[R] strange result with contrasts

Arne.Muller@aventis.com Arne.Muller at aventis.com
Tue Apr 20 15:44:41 CEST 2004


I'm trying to reproduce some SAS result wit R (after I got suspicious with the result in R). I struggle with the contrasts in a linear model.

I've got three factors

> d$dose <- as.factor(d$dose)   # 5 levels
> d$time <- as.factor(d$time)   # 2 levels
> d$batch <- as.factor(d$batch) # 3 levels

the data frame d contains 82 rows. There are 2 to 4 replicates of each dose within each time point and each batch. There's one dose completely missing from one batch.

I then generate Dunnett contrasts using the multicomp library:

> contrasts(d$dose) <- contr.Dunnett(levels(d$dose), 1)
> contrasts(d$time) <- contr.Dunnett(levels(d$time), 1)
> contrasts(d$batch) <- contr.Dunnett(levels(d$batch), 1)

For the moment I'm just looking at the dose effects of the complete model:

> summary(lm(value ~ dose * time * batch, data = d))$coefficients[1:5,]
                   Estimate Std. Error     t value      Pr(>|t|)
(Intercept)      6.80211741 0.01505426 451.8399839 1.962247e-101
dose010mM-000mM -0.03454211 0.04113846  -0.8396549  4.046723e-01
dose025mM-000mM -0.01972550 0.04288981  -0.4599111  6.473607e-01
dose050mM-000mM -0.12015983 0.05356935  -2.2430704  2.886726e-02 <- significant
dose100mM-000mM  0.01252061 0.04113846   0.3043529  7.619872e-01

A collegue of mine has run the same data through a SAS program (listed below)

proc glm data = dftest;
  class dose time batch;
  model value = dose|time|batch;
  means dose / dunnett ('000mM');
  lsmeans dose /pdiff singular=1; 

Giving the following p-values:
          dose010mM-000mM  0.4047
          dose025mM-000mM  0.6474
          dose050mM-000mM  0.5745 <---
          dose100mM-000mM  0.7620

The p-values are the same expect for the one indicated.

A stripchart for the data in R shows that "dose050mM-000mM" should not be significant (it doesn't look different from e.g. "dose025mM-000mM").

Do you've any suggestions what I'm doing wrong here (assuming that I believe the SAS result)? Any hints what I can do to further analyse this problem?

	Many thanks for your help,


Arne Muller, Ph.D.
Toxicogenomics, Aventis Pharma
arne dot muller domain=aventis com

More information about the R-help mailing list