arXiv subjects

Paul R. Rosenbaum

Publications and source records attributed to Paul R. Rosenbaum.

13 recordsLinked to original sources

arXiv2024

Effect Aliasing in Observational Studies

In experimental design, aliasing of effects occurs in fractional factorial experiments, where certain low order factorial effects are indistinguishable from certain high order interactions: low order contrasts may be orthogonal to one another, while their higher order interactions are aliased and not identified. In observational studies, aliasing occurs when certain combinations of covariates -- e.g., time period and various eligibility criteria for treatment -- perfectly predict the treatment that an individual will receive, so a covariate combination is aliased with a particular treatment. In this situation, when a contrast among several groups is used to estimate a treatment effect, collections of individuals defined by contrast weights may be balanced with respect to summaries of low-order interactions between covariates and treatments, but necessarily not balanced with respect to summaries of high-order interactions between covariates and treatments. We develop a theory of aliasing in observational studies, illustrate that theory in an observational study whose aliasing is more robust than conventional difference-in-differences, and develop a new form of matching to construct balanced confounded factorial designs from observational data.

stat.ME↗

arXiv2019

Causal Inference with Two Versions of Treatment

Causal effects are commonly defined as comparisons of the potential outcomes under treatment and control, but this definition is threatened by the possibility that the treatment or control condition is not well-defined, existing instead in more than one version. A simple, widely applicable analysis is proposed to address the possibility that the treatment or control condition exists in two versions with two different treatment effects. This analysis loses no power in the main comparison of treatment and control, provides additional information about version effects, and controls the family-wise error rate in several comparisons. The method is motivated and illustrated using an on-going study of the possibility that repeated head trauma in high school football causes an increase in risk of early on-set dementia.

stat.ME↗

arXiv2018

A powerful approach to the study of moderate effect modification in observational studies

Effect modification means the magnitude or stability of a treatment effect varies as a function of an observed covariate. Generally, larger and more stable treatment effects are insensitive to larger biases from unmeasured covariates, so a causal conclusion may be considerably firmer if this pattern is noted if it occurs. We propose a new strategy, called the submax-method, that combines exploratory and confirmatory efforts to determine whether there is stronger evidence of causality - that is, greater insensitivity to unmeasured confounding - in some subgroups of individuals. It uses the joint distribution of test statistics that split the data in various ways based on certain observed covariates. For $L$ binary covariates, the method splits the population $L$ times into two subpopulations, perhaps first men and women, perhaps then smokers and nonsmokers, computing a test statistic from each subpopulation, and appends the test statistic for the whole population, making $2L+1$ test statistics in total. Although $L$ binary covariates define $2^{L}$ interaction groups, only $2L+1$ tests are performed, and at least $L+1$ of these tests use at least half of the data. The submax-method achieves the highest design sensitivity and the highest Bahadur efficiency of its component tests. Moreover, the form of the test is sufficiently tractable that its large sample power may be studied analytically. The simulation suggests that the submax method exhibits superior performance, in comparison with an approach using CART, when there is effect modification of moderate size. Using data from the NHANES I Epidemiologic Follow-Up Survey, an observational study of the effects of physical activity on survival is used to illustrate the method. The method is implemented in the $\texttt{R}$ package $\texttt{submax}$ which contains the NHANES example.

stat.ME↗

arXiv2017

Cross-screening in observational studies that test many hypotheses

We discuss observational studies that test many causal hypotheses, either hypotheses about many outcomes or many treatments. To be credible an observational study that tests many causal hypotheses must demonstrate that its conclusions are neither artifacts of multiple testing nor of small biases from nonrandom treatment assignment. In a sense that needs to be defined carefully, hidden within a sensitivity analysis for nonrandom assignment is an enormous correction for multiple testing: in the absence of bias, it is extremely improbable that multiple testing alone would create an association insensitive to moderate biases. We propose a new strategy called "cross-screening", different from but motivated by recent work of Bogomolov and Heller on replicability. Cross-screening splits the data in half at random, uses the first half to plan a study carried out on the second half, then uses the second half to plan a study carried out on the first half, and reports the more favorable conclusions of the two studies correcting using the Bonferroni inequality for having done two studies. If the two studies happen to concur, then they achieve Bogomolov-Heller replicability; however, importantly, replicability is not required for strong control of the family-wise error rate, and either study alone suffices for firm conclusions. In randomized studies with a few hypotheses, cross-split screening is not an attractive method when compared with conventional methods of multiplicity control, but it can become attractive when hundreds or thousands of hypotheses are subjected to sensitivity analyses in an observational study. We illustrate the technique by comparing 46 biomarkers in individuals who consume large quantities of fish versus little or no fish.

