LINGUIST List 9.1438

Wed Oct 14 1998

Sum: GoldVarb

Editor for this issue: Brett Churchill <brettlinguistlist.org>


Directory

  1. Mario Cal Varela, Sum: GoldVarb

Message 1: Sum: GoldVarb

Date: Tue, 13 Oct 1998 18:37:20 +0200
From: Mario Cal Varela <iamariousc.es>
Subject: Sum: GoldVarb

Dear fellow linguists,
two weeks ago I posted a query on linguist list regarding the use of
GoldVarb. I want to thank the following people for their advice and comments:

Dennis R. Preston
Ron Smyth
Naomi Nagy
Hiroko Tajika
Barbara Avila-Shah

For those of you, who might be interested in using GoldVarb, I'll try to
summarize their points.

This was the original query:

>I'm right in the middle of an analysis of /t,d/ deletion in a second
language variety of English using Goldvarb 2.1. I have selected the step
up/step down option so I have got to the point where the program gives me a
probability value for each of the factors in those factor groups that
turned out to be significant. Now, in order to decide which of the factors
to collapse for further runs of the program, I need to know which of the
differences among the factors in each group are actually significant. But I
don't know how to proceed. Could anyone out there help me with this? Thank
you very much in advance.

Dennis Preston warns me that the right way to proceed is to run the step
up/ step down analysis first and discard the non-significant factor groups
(...)

>>You then go on to run VARBRUL itself (the so called-"binomial" run). When
>>you get the results from this you can investigate the probability for
>>"close-score" factors in individual factor groups. (A hint that there are
>>such things will be large (significant) chi-square values (which you don't
>>want).
>>
>>Of course, you can't combine close factors unless they "make sense"
>>linguistically. If, for example, you find that an environment beofre /l/
>>has a.62 probability and an environment before /r/ has a .58 probability,
>>then you are justified in collapsing them into a new (single) cateogry
>>within that factor group, called, say, "liquids." If two scores are similar
>>for low vowels and sibilants, there would be no such justification.
>>
>>Rememebr, you must test for significance AFTER collapse (comparing that run
>>to the previous one by using the log-likelihood method described in most of
>>the literature). If there is a significant difference between the two runs,
>>then you must go back to the pre-collapsed run.

The log-likelihood method referred to here seems to be well described in
the last chapter of R. Bailey & D. Preston. 1996. Second Language
Acquisition and Linguistic Variation. Benjamins. pp. 253-306. 

Barbara Avila kindly provided me with a short summary of the procedure:

-Each VARBRUL run gives you in the output a negative number which is
labeled log likelihood. For the two runs that are being compared, calculate
the difference between the two log likelihoods (Just take the absolute value).
-Multiply this number by two.
-Figure out the difference in degrees of freedom between the two runs. The
number in degrees of freedom for any run is the total number of factors in
the run minus the number of factor groups.
- Now look in a chi-square table. You have the number of degrees of freedom
that you just calculated as the difference between the two runs, and your
test statistic (chi-square value) is the doubled difference between the two
log-likelihoods. If your test statistic corresponds to a P of less than
.05, then the difference between the factors that you combined is
significant, and combining the factors obscures a significant influence on
the variation under study.

Finally, Ron Smyth calls my attention to two limitations of the variable
rule applications that could perhaps be commented on by more
statistically-oriented researches than myself. The first one has to do with
the fact that when the design has several factors, the output of the
program does not give any information about some of the interactions. The
other is that the program seems to handle nicely data with very few
subjects per cell, where other applications would not give out anything
significant. That is, GoldVarb does not keep track of subjects and seems to
disregard individual differences.


Mario Cal Varela
Departamento de Filoloxia Inglesa e Alemana, despacho 307
Facultade de Filoloxia
Universidade de Santiago de Compostela
c/ Burgo das Nacions s/n
Santiago 15705
ESPANA
tlf (981) 563100 ext. 11858
fax (981) 574646
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue