detecting disease clusters and can improve the accuracy of disease surveillance systems.

Biometrika Other · 发表于 2024-07-10

More power by using fewer permutations

Nick W Koning

摘要

It is conventionally believed that permutation-based testing methods should ideally use all permutations. We challenge this by showing that we can sometimes obtain dramatically more power by using a tiny subgroup. As the subgroup is tiny, this also comes at a much lower computational cost. Moreover, the method remains valid for the same hypotheses. We exploit this to improve the popular permutation-based Westfall and Young MaxT multiple testing method. We analyse the relative efficiency in a Gaussian location model, and find the largest gain in high dimensions.

JRSSB Causal Inference Networks & Graphs Survival Analysis · 发表于 2024-07-10

Graphical criteria for the identification of marginal causal effects in continuous-time survival and event-history analyses

Kjetil Røysland, Pål C Ryalen, Mari Nygård, Vanessa Didelez

摘要

We consider continuous-time survival and event-history settings, where our aim is to graphically represent causal structures allowing us to characterize when a causal parameter is identified from observational data. This causal parameter is formalized as the effect on an outcome event of a (possibly hypothetical) intervention on the intensity of a treatment process. To establish identifiability, we propose novel graphical rules indicating whether the observed information is sufficient to obtain the desired causal effect by suitable reweighting. This requires a different type of graph than in discrete time. We formally define causal semantics for the corresponding dynamic graphs that represent local independence models for multivariate counting processes. Importantly, our work highlights that causal inference from censored data relies on subtle structural assumptions on the censoring process beyond independent censoring; these can be verified graphically. Put together, our results are the first to establish graphical rules for nonparametric causal identifiability in event processes in this generality for the continuous-time case, not relying on particular parametric survival models. We conclude with a data example on Human papillomavirus (HPV) testing for cervical cancer screening, where the assumptions are illustrated graphically and the desired effect is estimated by reweighted cumulative incidence curves.

JRSSB Clinical Trials Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-09

Judith ter Schure’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Judith ter Schure

摘要

Once there was a randomized controlled trial (RCT) that studied the recurrence of stroke. It had assumed for the treatment of interest a 40% reduction in hazard, and had set the sample size accordingly (Kamel et al., 2024). The RCT found a 95% confidence interval of (−55%, 36%) and concluded no statistical significance. ‘Absence of evidence is not evidence of absence’ said the statisticians (Altman & Bland, 1995; Harrell, 2024). But the story goes that the sample size was justified based on an earlier study’s 95% confidence interval of (38%, 68%) (Connolly et al., 2011; Kaul, 2024). In the end, no-one knew whether there was a type-II error guarantee, or whether this result would discourage new similar RCTs. If 40% reduction in hazard had been a smallest effect-size of interest (SESOI), there was evidence against clinical relevance. But 40% is not actually small, and in practice, the use of such worst-case scenarios is rare in RCT power analysis. Setting a SESOI is difficult, especially since it interacts with the fundability of the research because small effects require trials that are impractically large and waste scarce resources (Mandrola, 2024). Hence, a principled use of Neyman–Pearson testing in RCTs is unpopular, which might be more attributed to the practicalities of power than of type-I error. I congratulate the authors for writing a reference paper on optimal e-values in concrete testing problems that also inspires discussing the practicalities of power. Doing so motivates e-values without SESOI-requirements, and comparisons with two-sided Neyman–Pearson tests that also leave an aspect of the alternative hypothesis unspecified at the price of power. ‘Unlimited-horizon’ (ter Schure et al., 2024) and ‘power-one’ (Robbins & Siegmund, 1974) can be emphasized by plotting cumulative power by sample size, for various true-effect/nmax-combinations. Continuously increasing functions visualize the suitable e-value tests, also for the small true-effects. Step functions visualize the Neyman–Pearson tests of fixed-sample size (and extensions to alpha-spending with an nmax)—highlighting the lack of power ‘pay-out’ before nmax in the absence of interim analyses, and complete lack of power increase after nmax where bad combinations are stuck at low power. Suitable e-values allow us to collect data, have a look, and take it from there. In some cases, a worst-case power analysis is impossible. If the question is whether the effect is not smaller than a SESOI, your SESOI specifies the null hypothesis and the worst-case alternative. A parallel situation occurs in non-inferiority testing. This impossibility is not front-and-center, but, in my experience, is apparent from the unpopularity of non-inferiority sample size calculations. No new data were generated or analysed in support of this research. *(abstract via OpenAlex)*

SiM Networks & Graphs · 发表于 2024-07-09

Exploiting relationship directionality to enhance statistical modeling of peer‐influence across social networks

Xin Ran, Nancy E. Morden, Ellen Meara, Erika L. Moen, Daniel N. Rockmore, A. James O'Malley

摘要

Risky‐prescribing is the excessive or inappropriate prescription of drugs that singly or in combination pose significant risks of adverse health outcomes. In the United States, prescribing of opioids and other “risky” drugs is a national public health concern. We use a novel data framework—a directed network connecting physicians who encounter the same patients in a sequence of visits—to investigate if risky‐prescribing diffuses across physicians through a process of peer‐influence. Using a shared‐patient network of 10 661 Ohio‐based physicians constructed from Medicare claims data over 2014‐2015, we extract information on the order in which patients encountered physicians to derive a directed patient‐sharing network. This enables the novel decomposition of peer‐effects of a medical practice such as risky‐prescribing into directional (outbound and inbound) and bidirectional (mutual) relationship components. Using this framework, we develop models of peer‐effects for contagion in risky‐prescribing behavior as well as spillover effects. The latter is measured in terms of adverse health events suspected to be related to risky‐prescribing in patients of peer‐physicians. Estimated peer‐effects were strongest when the patient‐sharing relationship was mutual as opposed to directional. Using simulations we confirmed that our modeling and estimation strategies allows simultaneous estimation of each type of peer‐effect (mutual and directional) with accuracy and precision. We also show that failing to account for these distinct mechanisms (a form of model mis‐specification) produces misleading results, demonstrating the importance of retaining directional information in the construction of physician shared‐patient networks. These findings suggest network‐based interventions for reducing risky‐prescribing.

