arXiv subjects

Elias Chaibub Neto

Publications and source records attributed to Elias Chaibub Neto.

At least 19 recordsLinked to original sources

arXiv2024

Computationally efficient permutation tests for the multivariate two-sample problem based on energy distance or maximum mean discrepancy statistics

Non-parametric two-sample tests based on energy distance or maximum mean discrepancy are widely used statistical tests for comparing multivariate data from two populations. While these tests enjoy desirable statistical properties, their test statistics can be expensive to compute as they require the computation of 3 distinct Euclidean distance (or kernel) matrices between samples, where the time complexity of each of these computations (namely, $O(n_{x}^2 p)$, $O(n_{y}^2 p)$, and $O(n_{x} n_{y} p)$) scales quadratically with the number of samples ($n_x$, $n_y$) and linearly with the number of variables ($p$). Since the standard permutation test requires repeated re-computations of these expensive statistics it's application to large datasets can become unfeasible. While several statistical approaches have been proposed to mitigate this issue, they all sacrifice desirable statistical properties to decrease the computational cost (e.g., trade computation speed by a decrease in statistical power). A better computational strategy is to first pre-compute the Euclidean distance (kernel) matrix of the concatenated data, and then permute indexes and retrieve the corresponding elements to compute the re-sampled statistics. While this strategy can reduce the computation cost relative to the standard permutation test, it relies on the computation of a larger Euclidean distance (kernel) matrix with complexity $O((n_x + n_y)^2 p)$. In this paper, we present a novel computationally efficient permutation algorithm which only requires the pre-computation of the 3 smaller matrices and achieves large computational speedups without sacrificing finite-sample validity or statistical power. We illustrate its computational gains in a series of experiments and compare its statistical power to the current state-of-the-art approach for balancing computational cost and statistical performance.

stat.CO↗

arXiv2023

Statistical disclosure control for numeric microdata via sequential joint probability preserving data shuffling

Traditional perturbative statistical disclosure control (SDC) approaches such as microaggregation, noise addition, rank swapping, etc, perturb the data in an ``ad-hoc" way in the sense that while they manage to preserve some particular aspects of the data, they end up modifying others. Synthetic data approaches based on the fully conditional specification data synthesis paradigm, on the other hand, aim to generate new datasets that follow the same joint probability distribution as the original data. These synthetic data approaches, however, rely either on parametric statistical models, or non-parametric machine learning models, which need to fit well the original data in order to generate credible and useful synthetic data. Another important drawback is that they tend to perform better when the variables are synthesized in the correct causal order (i.e., in the same order as the true data generating process), which is often unknown in practice. To circumvent these issues, we propose a fully non-parametric and model free perturbative SDC approach that approximates the joint distribution of the original data via sequential applications of restricted permutations to the numerical microdata (where the restricted permutations are guided by the joint distribution of a discretized version of the data). Empirical comparisons against popular SDC approaches, using both real and simulated datasets, suggest that the proposed approach is competitive in terms of the trade-off between confidentiality and data utility.

stat.AP↗

arXiv2023

A note on bounded distance-based information loss metrics for statistical disclosure control of numeric microdata

In the field of statistical disclosure control, the tradeoff between data confidentiality and data utility is measured by comparing disclosure risk and information loss metrics. Distance based metrics such as the mean absolute error (MAE), mean squared error (MSE), mean variation (IL1), and its scaled alternative (IL1s) are popular information loss measures for numerical microdata. However, the fact that these measures are unbounded makes it is difficult to compare them against disclosure risk measures which are usually bounded between 0 and 1. In this note, we propose rank-based versions of the MAE and MSE metrics that are bounded in the same range as the disclosure risk metrics. We empirically compare the proposed bounded metrics against the distance-based metrics in a series of experiments where the metrics are evaluated over multiple masked datasets, generated by the application of increasing amounts of perturbation (e.g., by adding increasing amounts of noise). Our results show that the proposed bounded metrics produce similar rankings as the traditional ones (as measured by Spearman correlation), suggesting that they are a viable additions to the toolbox of distance-based information loss metrics currently in use in the SDC literature.

