R help - Jun 2006 - ANCOVA in S-plus/R?

Dear R user:
   
  I have a question about doing ANCOVA in S-plus or R. 
   
  I know that many users use lm to do the regression and check the ANCOVA. But
is there a way to get the traditional Table form of the ANCOVA test through
S-plus (like what we would get from SPSS or SAS)?
   
  The problem I’m interested in is whether or not there is a treatment effect on
some medical measurement. I will have the endpoint (E) and the baseline (B) of
the measurement,   together with a dummy variable (T) of what treatment the
patients get (control, or new). I’m thinking of using lm( (E-B)  ~ B + T, data =
..)
   
  Question 1: if I get the regression results (E-B) = b0 + b1*B + b2*T and in
order to check whether or not there is a treatment effect, all I need to do is
to check whether the p-value associated with b2 is less than the significant
level I set or there’s more I need to do?
   
  Question 2: If I put  anova( lm( (E-B)  ~ B + T, data = ..) ), is the table I
get here the one I want for an ANCOVA? My concern is that ANCOVA is supposed to
work only on residual compared to ANOVA, but we are using the same command here.
If this gives me the ANCOVA table, does the order of the variables in regression
matter? Will
  anova( lm( (E-B)  ~  T + B, data = ..) ) gives the same result?
   
  Question 3: if I have another dummy variable to indicate the location of
hospital each patient goes to (L)  and want to include it to my model, should I
do
   lm( (E-B)  ~ B + T + L, data = ..)   or  I need to consider the interactions?
   
  Thanks a lot!
   
  Cindy

 __________________________________________________



	[[alternative HTML version deleted]]

To get the ANCOVA table you need to look at the anova() function.

The variables T and L must be factors to get the multi-degree-of-freedom
anova tables you are looking for.

Order matters.  You get the same residual, but the sequential sums of 
squares differ.

bt.aov <- aov(E ~ B + T)
anova(bt.aov)
gives you the (t-1)df test of common intercept vs variable intercept.
This is equivalent to comparing the adjusted means.


tb.aov <- aov(E ~ T + B)
anova(tb.aov)
gives you the 1df test of 0 slope (ordinary ANOVA) vs non-zero common slope.

lm() and aov() are essentially the same program.  The default output
is different.


When you bring in the blocking effect of hospital, you will most
likely want that sequentially first.

lbt.aov <- aov(E ~ L + B + T)
anova(lbt.aov)

If you want to test for interaction with the hospital effect, you can use

l.bt.aov <- aov(E ~ L * (B + T))
anova(l.bt.aov)