SiM Clinical Trials · 发表于 2024-07-09

Determining sample size in a personalized randomized controlled ( PRACTical ) trial

Rebecca M. Turner, Kim May Lee, A. Sarah Walker, Sally Ellis, Michael Sharland, Julia A. Bielicki, Wolfgang Stöhr, Ian R. White

摘要

In clinical settings with no commonly accepted standard‐of‐care, multiple treatment regimens are potentially useful, but some treatments may not be appropriate for some patients. A personalized randomized controlled trial (PRACTical) design has been proposed for this setting. For a network of treatments, each patient is randomized only among treatments which are appropriate for them. The aim is to produce treatment rankings that can inform clinical decisions about treatment choices for individual patients. Here we propose methods for determining sample size in a PRACTical design, since standard power‐based methods are not applicable. We derive a sample size by evaluating information gained from trials of varying sizes. For a binary outcome, we quantify how many adverse outcomes would be prevented by choosing the top‐ranked treatment for each patient based on trial results rather than choosing a random treatment from the appropriate personalized randomization list. In simulations, we evaluate three performance measures: mean reduction in adverse outcomes using sample information, proportion of simulated patients for whom the top‐ranked treatment performed as well or almost as well as the best appropriate treatment, and proportion of simulated trials in which the top‐ranked treatment performed better than a randomly chosen treatment. We apply the methods to a trial evaluating eight different combination antibiotic regimens for neonatal sepsis (NeoSep1), in which a PRACTical design addresses varying patterns of antibiotic choice based on disease characteristics and resistance. Our proposed approach produces results that are more relevant to complex decision making by clinicians and policy makers.

Biometrika Longitudinal Data High-Dimensional · 发表于 2024-07-08

Testing independence for sparse longitudinal data

Changbo Zhu, Junwen Yao, Jane-Ling Wang

摘要

With the advance of science and technology, more and more data are collected in the form of functions. A fundamental question for a pair of random functions is to test whether they are independent. This problem becomes quite challenging when the random trajectories are sampled irregularly and sparsely for each subject. In other words, each random function is only sampled at a few time-points, and these time-points vary with subjects. Furthermore, the observed data may contain noise. To the best of our knowledge, there exists no consistent test in the literature to test the independence of sparsely observed functional data. We show in this work that testing pointwise independence simultaneously is feasible. The test statistics are constructed by integrating pointwise distance covariances (Székely et al., 2007) and are shown to converge, at a certain rate, to their corresponding population counterparts, which characterize the simultaneous pointwise independence of two random functions. The performance of the proposed methods is further verified by Monte Carlo simulations and analysis of real data.

JASA Networks & Graphs · 发表于 2024-07-08

Node-Level Community Detection within Edge Exchangeable Models for Interaction Processes

Yuhua Zhang, Walter Dempsey

摘要

Scientists are increasingly interested in discovering community structure from modern relational data arising on large-scale social networks. While many methods have been proposed for learning community structure, few account for the fact that these modern networks arise from processes of interactions in the population. We introduce block edge exchangeable models (BEEM) for the study of interaction networks with latent node-level community structure. The block vertex components model (B-VCM) is derived as a canonical example. Several theoretical and practical advantages over traditional vertex-centric approaches are highlighted. In particular, BEEMs allow for sparse degree structure and power-law degree distributions within communities. Our theoretical analysis bounds the misspecification rate of block assignments while supporting simulations show the properties of the network can be recovered. A computationally tractable Gibbs algorithm is derived. We demonstrate the proposed model using post-comment interaction data from Talklife, a large-scale online peer-to-peer support network, and contrast the learned communities from those using standard algorithms including degree-corrected stochastic block models, popularity-adjusted block models, and weighted stochastic block models. *(abstract via OpenAlex)*

JRSSB Bayesian Methods Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-08

Luigi Pace and Alessandra Salvan’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Luigi Pace, Alessandra Salvan

摘要

We wish to congratulate the authors on their important paper. In this contribution, we will focus on parametric models, where e-values provide a welcome new perspective on likelihood ratios and Bayes factors, as well as a useful addition to the discussion about the foundational concept of statistical evidence. The guarantees sought for in safe testing are frequentist, but prior information may be injected in Bayesian terms, establishing an alliance between traditionally antagonistic inferential paradigms. The core ideas of safe testing are that error probabilities need to be constrained only under H0⁠, and that their evaluation is conservative and nonasymptotic, relying on Ville’s inequality, which is the basis for anytime-valid inference (Ramdas et al., 2023) in the form of tests or confidence sequences. Incidentally, the associated safe bound on the error probabilities may deserve a new name, different from the traditional significance/confidence level. Perhaps, strong significance/confidence level, or persistence level, as suggested in Pace and Salvan (2020) for confidence sequences. Using a likelihood ratio threshold as a universal yardstick, likelihood-based e-values obey the strong likelihood principle, in spite of the opinion that any frequentist desideratum is irreconcilable with such principle. A consequence of obeying the strong likelihood principle is the validity of e-values under conditioning on ancillary statistics, because likelihood factors not depending on the parameter cancel. Ancillary conditioning is a rather Fisherian concern, but a resemblance with e-values may be highlighted for simple H0 and H1⁠. Indeed, in a fixed n setting, when the likelihood ratio r has continuous distributions, Pace and Salvan (2014) show that there exists an ancillary statistic a=a(r)⁠, i.e. a statistic having the same distribution under H0 and under H1 and which, joint with the maximum likelihood estimate, is minimal sufficient. As in safe testing, the conditional Type I and Type II error probabilities, α(a) and β(a)⁠, satisfy the universal inequalities α(a)≤1/a=1/r⁠, when r>1⁠, and β(a)≤s(a)=r⁠, when r1⁠, so that α(a)≤α entails r≥(1−α)/α⁠. We wonder whether substitution of the threshold 1/α with (1−α)/α might be meaningful also for tests based on e-values. *(abstract via OpenAlex)*

