On Aug 6, 2011, at 03:23 , Walmes Zeviani wrote:
> Hi,
>
> I've data from an incomplete fatorial design. One level of a factor
doesn't
> has the levels of the other. When I use lm(), the summary() return NA for
> that non estimable parameters. Ok, I understant it. But I use
> contrast::contrast(), gmodels::estimable(), multcomp::glht() and all these
> fail when model has NA estimates. This is becouse vcov() and coef() has
> different dimensions. Is possible set lm() to omit NA? Below same toy data
> and code.
>
>> # toy data
>> adi <- expand.grid(cult=gl(1,3,la=LETTERS[1]), fert=101)
>> fat <- expand.grid(cult=gl(2,3,la=LETTERS[2:3]),
fert=seq(50,150,50))
>> da <- rbind(adi, fat)
>> da$y <- da$fert+rnorm(nrow(da),0,10)
>>
>> # plot
>> require(lattice)
>> xyplot(y~fert|cult, da)
>>
>> # adjust
>> m0 <- lm(y~cult*fert, da)
>> summary(m0)
>
> ...
> Coefficients: (1 not defined because of singularities)
> Estimate Std. Error t value Pr(>|t|)
> (Intercept) 7.55401 10.18956 0.741 0.469
> cultB -8.04486 13.54672 -0.594 0.561
> cultC -3.83644 6.74848 -0.568 0.578
> fert 0.87486 0.08265 10.586 1.24e-08 ***
> cultB:fert 0.13509 0.11688 1.156 0.265
> cultC:fert NA NA NA NA
> ...
>> require(gmodels)
>> estimable(m0, cm=c(1,0,0,100,0,0))
> Erro em estimable.default(m0, cm = c(1, 0, 0, 100, 0, 0)) :
> Dimension of structure(c(1, 0, 0, 100, 0, 0), .Dim = c(1L, 6L)): 1x6, not
> compatible with no of parameters in m0: 5
...and of course (you forgot to say) just removing the 6th element of cm
won't do either.
A generic trick is to modify the model matrix directly:
> M <- model.matrix(m0)[,-6]
> m1 <- update(m0, ~ M - 1)
> summary(m1)
Call:
lm(formula = y ~ M - 1, data = da)
Residuals:
Min 1Q Median 3Q Max
-27.165 -4.704 -2.446 6.666 25.328
Coefficients:
Estimate Std. Error t value Pr(>|t|)
M(Intercept) -1.8618 14.6718 -0.127 0.901
McultB 2.2102 19.5058 0.113 0.911
McultC 11.5308 9.7171 1.187 0.253
Mfert 0.9506 0.1190 7.989 5.65e-07 ***
McultB:fert 0.0758 0.1683 0.450 0.658
---
Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1
Residual standard error: 14.57 on 16 degrees of freedom
Multiple R-squared: 0.9866, Adjusted R-squared: 0.9824
F-statistic: 235.5 on 5 and 16 DF, p-value: 2.165e-14
> estimable(m1, cm=c(1,0,0,100,0))
Estimate Std. Error t value DF Pr(>|t|)
(1 0 0 100 0) 93.20357 8.415451 11.07529 16 6.515116e-09
Two things to notice: The coefficient names are not quite the same as the
original, and you need to fit without the intercept (or remove the 1st column as
well).
--
Peter Dalgaard, Professor,
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Email: pd.mes at cbs.dk Priv: PDalgd at gmail.com
"D?den skal tape!" --- Nordahl Grieg