Steve Yadlowsky
Research Subjects
Research Interests
Steve Yadlowsky has a total of 63 co-authors. Here are the top co-authors by number of shared publications:
- Hongseok Namkoong
Decision, Risk, and Operations Division, Columbia University, New York, New York 10027
3 shared publications - Alexander D'Amour
Google research
2 shared publications - Ethan Steinberg
Stanford University Stanford University Stanford University Stanford University Technology and Digital Solutions Stanford Healthcare
2 shared publications - John Duchi
Department of Statistics Stanford University
2 shared publications - Lu Tian
Department of Biomedical Data Science, Stanford University School of Medicine, Stanford, California
2 shared publications - Nigam Haresh Shah
Stanford University, USA
2 shared publications
Latest Publications
- Open Access
Pretraining Data Mixtures Enable Narrow Model Selection Capabilities in Transformer Models
Transformer models, notably large language models (LLMs), have the remarkable ability to perform in-context learning (ICL) -- to perform new tasks when prompted with unseen input-output examples without any explicit model training. In this work, we study how effectively transformers can bridge between their pretraining data mixture, comprised of multiple distinct task families, to identify and learn new tasks in-context which are both inside and outside the pretraining distribution. Building on previous work, we investigate this question in a controlled setting, where we study transformer models trained on sequences of $(x, f(x))$ pairs rather than natural language. Our empirical results show transformers demonstrate near-optimal unsupervised model selection capabilities, in their ability to first in-context identify different task families and in-context learn within them when the task families are well-represented in their pretraining data. However when presented with tasks or functions which are out-of-domain of their pretraining data, we demonstrate various failure modes of transformers and degradation of their generalization for even simple extrapolation tasks. Together our results highlight that the impressive ICL abilities of high-capacity sequence models may be more closely tied to the coverage of their pretraining data mixtures than inductive biases that create fundamental generalization capabilities.Open AccessPretraining Data Mixtures Enable Narrow Model Selection Capabilities in Transformer Models
Transformer models, notably large language models (LLMs), have the remarkable ability to perform in-context learning (ICL) -- to perform new tasks when prompted with unseen input-output examples without any explicit model training. In this work, we study how effectively transformers can bridge between their pretraining data mixture, comprised of multiple distinct task families, to identify and learn new tasks in-context which are both inside and outside the pretraining distribution. Building on previous work, we investigate this question in a controlled setting, where we study transformer models trained on sequences of $(x, f(x))$ pairs rather than natural language. Our empirical results show transformers demonstrate near-optimal unsupervised model selection capabilities, in their ability to first in-context identify different task families and in-context learn within them when the task families are well-represented in their pretraining data. However when presented with tasks or functions which are out-of-domain of their pretraining data, we demonstrate various failure modes of transformers and degradation of their generalization for even simple extrapolation tasks. Together our results highlight that the impressive ICL abilities of high-capacity sequence models may be more closely tied to the coverage of their pretraining data mixtures than inductive biases that create fundamental generalization capabilities. - Open Access
- 9 September 2023
- Published by Springer Nature in BMC Medical Research Methodology
- Vol. 23 (1) , 1-13
Using public clinical trial reports to probe non-experimental causal inference methods
Background: Non-experimental studies (also known as observational studies) are valuable for estimating the effects of various medical interventions, but are notoriously difficult to evaluate because the methods used in non-experimental studies require untestable assumptions. This lack of intrinsic verifiability makes it difficult both to compare different non-experimental study methods and to trust the results of any particular non-experimental study.Methods: We introduceTrialProbe, a data resource and statistical framework for the evaluation of non-experimental methods. We first collect a dataset of pseudo “ground truths” about the relative effects of drugs by using empirical Bayesian techniques to analyze adverse events recorded in public clinical trial reports. We then develop a framework for evaluating non-experimental methods against that ground truth by measuring concordance between the non-experimental effect estimates and the estimates derived from clinical trials. As a demonstration of our approach, we also perform an example methods evaluation between propensity score matching, inverse propensity score weighting, and an unadjusted approach on a large national insurance claims dataset.Results: From the 33,701 clinical trial records in our version of the ClinicalTrials.gov dataset, we are able to extract 12,967 unique drug/drug adverse event comparisons to form a ground truth set. During our corresponding methods evaluation, we are able to use that reference set to demonstrate that both propensity score matching and inverse propensity score weighting can produce estimates that have high concordance with clinical trial results and substantially outperform an unadjusted baseline.Conclusions: We find thatTrialProbeis an effective approach for probing non-experimental study methods, being able to generate large ground truth sets that are able to distinguish how well non-experimental methods perform in real world observational data.