stat.ME↗

arXiv2017

Inference with interference between units in an fMRI experiment of motor inhibition

An experimental unit is an opportunity to randomly apply or withhold a treatment. There is interference between units if the application of the treatment to one unit may also affect other units. In cognitive neuroscience, a common form of experiment presents a sequence of stimuli or requests for cognitive activity at random to each experimental subject and measures biological aspects of brain activity that follow these requests. Each subject is then many experimental units, and interference between units within an experimental subject is likely, in part because the stimuli follow one another quickly and in part because human subjects learn or become experienced or primed or bored as the experiment proceeds. We use a recent fMRI experiment concerned with the inhibition of motor activity to illustrate and further develop recently proposed methodology for inference in the presence of interference. A simulation evaluates the power of competing procedures.

stat.ME↗

arXiv2016

Discovering Effect Modification in an Observational Study of Surgical Mortality at Hospitals with Superior Nursing

There is effect modification if the magnitude or stability of a treatment effect varies systematically with the level of an observed covariate. A larger or more stable treatment effect is typically less sensitive to bias from unmeasured covariates, so it is important to recognize effect modification when it is present. We illustrate a recent proposal for conducting a sensitivity analysis that empirically discovers effect modification by exploratory methods, but controls the family-wise error rate in discovered groups. The example concerns a study of mortality and use of the intensive care unit in 23,715 matched pairs of two Medicare patients, one of whom underwent surgery at a hospital identified for superior nursing, the other at a conventional hospital. The pairs were matched exactly for 130 four-digit ICD-9 surgical procedure codes and balanced 172 observed covariates. The pairs were then split into five groups of pairs by CART in its effort to locate effect modification. The evidence of a beneficial effect of magnet hospitals on mortality is least sensitive to unmeasured biases in a large group of patients undergoing rather serious surgical procedures, but in the absence of other life-threatening conditions, such as a comorbidity of congestive heart failure or an emergency admission leading to surgery.

stat.AP↗

arXiv2015

Isolation in the construction of natural experiments

A natural experiment is a type of observational study in which treatment assignment, though not randomized by the investigator, is plausibly close to random. A process that assigns treatments in a highly nonrandom, inequitable manner may, in rare and brief moments, assign aspects of treatments at random or nearly so. Isolating those moments and aspects may extract a natural experiment from a setting in which treatment assignment is otherwise quite biased, far from random. Isolation is a tool that focuses on those rare, brief instances, extracting a small natural experiment from otherwise useless data. We discuss the theory behind isolation and illustrate its use in a reanalysis of a well-known study of the effects of fertility on workforce participation. Whether a woman becomes pregnant at a certain moment in her life and whether she brings that pregnancy to term may reflect her aspirations for family, education and career, the degree of control she exerts over her fertility, and the quality of her relationship with the father; moreover, these aspirations and relationships are unlikely to be recorded with precision in surveys and censuses, and they may confound studies of workforce participation. However, given that a women is pregnant and will bring the pregnancy to term, whether she will have twins or a single child is, to a large extent, simply luck. Given that a woman is pregnant at a certain moment, the differential comparison of two types of pregnancies on workforce participation, twins or a single child, may be close to randomized, not biased by unmeasured aspirations. In this comparison, we find in our case study that mothers of twins had more children but only slightly reduced workforce participation, approximately 5% less time at work for an additional child.

stat.AP↗

arXiv2014

Matching for balance, pairing for heterogeneity in an observational study of the effectiveness of for-profit and not-for-profit high schools in Chile

Conventionally, the construction of a pair-matched sample selects treated and control units and pairs them in a single step with a view to balancing observed covariates $\mathbf{x}$ and reducing the heterogeneity or dispersion of treated-minus-control response differences, $Y$. In contrast, the method of cardinality matching developed here first selects the maximum number of units subject to covariate balance constraints and, with a balanced sample for $\mathbf{x}$ in hand, then separately pairs the units to minimize heterogeneity in $Y$. Reduced heterogeneity of pair differences in responses $Y$ is known to reduce sensitivity to unmeasured biases, so one might hope that cardinality matching would succeed at both tasks, balancing $\mathbf{x}$, stabilizing $Y$. We use cardinality matching in an observational study of the effectiveness of for-profit and not-for-profit private high schools in Chile - a controversial subject in Chile - focusing on students who were in government run primary schools in 2004 but then switched to private high schools. By pairing to minimize heterogeneity in a cardinality match that has balanced covariates, a meaningful reduction in sensitivity to unmeasured biases is obtained.

