[R] Contrasts for ordered factors
lorenz.gygax at art.admin.ch
lorenz.gygax at art.admin.ch
Mon Jan 8 10:13:39 CET 2007
Dear all,
I do not seem to grasp how contrasts are set for ordered factors. Perhaps someone can elighten me?
When I work with ordered factors, I would often like to be able to reduce the used polynomial to a simpler one (where possible). Thus, I would like to explicetly code the polynomial but ideally, the intial model (thus, the full polynomial) would be identical to one with an ordered factor.
Here is a toy example with an explanatory variable (EV) with three distinct values (1 to 3) and a continuous response variable (RV):
options (contrasts= c ('contr.treatment', 'contr.poly'))
example.df < data.frame (EV= rep (1:3, 5))
set.seed (298)
example.df$RV < 2 * example.df$EV + rnorm (15)
I evaluate this data using either an ordered factor or a polynomial with a linear and a quadratic term:
lm.ord < lm (RV ~ ordered (EV), example.df)
lm.pol < lm (RV ~ EV + I(EV^2), example.df)
I then see that the estimated coefficients differ (and in other examples that I have come across, it is often even more extreme):
coef (lm.ord)
(Intercept) ordered(EV).L ordered(EV).Q
3.9497767 2.9740535 0.1580798
coef (lm.pol)
(Intercept) EV I(EV^2)
0.9015283 2.8774032 0.1936074
but the predictions are the same (except for some rounding):
table (round (predict (lm.ord), 6) == round (predict (lm.pol), 6))
TRUE
15
I thus conclude that the two models are the same and are just using a different parametrisation. I can easily interprete the parameters of the explicit polynomial but I started to wonder about the parametrisation of the ordered factor. In search of an answer, I did check the contrasts:
contr.poly (levels (ordered (example.df$EV)))
.L .Q
[1,] 7.071068e01 0.4082483
[2,] 9.073264e17 0.8164966
[3,] 7.071068e01 0.4082483
The linear part basically seems to be 0.707, 0 (apart for numerical rounding) and 0.707. I can understand that any evenspaced parametrisation is possible for the linear part. But does someone know where the value of 0.707 comes from (it seems to be the squareroot of 0.5, but why?) and why the middle term is not exactly 0?
I do not understand the quadratic part at all. Wouldn't that need the be the linear part to the power of 2?
Thank you for your thoughts! Lorenz

Lorenz Gygax
Swiss Federal Veterinary Office
Centre for proper housing of ruminants and pigs
Tänikon, CH8356 Ettenhausen / Switzerland
More information about the Rhelp
mailing list