stat.AP↗

arXiv2020

Towards causality-aware predictions in static anticausal machine learning tasks: the linear structural causal model case

We propose a counterfactual approach to train ``causality-aware" predictive models that are able to leverage causal information in static anticausal machine learning tasks (i.e., prediction tasks where the outcome influences the features). In applications plagued by confounding, the approach can be used to generate predictions that are free from the influence of observed confounders. In applications involving observed mediators, the approach can be used to generate predictions that only capture the direct or the indirect causal influences. Mechanistically, we train supervised learners on (counterfactually) simulated features which retain only the associations generated by the causal relations of interest. We focus on linear models, where analytical results connecting covariances, causal effects, and prediction mean squared errors are readily available. Quite importantly, we show that our approach does not require knowledge of the full causal graph. It suffices to know which variables represent potential confounders and/or mediators. We discuss the stability of the method with respect to dataset shifts generated by selection biases and validate the approach using synthetic data experiments.

stat.AP↗

arXiv2020

Causality-aware counterfactual confounding adjustment for feature representations learned by deep models

Causal modeling has been recognized as a potential solution to many challenging problems in machine learning (ML). Here, we describe how a recently proposed counterfactual approach developed to deconfound linear structural causal models can still be used to deconfound the feature representations learned by deep neural network (DNN) models. The key insight is that by training an accurate DNN using softmax activation at the classification layer, and then adopting the representation learned by the last layer prior to the output layer as our features, we have that, by construction, the learned features will fit well a (multi-class) logistic regression model, and will be linearly associated with the labels. As a consequence, deconfounding approaches based on simple linear models can be used to deconfound the feature representations learned by DNNs. We validate the proposed methodology using colored versions of the MNIST dataset. Our results illustrate how the approach can effectively combat confounding and improve model stability in the context of dataset shifts generated by selection biases.

cs.LG↗

arXiv2020

Stable predictions for health related anticausal prediction tasks affected by selection biases: the need to deconfound the test set features

In health related machine learning applications, the training data often corresponds to a non-representative sample from the target populations where the learners will be deployed. In anticausal prediction tasks, selection biases often make the associations between confounders and the outcome variable unstable across different target environments. As a consequence, the predictions from confounded learners are often unstable, and might fail to generalize in shifted test environments. Stable prediction approaches aim to solve this problem by producing predictions that are stable across unknown test environments. These approaches, however, are sometimes applied to the training data alone with the hope that training an unconfounded model will be enough to generate stable predictions in shifted test sets. Here, we show that this is insufficient, and that improved stability can be achieved by deconfounding the test set features as well. We illustrate these observations using both synthetic data and real world data from a mobile health study.

stat.ML↗

arXiv2020

Causality-aware counterfactual confounding adjustment as an alternative to linear residualization in anticausal prediction tasks based on linear learners

Linear residualization is a common practice for confounding adjustment in machine learning (ML) applications. Recently, causality-aware predictive modeling has been proposed as an alternative causality-inspired approach for adjusting for confounders. The basic idea is to simulate counterfactual data that is free from the spurious associations generated by the observed confounders. In this paper, we compare the linear residualization approach against the causality-aware confounding adjustment in anticausal prediction tasks, and show that the causality-aware approach tends to (asymptotically) outperform the residualization adjustment in terms of predictive performance in linear learners. Importantly, our results still holds even when the true model is not linear. We illustrate our results in both regression and classification tasks, where we compared the causality-aware and residualization approaches using mean squared errors and classification accuracy in synthetic data experiments where the linear regression model is mispecified, as well as, when the linear model is correctly specified. Furthermore, we illustrate how the causality-aware approach is more stable than residualization with respect to dataset shifts in the joint distribution of the confounders and outcome variables.

stat.ML↗

arXiv2019

Causality-based tests to detect the influence of confounders on mobile health diagnostic applications: a comparison with restricted permutations

