What is the best approach for my research?

[Benjamin Spivak (Med)]

Dear List,

 

Please help I am stuck. I am performing some jury research with some very strange result and I want to find the ideal statistical model to fit my data.

 

My study. Essentially a 2x3 jury simulation study measuring comprehension of law. I also have over 18 DV's (questions) for comprehension. I have considered condensing the questions into one variable. But Cronbach's alpha is quite low for any combination of questions that I can relate to one construct.

 

Distribution is extremely non-normal and there is a heterogeneity of variance between groups. Also, because the jury study uses deliberating groups I have violated the assumption of independence.

 

Considering all this, I am having real trouble determining what the best approach for analysis would be. I have tried multi-level linear modelling, but I get results that I cannot make sense of.

 

I am at my wits end. If somebody could help, I would be in their debt.

 

Regards,

 

Ben.

 

 

 

 I would like to analyze individual jury data, but clearly my model violates the assumption of independence. So I have to account for this.

[Benjamin Spivak (Med)]

Thanks Zack,

Most of the DV's are binary (yes/no), some are continuous. I would like to look for differences between conditions, much like you would for an ANOVA. I'm not sure that CSLOGISTIC or CSORDINAL are capable of doing this. Although I would be happy to be wrong.

As for Multilevel models, I tried but could not make sense of the results (Also there was a convergence failure for many of the DV's).

Thanks again, if you have further advice I would greatly appreciate it.

Ben. 

  

On 23 February 2012 23:11, Zack, Matthew M. (CDC/ONDIEH/NCCDPHP) <[hidden email]> wrote:

Are your dependent variables nominal, ordinal, or interval-ratio [=continuous] variables?

Since your deliberating groups are correlated, you should take this into account by using

either multilevel models (GENLINMIXED) or complex samples (CSLOGISTIC, CSORDINAL, or

CSGLM).  If the questions implied by the dependent variables are so different and unrelated,

you will have to analyze them separately; otherwise, you might consider factor analysis to

explore how these variables may be related, although this may be problematic with

non-normally distributed variables.

 

Matthew Zack

 

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Benjamin Spivak
Sent: Thursday, February 23, 2012 1:14 AM
To: [hidden email]
Subject: What is the best approach for my research?

 

Dear List,

 

Please help I am stuck. I am performing some jury research with some very strange result and I want to find the ideal statistical model to fit my data.

 

My study. Essentially a 2x3 jury simulation study measuring comprehension of law. I also have over 18 DV's (questions) for comprehension. I have considered condensing the questions into one variable. But Cronbach's alpha is quite low for any combination of questions that I can relate to one construct.

 

Distribution is extremely non-normal and there is a heterogeneity of variance between groups. Also, because the jury study uses deliberating groups I have violated the assumption of independence.

 

Considering all this, I am having real trouble determining what the best approach for analysis would be. I have tried multi-level linear modelling, but I get results that I cannot make sense of.

 

I am at my wits end. If somebody could help, I would be in their debt.

 

Regards,

 

Ben.

 

 

 

 I would like to analyze individual jury data, but clearly my model violates the assumption of independence. So I have to account for this.

[Ryan]

Benjamin,

The first thing I would consider would be whether there is an underlying unidimensional construct underlying comprehension. You could fit a Rasch model via Winsteps, which allows for items with varying response options.

Ryan

On Thu, Feb 23, 2012 at 1:14 AM, Benjamin Spivak <[hidden email]> wrote:

Dear List,

 

Please help I am stuck. I am performing some jury research with some very strange result and I want to find the ideal statistical model to fit my data.

 

My study. Essentially a 2x3 jury simulation study measuring comprehension of law. I also have over 18 DV's (questions) for comprehension. I have considered condensing the questions into one variable. But Cronbach's alpha is quite low for any combination of questions that I can relate to one construct.

 

Distribution is extremely non-normal and there is a heterogeneity of variance between groups. Also, because the jury study uses deliberating groups I have violated the assumption of independence.

 

Considering all this, I am having real trouble determining what the best approach for analysis would be. I have tried multi-level linear modelling, but I get results that I cannot make sense of.

 

I am at my wits end. If somebody could help, I would be in their debt.

 

Regards,

 

Ben.