- 9 September 2023
- Published by Springer Nature in BMC Medical Research Methodology
- Vol. 23 (1) , 1-13
Open AccessUsing public clinical trial reports to probe non-experimental causal inference methods
Background: Non-experimental studies (also known as observational studies) are valuable for estimating the effects of various medical interventions, but are notoriously difficult to evaluate because the methods used in non-experimental studies require untestable assumptions. This lack of intrinsic verifiability makes it difficult both to compare different non-experimental study methods and to trust the results of any particular non-experimental study.Methods: We introduceTrialProbe, a data resource and statistical framework for the evaluation of non-experimental methods. We first collect a dataset of pseudo “ground truths” about the relative effects of drugs by using empirical Bayesian techniques to analyze adverse events recorded in public clinical trial reports. We then develop a framework for evaluating non-experimental methods against that ground truth by measuring concordance between the non-experimental effect estimates and the estimates derived from clinical trials. As a demonstration of our approach, we also perform an example methods evaluation between propensity score matching, inverse propensity score weighting, and an unadjusted approach on a large national insurance claims dataset.Results: From the 33,701 clinical trial records in our version of the ClinicalTrials.gov dataset, we are able to extract 12,967 unique drug/drug adverse event comparisons to form a ground truth set. During our corresponding methods evaluation, we are able to use that reference set to demonstrate that both propensity score matching and inverse propensity score weighting can produce estimates that have high concordance with clinical trial results and substantially outperform an unadjusted baseline.Conclusions: We find thatTrialProbeis an effective approach for probing non-experimental study methods, being able to generate large ground truth sets that are able to distinguish how well non-experimental methods perform in real world observational data.
- Open Access
Diagnosing Model Performance Under Distribution Shift
Prediction models can perform poorly when deployed to target distributions different from the training distribution. To understand these operational failure modes, we develop a method, called DIstribution Shift DEcomposition (DISDE), to attribute a drop in performance to different types of distribution shifts. Our approach decomposes the performance drop into terms for 1) an increase in harder but frequently seen examples from training, 2) changes in the relationship between features and outcomes, and 3) poor performance on examples infrequent or unseen during training. These terms are defined by fixing a distribution on $X$ while varying the conditional distribution of $Y \mid X$ between training and target, or by fixing the conditional distribution of $Y \mid X$ while varying the distribution on $X$. In order to do this, we define a hypothetical distribution on $X$ consisting of values common in both training and target, over which it is easy to compare $Y \mid X$ and thus predictive performance. We estimate performance on this hypothetical distribution via reweighting methods. Empirically, we show how our method can 1) inform potential modeling improvements across distribution shifts for employment prediction on tabular census data, and 2) help to explain why certain domain adaptation methods fail to improve model performance for satellite image classification.Open AccessDiagnosing Model Performance Under Distribution Shift
Prediction models can perform poorly when deployed to target distributions different from the training distribution. To understand these operational failure modes, we develop a method, called DIstribution Shift DEcomposition (DISDE), to attribute a drop in performance to different types of distribution shifts. Our approach decomposes the performance drop into terms for 1) an increase in harder but frequently seen examples from training, 2) changes in the relationship between features and outcomes, and 3) poor performance on examples infrequent or unseen during training. These terms are defined by fixing a distribution on $X$ while varying the conditional distribution of $Y \mid X$ between training and target, or by fixing the conditional distribution of $Y \mid X$ while varying the distribution on $X$. In order to do this, we define a hypothetical distribution on $X$ consisting of values common in both training and target, over which it is easy to compare $Y \mid X$ and thus predictive performance. We estimate performance on this hypothetical distribution via reweighting methods. Empirically, we show how our method can 1) inform potential modeling improvements across distribution shifts for employment prediction on tabular census data, and 2) help to explain why certain domain adaptation methods fail to improve model performance for satellite image classification. - Top Cited Papers
- 1 October 2022
- Published by Institute of Mathematical Statistics in The Annals of Statistics
- Vol. 50 (5) , 2587-2615
Bounds on the conditional and average treatment effect with unobserved confounding factors
