My thanks to individuals who responded to my query, in particular:
Andy Sloggett
Catherine Thomson
Geoff Der
Stephen Senn
Nick Taub
The consensus answer seems to be that indicator contrasts are
preferable. However I did an experiment to see whether, as reported by
one person, the 'simple' contrast in SPSS was not doing what the
description said it should. I did logistic regression using a large
project data file that I am involved with. The odds ratios and limits for
their confidence intervals from using indicator contrasts with the first
value as the reference category (= level) were identical with those
obtained using the 'simple' contrast with the same reference category -
only the constant term differed. So, unless one is using the model for
prediction, it makes no difference which one uses. This does not just
apply to binary explanatory variables, because to answer the question
of a colleague, I recoded missing values for the explanatory variables
as a third category. As it happened, repeating the analysis after
removing missing value cases (70% of the cases in fact, in this
'experiment') made very little practical difference to the resulting odds
ratios and their CIs. I think I may advise my colleague eliminating
cases containing missing values (as SPSS does) will make little
difference to the result, though to be on the safe side I might compare
the eliminated cases with the others w.r.t. baseline characteristics.
REgards
Miland Joshi
Department of Epidemiology and Public Health
University of Leicester
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|