Machine learning practice is often impacted by confounders. Confounding can be particularly severe in remote digital health studies where the participants self-select to enter the study. While many different confounding adjustment approaches have been proposed in the literature, most of these methods rely on modeling assumptions, and it is unclear how robust they are to violations of these assumptions. This realization has recently motivated the development of restricted permutation methods to quantify the influence of observed confounders on the predictive performance of a machine learning models and evaluate if confounding adjustment methods are working as expected. In this paper we show, nonetheless, that restricted permutations can generate biased estimates of the contribution of the confounders to the predictive performance of a learner, and we propose an alternative approach to tackle this problem. By viewing a classification task from a causality perspective, we are able to leverage conditional independence tests between predictions and test set labels and confounders in order to detect confounding on the predictive performance of a classifier. We illustrate the application of our causality-based approach to data collected from mHealth study in Parkinson's disease.

stat.AP↗

arXiv2019

Indicators of retention in remote digital health studies: A cross-study evaluation of 100,000 participants

Digital technologies such as smartphones are transforming the way scientists conduct biomedical research using real-world data. Several remotely-conducted studies have recruited thousands of participants over a span of a few months. Unfortunately, these studies are hampered by substantial participant attrition, calling into question the representativeness of the collected data including generalizability of findings from these studies. We report the challenges in retention and recruitment in eight remote digital health studies comprising over 100,000 participants who participated for more than 850,000 days, completing close to 3.5 million remote health evaluations. Survival modeling surfaced several factors significantly associated(P < 1e-16) with increase in median retention time i) Clinician referral(increase of 40 days), ii) Effect of compensation (22 days), iii) Clinical conditions of interest to the study (7 days) and iv) Older adults(4 days). Additionally, four distinct patterns of daily app usage behavior that were also associated(P < 1e-10) with participant demographics were identified. Most studies were not able to recruit a representative sample, either demographically or regionally. Combined together these findings can help inform recruitment and retention strategies to enable equitable participation of populations in future digital health research.

stat.AP↗

arXiv2018

Using permutations to assess confounding in machine learning applications for digital health

Clinical machine learning applications are often plagued with confounders that can impact the generalizability and predictive performance of the learners. Confounding is especially problematic in remote digital health studies where the participants self-select to enter the study, thereby making it challenging to balance the demographic characteristics of participants. One effective approach to combat confounding is to match samples with respect to the confounding variables in order to balance the data. This procedure, however, leads to smaller datasets and hence impact the inferences drawn from the learners. Alternatively, confounding adjustment methods that make more efficient use of the data (e.g., inverse probability weighting) usually rely on modeling assumptions, and it is unclear how robust these methods are to violations of these assumptions. Here, rather than proposing a new approach to control for confounding, we develop novel permutation based statistical methods to detect and quantify the influence of observed confounders, and estimate the unconfounded performance of the learner. Our tools can be used to evaluate the effectiveness of existing confounding adjustment methods. We illustrate their application using real-life data from a Parkinson's disease mobile health study collected in an uncontrolled environment.

stat.AP↗

arXiv2018

Using permutations to detect, quantify and correct for confounding in machine learning predictions

Clinical machine learning applications are often plagued with confounders that are clinically irrelevant, but can still artificially boost the predictive performance of the algorithms. Confounding is especially problematic in mobile health studies run "in the wild", where it is challenging to balance the demographic characteristics of participants that self select to enter the study. An effective approach to remove the influence of confounders is to match samples in order to improve the balance in the data. The caveat is that we end-up with a smaller number of participants to train and evaluate the machine learning algorithm. Alternative confounding adjustment methods that make more efficient use of the data (e.g., inverse probability weighting) usually rely on modeling assumptions, and it is unclear how robust these methods are to violations of these assumptions. Here, rather than proposing a new approach to prevent/reduce the learning of confounding signals by a machine learning algorithm, we develop novel statistical tools to detect, quantify and correct for the influence of observed confounders. Our tools are based on restricted and standard permutation approaches and can be used to evaluate how well a confounding adjustment method is actually working. We use restricted permutations to test if an algorithm has learned disease signal in the presence of confounding signal, and to develop a novel statistical test to detect confounding learning per se. Furthermore, we prove that restricted permutations provide an alternative method to compute partial correlations, and use this result as a motivation to develop a novel approach to estimate the corrected predictive performance of a learner. We evaluate the statistical properties of our methods in simulation studies.

