Jonathan Rosenblatt
Research Subjects
Jonathan Rosenblatt has a total of 50 co-authors. Here are the top co-authors by number of shared publications:
Itai Kloog
Department of Geography and Environmental Development, Ben Gurion University of the Negev, Be'er Sheva 84105, Israel
3 shared publicationsAldo Solari
University of Milano Bicocca, Department of Economics, Management and Statistics, Italy; NeuroMI- Milan Center for Neuroscience, Italy
2 shared publications- Allan Carpenter Just
Department of Epidemiology, Brown University School of Public Health, Providence, RI, USA; Institute at Brown for Environment and Society, Brown University, Providence, RI, USA; Department of Environmental Medicine and Public Health, Icahn School of Medicine at Mount Sinai, New York, NY, USA
2 shared publications - Ian Hough
Univ. Grenoble Alpes, CNRS, INRAE, IRD, Grenoble INP, IGE
2 shared publications - Jelle J. Goeman
Department of Biomedical Data Sciences Leiden University Medical Center Leiden the Netherlands
2 shared publications - Livio Finos
Department of Statistical Sciences, University of Padua, 35121 Padua, Italy
2 shared publications
Latest Publications
- Open Access
- 1 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 35827-35843
Geometry Design for DOA Estimation in Seismic 2D-Arrays: Simulation Study
Sensor array geometry has a direct impact on the direction-of-arrival (DOA) estimation of a seismic signal. In this paper, we design a planar array that aims to optimize the DOA estimation of a narrowband signal in the sense of the minimum mean-squared-periodic-error (MSPE) obtained by the maximum a-posteriori (MAP) estimator of the DOA. We investigate the MSPE of the MAP estimator as a main design criterion and compare it with the criteria: 1) the cyclic Bayesian Cramér-Rao Bound (CBCRB); and 2) the expected log-likelihood (ELL). The theoretical properties of these criteria are discussed. We show that minimizing the CBCRB is equivalent to maximizing the expected Fisher information matrix. Additionally, maximizing the ELL under a uniform prior is equivalent to minimizing the Kullback-Leibler divergence between the posterior PDF and its estimation. The criteria are compared across three different array geometries, specifically: small arrays, uniform circular arrays (UCAs), and concentric circular arrays (CCAs). Simulation results show that 1) direct MAP-MSPE optimization notably exceeds CBCRB- and ELL-based designs, especially in small arrays; 2) UCAs have suboptimal performance compared to non-circular arrays in many scenarios; 3) under the MAP-MSPE criterion, CCAs match unconstrained design performance with lower computational complexity, making them preferable for smaller arrays; 4) for CCAs and larger UCAs, CBCRB and MAP-MSPE designs yield similar results, while the ELL design excels in the case of small UCAs. Our results highlight the need for selecting suitable array geometries and design criteria in accordance with the scenario and array size in order to achieve the best DOA estimation results.
- 1 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 35827-35843
Open AccessGeometry Design for DOA Estimation in Seismic 2D-Arrays: Simulation Study
Sensor array geometry has a direct impact on the direction-of-arrival (DOA) estimation of a seismic signal. In this paper, we design a planar array that aims to optimize the DOA estimation of a narrowband signal in the sense of the minimum mean-squared-periodic-error (MSPE) obtained by the maximum a-posteriori (MAP) estimator of the DOA. We investigate the MSPE of the MAP estimator as a main design criterion and compare it with the criteria: 1) the cyclic Bayesian Cramér-Rao Bound (CBCRB); and 2) the expected log-likelihood (ELL). The theoretical properties of these criteria are discussed. We show that minimizing the CBCRB is equivalent to maximizing the expected Fisher information matrix. Additionally, maximizing the ELL under a uniform prior is equivalent to minimizing the Kullback-Leibler divergence between the posterior PDF and its estimation. The criteria are compared across three different array geometries, specifically: small arrays, uniform circular arrays (UCAs), and concentric circular arrays (CCAs). Simulation results show that 1) direct MAP-MSPE optimization notably exceeds CBCRB- and ELL-based designs, especially in small arrays; 2) UCAs have suboptimal performance compared to non-circular arrays in many scenarios; 3) under the MAP-MSPE criterion, CCAs match unconstrained design performance with lower computational complexity, making them preferable for smaller arrays; 4) for CCAs and larger UCAs, CBCRB and MAP-MSPE designs yield similar results, while the ELL design excels in the case of small UCAs. Our results highlight the need for selecting suitable array geometries and design criteria in accordance with the scenario and array size in order to achieve the best DOA estimation results.