JASA Other · 发表于 2024-07-08

Generalizing the Intention-to-Treat Effect of an Active Control from Historical Placebo-Controlled Trials: A Case Study of the Efficacy of Daily Oral TDF/FTC in the HPTN 084 Study

Qijia He, Fei Gao, Oliver Dukes, Sinead Delany-Moretlwe, Bo Zhang

摘要

In many clinical settings, an active-controlled trial design (e.g., a non-inferiority or superiority design) is often used to compare an experimental medicine to an active control (e.g., an FDA-approved, standard therapy). One prominent example is a recent phase 3 efficacy trial, HIV Prevention Trials Network Study 084 (HPTN 084), comparing long-acting cabotegravir, a new HIV pre-exposure prophylaxis (PrEP) agent, to the FDA-approved daily oral tenofovir disoproxil fumarate plus emtricitabine (TDF/FTC) in a population of heterosexual women in 7 African countries. One key complication of interpreting study results in an active-controlled trial like HPTN 084 is that the placebo arm is not present and the efficacy of the active control (and hence the experimental drug) compared to the placebo can only be inferred by leveraging other data sources. In this article, we study statistical inference for the intention-to-treat (ITT) effect of the active control using relevant historical placebo-controlled trials data under the potential outcomes (PO) framework. We highlight the role of adherence and unmeasured confounding, discuss in detail identification assumptions and two modes of inference (point versus partial identification), propose estimators under identification assumptions permitting point identification, and lay out sensitivity analyses needed to relax identification assumptions. We applied our framework to estimating the intention-to-treat effect of daily oral TDF/FTC versus placebo in HPTN 084 using data from an earlier Phase 3, placebo-controlled trial of daily oral TDF/FTC (Partners PrEP). *(abstract via PubMed)*

JASA Testing & Uncertainty · 发表于 2024-07-08

False Discovery Rate Control For Structured Multiple Testing: Asymmetric Rules And Conformal Q -values

Zinan Zhao, Wenguang Sun

摘要

The effective utilization of structural information in data while ensuring statistical validity poses a significant challenge in false discovery rate (FDR) analyses. Conformal inference provides rigorous theory for grounding complex machine learning methods without relying on strong assumptions or highly idealized models. However, existing conformal methods have limitations in handling structured multiple testing, as their validity often requires the deployment of symmetric decision rules, which assume the exchangeability of data points and permutation-invariance of fitting algorithms. To overcome these limitations, we introduce the pseudo local index of significance (PLIS) procedure, which is capable of accommodating asymmetric rules and requires only pairwise exchangeability between the null conformity scores. We demonstrate that PLIS offers finite-sample guarantees in FDR control and the ability to assign higher weights to relevant data points. Numerical results confirm the effectiveness and robustness of PLIS and demonstrate improvements in power compared to existing model-free methods in various scenarios. *(abstract via OpenAlex)*

JASA Other · 发表于 2024-07-08

Efficient Stochastic Generators with Spherical Harmonic Transformation for High-Resolution Global Climate Simulations from CESM2-LENS2

Yan Song, Zubair Khalid, Marc G. Genton

摘要

Earth system models (ESMs) are fundamental for understanding Earth's complex climate system. However, the computational demands and storage requirements of ESM simulations limit their utility. For the newly published CESM2-LENS2 data, which suffer from this issue, we propose a novel stochastic generator (SG) as a practical complement to the CESM2, capable of rapidly producing emulations closely mirroring training simulations. Our SG leverages the spherical harmonic transformation (SHT) to shift from spatial to spectral domains, enabling efficient low-rank approximations that significantly reduce computational and storage costs. By accounting for axial symmetry and retaining distinct ranks for land and ocean regions, our SG captures intricate non-stationary spatial dependencies. Additionally, a modified Tukey g-and-h (TGH) transformation accommodates non-Gaussianity in high-temporal-resolution data. We apply the proposed SG to generate emulations for surface temperature simulations from the CESM2-LENS2 data across various scales, marking the first attempt of reproducing daily data. These emulations are then meticulously validated against training simulations. This work offers a promising complementary pathway for efficient climate modeling and analysis while overcoming computational and storage limitations. *(abstract via OpenAlex)*

JRSSB Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-08

Christine P. Chai’s contribution to the Discussion of ‘Safe testing’ by Grünwald, De Heide, and Koolen

Christine P Chai

摘要

This paper proposes the e-value for ‘safe testing’ as an alternative to the p-value in hypothesis testing. The e-value allows for optional continuation while preserving Type-I error guarantees. I hope to see lots of discussion at the Royal Statistical Society meeting, like the previous ‘Signs and sizes: Understanding and replicating statistical findings’ (Hutton et al., 2020). I have some questions regarding the descriptions: I see some comparisons between the e-value and several alternatives to the ordinary p-value. How does the e-value compare with the sceptical p-value proposed by Held (2020), which is designed to combine replication studies? This paper discusses both optional continuation (OC) and optional stopping (OS) settings. Are there any real-life examples where OC or OS is more appropriate? If we can choose either one, what are the guidelines to decide? In hypothesis testing, β is the probability of making a type II error, i.e. failing to reject the null when it is actually false (Banerjee et al., 2009). I am confused with the narrative, because some parts say the power is set to 1−β=0.8 and other parts say β=0.8⁠. Which is the one used in the experiments? To further improve the manuscript, I also have some comments: It is common to decide whether to perform additional tests depending on the outcome of previous tests, and the rationale is that we do not want to spend money on more testing unless it is necessary. For better clarification, we can state in the motivation that data collection is expensive (Coffey & Elliott, 2024; Jones et al., 2015). I agree with the consensus that we should be cautious in using p-values for hypothesis testing (Wasserstein & Lazar, 2016). There has also been a recent shift from hypothesis testing to estimation as ‘the new statistics’ (Cumming, 2014; Kruschke & Liddell, 2018), and I hope that e-values is a step to better quantification of the decision-making uncertainty. Finally, it is important to make the images accessible to everyone, including people with disabilities. Please make Figures 2 and 4 more readable by increasing the font size of the axis labels. I also recommend using different line shapes to represent each trend, because people with colour blindness may not be able to see the colours. No new data were generated or analysed in support of this research. *(abstract via OpenAlex)*