For observational studies, we study the sensitivity of causal inference when treatment assignments may depend on unobserved confounders. We develop a loss minimization approach for estimating bounds on the conditional average treatment effect (CATE) when unobserved confounders have a bounded effect on the odds ratio of treatment selection. Our approach is scalable and allows flexible use of model classes in estimation, including nonparametric and black-box machine learning methods. Based on these bounds for the CATE, we propose a sensitivity analysis for the average treatment effect (ATE). Our semiparametric estimator extends/bounds the augmented inverse propensity weighted (AIPW) estimator for the ATE under bounded unobserved confounding. By constructing a Neyman orthogonal score, our estimator of the bound for the ATE is a regular root-n estimator so long as the nuisance parameters are estimated at the op(n−1/4) rate. We complement our methodology with optimality results showing that our proposed bounds are tight in certain cases. We demonstrate our method on simulated and real data examples, and show accurate coverage of our confidence intervals in practical finite sample regimes with rich covariate information.
- 1 October 2022
- Published by Institute of Mathematical Statistics in The Annals of Statistics
- Vol. 50 (5) , 2587-2615
Top Cited PapersBounds on the conditional and average treatment effect with unobserved confounding factors
For observational studies, we study the sensitivity of causal inference when treatment assignments may depend on unobserved confounders. We develop a loss minimization approach for estimating bounds on the conditional average treatment effect (CATE) when unobserved confounders have a bounded effect on the odds ratio of treatment selection. Our approach is scalable and allows flexible use of model classes in estimation, including nonparametric and black-box machine learning methods. Based on these bounds for the CATE, we propose a sensitivity analysis for the average treatment effect (ATE). Our semiparametric estimator extends/bounds the augmented inverse propensity weighted (AIPW) estimator for the ATE under bounded unobserved confounding. By constructing a Neyman orthogonal score, our estimator of the bound for the ATE is a regular root-n estimator so long as the nuisance parameters are estimated at the op(n−1/4) rate. We complement our methodology with optimality results showing that our proposed bounds are tight in certain cases. We demonstrate our method on simulated and real data examples, and show accurate coverage of our confidence intervals in practical finite sample regimes with rich covariate information.
- Open Access
Using public clinical trial reports to evaluate observational study methods
Observational studies are valuable for estimating the effects of various medical interventions, but are notoriously difficult to evaluate because the methods used in observational studies require many untestable assumptions. This lack of verifiability makes it difficult both to compare different observational study methods and to trust the results of any particular observational study. In this work, we propose TrialVerify, a new approach for evaluating observational study methods based on ground truth sourced from clinical trial reports. We process trial reports into a denoised collection of known causal relationships that can then be used to estimate the precision and recall of various observational study methods. We then use TrialVerify to evaluate multiple observational study methods in terms of their ability to identify the known causal relationships from a large national insurance claims dataset. We found that inverse propensity score weighting is an effective approach for accurately reproducing known causal relationships and outperforms other observational study methods. TrialVerify is made freely available for others to evaluate observational study methods.
Open AccessUsing public clinical trial reports to evaluate observational study methods
Observational studies are valuable for estimating the effects of various medical interventions, but are notoriously difficult to evaluate because the methods used in observational studies require many untestable assumptions. This lack of verifiability makes it difficult both to compare different observational study methods and to trust the results of any particular observational study. In this work, we propose TrialVerify, a new approach for evaluating observational study methods based on ground truth sourced from clinical trial reports. We process trial reports into a denoised collection of known causal relationships that can then be used to estimate the precision and recall of various observational study methods. We then use TrialVerify to evaluate multiple observational study methods in terms of their ability to identify the known causal relationships from a large national insurance claims dataset. We found that inverse propensity score weighting is an effective approach for accurately reproducing known causal relationships and outperforms other observational study methods. TrialVerify is made freely available for others to evaluate observational study methods.
