arXiv subjects
Publications and source records attributed to Jesse Hemerik.
This text provides an introduction to multiple hypothesis testing. It covers various error criteria and testing procedures, and includes references to relevant R packages. An earlier version of this text served as the lecture notes for a PhD-level course on multiple testing.
MaxT is a highly popular resampling-based multiple testing procedure, which controls the Familywise Error Rate (FWER) and is powerful under dependence. This paper generalizes maxT to what we term ``multi-resolution'' False Discovery eXceedance (FDX) control. Basic FDX control means ensuring that the FDP -- the proportion of false discoveries among all rejections -- is at most $γ$ with probability at least $1-α$. Here $γ$ and $α$ are prespecified, small values between 0 and 1. The proposed method is in addition simultaneous, in the following way: the procedure outputs a single rejection threshold $q$, but ensures that with probability $1-α$, simultaneously over all stricter thresholds, the corresponding FDPs are also below $γ$. In particular, for a small set of hypotheses, the FDP bound is 0, i.e., the FWER is 0. Despite these additional, simultaneous guarantees, our method has power comparable to Romano-Wolf, the most powerful non-simultaneous FDX method. Further, our method is valid under the same assumptions. Thus, this paper shows that FDX methods can often be made simultaneous almost for free. The proposed method can be formulated as an extension of simultaneous approaches such as Hemerik, Solari and Goeman (2019), for the first time allowing for confidence envelopes with a data-dependent shape -- thus resolving a major limitation of such methods.
When testing many hypotheses, often we do not have strong expectations about the directions of the effects. In some situations however, the alternative hypotheses are that the parameters lie in a certain direction or interval, and it is in fact expected that most hypotheses are false. This is often the case when researchers perform multiple noninferiority or equivalence tests, e.g. when testing food safety with metabolite data. The goal is then to use data to corroborate the expectation that most hypotheses are false. We propose a nonparametric multiple testing approach that is powerful in such situations. If the user's expectations are wrong, our approach will still be valid but have low power. Of course all multiple testing methods become more powerful when appropriate one-sided instead of two-sided tests are used, but our approach often has superior power then. The proposed methods are not at all limited to safety testing and can be used for testing hypotheses about various kinds of parameters, such as coefficients of a model. The methods in this paper control the median of the false discovery proportion (FDP), which is the fraction of false discoveries among the rejected hypotheses. This approach is comparable to false discovery rate control, where one ensures that the mean rather than the median of the FDP is small. Our procedures make use of a symmetry property of the test statistics, do not require independence and have finite-sample properties.
Despite the versatility of generalized linear mixed models in handling complex experimental designs, they often suffer from misspecification and convergence problems. This makes inference on the values of coefficients problematic. To address these challenges, we propose a robust extension of the score-based statistical test using sign-flipping transformations. Our approach efficiently handles within-variance structure and heteroscedasticity, ensuring accurate regression coefficient testing. The approach is illustrated by analyzing the reduction of health issues over time for newly adopted children. The model is characterized by a binomial response with unbalanced frequencies and several categorical and continuous predictors. The proposed approach efficiently deals with critical problems related to longitudinal nonlinear models, surpassing common statistical approaches such as generalized estimating equations and generalized linear mixed models.
A fundamental assumption of classical hypothesis testing is that the significance threshold $α$ is chosen independently from the data. The validity of confidence intervals likewise relies on choosing $α$ beforehand. We point out that the independence of $α$ is guaranteed in practice because, in most fields, there exists one standard $α$ that everyone uses -- so that $α$ is automatically independent of everything. However, there have been recent calls to decrease $α$ from $0.05$ to $0.005$. We note that this may lead to multiple accepted standard thresholds within one scientific field. For example, different journals may require different significance thresholds. As a consequence, some researchers may be tempted to conveniently choose their $α$ based on their p-value. We use examples to illustrate that this severely invalidates hypothesis tests, and mention some potential solutions.
In many applied sciences a popular analysis strategy for high-dimensional data is to fit many multivariate generalized linear models in parallel. This paper presents a novel approach to address the resulting multiple testing problem by combining a recently developed sign-flip test with permutation-based multiple-testing procedures. Our method builds upon the univariate standardized flip-scores test which offers robustness against misspecified variances in generalized linear models, a crucial feature in high-dimensional settings where comprehensive model validation is particularly challenging. We extend this approach to the multivariate setting, enabling adaptation to unknown response correlation structures. This approach yields relevant power improvements over conventional multiple testing methods when correlation is present.