SiM Causal Inference Clinical Trials · 发表于 2024-07-08

A sequential, multiple assignment, randomized trial design with a tailoring function

Holly Hartman, Matthew Schipper, Kelley Kidwell

摘要

We present a trial design for sequential multiple assignment randomized trials (SMARTs) that use a tailoring function instead of a binary tailoring variable allowing for simultaneous development of the tailoring variable and estimation of dynamic treatment regimens (DTRs). We apply methods for developing DTRs from observational data: tree‐based regression learning and Q‐learning. We compare this to a balanced randomized SMART with equal re‐randomization probabilities and a typical SMART design where re‐randomization depends on a binary tailoring variable and DTRs are analyzed with weighted and replicated regression. This project addresses a gap in clinical trial methodology by presenting SMARTs where second stage treatment is based on a continuous outcome removing the need for a binary tailoring variable. We demonstrate that data from a SMART using a tailoring function can be used to efficiently estimate DTRs and is more flexible under varying scenarios than a SMART using a tailoring variable.

SiM Missing Data & Measurement Error Bayesian Methods Causal Inference · 发表于 2024-07-08

A Bayesian semi‐parametric scalar‐on‐function regression with measurement error using instrumental variables

Roger S Zoh, Yuanyuan Luan, Lan Xue, David B Allison, Carmen D Tekwe

摘要

Wearable devices such as the ActiGraph are now commonly used in research to monitor or track physical activity. This trend corresponds with the growing need to assess the relationships between physical activity and health outcomes, such as obesity, accurately. Device‐based physical activity measures are best treated as functions when assessing their associations with scalar‐valued outcomes such as body mass index. Scalar‐on‐function regression (SoFR) is a suitable regression model in this setting. Most estimation approaches in SoFR assume that the measurement error in functional covariates is white noise. Violating this assumption can lead to underestimating model parameters. There are limited approaches to correcting measurement errors for frequentist methods and none for Bayesian methods in this area. We present a non‐parametric Bayesian measurement error‐corrected SoFR model that relaxes all the constraining assumptions often involved with these models. Our estimation relies on an instrumental variable allowing a time‐varying biasing factor, a significant departure from the current generalized method of moment (GMM) approach. Our proposed method also permits model‐based grouping of the functional covariate following measurement error correction. This grouping of the measurement error‐corrected functional covariate allows additional ease of interpretation of how the different groups differ. Our method is easy to implement, and we demonstrate its finite sample properties in extensive simulations. Finally, we applied our method to data from the National Health and Examination Survey to assess the relationship between wearable device‐based measures of physical activity and body mass index in adults in the United States.

JRSSB Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-05

Glenn Shafer’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Glenn Shafer

摘要

It is a delight to see this path-breaking paper widely discussed. I congratulate the authors for its clarity, catholic spirit, and relevance to practice. On one important practical point, the authors concede too much. Testing with e-values, they tell us, requires more data than testing with p-values. Was this concession extracted by referees? To my mind, it misses the main point of Shafer et al. (2011), which is that interpreting a p-value as evidence on the same scale as a fixed significance level is cheating. A test with significance level α that rejects when E happens is equivalent to an e-variable equal to 1/α when E happens and zero otherwise. A p-value equal to α is weaker, because you chose in advance only a test statistic, not the exact rejection region E. The point of calibrating a p-value p is to adjust 1/p downward to compensate for the cheating, making it comparable to 1/α or any other e-value. As the authors mention, my favourite calibrator is the simple 1/p−1⁠. It reduces a p-value of 0.05⁠, for example, from an apparent e-value of 20 to a true e-value of 3.47⁠. When we make p-values honest, e-values do not need more data to catch up. As the authors suggest in Section 6, the value of testing with e-values will become clearer once we no longer “insist on power and significance requirements”. Statisticians have been trying to combine power not with fixed significance levels (which makes some sense) but with p-values (which makes no sense). e-values (or betting scores, as I prefer to say) generalize and are more useful than fixed significance, and this greater usefulness makes the authors’ GRO criterion more useful than power. As the authors know, I believe that the value of their work will become even clearer once mathematical statisticians no longer distort their ideas to fit them into measure theory. Most scientists and applied statisticians will accept the authors’ informal idea in Section 1.1: an e-value E(j+1) can be “independent” of E(1),…,E(j) even though the experiment that produces E(j+1) was never even imagined before the values of E(1),…,E(j) were observed. But when we get to Proposition 2, we learn that E(j+1) is conditional on what happened before, including presumably all the information and decision making that went into deciding on the (j+1)st experiment. Is an imaginary probability measure encompassing all these events what the scientist wants to test? I am grateful to the authors for putting their important work and these important issues on the table. *(abstract via OpenAlex)*

JRSSB Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-04

Maozai Tian, Keming Yu and Jiangfeng Wang’s contribution to the Discussion of ‘Safe testing’ by Grünwald, De Heide, and Koolen

Maozai Tian, Keming Yu, Jiangfeng Wang

摘要

We want to congratulate the authors on the publication of ‘Safe Testing’ on JRSS-B. It deepens our understanding of the statistical hypothesis test that is one particular method of statistical inference used to decide whether the data sufficiently support a particular hypothesis. The e-variables and safe tests offer several advantages. The authors develop the theory of hypothesis testing based on the e-value that behave excellently under optional continuation and are enhanced with interpretability and flexibility. It is particularly noteworthy that the authors pay careful attention to the issues involved in Type I error guarantees. They argue that e-variables can be based on Bayesian prior knowledge on earlier data but also on minimax performance considerations, in all cases preserving frequentist Type I error guarantees. Our main focus is on two issues: First, for Type I error as shown in this article, the minimax results based on e-variables for composite hypothesis, i.e. the parameter’s value is given by a set of numbers, are too pessimistic. The minimax approach is conservative because it only guarantees the best possible performance in the ‘worst case’ situation. Second, for the GRO criterion that is growth rate optimality (GRO) as an analogue of power in an optional continuation context: The asymptotic rate results are almost uninformative and cannot be used for practical choice of the tuning parameters in finite sample studies or for comparing different procedures. The authors thank the RSS Journals Manager Judith Shorten for her kind invitation to make presentation with discussion of the article. The work was the Fundamental Research Funds for the Central Universities, and the Research Funds of Renmin University of China (22XNL016). *(abstract via OpenAlex)*

