Methods you can count on: A simulation experiment showing the advantages generalized linear modeling (GLM) over the linear regression of log-transformed count data

Abstract

Count data are the oldest and most basic form of data in the field of archaeology. Count data, however, have long been known to defy the assumptions of many commonly used parametric statistical methods, such as linear regression, in generally failing to follow a normal distribution. Most archaeological analyses have dealt with this problem through the logarithmic transformation of count data for the purposes of normalization. Recently, other scientific fields have recognized major shortcomings associated with log transformation and they have adopted nonparametric alternatives, such as generalized linear modeling (GLM). With a few exceptions, the field of archaeology has not followed suit. This paper presents the results of a simulation experiment designed to compare the performance of linear regression of log-transformed data and GLM under some commonly occurring conditions for count data in the field of archaeology. The results of this simulation experiment confirm the striking superiority of GLM compared with the linear regression of log-transformed data, showing that the latter is particularly prone to false-negative results (i.e. Type II error). We argue that archaeologists, therefore, should join other scientific fields in more widely adopting GLM and its nonparametric counterparts in analyzing count data.