- 13 June 2022
- Published by Elsevier in Current Biology
- Vol. 32  (14) , 3203-3209.e3
A distinct route for efficient learning and generalization in autism
- 13 June 2022
- Published by Elsevier in Current Biology
- Vol. 32  (14) , 3203-3209.e3
A distinct route for efficient learning and generalization in autism
- 27 October 2021
- Published by Taylor & Francis
- p. 183-210
Prevalence Estimation
Given a multivariate parameter, ? and appropriate data, prevalence estimation deals with the counting of the number of entries in ? that depart from their hypothesized null values. The problem has two main motivations: First, in the case, a population consists of two sub-populations, we may want to know the prevalence of each. Examples include a sub-population of respondents and a sub-population of non-respondents in personalized medicine; active and inactive subjects in neuroimaging; associated and non-associated genes in genetic studies (GWAS); tolerant and intolerant animals in toxicology, etc. Second, various multiple testing algorithms may bene?t from knowledge of the effect’s prevalence. The Adaptive Benjamini-Hochberg algorithm is one of many such algorithms. The algorithm """adapts""" by estimating the signal’s prevalence before a multiple testing stage. In this chapter, we will cover the vast literature on prevalence estimators, try to organize it along design principles and statistical guarantees, with recommendations to the practitioner.
- 27 October 2021
- Published by Taylor & Francis
- p. 183-210
Prevalence Estimation
Given a multivariate parameter, ? and appropriate data, prevalence estimation deals with the counting of the number of entries in ? that depart from their hypothesized null values. The problem has two main motivations: First, in the case, a population consists of two sub-populations, we may want to know the prevalence of each. Examples include a sub-population of respondents and a sub-population of non-respondents in personalized medicine; active and inactive subjects in neuroimaging; associated and non-associated genes in genetic studies (GWAS); tolerant and intolerant animals in toxicology, etc. Second, various multiple testing algorithms may bene?t from knowledge of the effect’s prevalence. The Adaptive Benjamini-Hochberg algorithm is one of many such algorithms. The algorithm """adapts""" by estimating the signal’s prevalence before a multiple testing stage. In this chapter, we will cover the vast literature on prevalence estimators, try to organize it along design principles and statistical guarantees, with recommendations to the practitioner.
- 13 August 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 59  (6) , 5197-5205
A Domain Adaptation Approach for Performance Estimation of Spatial Predictions
Spatial predictions, like other supervised learning tasks, require some criterion for a predictor's quality. Typical data-splitting schemes, such as holdouts and k-fold cross-validation, ignore the fact that the training data are usually not available where predictions are being made. The common data-splitting schemes are thus biased estimates of a predictor's performance, which in turn may lead to choosing suboptimal predictors. In this contribution, we borrow ideas from the domain adaptation machine-learning literature, to suggest the importance-weighted source risk (IWSR). IWSR is a principled approach for weighting the prediction risk, which allows the practitioner to explicitly state the target locations for prediction. IWSR essentially consists of down-weighting training locations and up-weighting target locations. We show that, unlike the usual (unweighted) empirical risk, IWSR is an unbiased estimator of the prediction error. Equipped with this risk estimator, we use it to learn a model in the empirical risk minimization framework and to evaluate the existing predictors. We show the superiority of this weighted risk, using both simulated data and an empirical control: air-temperature prediction in France.
- 13 August 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 59  (6) , 5197-5205
A Domain Adaptation Approach for Performance Estimation of Spatial Predictions
Spatial predictions, like other supervised learning tasks, require some criterion for a predictor's quality. Typical data-splitting schemes, such as holdouts and k-fold cross-validation, ignore the fact that the training data are usually not available where predictions are being made. The common data-splitting schemes are thus biased estimates of a predictor's performance, which in turn may lead to choosing suboptimal predictors. In this contribution, we borrow ideas from the domain adaptation machine-learning literature, to suggest the importance-weighted source risk (IWSR). IWSR is a principled approach for weighting the prediction risk, which allows the practitioner to explicitly state the target locations for prediction. IWSR essentially consists of down-weighting training locations and up-weighting target locations. We show that, unlike the usual (unweighted) empirical risk, IWSR is an unbiased estimator of the prediction error. Equipped with this risk estimator, we use it to learn a model in the empirical risk minimization framework and to evaluate the existing predictors. We show the superiority of this weighted risk, using both simulated data and an empirical control: air-temperature prediction in France.