SiM Other · 发表于 2024-07-04

Covariate‐adjusted generalized pairwise comparisons in small samples

Stijn Jaspers, Johan Verbeeck, Olivier Thas

摘要

Semiparametric probabilistic index models allow for the comparison of two groups of observations, whilst adjusting for covariates, thereby fitting nicely within the framework of generalized pairwise comparisons (GPC). As with most regression approaches in this setting, the limited amount of data results in invalid inference as the asymptotic normality assumption is not met. In addition, separation issues might arise when considering small samples. In this article, we show that the parameters of the probabilistic index model can be estimated using generalized estimating equations, for which adjustments exist that lead to estimators of the sandwich variance‐covariance matrix with improved finite sample properties and that can deal with bias due to separation. In this way, appropriate inference can be performed as is shown through extensive simulation studies. The known relationships between the probabilistic index and other GPC statistics allow to also provide valid inference for example, the net treatment benefit or the success odds.

SiM Survival Analysis Bayesian Methods · 发表于 2024-07-04

BHAFT: Bayesian heredity‐constrained accelerated failure time models for detecting gene‐environment interactions in survival analysis

Na Sun, Jiadong Chu, Qida He, Yu Wang, Qiang Han, Nengjun Yi, Ruyang Zhang, Yueping Shen

摘要

In addition to considering the main effects, understanding gene‐environment (G × E) interactions is imperative for determining the etiology of diseases and the factors that affect their prognosis. In the existing statistical framework for censored survival outcomes, there are several challenges in detecting G × E interactions, such as handling high‐dimensional omics data, diverse environmental factors, and algorithmic complications in survival analysis. The effect heredity principle has widely been used in studies involving interaction identification because it incorporates the dependence of the main and interaction effects. However, Bayesian survival models that incorporate the assumption of this principle have not been developed. Therefore, we propose Bayesian heredity‐constrained accelerated failure time (BHAFT) models for identifying main and interaction (M‐I) effects with novel spike‐and‐slab or regularized horseshoe priors to incorporate the assumption of effect heredity principle. The R package rstan was used to fit the proposed models. Extensive simulations demonstrated that BHAFT models had outperformed other existing models in terms of signal identification, coefficient estimation, and prognosis prediction. Biologically plausible G × E interactions associated with the prognosis of lung adenocarcinoma were identified using our proposed model. Notably, BHAFT models incorporating the effect heredity principle could identify both main and interaction effects, which are highly useful in exploring G × E interactions in high‐dimensional survival analysis. The code and data used in our paper are available at https://github.com/SunNa‐bayesian/BHAFT.

JRSSB Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-04

Andrej Srakar’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Andrej Srakar

摘要

Paper presented by Grünwald, de Heide, and Koolen presents an extension of e-values to the testing of composite hypotheses in a context they label ‘safe testing’. It discusses what should be a merging of the Fisherian, Neymanian, and Jeffreys-Bayesian interpretations to the testing of hypotheses. Proposed tests based on e-values are safe, i.e. they preserve type-I error guarantees, under proposed optional continuation where the decision to perform a new study may depend on previous outcomes. I provide short remarks related to probability aspects and possibilities in the proposed approach. Namely, construction of the approach is based on supermartingales and relation to previously studied e-processes (Ramdas et al., 2021). Brownian motion is mentioned at the end of the article appendix but not contextualized in more detail. What I question (this is also based on the illustrative Figure 1) is about possible additional extensions in a probability context. For example, when one would want to address the context of composite hypotheses in their dependencies, approaches and stochastic models based on interacting particle systems (such as simple exclusion or simple inclusion processes, or many extensions therein) could be interesting and useful to include. Literature in this area has previously studied relationship to martingales (e.g. Delgadino et al., 2021; Oelschläger, 1984). Other probability theory contexts could also be addressed for future research on e-values. *(abstract via OpenAlex)*

JASA Functional Data Diagnostics & Prediction · 发表于 2024-07-02

Weighted Functional Data Analysis for the Calibration of a Ground Motion Model in Italy

Teresa Bortolotti, Riccardo Peli, Giovanni Lanzano, Sara Sgobba, Alessandra Menafoglio

摘要

Motivated by the crucial implications of Ground Motion Models in terms of seismic hazard analysis and civil protection planning, this work extends a scalar Ground Motion Model for Italy to the framework of Functional Data Analysis. The inherent characteristic of seismic data to be incomplete over the observation domain of oscillation periods entails embedding the analysis in the context of partially observed functional data and performing data reconstruction. This work proposes a novel methodology that accounts for the fact that parts of the curves are directly observed and other parts are reconstructed, thus, characterized by greater uncertainty. The method defines observation-specific functional weights, which enter the estimation process to reduce the impact that the less reliable portions of the curves have on the final estimates. The classical methods of smoothing and concurrent functional regression are extended to include weights. The advantages of the proposed methodology are assessed on synthetic data. Eventually, the weighted functional analysis performed on seismological data is shown to provide a natural smoothing and stabilization of the spectral estimates of the Ground Motion Model considered. Supplementary materials for this article are available online. *(abstract via OpenAlex)*

JRSSB Clinical Trials Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-02

Marco Cattaneo's contribution to the Discussion of “Safe testing” by Grünwald, de Heide, and Koolen