- Open Access
Explaining Practical Differences Between Treatment Effect Estimators with High Dimensional Asymptotics
We revisit the classical causal inference problem of estimating the average treatment effect in the presence of fully observed confounding variables using two-stage semiparametric methods. In existing theoretical studies of methods such as G-computation, inverse propensity weighting (IPW), and two common doubly robust estimators -- augmented IPW (AIPW) and targeted maximum likelihood estimation (TMLE) -- they are either bias-dominated, or have similar asymptotic statistical properties. However, when applied to real datasets, they often appear to have notably different variance. We compare these methods when using a machine learning (ML) model to estimate the nuisance parameters of the semiparametric model, and highlight some of the important differences. When the outcome model estimates have little bias, which is common among some key ML models, G-computation and the TMLE outperforms the other estimators in both bias and variance. We show that the differences can be explained using high-dimensional statistical theory, where the number of confounders $d$ is of the same order as the sample size $n$. To make this theoretical problem tractable, we posit a generalized linear model for the effect of the confounders on the treatment assignment and outcomes. Despite making parametric assumptions, this setting is a useful surrogate for some machine learning methods used to adjust for confounding in two-stage semiparametric methods. In particular, the estimation of the first stage adds variance that does not vanish, forcing us to confront terms in the asymptotic expansion that normally are brushed aside as finite sample defects. However, our model emphasizes differences in performance between these estimators beyond first-order asymptotics.Open AccessExplaining Practical Differences Between Treatment Effect Estimators with High Dimensional Asymptotics
We revisit the classical causal inference problem of estimating the average treatment effect in the presence of fully observed confounding variables using two-stage semiparametric methods. In existing theoretical studies of methods such as G-computation, inverse propensity weighting (IPW), and two common doubly robust estimators -- augmented IPW (AIPW) and targeted maximum likelihood estimation (TMLE) -- they are either bias-dominated, or have similar asymptotic statistical properties. However, when applied to real datasets, they often appear to have notably different variance. We compare these methods when using a machine learning (ML) model to estimate the nuisance parameters of the semiparametric model, and highlight some of the important differences. When the outcome model estimates have little bias, which is common among some key ML models, G-computation and the TMLE outperforms the other estimators in both bias and variance. We show that the differences can be explained using high-dimensional statistical theory, where the number of confounders $d$ is of the same order as the sample size $n$. To make this theoretical problem tractable, we posit a generalized linear model for the effect of the confounders on the treatment assignment and outcomes. Despite making parametric assumptions, this setting is a useful surrogate for some machine learning methods used to adjust for confounding in two-stage semiparametric methods. In particular, the estimation of the first stage adds variance that does not vanish, forcing us to confront terms in the asymptotic expansion that normally are brushed aside as finite sample defects. However, our model emphasizes differences in performance between these estimators beyond first-order asymptotics. - Open Access
The MultiBERTs: BERT Reproductions for Robustness Analysis
Experiments with pre-trained models such as BERT are often based on a single checkpoint. While the conclusions drawn apply to the artifact tested in the experiment (i.e., the particular instance of the model), it is not always clear whether they hold for the more general procedure which includes the architecture, training data, initialization scheme, and loss function. Recent work has shown that repeating the pre-training process can lead to substantially different performance, suggesting that an alternate strategy is needed to make principled statements about procedures. To enable researchers to draw more robust conclusions, we introduce the MultiBERTs, a set of 25 BERT-Base checkpoints, trained with similar hyper-parameters as the original BERT model but differing in random weight initialization and shuffling of training data. We also define the Multi-Bootstrap, a non-parametric bootstrap method for statistical inference designed for settings where there are multiple pre-trained models and limited test data. To illustrate our approach, we present a case study of gender bias in coreference resolution, in which the Multi-Bootstrap lets us measure effects that may not be detected with a single checkpoint. We release our models and statistical library along with an additional set of 140 intermediate checkpoints captured during pre-training to facilitate research on learning dynamics.