stat.AP↗

arXiv2013

Stronger instruments via integer programming in an observational study of late preterm birth outcomes

In an optimal nonbipartite match, a single population is divided into matched pairs to minimize a total distance within matched pairs. Nonbipartite matching has been used to strengthen instrumental variables in observational studies of treatment effects, essentially by forming pairs that are similar in terms of covariates but very different in the strength of encouragement to accept the treatment. Optimal nonbipartite matching is typically done using network optimization techniques that can be quick, running in polynomial time, but these techniques limit the tools available for matching. Instead, we use integer programming techniques, thereby obtaining a wealth of new tools not previously available for nonbipartite matching, including fine and near-fine balance for several nominal variables, forced near balance on means and optimal subsetting. We illustrate the methods in our on-going study of outcomes of late-preterm births in California, that is, births of 34 to 36 weeks of gestation. Would lengthening the time in the hospital for such births reduce the frequency of rapid readmissions? A straightforward comparison of babies who stay for a shorter or longer time would be severely biased, because the principal reason for a long stay is some serious health problem. We need an instrument, something inconsequential and haphazard that encourages a shorter or a longer stay in the hospital. It turns out that babies born at certain times of day tend to stay overnight once with a shorter length of stay, whereas babies born at other times of day tend to stay overnight twice with a longer length of stay, and there is nothing particularly special about a baby who is born at 11:00 pm.

stat.AP↗

arXiv2012

An exact adaptive test with superior design sensitivity in an observational study of treatments for ovarian cancer

A sensitivity analysis in an observational study determines the magnitude of bias from nonrandom treatment assignment that would need to be present to alter the qualitative conclusions of a na\"ıve analysis that presumes all biases were removed by matching or by other analytic adjustments. The power of a sensitivity analysis and the design sensitivity anticipate the outcome of a sensitivity analysis under an assumed model for the generation of the data. It is known that the power of a sensitivity analysis is affected by the choice of test statistic, and, in particular, that a statistic with good Pitman efficiency in a randomized experiment, such as Wilcoxon's signed rank statistic, may have low power in a sensitivity analysis and low design sensitivity when compared to other statistics. For instance, for an additive treatment effect and errors that are Normal or logistic or $t$-distributed with 3 degrees of freedom, Brown's combined quantile average test has Pitman efficiency close to that of Wilcoxon's test but has higher power in a sensitivity analysis, while a version of Noether's test has poor Pitman efficiency in a randomized experiment but much higher design sensitivity so it is vastly more powerful than Wilcoxon's statistic in a sensitivity analysis if the sample size is sufficiently large.

stat.AP↗

arXiv2009

Error-free milestones in error prone measurements

A predictor variable or dose that is measured with substantial error may possess an error-free milestone, such that it is known with negligible error whether the value of the variable is to the left or right of the milestone. Such a milestone provides a basis for estimating a linear relationship between the true but unknown value of the error-free predictor and an outcome, because the milestone creates a strong and valid instrumental variable. The inferences are nonparametric and robust, and in the simplest cases, they are exact and distribution free. We also consider multiple milestones for a single predictor and milestones for several predictors whose partial slopes are estimated simultaneously. Examples are drawn from the Wisconsin Longitudinal Study, in which a BA degree acts as a milestone for sixteen years of education, and the binary indicator of military service acts as a milestone for years of service.

stat.AP↗

arXiv2007

R-Estimates vs. GMM: A Theoretical Case Study of Validity and Efficiency

What role should assumptions play in inference? We present a small theoretical case study of a simple, clean case, namely the nonparametric comparison of two continuous distributions using (essentially) information about quartiles, that is, the central information displayed in a pair of boxplots. In particular, we contrast a suggestion of John Tukey--that the validity of inferences should not depend on assumptions, but assumptions have a role in efficiency--with a competing suggestion that is an aspect of Hansen's generalized method of moments--that methods should achieve maximum asymptotic efficiency with fewer assumptions. In our case study, the practical performance of these two suggestions is strikingly different. An aspect of this comparison concerns the unification or separation of the tasks of estimation assuming a model and testing the fit of that model. We also look at a method (MERT) that aims not at best performance, but rather at achieving reasonable performance across a set of plausible models.