Marco E G V Cattaneo

摘要

This is an excellent paper, neatly summarizing previous results about e-variables, as well as presenting important advances. The possibility of obtaining conservative p-values and confidence intervals even under optional continuation is a substantial advantage of e-values over alternative measures of statistical evidence. Criticisms of the use of p-values often include the advice that p-values should only be interpreted jointly with an estimate of the effect size of interest. Under optional continuation, effect size estimation may be difficult. With preplanned sequential testing, standard estimators (such as maximum likelihood estimators) can be severely biased, but it may be possible to estimate their bias and improve the effect size estimates. In the optional continuation scenario, even the concept of bias seems problematic, since the sequential analysis is not preplanned and unconditional expectations cannot be easily evaluated. Given an effect size of interest δ and an e-variable Eδ0 for each null hypothesis δ=δ0⁠, the convex hull of {δ:Eδ≤1/α} is a conservative (1−α) confidence interval for δ. In analogy with likelihood ratios, it seems quite natural to consider argminδEδ as an estimate of the effect size δ. Have the properties of such ‘minimum e-value’ estimators been studied, in particular under optional continuation? The end of Section 6 discusses the sample size requirements in situations in which stopping is possible after each data point. It is shown that in order to achieve the desired power, hypothesis test based on e-variables may need on average less data than standard tests with fixed sample size. It would be interesting to include also standard sequential tests in the comparison, since they would also be possible in such situations and may also need on average less data than tests with fixed sample size. *(abstract via OpenAlex)*

Biometrika Spatial & Time Series · 发表于 2024-07-02

Debiasing Welch’s method for spectral density estimation

Lachlan C Astfalck, Adam M Sykulski, Edward J Cripps

摘要

Welch’s method provides an estimator of the power spectral density that is statistically consistent. This is achieved by averaging over periodograms calculated from overlapping segments of a time series. For a finite-length time series, while the variance of the estimator decreases as the number of segments increases, the magnitude of the estimator’s bias increases: a bias-variance trade-off ensues when setting the segment number. We address this issue by providing a novel method for debiasing Welch’s method that maintains the computational complexity and asymptotic consistency, and leads to improved finite-sample performance. Theoretical results are given for fourth-order stationary processes with finite fourth-order moments and an absolutely convergent fourth-order cumulant function. The significant bias reduction is demonstrated with numerical simulation and an application to real-world data. Our estimator also permits irregular spacing over frequency and we demonstrate how this may be employed for signal compression and further variance reduction. The code accompanying this work is available in R and python.

SiM Spatial & Time Series Bayesian Methods · 发表于 2024-07-02

A Bayesian non‐stationary heteroskedastic time series model for multivariate critical care data

Zayd Omar, David A. Stephens, Alexandra M. Schmidt, David L. Buckeridge

摘要

We propose a multivariate GARCH model for non‐stationary health time series by modifying the observation‐level variance of the standard state space model. The proposed model provides an intuitive and novel way of dealing with heteroskedastic data using the conditional nature of state‐space models. We follow the Bayesian paradigm to perform the inference procedure. In particular, we use Markov chain Monte Carlo methods to obtain samples from the resultant posterior distribution. We use the forward filtering backward sampling algorithm to efficiently obtain samples from the posterior distribution of the latent state. The proposed model also handles missing data in a fully Bayesian fashion. We validate our model on synthetic data and analyze a data set obtained from an intensive care unit in a Montreal hospital and the MIMIC dataset. We further show that our proposed models offer better performance, in terms of WAIC than standard state space models. The proposed model provides a new way to model multivariate heteroskedastic non‐stationary time series data. Model comparison can then be easily performed using the WAIC.

JRSSB Testing & Uncertainty Bayesian Methods 讨论 / 评论 · 发表于 2024-07-01

Samuel Pawel and Leonhard Held’s contribution to the Discussion of ‘Safe Testing’ by Grünwald, de Heide, and Koolen

Samuel Pawel, Leonhard Held

摘要

We congratulate the authors on their paper introducing a ‘safe’ inference framework. However, despite their supposed safety, we have reservations about whether ‘safe’ tests can make the scientific enterprise more reliable. For instance, researchers may still prefer classical non-sequential over sequential study designs, as the latter can be difficult to implement in practice. For example, interim analyses in randomized clinical trials will require unblinding and may as such threaten the integrity of the trial (Ellenberg et al., 2019, Chapter 5). In addition, it has been argued that practical problems in the use of traditional p-values (rather than e-values) stem more from cognitive factors than from the choice of inferential statistics (Gelman, 2016). For example, an excessive focus on testing parameter values of ‘no effect’ while ignoring other relevant parameter values (‘nullism’), unnecessary dichotomization of quantitative information (‘dichotomania’), or treating statistical models as known physical laws rather than speculative assumptions (‘statistical reification’) can all lead to distorted interpretations of study results (Greenland, 2017). Especially with respect to the last point, we believe that naming a statistical test as ‘safe’ can mislead researchers into thinking that the method is ‘always valid’, when this is only true in a certain technical sense and under certain speculative assumptions (e.g. that the underlying data model is correctly specified). For instance, in meta-analysis, a promising application according to the authors, inferences based on e-values will not be ‘safe’ unless the analysis accounts for possible bias, for example, publication bias (Cooper et al., 2019, Chapter 18). In this case, can e-values be adjusted for potential model misspecification (Copas & Eguchi, 2005)? The evidence interpretation of e-values as Bayes factors is more aligned with our view, since Bayes factors quantify relative predictive performance of competing hypotheses without claims about ‘safe’ error rates, yet challenges with nullistic and dichotomous thinking remain. If conventional p-values and confidence intervals were replaced by e-value counterparts, how should results be reported to address these challenges? For testing of one- and two-dimensional parameters, we propose a graphical display inspired by p-value functions and confidence curves (Birnbaum, 1961; Cox, 1958), and their Bayes factor analogues (Pawel et al., 2024). The idea is to plot the e-value as a function of the tested parameter value, see Figure 1 for an illustration. Like the p-value function, this shifts emphasis away from nullistic-dichotomous to gradual-quantitative inferences. Example of an e-value function, i.e. the e-value displayed as a function of the tested parameter value. The graph provides a wealth of information as e-values and conservative p-values (for any tested parameter value), conservative confidence intervals (at any level of interest), and point estimates (the parameter values with the least evidence against them) can be easily read off. In summary, while ‘safe’ tests offer potential advantages, they are not always safe to use and can be affected by the same cognitive issues as other methods. Careful application and contextual interpretation are paramount to their success. *(abstract via OpenAlex)*