- Open Access
- 20 June 2020
- Published by Center for Open Science in PsyArXiv
The Predictive Five: A supervised learning approach to personality psychology
In the current study, we set out to examine the viability of a novel approach to modeling human personality. Current research in psychology suggests that people’s personalities can be effectively described using five broad dimensions (the Five-Factor Model; FFM); however, the FFM has been criticized for its relatively limited predictive ability. We propose a novel approach to modeling human personality that is based on the maximization of the model’s predictive accuracy. Unlike the FFM, which performs unsupervised dimensionality reduction, we utilized supervised machine learning techniques for dimensionality reduction of questionnaire data, using numerous psychologically meaningful outcomes as data labels (e.g., intelligence, well-being, sociability). The results showed that our five dimensional personality summary, which we term the Predictive Five (PF), provides predictive performance that is superior to the FFM in independent validation datasets, and on a new set of outcome variables selected by an independent group of psychologists. Furthermore, we examine the between-participants’ replicability of the PF representation and show that the PF has good test-retest reliability, and as such provides an important addition to the psychologists' toolbox. The approach described herein has the promise of providing an interpretable low-dimensional personality representation, which is also predictive of behaviour.
- 20 June 2020
- Published by Center for Open Science in PsyArXiv
Open AccessThe Predictive Five: A supervised learning approach to personality psychology
In the current study, we set out to examine the viability of a novel approach to modeling human personality. Current research in psychology suggests that people’s personalities can be effectively described using five broad dimensions (the Five-Factor Model; FFM); however, the FFM has been criticized for its relatively limited predictive ability. We propose a novel approach to modeling human personality that is based on the maximization of the model’s predictive accuracy. Unlike the FFM, which performs unsupervised dimensionality reduction, we utilized supervised machine learning techniques for dimensionality reduction of questionnaire data, using numerous psychologically meaningful outcomes as data labels (e.g., intelligence, well-being, sociability). The results showed that our five dimensional personality summary, which we term the Predictive Five (PF), provides predictive performance that is superior to the FFM in independent validation datasets, and on a new set of outcome variables selected by an independent group of psychologists. Furthermore, we examine the between-participants’ replicability of the PF representation and show that the PF has good test-retest reliability, and as such provides an important addition to the psychologists' toolbox. The approach described herein has the promise of providing an interpretable low-dimensional personality representation, which is also predictive of behaviour.
- Open Access
- 28 May 2020
- Published by MDPI AG in Remote Sensing
- Vol. 12  (11) , 1741
Estimation of Hourly near Surface Air Temperature Across Israel Using an Ensemble Model
Mapping of near-surface air temperature (Ta) at high spatio-temporal resolution is essential for unbiased assessment of human health exposure to temperature extremes, not least given the observed trend of urbanization and global climate change. Data constraints have led previous studies to focus merely on daily Ta metrics, rather than hourly ones, making them insufficient for intra-day assessment of health exposure. In this study, we present a three-stage machine learning-based ensemble model to estimate hourly Ta at a high spatial resolution of 1 × 1 km2, incorporating remotely sensed surface skin temperature (Ts) from geostationary satellites, reanalysis synoptic variables, and observations from weather stations, as well as auxiliary geospatial variables, which account for spatio-temporal variability of Ta. The Stage 1 model gap-fills hourly Ts at 4 × 4 km2 from the Spinning Enhanced Visible and InfraRed Imager (SEVIRI), which are subsequently fed into the Stage 2 model to estimate hourly Ta at the same spatio-temporal resolution. The Stage 3 model downscales the residuals between estimated and measured Ta to a grid of 1 × 1 km2, taking into account additionally the monthly diurnal pattern of Ts derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) data. In each stage, the ensemble model synergizes estimates from the constituent base learners—random forest (RF) and extreme gradient boosting (XGBoost)—by applying a geographically weighted generalized additive model (GAM), which allows the weights of results from individual models to vary over space and time. Demonstrated for Israel for the period 2004–2017, the proposed ensemble model outperformed each of the two base learners. It also attained excellent five-fold cross-validated performance, with overall root mean square error (RMSE) of 0.8 and 0.9 °C, mean absolute error (MAE) of 0.6 and 0.7 °C, and R2 of 0.95 and 0.98 in Stage 1 and Stage 2, respectively. The Stage 3 model for downscaling Ta residuals to 1 km MODIS grids achieved overall RMSE of 0.3 °C, MAE of 0.5 °C, and R2 of 0.63. The generated hourly 1 × 1 km2 Ta thus serves as a foundation for monitoring and assessing human health exposure to temperature extremes at a larger geographical scale, helping to further minimize exposure misclassification in epidemiological studies.
