Searcharxiv⌕ Search

arXiv subjects

Dylan S. Small

Publications and source records attributed to Dylan S. Small.

89 records · Page 5Linked to original sources
arXiv2016

Using an Instrumental Variable to Test for Unmeasured Confounding

An important concern in an observational study is whether or not there is unmeasured confounding, i.e., unmeasured ways in which the treatment and control groups differ before treatment that affect the outcome. We develop a test of whether there is unmeasured confounding when an instrumental variable (IV) is available. An IV is a variable that is independent of the unmeasured confounding and encourages a subject to take one treatment level vs. another, while having no effect on the outcome beyond its encouragement of a certain treatment level. We show what types of unmeasured confounding can be tested for with an IV and develop a test for this type of unmeasured confounding that has correct type I error rate. We show that the widely used Durbin-Wu-Hausman (DWH) test can have inflated type I error rates when there is treatment effect heterogeneity. Additionally, we show that our test provides more insight into the nature of the unmeasured confounding than the DWH test. We apply our test to an observational study of the effect of a premature infant being delivered in a high-level neonatal intensive care unit (one with mechanical assisted ventilation and high volume) vs. a lower level unit, using the excess travel time a mother lives from the nearest high-level unit to the nearest lower-level unit as an IV.

stat.ME↗
arXiv2015

Full Matching Approach to Instrumental Variables Estimation with Application to the Effect of Malaria on Stunting

Most previous studies of the causal relationship between malaria and stunting have been studies where potential confounders are controlled via regression-based methods, but these studies may have been biased by unobserved confounders. Instrumental variables (IV) regression offers a way to control for unmeasured confounders where, in our case, the sickle cell trait can be used as an instrument. However, for the instrument to be valid, it may still be important to account for measured confounders. The most commonly used instrumental variable regression method, two-stage least squares, relies on parametric assumptions on the effects of measured confounders to account for them. Additionally, two-stage least squares lacks transparency with respect to covariate balance and weighing of subjects and does not blind the researcher to the outcome data. To address these drawbacks, we propose an alternative method for IV estimation based on full matching. We evaluate our new procedure on simulated data and real data concerning the causal effect of malaria on stunting among children. We estimate that the risk of stunting among children with the sickle cell trait decrease by 0.22 times the average number of malaria episodes prevented by the sickle cell trait, a substantial effect of malaria on stunting (p-value: 0.011, 95% CI: 0.044, 1).

stat.AP↗
arXiv2015

Sensitivity Analysis for Multiple Comparisons in Matched Observational Studies through Quadratically Constrained Linear Programming

A sensitivity analysis in an observational study assesses the robustness of significant findings to unmeasured confounding. While sensitivity analyses in matched observational studies have been well addressed when there is a single outcome variable, accounting for multiple comparisons through the existing methods yields overly conservative results when there are multiple outcome variables of interest. This stems from the fact that unmeasured confounding cannot affect the probability of assignment to treatment differently depending on the outcome being analyzed. Existing methods allow this to occur by combining the results of individual sensitivity analyses to assess whether at least one hypothesis is significant, which in turn results in an overly pessimistic assessment of a study's sensitivity to unobserved biases. By solving a quadratically constrained linear program, we are able to perform a sensitivity analysis while enforcing that unmeasured confounding must have the same impact on the treatment assignment probabilities across outcomes for each individual in the study. We show that this allows for uniform improvements in the power of a sensitivity analysis not only for testing the overall null of no effect, but also for null hypotheses on \textit{specific} outcome variables while strongly controlling the familywise error rate. We illustrate our method through an observational study on the effect of smoking on naphthalene exposure.

stat.ME↗
arXiv2015

Discrete Optimization for Interpretable Study Populations and Randomization Inference in an Observational Study of Severe Sepsis Mortality

Motivated by an observational study of the effect of hospital ward versus intensive care unit admission on severe sepsis mortality, we develop methods to address two common problems in observational studies: (1) when there is a lack of covariate overlap between the treated and control groups, how to define an interpretable study population wherein inference can be conducted without extrapolating with respect to important variables; and (2) how to use randomization inference to form confidence intervals for the average treatment effect with binary outcomes. Our solution to problem (1) incorporates existing suggestions in the literature while yielding a study population that is easily understood in terms of the covariates themselves, and can be solved using an efficient branch-and-bound algorithm. We address problem (2) by solving a linear integer program to utilize the worst case variance of the average treatment effect among values for unobserved potential outcomes that are compatible with the null hypothesis. Our analysis finds no evidence for a difference between the sixty day mortality rates if all individuals were admitted to the ICU and if all patients were admitted to the hospital ward among less severely ill patients and among patients with cryptic septic shock. We implement our methodology in R, providing scripts in the supplementary material.

