Sensitivity analysis for inverse probability weighting estimators via the percentile bootstrap

Qingyuan Zhao,Dylan S. Small,Bhaswar B. Bhattacharya

DOI: https://doi.org/10.1111/rssb.12327

2019-06-05

Abstract:<p>To identify the estimand in missing data problems and observational studies, it is common to base the statistical estimation on the 'missingness at random' and 'no unmeasured confounder' assumptions. However, these assumptions are unverifiable by using empirical data and pose serious threats to the validity of the qualitative conclusions of statistical inference. A sensitivity analysis asks how the conclusions may change if the unverifiable assumptions are violated to a certain degree. We consider a marginal sensitivity model which is a natural extension of Rosenbaum's sensitivity model that is widely used for matched observational studies. We aim to construct confidence intervals based on inverse probability weighting estimators, such that asymptotically the intervals have at least nominal coverage of the estimand whenever the data‐generating distribution is in the collection of marginal sensitivity models. We use a percentile bootstrap and a generalized minimax–maximin inequality to transform this intractable problem into a linear fractional programming problem, which can be solved very efficiently. We illustrate our method by using a real data set to estimate the causal effect of fish consumption on blood mercury level.</p>

statistics & probability

What problem does this paper attempt to address?

This paper aims to address the threats to the validity of statistical inferences in the problems of missing data and observational studies due to unverifiable assumptions (such as the "missing at random" and "no unmeasured confounders" assumptions). Specifically, the authors are concerned with how to assess the changes in statistical conclusions when these assumptions are violated to a certain extent. To this end, they propose a marginal sensitivity model, which is a natural extension of Rosenbaum's sensitivity model for matched observational studies.

The main contribution of the paper lies in constructing confidence intervals based on inverse - probability - weighted estimators, ensuring that these confidence intervals have at least a nominal coverage probability when the data - generating distribution belongs to a family of marginal sensitivity models. To achieve this goal, the authors use the percentile bootstrap method and the generalized minimax inequality to transform the originally intractable problem into a linear fractional programming problem, which can be solved efficiently.

### Specific problems the paper attempts to solve:

1. **Unverifiability of assumptions**: In the problems of missing data and observational studies, the commonly used "missing at random" (MAR) and "no unmeasured confounders" (NUC) assumptions cannot be verified by empirical data, which seriously threatens the validity of statistical inferences.

2. **Sensitivity analysis**: How to assess the changes in statistical conclusions when the above assumptions are violated to a certain extent. Specifically, the authors hope to still be able to provide valid confidence intervals within a certain range of violations of these assumptions to ensure the robustness of statistical inferences.

3. **Computational efficiency**: Traditional sensitivity analysis methods are often computationally intractable, especially when multiple possible assumption - violation scenarios need to be considered. The method proposed by the authors improves computational efficiency by transforming the problem into a linear fractional programming problem.

### Key points of the solution:

- **Marginal sensitivity model**: This is a non - parametric sensitivity model that quantifies the degree of violation of the "no unmeasured confounders" assumption through the ratio of conditional probabilities (i.e., the odds ratio).

- **Percentile bootstrap method**: Generate a large number of samples through the bootstrap method and calculate the point - estimate range under each sample to construct confidence intervals.

- **Linear fractional programming**: Transform the optimization problem into a linear fractional programming problem and use techniques such as the Charnes - Cooper transformation to improve computational efficiency.

### Application example:

The authors estimate the causal effect of fish consumption on blood mercury levels through an actual data set, demonstrating the effectiveness and practicality of the proposed method.

In summary, this paper solves the statistical inference problems caused by the unverifiability of assumptions in missing data and observational studies by proposing a new framework and method, and improves the practicality and computational efficiency of sensitivity analysis.