Dear all,
one way of coding an indicator variable for inclusion in a regression model
is effect coding. Given an indicator variable X with 3 categories, the
effect coding procedure creates 2 dummy variables (X1 and X2) as follows:
X X1 X2
1 1 0
2 0 1
3 -1 -1
Using the above coding scheme, the intercept of the regression model is
equal to the unweighted average of the category means (of Y) of categories
1,2 and 3. And the slopes of X1 and X2 denote the difference between this
unweighted average and the means of categories 1 and 2, respectively. Would
it be possible to change the coding scheme such that the intercept always
(even in the unbalanced case) represents the grand mean, i.e. a weighted
average of the category means?
Any suggestions or references would be welcome.
Jerry
********************************************
Jerry Welkenhuysen-Gybels
Center for Survey Methodology
Faculty of Social Science
E. Van Evenstraat 2B
3000 Leuven
Belgium
Tel.: +32 16 32.31.92
Fax: +32 16 32.33.65
|