- 28 May 2020
- Published by MDPI AG in Remote Sensing
- Vol. 12  (11) , 1741
Open AccessEstimation of Hourly near Surface Air Temperature Across Israel Using an Ensemble Model
Mapping of near-surface air temperature (Ta) at high spatio-temporal resolution is essential for unbiased assessment of human health exposure to temperature extremes, not least given the observed trend of urbanization and global climate change. Data constraints have led previous studies to focus merely on daily Ta metrics, rather than hourly ones, making them insufficient for intra-day assessment of health exposure. In this study, we present a three-stage machine learning-based ensemble model to estimate hourly Ta at a high spatial resolution of 1 × 1 km2, incorporating remotely sensed surface skin temperature (Ts) from geostationary satellites, reanalysis synoptic variables, and observations from weather stations, as well as auxiliary geospatial variables, which account for spatio-temporal variability of Ta. The Stage 1 model gap-fills hourly Ts at 4 × 4 km2 from the Spinning Enhanced Visible and InfraRed Imager (SEVIRI), which are subsequently fed into the Stage 2 model to estimate hourly Ta at the same spatio-temporal resolution. The Stage 3 model downscales the residuals between estimated and measured Ta to a grid of 1 × 1 km2, taking into account additionally the monthly diurnal pattern of Ts derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) data. In each stage, the ensemble model synergizes estimates from the constituent base learners—random forest (RF) and extreme gradient boosting (XGBoost)—by applying a geographically weighted generalized additive model (GAM), which allows the weights of results from individual models to vary over space and time. Demonstrated for Israel for the period 2004–2017, the proposed ensemble model outperformed each of the two base learners. It also attained excellent five-fold cross-validated performance, with overall root mean square error (RMSE) of 0.8 and 0.9 °C, mean absolute error (MAE) of 0.6 and 0.7 °C, and R2 of 0.95 and 0.98 in Stage 1 and Stage 2, respectively. The Stage 3 model for downscaling Ta residuals to 1 km MODIS grids achieved overall RMSE of 0.3 °C, MAE of 0.5 °C, and R2 of 0.63. The generated hourly 1 × 1 km2 Ta thus serves as a foundation for monitoring and assessing human health exposure to temperature extremes at a larger geographical scale, helping to further minimize exposure misclassification in epidemiological studies.
- 29 October 2019
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 116  (47) , 23382-23383
The harmonic mean p -value: Strong versus weak control, and the assumption of independence
- 29 October 2019
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 116  (47) , 23382-23383
The harmonic mean p -value: Strong versus weak control, and the assumption of independence
- 14 October 2019
- Published by Oxford University Press (OUP) in Biostatistics
- Vol. 22  (2) , 365-380
Better-than-chance classification for signal detection
Summary: The estimated accuracy of a classifier is a random quantity with variability. A common practice in supervised machine learning, is thus to test if the estimated accuracy is significantly better than chance level. This method of signal detection is particularly popular in neuroimaging and genetics. We provide evidence that using a classifier’s accuracy as a test statistic can be an underpowered strategy for finding differences between populations, compared to a bona fide statistical test. It is also computationally more demanding than a statistical test. Via simulation, we compare test statistics that are based on classification accuracy, to others based on multivariate test statistics. We find that the probability of detecting differences between two distributions is lower for accuracy-based statistics. We examine several candidate causes for the low power of accuracy-tests. These causes include: the discrete nature of the accuracy-test statistic, the type of signal accuracy-tests are designed to detect, their inefficient use of the data, and their suboptimal regularization. When the purpose of the analysis is the evaluation of a particular classifier, not signal detection, we suggest several improvements to increase power. In particular, to replace V-fold cross-validation with the Leave-One-Out Bootstrap.