Generalized linear models usually assume a common dispersion parameter, an assumption that is seldom true in practice. Consequently, standard parametric methods may suffer appreciable loss of type I error control. As an alternative, we present a semi-parametric group-invariance method based on sign flipping of score contributions. Our method requires only the correct specification of the mean model, but is robust against any misspecification of the variance. We present tests for single as well as multiple regression coefficients. The test is asymptotically valid but shows excellent performance in small samples. We illustrate the method using RNA sequencing count data, for which it is difficult to model the overdispersion correctly. The method is available in the R library flipscores.
We introduce a multiple testing procedure that controls the median of the proportion of false discoveries (FDP) in a flexible way. The procedure only requires a vector of p-values as input and is comparable to the Benjamini-Hochberg method, which controls the mean of the FDP. Our method allows freely choosing one or several values of alpha after seeing the data -- unlike Benjamini-Hochberg, which can be very liberal when alpha is chosen post hoc. We prove these claims and illustrate them with simulations. Our procedure is inspired by a popular estimator of the total number of true hypotheses. We adapt this estimator to provide simultaneously median unbiased estimators of the FDP, valid for finite samples. This simultaneity allows for the claimed flexibility. Our approach does not assume independence. The time complexity of our method is linear in the number of hypotheses, after sorting the p-values.
There exists no consensus on the meaning of the term "randomization test". Contradicting uses of the term are leading to confusion, misunderstandings and indeed invalid data analyses. As we point out, a main source of the confusion is that the term was not explicitly defined when it was first used in the 1930's. Later authors made clear proposals to reach a consensus regarding the term. This resulted in some level of agreement around the 1970's. However, in the last few decades, the term has often been used in ways that contradict these proposals. This paper provides an overview of the history of the term per se, for the first time tracing it back to 1937. This will hopefully lead to more agreement on terminology and less confusion on the related fundamental concepts.
We propose a permutation-based method for testing a large collection of hypotheses simultaneously. Our method provides lower bounds for the number of true discoveries in any selected subset of hypotheses. These bounds are simultaneously valid with high confidence. The methodology is particularly useful in functional Magnetic Resonance Imaging cluster analysis, where it provides a confidence statement on the percentage of truly activated voxels within clusters of voxels, avoiding the well-known spatial specificity paradox. We offer a user-friendly tool to estimate the percentage of true discoveries for each cluster while controlling the family-wise error rate for multiple testing and taking into account that the cluster was chosen in a data-driven way. The method adapts to the spatial correlation structure that characterizes functional Magnetic Resonance Imaging data, gaining power over parametric approaches.
Non-parametric tests based on permutation, rotation or sign-flipping are examples of group-invariance tests. These tests test invariance of the null distribution under a set of transformations that has a group structure, in the algebraic sense. Such groups are often huge, which makes it computationally infeasible to test using the entire group. Hence, it is standard practice to test using a randomly sampled set of transformations from the group. This random sample still needs to be substantial to obtain good power and replicability. We improve upon this standard practice by using a well-designed subgroup of transformations instead of a random sample. The resulting subgroup-invariance test is still exact, as invariance under a group implies invariance under its subgroups. We illustrate this in a generalized location model and obtain more powerful tests based on the same number of transformations. In particular, we show that a subgroup-invariance test is consistent for lower signal-to-noise ratios than a test based on a random sample. For the special case of a normal location model and a particular design of the subgroup, we show that the power improvement is equivalent to the power difference between a Monte Carlo $Z$-test and a Monte Carlo $t$-test.
We consider the class of all multiple testing methods controlling tail probabilities of the false discovery proportion, either for one random set or simultaneously for many such sets. This class encompasses methods controlling familywise error rate, generalized familywise error rate, false discovery exceedance, joint error rate, simultaneous control of all false discovery proportions, and others, as well as seemingly unrelated methods such as gene set testing in genomics and cluster inference methods in neuroimaging. We show that all such methods are either equivalent to a closed testing method, or are uniformly improved by one. Moreover, we show that a closed testing method is admissible as a method controlling tail probabilities of false discovery proportions if and only if all its local tests are admissible. This implies that, when designing such methods, it is sufficient to restrict attention to closed testing methods only. We demonstrate the practical usefulness of this design principle by constructing a uniform improvement of a recently proposed method.
