[R] checkConv problems in R

[Saudi Sadiq]

Hi All,

I hope you will give me a hand with the checkConv problems.  have two
datasets, vowels and qaaf, and both have many columns. I am interested in
these 8 columns clarified as follows:

1.         convergence: DV (whether participants succeeded to use CA (Cairo
Arabic) instead of MA (Minia Arabic)

2.         speaker: 62 participants

3.         item: as pronounced

4.         style: careful/casual

5.         gender: males/females

6.         age: continues variable

7.         residence: urbanite/rural migrant/villager

8.         education: secondary or below/university/postgraduate



The only difference between the two datasets is the number of items. With
the vowels dataset, there are 1339 items; in the qaaf dataset there are
4064 items.

The aim of the test done was to know which independent variable is more
responsible for using CA forms. I used the lme4 package, function glmer.

I ran the model:

A.     modelvowels <- glmer(convergence ~ gender + age + residence +
education +  style+ (1|lexical.item) + (1|speaker), data=vowels,
family='binomial')

The message came on the screen:

B.     Warning message:

In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :

  Model failed to converge with max|grad| = 0.00210845 (tol = 0.001,
component 1)

Then I ran the model after removing STYLE as follows:

C.     modelvowels <- glmer(convergence ~ gender + age + residence +
education + (1|lexical.item) + (1|speaker), data=vowels, family='binomial')

This produced a result. Then, I ran

D.     plot(allEffects(modelvowels))

and this gave four charts (for the four independent variables: gender, age,
residence and education).

Then, I moved to the qaaf dataset (4064 items) and ran the same model

E.      modelqaaf <- glmer(convergence ~ gender + real.age + residence +
education +

                            (1|lexical.item) + (1|speaker), data=qaaf,
family='binomial')

which gave results with the vowels dataset but there was a warning message
this time

F.      Warning message:

            In checkConv(attr(opt, "derivs"), opt$par, ctrl =
control$checkConv,  :

            Model failed to converge with max|grad| = 0.429623 (tol =
0.001, component 8)

So, I removed one independent variable (residence) and ran this model again:

G.     modelqaaf <- glmer(convergence ~ gender + real.age + education +

                            (1|lexical.item) + (1|speaker), data=qaaf,
family='binomial')

This gave a result. I removed another independent variable (gender) after
returning (residence) and ran the model:

H.     modelqaaf1<- glmer(convergence ~  real.age + residence + education +

                            (1|lexical.item) + (1|speaker), data=qaaf,
family='binomial')

It also worked well and produced a result.


Now, my questions:

·     Why did not A work, why did C work, why did not E work though it has
the same four predictors of C,  why G and H worked with only three
predictors?

·     What are the packages that must be installed with, before or after
the lme4 package?

 Please, find attached the datasets.


 Best

-- 
Saudi Sadiq,
Assistant Lecturer, English Department,
Faculty of Al-Alsun,Minia University,
Minia City, Egypt &
PhD Student, Language and Linguistic Science Department,
University of York, York, North Yorkshire, UK,
YO10 5DD
http://york.academia.edu/SaudiSadiq
https://www.researchgate.net/profile/Saudi_Sadiq
Certified Interpreter by Pearl Linguistics

Forum for Arabic Linguistics conference رواق العربية
28-30th July 2015 - call for papers now open
https://sites.google.com/a/york.ac.uk/fal2015/