Generating multivariate continuous data via the notion of nearest neighbors

Abstract

Taylor and Thompson [15] introduced a clever algorithm for simulating multivariate continuous data sets that resemble the original data. Their approach is predicated upon determining a few nearest neighbors of a given row of data through a statistical distance measure, and subsequently combining the observations by stochastic multipliers that are drawn from a uniform distribution to generate simulated data that essentially maintain the original data trends. The newly drawn values are assumed to come from the same underlying hypothetical process that governs the mechanism of how the data are formed. This technique is appealing in that no density estimation is required. We believe that this data-based simulation method has substantial potential in multivariate data generation due to the local nature of the generation scheme, which does not have strict specification requirements as in most other algorithms. In this work, we provide two R routines: one has a built-in simulator for finding the optimal number of nearest neighbors for any given data set, and the other generates pseudo-random data using this optimal number.

Keywords:

• simulation
• random number generation
• density estimation
• bootstrap
• nearest neighbors