stat.AP↗

arXiv2018

Learning Disease vs Participant Signatures: a permutation test approach to detect identity confounding in machine learning diagnostic applications

Recently, Saeb et al (2017) showed that, in diagnostic machine learning applications, having data of each subject randomly assigned to both training and test sets (record-wise data split) can lead to massive underestimation of the cross-validation prediction error, due to the presence of "subject identity confounding" caused by the classifier's ability to identify subjects, instead of recognizing disease. To solve this problem, the authors recommended the random assignment of the data of each subject to either the training or the test set (subject-wise data split). The adoption of subject-wise split has been criticized in Little et al (2017), on the basis that it can violate assumptions required by cross-validation to consistently estimate generalization error. In particular, adopting subject-wise splitting in heterogeneous data-sets might lead to model under-fitting and larger classification errors. Hence, Little et al argue that perhaps the overestimation of prediction errors with subject-wise cross-validation, rather than underestimation with record-wise cross-validation, is the reason for the discrepancies between prediction error estimates generated by the two splitting strategies. In order to shed light on this controversy, we focus on simpler classification performance metrics and develop permutation tests that can detect identity confounding. By focusing on permutation tests, we are able to evaluate the merits of record-wise and subject-wise data splits under more general statistical dependencies and distributional structures of the data, including situations where cross-validation breaks down. We illustrate the application of our tests using synthetic and real data from a Parkinson's disease study.

stat.AP↗

arXiv2018

Detecting Learning vs Memorization in Deep Neural Networks using Shared Structure Validation Sets

The roles played by learning and memorization represent an important topic in deep learning research. Recent work on this subject has shown that the optimization behavior of DNNs trained on shuffled labels is qualitatively different from DNNs trained with real labels. Here, we propose a novel permutation approach that can differentiate memorization from learning in deep neural networks (DNNs) trained as usual (i.e., using the real labels to guide the learning, rather than shuffled labels). The evaluation of weather the DNN has learned and/or memorized, happens in a separate step where we compare the predictive performance of a shallow classifier trained with the features learned by the DNN, against multiple instances of the same classifier, trained on the same input, but using shuffled labels as outputs. By evaluating these shallow classifiers in validation sets that share structure with the training set, we are able to tell apart learning from memorization. Application of our permutation approach to multi-layer perceptrons and convolutional neural networks trained on image data corroborated many findings from other groups. Most importantly, our illustrations also uncovered interesting dynamic patterns about how DNNs memorize over increasing numbers of training epochs, and support the surprising result that DNNs are still able to learn, rather than only memorize, when trained with pure Gaussian noise as input.

stat.ML↗

arXiv2017

Towards personalized causal inference of medication response in mobile health: an instrumental variable approach for randomized trials with imperfect compliance

Mobile health studies can leverage longitudinal sensor data from smartphones to guide the application of personalized medical interventions. In this paper, we propose that adoption of an instrumental variable approach for randomized trials with imperfect compliance provides a natural framework for personalized causal inference of medication response in mobile health studies. Randomized treatment suggestions can be easily delivered to the study participants via electronic messages popping up on the smart-phone screen. Under quite general assumptions we can identify the causal effect of the actual treatment on the response in the presence of unobserved confounders. We implement a personalized randomization test of the null hypothesis of no causal effect of treatment on response, and evaluate its performance in a large scale simulation study encompassing data generated from linear and non-linear time series models under several simulation conditions. In particular, we evaluate the empirical power of the proposed test under varying degrees of compliance between the suggested and actual treatment adopted by the participant. Our investigations provide encouraging results in terms of power and control of type I error rates. Finally, we compare the proposed instrumental variable approach to a simple intent-to-treat strategy, and develop randomization confidence intervals for the causal effects.

