اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدInteractive Martingale Tests for the Global Null
69
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Global null testing is a classical problem going back about a century to Fishers and Stouffers combination tests. In this work, we present simple martingale analogs of these classical tests, which are applicable in two distinct settings: (a) the online setting in which there is a possibly infinite sequence of $p$-values, and (b) the batch setting, where one uses prior knowledge to preorder the hypotheses. Through theory and simulations, we demonstrate that our martingale variants have higher power than their classical counterparts even when the preordering is only weakly informative. Finally, using a recent idea of masking $p$-values, we develop a novel interactive test for the global null that can take advantage of covariates and repeated user guidance to create a data-adaptive ordering that achieves higher detection power against structured alternatives.
قيم البحث
اقرأ أيضاً
Two-sample testing is a fundamental problem in statistics. Despite its long history, there has been renewed interest in this problem with the advent of high-dimensional and complex data. Specifically, in the machine learning literature, there have be
en recent methodological developments such as classification accuracy tests. The goal of this work is to present a regression approach to comparing multivariate distributions of complex data. Depending on the chosen regression model, our framework can efficiently handle different types of variables and various structures in the data, with competitive power under many practical scenarios. Whereas previous work has been largely limited to global tests which conceal much of the local information, our approach naturally leads to a local two-sample testing framework in which we identify local differences between multivariate distributions with statistical confidence. We demonstrate the efficacy of our approach both theoretically and empirically, under some well-known parametric and nonparametric regression methods. Our proposed methods are applied to simulated data as well as a challenging astronomy data set to assess their practical usefulness.
The prior distribution on parameters of a likelihood is the usual starting point for Bayesian uncertainty quantification. In this paper, we present a different perspective. Given a finite data sample $Y_{1:n}$ of size $n$ from an infinite population,
we focus on the missing $Y_{n+1:infty}$ as the source of statistical uncertainty, with the parameter of interest being known precisely given $Y_{1:infty}$. We argue that the foundation of Bayesian inference is to assign a predictive distribution on $Y_{n+1:infty}$ conditional on $Y_{1:n}$, which then induces a distribution on the parameter of interest. Demonstrating an application of martingales, Doob shows that choosing the Bayesian predictive distribution returns the conventional posterior as the distribution of the parameter. Taking this as our cue, we relax the predictive machine, avoiding the need for the predictive to be derived solely from the usual prior to posterior to predictive density formula. We introduce the martingale posterior distribution, which returns Bayesian uncertainty directly on any statistic of interest without the need for the likelihood and prior, and this distribution can be sampled through a computational scheme we name predictive resampling. To that end, we introduce new predictive methodologies for multivariate density estimation, regression and classification that build upon recent work on bivariate copulas.
Randomization (a.k.a. permutation) inference is typically interpreted as testing Fishers sharp null hypothesis that all effects are exactly zero. This hypothesis is often criticized as uninteresting and implausible. We show, however, that many random
ization tests are also valid for a bounded null hypothesis under which effects are all negative (or positive) for all units but otherwise heterogeneous. The bounded null is closely related to important concepts such as monotonicity and Pareto efficiency. Inverting tests of this hypothesis yields confidence intervals for the maximum (or minimum) individual treatment effect. We then extend randomization tests to infer other quantiles of individual effects, which can be used to infer the proportion of units with effects larger (or smaller) than any threshold. The proposed confidence intervals for all quantiles of individual effects are simultaneously valid, in the sense that no correction due to multiple analyses is needed. In sum, we provide a broader justification for Fisher randomization tests, and develop exact nonparametric inference for quantiles of heterogeneous individual effects. We illustrate our methods with simulations and applications, where we find that Stephenson rank statistics often provide the most informative results.
The standard concordance model of the Universe is based on the cosmological constant as the driver of accelerating expansion. This concordance model is being subjected to a growing range of inter-locking observations. In addition to using generic obs
ervational tests, one can also design tests that target the specific properties of the cosmological constant. These null tests do not rely on parametrisations of observables, but focus on quantities that are constant only if dark energy is a cosmological constant. We use supernova data in null tests that are based on the luminosity distance. In order to extract derivatives of the distance in a model-independent way, we use Gaussian Processes. We find that the concordance model is compatible with the Union 2.1 data, but the error bars are fairly large. Simulated datasets are generated for the DES supernova survey and we show that this survey will allow for a sharper null test of the cosmological constant if we assume the Universe is flat. Allowing for spatial curvature degrades the power of the null test.
It is possible to obtain a large Bayes Factor (BF) favoring the null hypothesis when both the null and alternative hypotheses have low likelihoods, and there are other hypotheses being ignored that are much more strongly supported by the data. As sam
ple sizes become large it becomes increasingly probable that a strong BF favouring a point null against a conventional Bayesian vague alternative co-occurs with a BF favouring various specific alternatives against the null. For any BF threshold q and sample mean, there is a value n such that sample sizes larger than n guarantee that although the BF comparing H0 against a conventional (vague) alternative exceeds q, nevertheless for some range of hypothetical {mu}, a BF comparing H0 against {mu} in that range falls below 1/q. This paper discusses the conditions under which this conundrum occurs and investigates methods for resolving it.
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