stat.AP↗
arXiv2015

Equivalence testing for functional data with an application to comparing pulmonary function devices

Equivalence testing for scalar data has been well addressed in the literature, however, the same cannot be said for functional data. The resultant complexity from maintaining the functional structure of the data, rather than using a scalar transformation to reduce dimensionality, renders the existing literature on equivalence testing inadequate for the desired inference. We propose a framework for equivalence testing for functional data within both the frequentist and Bayesian paradigms. This framework combines extensions of scalar methodologies with new methodology for functional data. Our frequentist hypothesis test extends the Two One-Sided Testing (TOST) procedure for equivalence testing to the functional regime. We conduct this TOST procedure through the use of the nonparametric bootstrap. Our Bayesian methodology employs a functional analysis of variance model, and uses a flexible class of Gaussian Processes for both modeling our data and as prior distributions. Through our analysis, we introduce a model for heteroscedastic variances within a Gaussian Process by modeling variance curves via Log-Gaussian Process priors. We stress the importance of choosing prior distributions that are commensurate with the prior state of knowledge and evidence regarding practical equivalence. We illustrate these testing methods through data from an ongoing method comparison study between two devices for pulmonary function testing. In so doing, we provide not only concrete motivation for equivalence testing for functional data, but also a blueprint for researchers who hope to conduct similar inference.

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

Instrumental Variables Estimation with Some Invalid Instruments and its Application to Mendelian Randomization

Instrumental variables have been widely used for estimating the causal effect between exposure and outcome. Conventional estimation methods require complete knowledge about all the instruments' validity; a valid instrument must not have a direct effect on the outcome and not be related to unmeasured confounders. Often, this is impractical as highlighted by Mendelian randomization studies where genetic markers are used as instruments and complete knowledge about instruments' validity is equivalent to complete knowledge about the involved genes' functions. In this paper, we propose a method for estimation of causal effects when this complete knowledge is absent. It is shown that causal effects are identified and can be estimated as long as less than $50$% of instruments are invalid, without knowing which of the instruments are invalid. We also introduce conditions for identification when the 50% threshold is violated. A fast penalized $\ell_1$ estimation method, called sisVIVE, is introduced for estimating the causal effect without knowing which instruments are valid, with theoretical guarantees on its performance. The proposed method is demonstrated on simulated data and a real Mendelian randomization study concerning the effect of body mass index on health-related quality of life index. An R package \emph{sisVIVE} is available online.

stat.ME↗
arXiv2014

Estimation of causal effects using instrumental variables with nonignorable missing covariates: Application to effect of type of delivery NICU on premature infants

Understanding how effective high-level NICUs (neonatal intensive care units that have the capacity for sustained mechanical assisted ventilation and high volume) are compared to low-level NICUs is important and valuable for both individual mothers and for public policy decisions. The goal of this paper is to estimate the effect on mortality of premature babies being delivered in a high-level NICU vs. a low-level NICU through an observational study where there are unmeasured confounders as well as nonignorable missing covariates. We consider the use of excess travel time as an instrumental variable (IV) to control for unmeasured confounders. In order for an IV to be valid, we must condition on confounders of the IV---outcome relationship, for example, month prenatal care started must be conditioned on for excess travel time to be a valid IV. However, sometimes month prenatal care started is missing, and the missingness may be nonignorable because it is related to the not fully measured mother's/infant's risk of complications. We develop a method to estimate the causal effect of a treatment using an IV when there are nonignorable missing covariates as in our data, where we allow the missingness to depend on the fully observed outcome as well as the partially observed compliance class, which is a proxy for the unmeasured risk of complications. A simulation study shows that under our nonignorable missingness assumption, the commonly used estimation methods, complete-case analysis and multiple imputation by chained equations assuming missingness at random, provide biased estimates, while our method provides approximately unbiased estimates. We apply our method to the NICU study and find evidence that high-level NICUs significantly reduce deaths for babies of small gestational age, whereas for almost mature babies like 37 weeks, the level of NICUs makes little difference. A sensitivity analysis is conducted to assess the sensitivity of our conclusions to key assumptions about the missing covariates. The method we develop in this paper may be useful for many observational studies facing similar issues of unmeasured confounders and nonignorable missing data as ours.

