why different results with chi-square, wald and likelihood ratio tests?

[samantha lee]

Dear listers,

I am new here. I am running some categorical data type analyses where both
my dependent and independent variables are categorical data. Must admit I am
not familiar with these analyses and would greatly appreciate some advice on
why I might be getting different results with different tests.

I have one dependent (binary) variable and two independent (binary)
variables (V1 and V2) with a sample size of more than 190.

With chi-square tests, I get V1 = 6.8 (p=0.009), V2=15.3(p=0.000) and V1xV2
= 2.2 (p=0.136).

I tried binary logistic regression analysis as well in SPSS and received the
following Wald stats: V1 = 10.8(p=0.001), V2 = 0.9 (p=0.34), V1xV2 = 7.7
(p=0.005).

Reading some textbooks, they advised that Wald test is not as reliable as
Likelihood ratio test. The only likelihood ratio test function I found was
in SPSS multinomial logistic regression analysis. Tried that as well and
received the following results: V1=6.1 (p=0.013), V2=15.1(p=0000), V1xV2 =
8.1 (p=0.005).

Would anyone be able to advise why I am getting different results with these
tests and which would be the most appropriate test for categorical data?

Many thanks!

Cheers
Samantha

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Albert-Jan Roskam]

Dear all,

Could anybody point me to a good (peer-reviewed) reference about Iterative Proportional Fitting (IPF)?

Thanks in advance!

Cheers!!
Albert-Jan

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[vlad simion]

Hi Albert,

You can start with 'On least square adjustment of sampled frequency tables
when the expected marginal totals are known.'
Deming, W. and Stephan, F. (1940) Ann. Math. Statist., 6, 427–444.
Also, there are a couple of articles from Spss site regarding the RIM
weighting algorithm that Quantum uses to calculate weights.

Hth,
Vlad


On Thu, Jan 8, 2009 at 12:28 PM, Albert-jan Roskam <[hidden email]> wrote:

> Dear all,
>
> Could anybody point me to a good (peer-reviewed) reference about Iterative
> Proportional Fitting (IPF)?
>
> Thanks in advance!
>
> Cheers!!
> Albert-Jan
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>

====================To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[David Hitchin]

Quoting samantha lee <[hidden email]>:
> I have one dependent (binary) variable and two independent (binary)
> variables (V1 and V2) with a sample size of more than 190.
>
Chi-squared tests on tables treat both variables as if they had the same
status, and the simply provide a test of whether one variable is
unrelated to the other. They are not the right tests to use when you
have a dependent and independent variables.

It is easy to find examples where variable A predicts variable B very
well, but B is a poor predictor of A.

Logistic regression is one of the ways of dealing with the relationship
between a binary dependent variable and independents whether they are
binary, other categorical, ordinal or scale variables.

The reason that chi-squared tests and logistic regression produce
different results is that they are testing for different kinds of
relationships.

David Hitchin

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Albert-Jan Roskam]

Hi Vlad,

Thank you! The article you mentioned turned out to be freely available on www.projecteuclid.org:

http://projecteuclid.org/DPubS?service=UI&version=1.0&verb=Display&handle=euclid.aoms
http://tinyurl.com/725ze2

Cheers!!
Albert-Jan


--- On Thu, 1/8/09, vlad simion <[hidden email]> wrote:

> From: vlad simion <[hidden email]>
> Subject: Re: Reference about Iterative Proportional Fitting (IPF)
> To: [hidden email]
> Date: Thursday, January 8, 2009, 12:11 PM
> Hi Albert,
>
> You can start with 'On least square adjustment of
> sampled frequency tables
> when the expected marginal totals are known.'
> Deming, W. and Stephan, F. (1940) Ann. Math. Statist., 6,
> 427–444.
> Also, there are a couple of articles from Spss site
> regarding the RIM
> weighting algorithm that Quantum uses to calculate weights.
>
> Hth,
> Vlad
>
>
> On Thu, Jan 8, 2009 at 12:28 PM, Albert-jan Roskam
> <[hidden email]> wrote:
>
> > Dear all,
> >
> > Could anybody point me to a good (peer-reviewed)
> reference about Iterative
> > Proportional Fitting (IPF)?
> >
> > Thanks in advance!
> >
> > Cheers!!
> > Albert-Jan
> >
> > =====================
> > To manage your subscription to SPSSX-L, send a message
> to
> > [hidden email] (not to SPSSX-L), with no
> body text except the
> > command. To leave the list, send the command
> > SIGNOFF SPSSX-L
> > For a list of commands to manage subscriptions, send
> the command
> > INFO REFCARD
> >
>
> ====================To manage your subscription to SPSSX-L,
> send a message to
> [hidden email] (not to SPSSX-L), with no body
> text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the
> command
> INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Bob Schacht-3]

At 03:24 PM 1/7/2009, samantha lee wrote:
>Dear listers,
>
>I am new here. I am running some categorical data type analyses where both
>my dependent and independent variables are categorical data. Must admit I am
>not familiar with these analyses and would greatly appreciate some advice on
>why I might be getting different results with different tests.

In addition to the answers given by others, different tests have different
"power-efficiency" and "robustness" that vary with a number of factors,
including sample size. Tests probably also differ in the probability of
Type I and Type II errors.

It is also important to match your data types with the right kind of test.
If your data is all binary, why are you using the Wald test at all? If by
"Wald" you mean the Wald-Wolfowitz runs test, doesn't that require that
your data be on at least an ordinal scale? Or is there a Wald test other
than the Wald-Wolfowitz?

Bob Schacht

>I have one dependent (binary) variable and two independent (binary)
>variables (V1 and V2) with a sample size of more than 190.
>
>With chi-square tests, I get V1 = 6.8 (p=0.009), V2=15.3(p=0.000) and V1xV2
>= 2.2 (p=0.136).
>
>I tried binary logistic regression analysis as well in SPSS and received the
>following Wald stats: V1 = 10.8(p=0.001), V2 = 0.9 (p=0.34), V1xV2 = 7.7
>(p=0.005).
>
>Reading some textbooks, they advised that Wald test is not as reliable as
>Likelihood ratio test. The only likelihood ratio test function I found was
>in SPSS multinomial logistic regression analysis. Tried that as well and
>received the following results: V1=6.1 (p=0.013), V2=15.1(p=0000), V1xV2 =
>8.1 (p=0.005).
>
>Would anyone be able to advise why I am getting different results with these
>tests and which would be the most appropriate test for categorical data?
>
>Many thanks!
>
>Cheers
>Samantha
>
>=====================
>To manage your subscription to SPSSX-L, send a message to
>[hidden email] (not to SPSSX-L), with no body text except the
>command. To leave the list, send the command
>SIGNOFF SPSSX-L
>For a list of commands to manage subscriptions, send the command
>INFO REFCARD

Robert M. Schacht, Ph.D. <[hidden email]>
Pacific Basin Rehabilitation Research & Training Center
1268 Young Street, Suite #204
Research Center, University of Hawaii
Honolulu, HI 96814

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD