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

A second evidence factor for a second control group

In an observational study of the effects caused by a treatment, a second control group is used in an effort to detect bias from unmeasured covariates, and the investigator is content if no evidence of bias is found. This strategy is not entirely satisfactory: two control groups may differ significantly, yet the difference may be too small to invalidate inferences about the treatment, or the control groups may not differ yet nonetheless fail to provide a tangible strengthening of the evidence of a treatment effect. Is a firmer conclusion possible? Is there a way to analyze a second control group such that the data might report measurably strengthened evidence of cause and effect, that is, insensitivity to larger unmeasured biases? Evidence factor analyses are not commonly used with a second control group: most analyses compare the treated group to each control group, but analyses of that kind are partially redundant; so, they do not constitute evidence factors. An alternative analysis is proposed here, one that does yield two evidence factors, and with a carefully designed test statistic, is capable of extracting strong evidence from the second factor. The new technical work here concerns the development of a test statistic with high design sensitivity and high Bahadur efficiency in a sensitivity analysis for the second factor. A study of binge drinking as a cause of high blood pressure is used as an illustration.

Sensitivity analyses informed by tests for bias in observational studies

In an observational study, the treatment received and the outcome exhibited may be associated in the absence of an effect caused by the treatment, even after controlling for observed covariates. Two tactics are common: (i) a test for unmeasured bias may be obtained using a secondary outcome for which the effect is known and (ii) a sensitivity analysis may explore the magnitude of unmeasured bias that would need to be present to explain the observed association as something other than an effect caused by the treatment. Can such a test for unmeasured bias inform the sensitivity analysis? If the test for bias does not discover evidence of unmeasured bias, then ask: Are conclusions therefore insensitive to larger unmeasured biases? Conversely, if the test for bias does find evidence of bias, then ask: What does that imply about sensitivity to biases? This problem is formulated in a new way as a convex quadratically constrained quadratic program and solved on a large scale using interior point methods by a modern solver. That is, a convex quadratic function of N variables is minimized subject to constraints on linear and convex quadratic functions of these variables. The quadratic function that is minimized is a statistic for the primary outcome that is a function of the unknown treatment assignment probabilities. The quadratic function that constrains this minimization is a statistic for subsidiary outcome that is also a function of these same unknown treatment assignment probabilities. In effect, the first statistic is minimized over a confidence set for the unknown treatment assignment probabilities supplied by the unaffected outcome. This process avoids the mistake of interpreting the failure to reject a hypothesis as support for the truth of that hypothesis. The method is illustrated by a study of the effects of light daily alcohol consumption on high-density lipoprotein (HDL) cholesterol levels. In this study, the method quickly optimizes a nonlinear function of N = 800 $N=800$ variables subject to linear and quadratic constraints. In the example, strong evidence of unmeasured bias is found using the subsidiary outcome, but, perhaps surprisingly, this finding makes the primary comparison insensitive to larger biases.

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 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.

Multivariate one-sided testing in matched observational studies as an adversarial game

We present a multivariate one-sided sensitivity analysis for matched observational studies, appropriate when the researcher has specified that a given causal mechanism should manifest itself in effects on multiple outcome variables in a known direction. The test statistic can be thought of as the solution to an adversarial game, where the researcher determines the best linear combination of test statistics to combat nature’s presentation of the worst-case pattern of hidden bias. The corresponding optimization problem is convex, and can be solved efficiently even for reasonably sized observational studies. Asymptotically, the test statistic converges to a chi-bar-squared distribution under the null, a common distribution in order-restricted statistical inference. The test attains the largest possible design sensitivity over a class of coherent test statistics, and facilitates one-sided sensitivity analyses for individual outcome variables while maintaining familywise error control through its incorporation into closed testing procedures.

Combining planned and discovered comparisons in observational studies

In observational studies of treatment effects, it is common to have several outcomes, perhaps of uncertain quality and relevance, each purporting to measure the effect of the treatment. A single planned combination of several outcomes may increase both power and insensitivity to unmeasured bias when the plan is wisely chosen, but it may miss opportunities in other cases. A method is proposed that uses one planned combination with only a mild correction for multiple testing and exhaustive consideration of all possible combinations fully correcting for multiple testing. The method works with the joint distribution of $\kappa^{T}\left( \mathbf{T}-\boldsymbol{\mu}\right) /\sqrt {\boldsymbol{\kappa}^{T}\boldsymbol{\Sigma\boldsymbol{\kappa}}}$ and $max_{\boldsymbol{\lambda}\neq\mathbf{0}}$$\,\lambda^{T}\left( \mathbf{T} -\boldsymbol{\mu}\right) /$$\sqrt{\boldsymbol{\lambda}^{T}\boldsymbol{\Sigma \lambda}}$ where $\kappa$ is chosen a priori and the test statistic $\mathbf{T}$ is asymptotically $N_{L}\left( \boldsymbol{\mu},\boldsymbol{\Sigma}\right) $. The correction for multiple testing has a smaller effect on the power of $\kappa^{T}\left( \mathbf{T}-\boldsymbol{\mu }\right) /\sqrt{\boldsymbol{\kappa}^{T}\boldsymbol{\Sigma\boldsymbol{\kappa} }}$ than does switching to a two-tailed test, even though the opposite tail does receive consideration when $\lambda=-\kappa$. In the application, there are three measures of cognitive decline, and the a priori comparison $\kappa$ is their first principal component, computed without reference to treatment assignments. The method is implemented in an R package sensitivitymult.

An exact test with high power and robustness to unmeasured confounding effects

In observational studies, it is agreed that the sensitivity of the findings to unmeasured confounders needs to be assessed. The issue is that a poor choice of test statistic can result in overstated sensitivity to hidden bias of this kind. In this article, a new adaptive test is proposed, guided by considerations of low sensitivity to hidden bias: it is tailored so that its power is greater than other leading tests, both in finite and infinite samples. One way of defining power in case of possible confounders is as the probability of reporting robustness (ie, insensitivity) of a true discovery to potential bias. In case of finite samples, we compute the power by simulations. When sample size approaches infinity, a meaningful indicator of the power is the design sensitivity, which is computed analytically and found to be better in the new test than in existing tests. Another asymptotic criterion for comparing tests when there is concern for confounders is Bahadur efficiency. The proposed test outperforms commonly used tests in terms of Bahadur efficiency in most sampling situations. The advantages of the new test mainly stem from its adaptivity: it combines two test statistics and consequently achieves the best design sensitivity and the best Bahadur efficiency of the two. As a "real-world" examination, we compare 441 daily smokers to 441 nonsmokers, to test the effect of smoking on periodontal disease. The new test is more robust to unmeasured confounders than both the Wilcoxon signed rank test and the paired t-test.

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 <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mn>2</mml:mn> <mml:mi>L</mml:mi> <mml:mo>+</mml:mo> <mml:mn>1</mml:mn></mml:math> test statistics in total. Although L binary covariates define <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msup><mml:mn>2</mml:mn> <mml:mi>L</mml:mi></mml:msup> </mml:math> interaction groups, only <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mn>2</mml:mn> <mml:mi>L</mml:mi> <mml:mo>+</mml:mo> <mml:mn>1</mml:mn></mml:math> tests are performed, and at least <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>L</mml:mi> <mml:mo>+</mml:mo> <mml:mn>1</mml:mn></mml:math> 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 <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>R</mml:mi></mml:math> package <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>submax</mml:mi></mml:math> which contains the NHANES example. An online Appendix provides simulation results and further analysis of the example.