Blackwell Reference Online

Extract

Log-linear modeling is a data analysis technique used to explore relationships among categorical variables. Log-linear models express the logarithms of the expected cell frequencies from a multiway contingency table as a linear combination of the variables and their interactions. The simplest models yield a test equivalent to the chi square test of independence, but the technique allows exploration of relationships among more than two variables. Since log-linear models are additive with respect to the logarithm of the cell frequencies, they easily generate estimates of odds ratios. Although statisticians such as Pearson and Yule addressed the association between two categorical variables early in the twentieth century, development of techniques similar to the analysis of variance and linear regression for continuous outcome variables was slow. Birch (1963) proposed the log-linear model. Goodman and others made log-linear modeling popular in the 1960s and 1970s ( Goodman 1970 ; Bishop et al. 1975 ). Goodman and his colleagues made a computer program available in the 1970s. A log-linear model assuming that two cross-classified categorical variables, A and B, are associated is denoted AB and has the form where i = 1 to I indicates the level of variable A, j = 1 to J indicates the level of variable B, e ij is the expected frequency for the ijth cell of the table, and the ls are ... log in or subscribe to read full text