JRSSB Bayesian Methods Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-01

Joshua Bon and Christian P. Robert’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Joshua Bon, Christian P Robert

摘要

The central proposal found in this quite original paper could prove relevant for the so-called objective Bayes approach (Berger, 1985; Robert, 2001) as it may open a way to deriving prior distributions from (frequentist) first principles. While some Bayesian schools refuse to consider the concept of testing the veracity of a null hypothesis, most do propose approaches and methodologies to address this challenge. These mainly rely on Jeffreys’ formalization via two prior distributions (Jeffreys, 1939; Wrinch & Jeffreys, 1919) via what would later come to be the Bayes Factor (Ly et al., 2016; Robert et al., 2009; Wagenmakers & Ly, 2022). In his Theory of Probability, Jeffreys constructs the equivalent of his point estimation prior for tests, although it had little impact on the field until rediscovered by Bayarri and Garcia-Donato (2007). We commend the points put forward to support e-values, behaviour under optional continuation [by a martingale reasoning] interpretation as ‘evidence against the null’ as gambling in all cases preserving frequentist Type I error guarantees e-variables turn out to be Bayes factors based on the right Haar prior (rather than at times on highly unusual (e.g. degenerate) priors?) e-variables need more extreme data than p-values in order to reject the null since they are mostly worthwhile, even though (b) is vague and (c) is frequentist. The link of the theory of e-values with right Haar priors reminds us of Pitman’s derivation of the best equivariant estimator (Robert, 2001), even though the fact that Haar priors are improper for the most part proves a major hindrance in a testing situation (DeGroot, 1973). The optional continuation in (a) is an even stronger argument—if from a Bayesian viewpoint—since it is been used to defend the Bayesian paradigm. On the other hand, if sequentially testing e-values, only the first e-value will be interpretable as a coherent Bayes factor. That is, in sequential testing, the product of the e-values does not share the same properties as a product of Bayes factors (i.e. cancellation of terms as a telescoping product). Central to our interest, point (d) above, despite the cautionary added label of an ‘Almost Bayesian Case’, brings a formal way to define least favourable priors (Berger, 1985) in a testing environment, albeit in no clear connection with the reference priors of Bayarri and Garcia-Donato (2007). This derivation, formalized as Theorem 1, provides a reversal of the typical prior choice in testing. A more common stance is to hold the (subjectively chosen) prior distribution on H0 as a starting point—this is an obviously known entity. One may wonder if a dual perspective, i.e. when starting from the prior distribution on H0⁠, would prove fruitful, e.g. in leading to an optimal prior on H1⁠. The paper seems to conclude at a perfect duality with the minimax–maximin result. This also relates to the projection priors we defined in Goutis and Robert (1998) and Dupuis and Robert (2003). As a plus, however, the approach found therein allows for improper priors on the nuisance parameters and thus somehow brings a stronger justification than Berger et al. (1998) in favour of the permanence of an identical measure (on the nuisance parameters) across hypotheses and models. In conclusion, we congratulate the authors on this endeavour but it remains unclear to us (as Bayesians) (i) how to construct the least favourable prior on H0 on a general basis, especially from a computational viewpoint, and, more importantly and (ii) whether it is at all of inferential interest (i.e. whether it degenerates into a point mass). With respect to the sequential directions of the article, we also wonder at the potential connections with sequential Monte Carlo, for instance, towards conducting sequential model choice by constructing efficiently an amalgamated evidence value when the product of Bayes factors is not a Bayes factor (see e.g. Buchholz et al., 2023). Funded by a Prairie chair from the Agence Nationale de la Recherche (ANR-19-P3IA-0001) and by the European Union (ERC-2022-SYG-OCEAN-101071601). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the a Prairie chair from the Agence Nationale de la Recherche (ANR-19-P3IA-0001), the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. *(abstract via OpenAlex)*

JRSSB Bayesian Methods Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-01

Joris Mulder’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Joris Mulder

摘要

The potential of e-values (or e-variables) to unite different schools of statistics and allowing optional stopping/continuation in a ‘safe’ manner is a major accomplishment of the authors. The theory implies a change in focus from classical error rates to the expected evidence against the null under the respective hypotheses. Under H0⁠, it is required that the expectation of the evidence against the null, E, should not exceed 1, i.e. EH0[E]≤1⁠, as in formula (1) of ‘Safe testing’. Clearly, using the reciprocal of p-values as a measure of evidence against the null severely violates this criterion as EH0[1/p]=∞⁠. Thus, we see (via a new route) that the routine use of p-values in scientific practice seriously harms the trustworthiness of research findings. Under H1⁠, Bayes factors using a right Haar prior on the scale behave as e-values with an optimal growth rate of the evidence. As stated by the authors, more (nontrivial) e-values exist. As an alternative, I consider fractional Bayes factors (FBFs; O’Hagan, 1995) which are easy to compute and easy to apply as no proper priors need to be formulated. In fractional Bayes methodology, a (minimal) fraction of the data, b, is used to update an improper prior to obtain a proper ‘data-adapted prior’ while the remaining (maximal) fraction is used for computing the evidence (Gilks, 1995; Mulder, 2014). For a simple t-test, the FBF of H1 against H0 is also scale-invariant and given by Γ(nb2)Γ(n−12)Γ(nb−12)Γ(n2)(1+(n−1)−1t2)−n(b−1)/2⁠, where t is the regular t-statistic. For n=20⁠, Figure 1 shows the expected evidence based on a FBF with a minimal fraction of b=2n=0.1⁠, computed with ‘BFpack’ in R (Mulder et al., 2021), and larger fractions, and using a Bayes factor based on the right Haar prior on the scale (Rouder et al., 2009) computed with the ‘BayesFactor’ package (Morey et al., 2024), all on a logarithmic scale. The figure illustrates that FBFs (i) are ‘safe’ as requirement (1) under H0 is satisfied, and (ii) behave competitively to a Bayes factor with optimal growth rate under H1 when using a minimal fraction, with even higher growth rates for large (fixed) nonzero effects. Formal proofs are left for future work. Finally note that for safe sequential testing, FBFs need to be recomputed on the entire (combined) data in order to be coherent. This can simply be done by storing the sufficient statistics or the raw data. Logarithm of the expected evidence E for H1:δ≠0 against H0:δ=0 using FBFs with different fractions b and the Bayes factor with the right Haar prior in case n=20 and varying standardized effects δ. The expected values were computed numerically using 1e7 randomly generated data sets. The horizontal dotted line denotes requirement (1) on a log-scale. *(abstract via OpenAlex)*