Permutation testing in linear models, where the number of nuisance coefficients is smaller than the sample size, is a well-studied topic. The common approach of such tests is to permute residuals after regressing on the nuisance covariates. Permutation-based tests are valuable in particular because they can be highly robust to violations of the standard linear model, such as non-normality and heteroscedasticity. Moreover, in some cases they can be combined with existing, powerful permutation-based multiple testing methods. Here, we propose permutation tests for models where the number of nuisance coefficients exceeds the sample size. The performance of the novel tests is investigated with simulations. In a wide range of simulation scenarios our proposed permutation methods provided appropriate type I error rate control, unlike some competing tests, while having good power.
The statistical literature is known to be inconsistent in the use of the terms "permutation test" and "randomization test". Several authors succesfully argue that these terms should be used to refer to two distinct classes of tests and that there are major conceptual differences between these classes. The present paper explains an important difference in mathematical reasoning between these classes: a permutation test fundamentally requires that the set of permutations has a group structure, in the algebraic sense; the reasoning behind a randomization test is not based on such a group structure and it is possible to use an experimental design that does not correspond to a group. In particular, we can use a randomization scheme where the number of possible treatment patterns is larger than in standard experimental designs. This leads to exact \emph{p}-values of improved resolution, providing increased power for very small significance levels, at the cost of decreased power for larger significance levels. We discuss applications in randomized trials and elsewhere. Further, we explain that Fisher's famous Lady Tasting Tea experiment, which is commonly referred to as the first permutation test, is in fact a randomization test. This distinction is important to avoid confusion and invalid tests.
Mediation analysis in high-dimensional settings often involves identifying potential mediators among a large number of measured variables. For this purpose, a two-step familywise error rate procedure called ScreenMin has been recently proposed (Djordjilović et al. 2019). In ScreenMin, variables are first screened and only those that pass the screening are tested. The proposed threshold for selection has been shown to guarantee asymptotic familywise error rate. In this work, we investigate the impact of the selection threshold on the finite sample familywise error rate. We derive a power maximizing selection threshold and show that it is well approximated by an adaptive threshold of Wang et al. (2016). We illustrate the investigated procedures on a case-control study examining the effect of fish intake on the risk of colorectal adenoma.
Generalized linear models are often misspecified due to overdispersion, heteroscedasticity and ignored nuisance variables. Existing quasi-likelihood methods for testing in misspecified models often do not provide satisfactory type-I error rate control. We provide a novel semi-parametric test, based on sign-flipping individual score contributions. The tested parameter is allowed to be multi-dimensional and even high-dimensional. Our test is often robust against the mentioned forms of misspecification and provides better type-I error control than its competitors. When nuisance parameters are estimated, our basic test becomes conservative. We show how to take nuisance estimation into account to obtain an asymptotically exact test. Our proposed test is asymptotically equivalent to its parametric counterpart.
Mediation analysis in high-dimensional settings often involves identifying potential mediators among a large number of measured variables. For this purpose, a two step familywise error rate (FWER) procedure called ScreenMin has been recently proposed (Djordjilović et al. 2019). In ScreenMin, variables are first screened and only those that pass the screening are tested. The proposed threshold for selection has been shown to guarantee asymptotic FWER. In this work, we investigate the impact of the selection threshold on the finite sample FWER. We derive power maximizing selection threshold and show that it is well approximated by an adaptive threshold of Wang et al. (2016). We study the performance of the proposed procedures in a simulation study, and apply them to a case-control study examining the effect of fish intake on the risk of colorectal adenoma.
When permutation methods are used in practice, often a limited number of random permutations are used to decrease the computational burden. However, most theoretical literature assumes that the whole permutation group is used, and methods based on random permutations tend to be seen as approximate. There exists a very limited amount of literature on exact testing with random permutations and only recently a thorough proof of exactness was given. In this paper we provide an alternative proof, viewing the test as a "conditional Monte Carlo test" as it has been called in the literature. We also provide extensions of the result. Importantly, our results can be used to prove properties of various multiple testing procedures based on random permutations.