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

Mediation Analysis Without Sequential Ignorability: Using Baseline Covariates Interacted with Random Assignment as Instrumental Variables

In randomized trials, researchers are often interested in mediation analysis to understand how a treatment works, in particular how much of a treatment's effect is mediated by an intermediated variable and how much the treatment directly affects the outcome not through the mediator. The standard regression approach to mediation analysis assumes sequential ignorability of the mediator, that is that the mediator is effectively randomly assigned given baseline covariates and the randomized treatment. Since the experiment does not randomize the mediator, sequential ignorability is often not plausible. Ten Have et al. (2007, Biometrics), Dunn and Bentall (2007, Statistics in Medicine) and Albert (2008, Statistics in Medicine) presented methods that use baseline covariates interacted with random assignment as instrumental variables, and do not require sequential ignorability. We make two contributions to this approach. First, in previous work on the instrumental variable approach, it has been assumed that the direct effect of treatment and the effect of the mediator are constant across subjects; we allow for variation in effects across subjects and show what assumptions are needed to obtain consistent estimates for this setting. Second, we develop a method of sensitivity analysis for violations of the key assumption that the direct effect of the treatment and the effect of the mediator do not depend on the baseline covariates.

stat.ME↗
arXiv2011

The effect of winning an Oscar Award on survival: Correcting for healthy performer survivor bias with a rank preserving structural accelerated failure time model

We study the causal effect of winning an Oscar Award on an actor or actress's survival. Does the increase in social rank from a performer winning an Oscar increase the performer's life expectancy? Previous studies of this issue have suffered from healthy performer survivor bias, that is, candidates who are healthier will be able to act in more films and have more chance to win Oscar Awards. To correct this bias, we adapt Robins' rank preserving structural accelerated failure time model and $g$-estimation method. We show in simulation studies that this approach corrects the bias contained in previous studies. We estimate that the effect of winning an Oscar Award on survival is 4.2 years, with a 95% confidence interval of $[-0.4,8.4]$ years. There is not strong evidence that winning an Oscar increases life expectancy.

stat.AP↗
arXiv2011

Causal inference for continuous-time processes when covariates are observed only at discrete times

Most of the work on the structural nested model and g-estimation for causal inference in longitudinal data assumes a discrete-time underlying data generating process. However, in some observational studies, it is more reasonable to assume that the data are generated from a continuous-time process and are only observable at discrete time points. When these circumstances arise, the sequential randomization assumption in the observed discrete-time data, which is essential in justifying discrete-time g-estimation, may not be reasonable. Under a deterministic model, we discuss other useful assumptions that guarantee the consistency of discrete-time g-estimation. In more general cases, when those assumptions are violated, we propose a controlling-the-future method that performs at least as well as g-estimation in most scenarios and which provides consistent estimation in some cases where g-estimation is severely inconsistent. We apply the methods discussed in this paper to simulated data, as well as to a data set collected following a massive flood in Bangladesh, estimating the effect of diarrhea on children's height. Results from different methods are compared in both simulation and the real application.

math.ST↗
arXiv2010

Hidden Markov models for alcoholism treatment trial data

In a clinical trial of a treatment for alcoholism, a common response variable of interest is the number of alcoholic drinks consumed by each subject each day, or an ordinal version of this response, with levels corresponding to abstinence, light drinking and heavy drinking. In these trials, within-subject drinking patterns are often characterized by alternating periods of heavy drinking and abstinence. For this reason, many statistical models for time series that assume steady behavior over time and white noise errors do not fit alcohol data well. In this paper we propose to describe subjects' drinking behavior using Markov models and hidden Markov models (HMMs), which are better suited to describe processes that make sudden, rather than gradual, changes over time. We incorporate random effects into these models using a hierarchical Bayes structure to account for correlated responses within subjects over time, and we estimate the effects of covariates, including a randomized treatment, on the outcome in a novel way. We illustrate the models by fitting them to a large data set from a clinical trial of the drug Naltrexone. The HMM, in particular, fits this data well and also contains unique features that allow for useful clinical interpretations of alcohol consumption behavior.

stat.AP↗
arXiv2009

Bayesian testing of many hypotheses $\times$ many genes: A study of sleep apnea

Substantial statistical research has recently been devoted to the analysis of large-scale microarray experiments which provide a measure of the simultaneous expression of thousands of genes in a particular condition. A typical goal is the comparison of gene expression between two conditions (e.g., diseased vs. nondiseased) to detect genes which show differential expression. Classical hypothesis testing procedures have been applied to this problem and more recent work has employed sophisticated models that allow for the sharing of information across genes. However, many recent gene expression studies have an experimental design with several conditions that requires an even more involved hypothesis testing approach. In this paper, we use a hierarchical Bayesian model to address the situation where there are many hypotheses that must be simultaneously tested for each gene. In addition to having many hypotheses within each gene, our analysis also addresses the more typical multiple comparison issue of testing many genes simultaneously. We illustrate our approach with an application to a study of genes involved in obstructive sleep apnea in humans.

stat.ME↗
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.

math.ST↗