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2024-07-01

Christian P. Robert and Joshua Bon’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Christian P Robert, Joshua Bon

摘要

Central to our interest in Bayesian foundations and reference priors, the connection made in this paper between e-values and Bayes factors, despite the cautionary added label of an ‘Almost Bayesian Case’, could bring a formal way to define least favourable priors (Berger, 1985) in a testing environment, albeit in no clear connection with the reference priors of Bayarri and Garcia-Donato (2007). This derivation, formalized as Theorem 1, provides a reversal of the typical prior choice in testing. A more common stance is to hold the (subjectively chosen) prior distribution on H0 as a starting point—this is an obviously known entity. One may wonder if a dual perspective, i.e. when starting from the prior distribution on H0⁠, would prove fruitful, e.g. in leading to an optimal prior on H1⁠. As a plus, however, the approach found therein allows for improper priors on the nuisance parameters and thus somehow brings a stronger justification than (Berger et al., 1998) in favour of the permanence of an identical measure (on the nuisance parameters) across hypotheses and models. Regarding the sequential directions of the paper, we wonder at its potential connections with sequential Monte Carlo, for instance, towards conducting sequential model choice by constructing efficiently an amalgamated evidence value when the product of Bayes factors is not a Bayes factor (see, e.g. Buchholz et al., 2023). In conclusion, we congratulate the authors on this worthwhile endeavour but remain agnostic as to whether or not the proposal may prove fruitful in the derivation of reference testing procedures. Funded by the European Union (ERC-2022-SYG-OCEAN-101071601) and by a Prairie chair from the Agence Nationale de la Recherche (ANR-19-P3IA-0001). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the Agence Nationale de la Recherche, the European Union, or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. *(abstract via OpenAlex)*

JRSSB Testing & Uncertainty 讨论 / 评论 · 发表于 2024-07-01

Christian Hennig’s contribution to the Discussion of ‘Safe testing’ by Grünwald, de Heide, and Koolen

Christian Hennig

摘要

I appreciate the elegance of the theory regarding (RE)GRO(W) e-variables with optional continuation for combining evidence. Here are some rather problematic aspects. (1) The authors state that ‘there seems to be growing consensus that, in the context of hypothesis testing, p-values should either not be used at all, or at least, with utter care’. I see little consensus that p-values should not be used at all. The request for ‘utter care’ is sensible, but I do not believe that the use of any statistical inference without ‘utter care’ should be advertised. This is exactly what the authors are doing, calling their paper ‘Safe testing’. This is marketing language. Users may trust the supposed ‘safety’ of the approach and may not understand that care is still needed. Actually the term ‘safety’ refers to a specific feature of the proposed approach that is indeed desirable, but many other issues with statistical inference apply to the presented approach just as much (Hennig, 2022). (2) Model assumptions are an issue with p-values and parametric inference in general. In many standard situations, outliers can cause trouble.Considering a standard t-statistic, an extreme outlier will affect the standard error in a stronger way than the mean. The resulting t-statistic will have a low absolute value, potentially leading to p≈1⁠. Regarding traditional approaches to combine evidence, both Fisher’s method (considering −2∑i=1klnpi⁠, where p1,…,pk are p-values from k different studies) and meta-analysis based on standard errors will give a single study with p≈1 and large standard error little influence on the overall result. The situation with e-values is different. One e-value ≈0 can bring down the combining product of e-values. I suspect that outliers can be more problematic for combining e-values than for combining p-values. (3) Although the authors state that e-values can be interpreted in an attractive manner, I am not convinced that their interpretation is clearer than the interpretation of p-values. Probabilities (and consequently p-values) have been given monetary interpretations as well in the literature. When interpreting how large a payout should be seen as large enough for rejection, the authors recur to rejection probabilities, even though the relation between e-values and rejection probabilities is only an inequality, and despite criticizing the choices of rejection thresholds in standard testing as ‘arbitrary’. (4) The GRO-criterion, namely optimizing E(logE)⁠, may not look so attractive to a gambler who is interested in a payout of E rather than logE in a real gamble. *(abstract via OpenAlex)*

Biometrics Bayesian Methods Latent Variables & Mixtures · 发表于 2024-07-01

Bayesian inference for multivariate probit model with latent envelope

Kwangmin Lee, Yeonhee Park

摘要

The response envelope model proposed by Cook et al. (2010) is an efficient method to estimate the regression coefficient under the context of the multivariate linear regression model. It improves estimation efficiency by identifying material and immaterial parts of responses and removing the immaterial variation. The response envelope model has been investigated only for continuous response variables. In this paper, we propose the multivariate probit model with latent envelope, in short, the probit envelope model, as a response envelope model for multivariate binary response variables. The probit envelope model takes into account relations between Gaussian latent variables of the multivariate probit model by using the idea of the response envelope model. We address the identifiability of the probit envelope model by employing the essential identifiability concept and suggest a Bayesian method for the parameter estimation. We illustrate the probit envelope model via simulation studies and real-data analysis. The simulation studies show that the probit envelope model has the potential to gain efficiency in estimation compared to the multivariate probit model. The real data analysis shows that the probit envelope model is useful for multi-label classification.

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