Open AccessThe MultiBERTs: BERT Reproductions for Robustness Analysis
Experiments with pre-trained models such as BERT are often based on a single checkpoint. While the conclusions drawn apply to the artifact tested in the experiment (i.e., the particular instance of the model), it is not always clear whether they hold for the more general procedure which includes the architecture, training data, initialization scheme, and loss function. Recent work has shown that repeating the pre-training process can lead to substantially different performance, suggesting that an alternate strategy is needed to make principled statements about procedures. To enable researchers to draw more robust conclusions, we introduce the MultiBERTs, a set of 25 BERT-Base checkpoints, trained with similar hyper-parameters as the original BERT model but differing in random weight initialization and shuffling of training data. We also define the Multi-Bootstrap, a non-parametric bootstrap method for statistical inference designed for settings where there are multiple pre-trained models and limited test data. To illustrate our approach, we present a case study of gender bias in coreference resolution, in which the Multi-Bootstrap lets us measure effects that may not be detected with a single checkpoint. We release our models and statistical library along with an additional set of 140 intermediate checkpoints captured during pre-training to facilitate research on learning dynamics.
- Open Access
Bounds on the conditional and average treatment effect with unobserved confounding factors
For observational studies, we study the sensitivity of causal inference when treatment assignments may depend on unobserved confounders. We develop a loss minimization approach for estimating bounds on the conditional average treatment effect (CATE) when unobserved confounders have a bounded effect on the odds ratio of treatment selection. Our approach is scalable and allows flexible use of model classes in estimation, including nonparametric and black-box machine learning methods. Based on these bounds for the CATE, we propose a sensitivity analysis for the average treatment effect (ATE). Our semi-parametric estimator extends/bounds the augmented inverse propensity weighted (AIPW) estimator for the ATE under bounded unobserved confounding. By constructing a Neyman orthogonal score, our estimator of the bound for the ATE is a regular root-$n$ estimator so long as the nuisance parameters are estimated at the $o_p(n^{-1/4})$ rate. We complement our methodology with optimality results showing that our proposed bounds are tight in certain cases. We demonstrate our method on simulated and real data examples, and show accurate coverage of our confidence intervals in practical finite sample regimes with rich covariate information.Open AccessBounds on the conditional and average treatment effect with unobserved confounding factors
For observational studies, we study the sensitivity of causal inference when treatment assignments may depend on unobserved confounders. We develop a loss minimization approach for estimating bounds on the conditional average treatment effect (CATE) when unobserved confounders have a bounded effect on the odds ratio of treatment selection. Our approach is scalable and allows flexible use of model classes in estimation, including nonparametric and black-box machine learning methods. Based on these bounds for the CATE, we propose a sensitivity analysis for the average treatment effect (ATE). Our semi-parametric estimator extends/bounds the augmented inverse propensity weighted (AIPW) estimator for the ATE under bounded unobserved confounding. By constructing a Neyman orthogonal score, our estimator of the bound for the ATE is a regular root-$n$ estimator so long as the nuisance parameters are estimated at the $o_p(n^{-1/4})$ rate. We complement our methodology with optimality results showing that our proposed bounds are tight in certain cases. We demonstrate our method on simulated and real data examples, and show accurate coverage of our confidence intervals in practical finite sample regimes with rich covariate information. - Open Access
SLOE: A Faster Method for Statistical Inference in High-Dimensional Logistic Regression
Logistic regression remains one of the most widely used tools in applied statistics, machine learning and data science. However, in moderately high-dimensional problems, where the number of features $d$ is a non-negligible fraction of the sample size $n$, the logistic regression maximum likelihood estimator (MLE), and statistical procedures based the large-sample approximation of its distribution, behave poorly. Recently, Sur and Cand\`es (2019) showed that these issues can be corrected by applying a new approximation of the MLE's sampling distribution in this high-dimensional regime. Unfortunately, these corrections are difficult to implement in practice, because they require an estimate of the \emph{signal strength}, which is a function of the underlying parameters $\beta$ of the logistic regression. To address this issue, we propose SLOE, a fast and straightforward approach to estimate the signal strength in logistic regression. The key insight of SLOE is that