Improving Point and Interval Estimates of Monotone Functions by Rearrangement

Victor Chernozhukov, Iván Fernández‐Val, Alfred Galichon2008

TL;DR

This paper introduces a powerful "rearrangement" technique to improve statistical estimates of functions that are known to be monotonic (e.g., increasing with age). By applying this method, the authors show that both point estimates and confidence intervals become more accurate and shorter, respectively, without sacrificing coverage. This offers a principled way to enforce monotonicity constraints in statistical modeling.

What Problem Does This Paper Address?

Many real-world phenomena are inherently monotonic. For instance, a person's height generally increases with age, and a demand function typically shows that as price increases, demand decreases. In statistics, we often try to estimate these underlying monotonic relationships using data. However, standard estimation methods (like regression or kernel smoothing) can produce estimates that are not monotonic, even if the true underlying function is. This "non-monotonicity" can arise from noise in the data or limitations of the estimation method. These non-monotonic estimates are not only theoretically unsatisfying but can also lead to incorrect conclusions and less precise inferences. The challenge is to develop methods that can reliably produce monotonic estimates while also improving their statistical accuracy and the precision of associated confidence intervals.

Methods

The core of this paper is the rearrangement operation, a concept from variational analysis. For a univariate function $f(x)$, its increasing rearrangement $f^*(x)$ is essentially its quantile function. Intuitively, if you have a set of function values, rearranging them means sorting them. The paper extends this to multivariate functions by applying the rearrangement sequentially to each argument.

The key theoretical contribution is proving that this rearrangement operation has two crucial properties:

  1. Monotonization: It transforms any function into a monotonic one.
  2. Improvement: The rearranged function $f^*$ is always "closer" to the true underlying monotonic function $f_0$ in terms of $L^p$ norms than the original estimate $f$, especially if $f$ itself violates monotonicity. This improvement is often strict.

The paper also shows how this rearrangement can be applied to confidence intervals. By rearranging the lower and upper bounds of a confidence interval, the authors create a new, monotonic confidence interval that is shorter and has at least the same coverage probability as the original.

Datasets

The paper uses data from the 2003-2004 US National Health and Nutrition Survey. Specifically, they analyze a subsample of US-born white males aged 2-20 to estimate age-height growth charts. This dataset is used for empirical illustrations of their methods.

Key Insights and Ideas

  • Monotonicity as a Constraint: The paper frames monotonicity not just as a desirable property but as a constraint that can be actively enforced to improve statistical inference.
  • Rearrangement as a Projection: Geometrically, the rearrangement can be viewed as a projection onto the space of monotonic functions that share the same distribution as the original estimate.
  • Universal Improvement: The improvement property of rearrangement is remarkably general. It applies to both univariate and multivariate functions and doesn't depend on the specific estimation method used to obtain the initial (potentially non-monotonic) estimate.
  • Interval Shrinkage: Rearranging confidence intervals not only makes them monotonic but also demonstrably reduces their length, leading to more precise inference without sacrificing confidence.
  • Comparison to Isotonization: The paper contrasts rearrangement with isotonization (a common method for enforcing monotonicity). While both improve estimates, rearrangement is shown to be more generally applicable and often leads to better results, especially when the target function is steep.

Key Results

  • Point Estimates: The rearranged estimates $f^*$ are proven to be closer to the true monotonic function $f_0$ than the original estimates $f$ in $L^p$ norms. This improvement is strict if the original estimate violates monotonicity on a subset of positive measure.
  • Confidence Intervals: Rearranged confidence intervals $[l^*, u^*]$ are shown to be weakly shorter than original intervals $[l, u]$ in $L^p$ length. They also maintain or improve coverage probabilities.
  • Empirical Illustration: In the age-height chart example, rearrangement successfully produced monotonic estimates and confidence intervals, correcting issues like the slowdown in growth at older ages that non-monotonic estimates struggled with.
  • Monte Carlo Simulations: Tables in the paper demonstrate significant improvements in $L^p$ errors for rearranged estimates compared to original ones across various methods (kernel, local linear, splines, Fourier series). For confidence intervals, rearrangement led to substantial reductions in integrated length and increases in coverage, particularly for problematic Fourier series estimates.

Limitations and Future Work

  • No Single Best Method: The paper notes that while rearrangement generally improves estimates, there isn't a single "best" method among rearrangement, isotonization, or their combinations. The optimal choice can depend on the specific characteristics of the target function (e.g., its steepness).
  • Computational Aspects: While rearrangement is described as computationally tractable, the paper doesn't delve deeply into its computational complexity for high-dimensional cases.
  • Future Work: The authors suggest that their methods can be combined with other monotone estimation procedures (like isotonization) to achieve potentially even better results. The paper also hints at further applications in areas like econometrics and biometrics.

Summary updated

Abstract

Suppose that a target function f0 : Rd -> R is monotonic, namely weakly increasing, and an original estimate f of this target function is available, which is not weakly increasing. Many common estimation methods used in statistics produce such estimates f. We show that these estimates can always be improved with no harm by using rearrangement techniques: The rearrangement methods, univariate and multivariate, transform the original estimate to a monotonic estimate f*, and the resulting estimate is closer to the true curve f0 in common metrics than the original estimate f. The improvement property of the rearrangement also extends to the construction of confidence bands for monotone functions. Let l and u be the lower and upper endpoint functions of a simultaneous confidence interval [l,u] that covers the true function f0 with probability (1-a), then the rearranged confidence interval [l*,u*], defined by the rearranged lower and upper end-point functions l* and u*, is shorter in length in common norms than the original interval and covers the true function f0 with probability greater or equal to (1-a). We illustrate the results with a computational example and an empirical example dealing with age-height growth charts.
BibTeX
@misc{chernozhukov2008improving,
  author = {Victor Chernozhukov and Iván Fernández‐Val and Alfred Galichon},
  title = {{Improving Point and Interval Estimates of Monotone Functions by Rearrangement}},
  year = {2008},
}

Loading related content...

Figure