Digital › Software2 min read

Skew-normal framework models triangular-valued data in symbolic data analysis, tested on climate data

Authors propose Bayesian and frequentist estimation for triangular-valued data, which carry a mode as well as bounds, and test it in simulations and on real climate data.

Skew-normal framework models triangular-valued data in symbolic data analysis, tested on climate data
Photo: Authors of the study: Jacqueline Peel and Hari M. Osofsky, Wikimedia Commons, CC BY 4.0 · License: CC BY 4.0

A paper covered by Discover Data describes a way to model triangular-valued data within the inferential framework of Symbolic Data Analysis (SDA). Triangular-valued data extend ordinary interval data by adding the mode of the values. The authors estimate parameters with both Bayesian Markov Chain Monte Carlo (MCMC) and frequentist methods, using skew-normal and generalized Wishart distributions. They evaluated the estimators in simulation studies and applied the framework to real climate data springer.comDiscover Dataspringer.comOpen source ↗.

Why the mode matters

Earlier inferential work on interval data mostly relied on midpoints and variances within a multivariate normal-Wishart framework. These models assume the values inside an interval are uniformly distributed and symmetric, which, according to the authors, often fails to capture the real structure of intervals. Daily weather shifts and financial returns are examples of asymmetric phenomena.

A triangular-valued observation is a triplet: minimum, mode and maximum. The triangular distribution is bounded and unimodal, and its lower-order moments have analytical expressions, which allows likelihood functions and Bayesian posteriors to be derived directly. The triplet is observed directly, so the approach avoids the identifiability problems of estimating an internal density from two bounds alone. Each triplet is converted into an internal mean, equal to one third of the sum of the three values, and a variance-covariance matrix, whose diagonal terms are computed from squares and pairwise products of the three values divided by eighteen. If the mode lies exactly midway between the bounds, the distribution is symmetric and classical normal and Wishart models may be adequate.

Model and estimation

The mean vectors are described by a multivariate skew-normal distribution with a location vector, a full symmetric positive-definite scale matrix and a skewness vector. The non-diagonal scale matrix captures dependencies between variables. The dispersion matrices are described by a generalized Wishart distribution, in which the degrees-of-freedom parameter d may be any positive real number rather than an integer springer.comDiscover Dataspringer.comOpen source ↗.

The joint log-likelihood has no closed-form solution, so maximum likelihood estimates are found numerically; the authors formulated a specialized optimization algorithm for this. Models are compared using the Bayes factor. The authors report that the analysis of real climate data demonstrates the versatility and precision of the model. The source material does not give numerical results of the comparison.

Sources · 2

Found a mistake?
  1. ↩

Read also

Up next

The next story opens automatically

More
Follow us

Type at least two letters — results appear instantly

↑↓to choose↵to openescto close