R help - Jul 2016 - R council with regard to the analysis needd

Dear R community,

I am a PhD student at the University of M?nster writing my thesis in
psycholinguistics.

I am currently running statistical analysis on a dataset with 3 fixed
factors (syntax, position), each with 2 levels and their interaction
(syntax*position). The accuracy rate is the dependent variable.

I first created a general linear mixed effects model (lme4), including both
the factors and their interaction and requested its summary.

 *M1 = glmer (correctness ~ syntax+position + syntax*position +(1|subj_nr)
+ (1|item_sf), data = df.rus2, family = binomial) *

I now want to test the contribution of the factors to the power of the
model, taken separately, by creating further models with individual factors
dropped and comparing the models to the baseline model.

My question is: should I drop the factors from the original model?

e.g. for the effect of syntax

*M2 = glmer (correctness ~ syntax + syntax*position +(1|subj_nr) +
(1|item_sf), data = df.rus2, family = binomial) *

Or should I test the main effects separately from the interaction?

The baseline model:

*M0 = glmer (correctness ~ syntax + position +(1|subj_nr) + (1|item_sf),
data = df.rus2, family = binomial)*

Models for comparison:

*M1 = glmer (correctness ~ syntax  +(1|subj_nr) + (1|item_sf), data df.rus2,
family = binomial)*


*M2 = glmer (correctness ~ position +(1|subj_nr) + (1|item_sf), data df.rus2,
family = binomial)*


*M3 = glmer (correctness ~ syntax * position +(1|subj_nr) + (1|item_sf),
data = df.rus2, family = binomial)*


Thank you in advance.
Sincerely
Julia Edeleva

	[[alternative HTML version deleted]]

> On Jul 9, 2016, at 9:12 AM, Julia Edeleva <psychexpert1992 at
gmail.com> wrote:
> 
> Dear R community,
> 
> I am a PhD student at the University of M?nster writing my thesis in
> psycholinguistics.
> 
> I am currently running statistical analysis on a dataset with 3 fixed
> factors (syntax, position), each with 2 levels and their interaction
> (syntax*position). The accuracy rate is the dependent variable.
> 
> I first created a general linear mixed effects model (lme4), including both
> the factors and their interaction and requested its summary.
> 
> *M1 = glmer (correctness ~ syntax+position + syntax*position +(1|subj_nr)
> + (1|item_sf), data = df.rus2, family = binomial) *
> 
> I now want to test the contribution of the factors to the power of the
> model, taken separately, by creating further models with individual factors
> dropped and comparing the models to the baseline model.
> 
In English statistical parlance power generally refers to the ability of
different statistical test to avoid so-called type 2 errors. The term goodness
of fit is usually used to describe the overall capacity of a model to capture or
apportion the sources of variation in a process.

> My question is: should I drop the factors from the original model?
> 
> e.g. for the effect of syntax
> 
> *M2 = glmer (correctness ~ syntax + syntax*position +(1|subj_nr) +
> (1|item_sf), data = df.rus2, family = binomial) *
> 
> Or should I test the main effects separately from the interaction?
> 
> The baseline model:
> 
> *M0 = glmer (correctness ~ syntax + position +(1|subj_nr) + (1|item_sf),
> data = df.rus2, family = binomial)*
> 
> Models for comparison:
> 
> *M1 = glmer (correctness ~ syntax  +(1|subj_nr) + (1|item_sf), data >
df.rus2, family = binomial)*
Let's call that one 'M4' since you already have an M1 defined above.
> 
> 
> *M2 = glmer (correctness ~ position +(1|subj_nr) + (1|item_sf), data >
df.rus2, family = binomial)*
And call that one 'M5' since a different M2 model was defined above.
> 
> 
> *M3 = glmer (correctness ~ syntax * position +(1|subj_nr) + (1|item_sf),
> data = df.rus2, family = binomial)*
M1, and M3 and M2 are the same models. Read: ?formula

M0 is nested in the M1 (or equivalent M3 or M2) model

M4 and M5 are nested in M0.

You should clarify with your research professor what your research hypothesis
(or hypotheses) might be. That is what should determine your model comparison
hierarchy.

-- 


David Winsemius
Alameda, CA, USA