Weighted-bootstrap alignment of explanatory variables

Peter A. Hall, Xiaoyan Leng, Hans‐Georg MüllerJournal of Statistical Planning and Inference 2007 · DOI

TL;DR

This paper introduces a novel "weighted-bootstrap alignment" method to adjust for differences in explanatory variables between two groups, particularly when the relationship between the explanatory variable and the response is nonlinear or unknown. This method aligns the distributions of explanatory variables, making group comparisons more robust to confounding by these variables.

What Problem Does This Paper Address?

In statistical analysis, when comparing a response variable between two groups (e.g., treatment vs. control, or different populations), observed differences in the response might be due to underlying differences in other variables, known as covariates or explanatory variables. If these covariates are imbalanced between the groups, they can confound the results, making it difficult to determine the true effect of the group difference. Traditional methods like including covariates in a linear regression model work well when the relationship is linear, but fail when it's nonlinear or the model is misspecified. This paper addresses the challenge of adjusting for covariate imbalances in a more general setting, where the underlying model might be unknown or nonlinear.

Methods

The core of the paper is the "weighted-bootstrap alignment" algorithm. The key idea is to adjust the distribution of an explanatory variable in one group to match that of another group, or a pre-specified target distribution. This is achieved by constructing bootstrap weights for each group. These weights are chosen to minimize a distance measure between the weighted empirical distribution functions of the explanatory variable in the two groups. The paper proposes using divergence measures from the power divergence class to quantify this distance.

The implementation involves solving a constrained optimization problem to find these weights. The authors then use these weighted samples to perform subsequent analyses, effectively performing an adjustment. This approach is model-free in the sense that it doesn't require specifying a particular functional form for the relationship between the explanatory variable and the response.

Datasets

The paper uses two main sources of data:

  • Simulated data: For the simulation study, responses were generated from a logistic regression model with a quadratic predictor component. The explanatory variable X was drawn from two normal distributions with different means but the same variance.
  • Historical French-Canadian women data: This dataset concerns the fertility and mortality of women from the 17th-18th century, with information on immigration status and age-at-death. This data is used to illustrate the application of the weighted-bootstrap alignment in analyzing the relationship between number of children and age-at-death.

Key Insights and Ideas

  • Model-free adjustment: The weighted-bootstrap alignment offers a powerful way to adjust for covariate imbalances without needing to specify a complex statistical model for the relationship between variables. This is particularly useful when the underlying relationships are unknown or nonlinear.
  • Flexibility in alignment: The method allows for aligning one group's covariate distribution to another's, or to an arbitrary target distribution, offering flexibility in how the adjustment is performed.
  • Improved robustness: By aligning covariate distributions, the method aims to ensure that observed differences in the response are more likely attributable to the group effect itself, rather than confounding by covariates.
  • Generalizability: The authors suggest that this method can be adapted to various statistical models where estimators or tests can be represented as functionals of empirical distributions.

Key Results

  • Simulation Study: In a simulation with a misspecified logistic regression model, the weighted-bootstrap adjusted model (Model IV) performed well, yielding log odds ratios with a mean close to zero and good coverage rates for confidence intervals (93.2%), comparable to the true model (93.8%). This contrasts sharply with misspecified models that did not account for covariate imbalance, which showed significant biases and poor coverage (e.g., Model III had 53.1% coverage).
  • French-Canadian Women Data: Applying the method to historical fertility data, the weighted-bootstrap adjustment for age-at-death led to a larger estimated difference in the number of children between immigrant and native-born women. Specifically, the estimated difference decreased from -0.0675 to -0.0929, with a highly significant p-value (7.24 x 10^-40). This suggests that after adjusting for age-at-death, immigrant women had a greater reduction in the number of children compared to native-born women.

Limitations and Future Work

  • Overlap of Supports: The weighted-bootstrap alignment procedure requires that the supports of the observed marginal distributions of the covariates for the two groups overlap. In cases where overlap is limited, the analysis might need to be restricted to the overlapping range, potentially reducing efficiency. The authors suggest strategies like sampling more data or retaining subjects with similar covariate ranges to address this.
  • Computational Aspects: While the method is general, the implementation involves solving constrained optimization problems, which can be computationally intensive.
  • Future Research: The authors suggest further research into the method's behavior under alternative hypotheses and for a wider range of situations. They also propose exploring its use in the exploratory phase of model selection, where data can be aligned first, followed by model fitting on the remainder.

Summary updated

BibTeX
@article{hall2007weighted,
  author = {Peter A. Hall and Xiaoyan Leng and Hans‐Georg Müller},
  title = {{Weighted-bootstrap alignment of explanatory variables}},
  journal = {Journal of Statistical Planning and Inference},
  year = {2007},
  doi = {10.1016/j.jspi.2007.06.032},
}
Figure