Modelling trends in digit preference patterns

Abstract

Digit preference is the habit of reporting certain end digits more often than others. If such a misreporting pattern is a concern, then measures to reduce digit preference can be taken and monitoring changes in digit preference becomes important. We propose a two-dimensional penalized composite link model to estimate the true distributions unaffected by misreporting, the digit preference pattern and a trend in the preference pattern simultaneously. A transfer pattern is superimposed on a series of smooth latent distributions and is modulated along a second dimension. Smoothness of the latent distributions is enforced by a roughness penalty. Ridge regression with an L1 -penalty is used to extract the misreporting pattern, and an additional weighted least squares regression estimates the modulating trend vector. Smoothing parameters are selected by the Akaike information criterion. We present a simulation study and apply the model to data on birth weight and on self-reported weight of adults.