- 14 October 2019
- Published by Oxford University Press (OUP) in Biostatistics
- Vol. 22  (2) , 365-380
Better-than-chance classification for signal detection
Summary: The estimated accuracy of a classifier is a random quantity with variability. A common practice in supervised machine learning, is thus to test if the estimated accuracy is significantly better than chance level. This method of signal detection is particularly popular in neuroimaging and genetics. We provide evidence that using a classifier’s accuracy as a test statistic can be an underpowered strategy for finding differences between populations, compared to a bona fide statistical test. It is also computationally more demanding than a statistical test. Via simulation, we compare test statistics that are based on classification accuracy, to others based on multivariate test statistics. We find that the probability of detecting differences between two distributions is lower for accuracy-based statistics. We examine several candidate causes for the low power of accuracy-tests. These causes include: the discrete nature of the accuracy-test statistic, the type of signal accuracy-tests are designed to detect, their inefficient use of the data, and their suboptimal regularization. When the purpose of the analysis is the evaluation of a particular classifier, not signal detection, we suggest several improvements to increase power. In particular, to replace V-fold cross-validation with the Leave-One-Out Bootstrap.
- 1 May 2019
- Published by Elsevier in Atmospheric Environment
- Vol. 205, 30-35
Gaussian Markov Random Fields versus Linear Mixed Models for satellite-based PM2.5 assessment: Evidence from the Northeastern USA
- 1 May 2019
- Published by Elsevier in Atmospheric Environment
- Vol. 205, 30-35
Gaussian Markov Random Fields versus Linear Mixed Models for satellite-based PM2.5 assessment: Evidence from the Northeastern USA
- Open Access
Gaussian Markov Random Fields versus Linear Mixed Models for satellite-based PM2.5 assessment: Evidence from the Northeastern USA
Studying the effects of air-pollution on health is a key area in environmental epidemiology. An accurate estimation of air-pollution effects requires spatio-temporally resolved datasets of air-pollution, especially, Fine Particulate Matter (PM). Satellite-based technology has greatly enhanced the ability to provide PM assessments in locations where direct measurement is impossible. Indirect PM measurement is a statistical prediction problem. The spatio-temporal statistical literature offer various predictive models: Gaussian Random Fields (GRF) and Linear Mixed Models (LMM), in particular. GRF emphasize the spatio-temporal structure in the data, but are computationally demanding to fit. LMMs are computationally easier to fit, but require some tampering to deal with space and time. Recent advances in the spatio-temporal statistical literature propose to alleviate the computation burden of GRFs by approximating them with Gaussian Markov Random Fields (GMRFs). Since LMMs and GMRFs are both computationally feasible, the question arises: which is statistically better? We show that despite the great popularity of LMMs in environmental monitoring and pollution assessment, LMMs are statistically inferior to GMRF for measuring PM in the Northeastern USA.
Open AccessGaussian Markov Random Fields versus Linear Mixed Models for satellite-based PM2.5 assessment: Evidence from the Northeastern USA
Studying the effects of air-pollution on health is a key area in environmental epidemiology. An accurate estimation of air-pollution effects requires spatio-temporally resolved datasets of air-pollution, especially, Fine Particulate Matter (PM). Satellite-based technology has greatly enhanced the ability to provide PM assessments in locations where direct measurement is impossible. Indirect PM measurement is a statistical prediction problem. The spatio-temporal statistical literature offer various predictive models: Gaussian Random Fields (GRF) and Linear Mixed Models (LMM), in particular. GRF emphasize the spatio-temporal structure in the data, but are computationally demanding to fit. LMMs are computationally easier to fit, but require some tampering to deal with space and time. Recent advances in the spatio-temporal statistical literature propose to alleviate the computation burden of GRFs by approximating them with Gaussian Markov Random Fields (GMRFs). Since LMMs and GMRFs are both computationally feasible, the question arises: which is statistically better? We show that despite the great popularity of LMMs in environmental monitoring and pollution assessment, LMMs are statistically inferior to GMRF for measuring PM in the Northeastern USA.