stat.AP↗

arXiv2017

On the analysis of personalized medication response and classification of case vs control patients in mobile health studies: the mPower case study

In this work we provide a couple of contributions to the analysis of longitudinal data collected by smartphones in mobile health applications. First, we propose a novel statistical approach to disentangle personalized treatment and "time-of-the-day" effects in observational studies. Under the assumption of no unmeasured confounders, we show how to use conditional independence relations in the data in order to determine if a difference in performance between activity tasks performed before and after the participant has taken medication, are potentially due to an effect of the medication or to a "time-of-the-day" effect (or still to both). Second, we show that smartphone data collected from a given study participant can represent a "digital fingerprint" of the participant, and that classifiers of case/control labels, constructed using longitudinal data, can show artificially improved performance when data from each participant is included in both training and test sets. We illustrate our contributions using data collected during the first 6 months of the mPower study.

stat.AP↗

arXiv2016

Reducing overfitting in challenge-based competitions

Over-fitting is a dreaded foe in challenge-based competitions. Because participants rely on public leaderboards to evaluate and refine their models, there is always the danger they might over-fit to the holdout data supporting the leaderboard. The recently published Ladder algorithm aims to address this problem by preventing the participants from exploiting willingly or inadvertently minor fluctuations in public leaderboard scores during model refinement. In this paper, we report a vulnerability of the Ladder that induces severe over-fitting of the leaderboard when the sample size is small. To circumvent this attack, we propose a variation of the Ladder that releases a bootstrapped estimate of the public leaderboard score instead of providing participants with a direct measure of performance. We also extend the scope of the Ladder to arbitrary performance metrics by relying on a more broadly applicable testing procedure based on the Bayesian bootstrap. Our method makes it possible to use a leaderboard, with the technical and social advantages that it provides, even in cases where data is scant.

stat.AP↗

arXiv2016

Using instrumental variables to disentangle treatment and placebo effects in blinded and unblinded randomized clinical trials influenced by unmeasured confounders

Clinical trials traditionally employ blinding as a design mechanism to reduce the influence of placebo effects. In practice, however, it can be difficult or impossible to blind study participants and unblinded trials are common in medical research. Here we show how instrumental variables can be used to quantify and disentangle treatment and placebo effects in randomized clinical trials comparing control and active treatments in the presence of confounders. The key idea is to use randomization to separately manipulate treatment assignment and psychological encouragement messages that increase the participants' desire for improved symptoms. The proposed approach is able to improve the estimation of treatment effects in blinded studies and, most importantly, opens the doors to account for placebo effects in unblinded trials.

stat.AP↗

arXiv2015

Identifying robust communities and multi-community nodes by combining top-down and bottom-up approaches to clustering

Biological functions are carried out by groups of interacting molecules, cells or tissues, known as communities. Membership in these communities may overlap when biological components are involved in multiple functions. However, traditional clustering methods detect non-overlapping communities. These detected communities may also be unstable and difficult to replicate, because traditional methods are sensitive to noise and parameter settings. These aspects of traditional clustering methods limit our ability to detect biological communities, and therefore our ability to understand biological functions. To address these limitations and detect robust overlapping biological communities, we propose an unorthodox clustering method called SpeakEasy which identifies communities using top-down and bottom-up approaches simultaneously. Specifically, nodes join communities based on their local connections, as well as global information about the network structure. This method can quantify the stability of each community, automatically identify the number of communities, and quickly cluster networks with hundreds of thousands of nodes. SpeakEasy shows top performance on synthetic clustering benchmarks and accurately identifies meaningful biological communities in a range of datasets, including: gene microarrays, protein interactions, sorted cell populations, electrophysiology and fMRI brain imaging.

cs.CE↗