COMPGROUPS.NET | Search | Post Question | Groups | Stream | About | Register

### TO STATS: Does a generalized Dunnet's test exist ?

• Follow

```Hi there,

I need to compare 2 dose groups to a control group and initially
thought to conduct two separate ANCOVAS, controlling for a few
covariates in each model.

Now, I thought about the Dunnet's test, that allows to perform both
comparisons in a single test and thus, avoids to deal with
for covariates and the Dunnet's test compares means, not adjusted
Least Square Means.

The Dunnet's test takes, for comparison of group i versus placebo, the
following test statistics:  t(i )= (mean i - mean P) / sqrt(2 MSE/n),
where:
- MSE is the mean square error yielded by the ANOVA model Y = trtgroup
and
- n is the number of patients per trt group (assumed equal-sized in
this formula).
t(i) is then compared to pivotal values from the Dunnet's table.

Does one know whether there has been some generalization of that test,
with, for example

t(i) = (LSMean i - LSMean Placebo) / sqrt(2 MSE/n),

and
MSE being the mean square error yielded by an ANCOVA model Y =
trtgroup + covariates, and
n is the number of patients by trt group ?

Catherine.
```
 0
Reply cat.b41 (102) 4/29/2010 3:21:10 PM

```On Apr 29, 10:21=A0am, "cat.." <cat....@gmail.com> wrote:
> Hi there,
>
> I need to compare 2 dose groups to a control group and initially
> thought to conduct two separate ANCOVAS, controlling for a few
> covariates in each model.
>
> Now, I thought about the Dunnet's test, that allows to perform both
> comparisons in a single test and thus, avoids to deal with
> for covariates and the Dunnet's test compares means, not adjusted
> Least Square Means.
>
> The Dunnet's test takes, for comparison of group i versus placebo, the
> following test statistics: =A0t(i )=3D (mean i - mean P) / sqrt(2 MSE/n),
> where:
> - MSE is the mean square error yielded by the ANOVA model Y =3D trtgroup
> and
> - n is the number of patients per trt group (assumed equal-sized in
> this formula).
> t(i) is then compared to pivotal values from the Dunnet's table.
>
> Does one know whether there has been some generalization of that test,
> with, for example
>
> t(i) =3D (LSMean i - LSMean Placebo) / sqrt(2 MSE/n),
>
> and
> MSE being the mean square error yielded by an ANCOVA model Y =3D
> trtgroup + covariates, and
> n is the number of patients by trt group ?
>
>
> Catherine.

According to the documentation for PROC MIXED v9.1.3 the LSMEANS
statement does have Dunnett's

You can specify the following options in the LSMEANS statement after a
slash (/).
```
 0

```On 29 avr, 17:43, "data _null_;" <datan...@gmail.com> wrote:
> On Apr 29, 10:21=A0am, "cat.." <cat....@gmail.com> wrote:
>
>
>
>
>
> > Hi there,
>
> > I need to compare 2 dose groups to a control group and initially
> > thought to conduct two separate ANCOVAS, controlling for a few
> > covariates in each model.
>
> > Now, I thought about the Dunnet's test, that allows to perform both
> > comparisons in a single test and thus, avoids to deal with
> > multiplicity-adjusted p-values. My concern is that I need to adjust
> > for covariates and the Dunnet's test compares means, not adjusted
> > Least Square Means.
>
> > The Dunnet's test takes, for comparison of group i versus placebo, the
> > following test statistics: =A0t(i )=3D (mean i - mean P) / sqrt(2 MSE/n=
),
> > where:
> > - MSE is the mean square error yielded by the ANOVA model Y =3D trtgrou=
p
> > and
> > - n is the number of patients per trt group (assumed equal-sized in
> > this formula).
> > t(i) is then compared to pivotal values from the Dunnet's table.
>
> > Does one know whether there has been some generalization of that test,
> > with, for example
>
> > t(i) =3D (LSMean i - LSMean Placebo) / sqrt(2 MSE/n),
>
> > and
> > MSE being the mean square error yielded by an ANCOVA model Y =3D
> > trtgroup + covariates, and
> > n is the number of patients by trt group ?
>
>
> > Catherine.
>
> According to the documentation for PROC MIXED v9.1.3 the LSMEANS
> statement does have Dunnett's
>
> You can specify the following options in the LSMEANS statement after a
> slash (/).

That sounds interesting, I'll into the SAS doc about proc mixed. Many
thanks.
```
 0

2 Replies
460 Views

Similiar Articles:

7/27/2012 2:59:27 AM