the Sur and Cand\`es (2019) correction can be reparameterized in terms of the \emph{corrupted signal strength}, which is only a function of the estimated parameters $\widehat \beta$. We propose an estimator for this quantity, prove that it is consistent in the relevant high-dimensional regime, and show that dimensionality correction using SLOE is accurate in finite samples. Compared to the existing ProbeFrontier heuristic, SLOE is conceptually simpler and orders of magnitude faster, making it suitable for routine use. We demonstrate the importance of routine dimensionality correction in the Heart Disease dataset from the UCI repository, and a genomics application using data from the UK Biobank. We provide an open source package for this method, available at \url{https://github.com/google-research/sloe-logistic}.Open AccessSLOE: A Faster Method for Statistical Inference in High-Dimensional Logistic Regression
Logistic regression remains one of the most widely used tools in applied statistics, machine learning and data science. However, in moderately high-dimensional problems, where the number of features $d$ is a non-negligible fraction of the sample size $n$, the logistic regression maximum likelihood estimator (MLE), and statistical procedures based the large-sample approximation of its distribution, behave poorly. Recently, Sur and Cand\`es (2019) showed that these issues can be corrected by applying a new approximation of the MLE's sampling distribution in this high-dimensional regime. Unfortunately, these corrections are difficult to implement in practice, because they require an estimate of the \emph{signal strength}, which is a function of the underlying parameters $\beta$ of the logistic regression. To address this issue, we propose SLOE, a fast and straightforward approach to estimate the signal strength in logistic regression. The key insight of SLOE is that the Sur and Cand\`es (2019) correction can be reparameterized in terms of the \emph{corrupted signal strength}, which is only a function of the estimated parameters $\widehat \beta$. We propose an estimator for this quantity, prove that it is consistent in the relevant high-dimensional regime, and show that dimensionality correction using SLOE is accurate in finite samples. Compared to the existing ProbeFrontier heuristic, SLOE is conceptually simpler and orders of magnitude faster, making it suitable for routine use. We demonstrate the importance of routine dimensionality correction in the Heart Disease dataset from the UCI repository, and a genomics application using data from the UK Biobank. We provide an open source package for this method, available at \url{https://github.com/google-research/sloe-logistic}. - Open Access
Underspecification Presents Challenges for Credibility in Modern Machine Learning
ML models often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification as a key reason for these failures. An ML pipeline is underspecified when it can return many predictors with equivalently strong held-out performance in the training domain. Underspecification is common in modern ML pipelines, such as those based on deep learning. Predictors returned by underspecified pipelines are often treated as equivalent based on their training domain performance, but we show here that such predictors can behave very differently in deployment domains. This ambiguity can lead to instability and poor model behavior in practice, and is a distinct failure mode from previously identified issues arising from structural mismatch between training and deployment domains. We show that this problem appears in a wide variety of practical ML pipelines, using examples from computer vision, medical imaging, natural language processing, clinical risk prediction based on electronic health records, and medical genomics. Our results show the need to explicitly account for underspecification in modeling pipelines that are intended for real-world deployment in any domain.
Open AccessUnderspecification Presents Challenges for Credibility in Modern Machine Learning
ML models often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification as a key reason for these failures. An ML pipeline is underspecified when it can return many predictors with equivalently strong held-out performance in the training domain. Underspecification is common in modern ML pipelines, such as those based on deep learning. Predictors returned by underspecified pipelines are often treated as equivalent based on their training domain performance, but we show here that such predictors can behave very differently in deployment domains. This ambiguity can lead to instability and poor model behavior in practice, and is a distinct failure mode from previously identified issues arising from structural mismatch between training and deployment domains. We show that this problem appears in a wide variety of practical ML pipelines, using examples from computer vision, medical imaging, natural language processing, clinical risk prediction based on electronic health records, and medical genomics. Our results show the need to explicitly account for underspecification in modeling pipelines that are intended for real-world deployment in any domain.