 

 

 

 I would like to analyze individual jury data, but clearly my model violates the assumption of independence. So I have to account for this.

[Benjamin Spivak (Med)]

Hi Ryan,

Thank you very much, I will look at Rasch models and attempt to apply it to my work. Could I trouble you with another question. 

Say, I do find that there is a uni-dimensional construct underlying comprehension. How could this assist me to look for differences between conditions?

Regards,

Ben

On 23 February 2012 23:41, R B <[hidden email]> wrote:

Benjamin,

The first thing I would consider would be whether there is an underlying unidimensional construct underlying comprehension. You could fit a Rasch model via Winsteps, which allows for items with varying response options.

Ryan

On Thu, Feb 23, 2012 at 1:14 AM, Benjamin Spivak <[hidden email]> wrote:

Dear List,

 

Please help I am stuck. I am performing some jury research with some very strange result and I want to find the ideal statistical model to fit my data.

 

My study. Essentially a 2x3 jury simulation study measuring comprehension of law. I also have over 18 DV's (questions) for comprehension. I have considered condensing the questions into one variable. But Cronbach's alpha is quite low for any combination of questions that I can relate to one construct.

 

Distribution is extremely non-normal and there is a heterogeneity of variance between groups. Also, because the jury study uses deliberating groups I have violated the assumption of independence.

 

Considering all this, I am having real trouble determining what the best approach for analysis would be. I have tried multi-level linear modelling, but I get results that I cannot make sense of.

 

I am at my wits end. If somebody could help, I would be in their debt.

 

Regards,

 

Ben.

 

 

 

 I would like to analyze individual jury data, but clearly my model violates the assumption of independence. So I have to account for this.

[Ryan]

Well, instead of dealing with 18 DVs, you would have a single "measure" of comprehension on which you can compare differences between conditions. 

There are certain details to this study that make it particularly challenging, not the least of which is the clustering of subjects in juries. This clustering (a.k.a. nesting) should be considered in order to arrive at the optimal data analytic strategy for your study. No more time to discuss at the moment. Hope this gives you some direction. Perhaps others can provide additional feedback.

Ryan

On Thu, Feb 23, 2012 at 8:13 AM, Benjamin Spivak <[hidden email]> wrote:

Hi Ryan,

Thank you very much, I will look at Rasch models and attempt to apply it to my work. Could I trouble you with another question. 

Say, I do find that there is a uni-dimensional construct underlying comprehension. How could this assist me to look for differences between conditions?

Regards,

Ben

On 23 February 2012 23:41, R B <[hidden email]> wrote:

Benjamin,

The first thing I would consider would be whether there is an underlying unidimensional construct underlying comprehension. You could fit a Rasch model via Winsteps, which allows for items with varying response options.

Ryan

On Thu, Feb 23, 2012 at 1:14 AM, Benjamin Spivak <[hidden email]> wrote:

Dear List,

 

Please help I am stuck. I am performing some jury research with some very strange result and I want to find the ideal statistical model to fit my data.

 

My study. Essentially a 2x3 jury simulation study measuring comprehension of law. I also have over 18 DV's (questions) for comprehension. I have considered condensing the questions into one variable. But Cronbach's alpha is quite low for any combination of questions that I can relate to one construct.

 

Distribution is extremely non-normal and there is a heterogeneity of variance between groups. Also, because the jury study uses deliberating groups I have violated the assumption of independence.

 

Considering all this, I am having real trouble determining what the best approach for analysis would be. I have tried multi-level linear modelling, but I get results that I cannot make sense of.

 

I am at my wits end. If somebody could help, I would be in their debt.

 

Regards,

 

Ben.

 

 

 

 I would like to analyze individual jury data, but clearly my model violates the assumption of independence. So I have to account for this.

[Benjamin Spivak (Med)]

Thank you for your help Ryan.

On 24 February 2012 00:36, R B <[hidden email]> wrote:

Well, instead of dealing with 18 DVs, you would have a single "measure" of comprehension on which you can compare differences between conditions. 

There are certain details to this study that make it particularly challenging, not the least of which is the clustering of subjects in juries. This clustering (a.k.a. nesting) should be considered in order to arrive at the optimal data analytic strategy for your study. No more time to discuss at the moment. Hope this gives you some direction. Perhaps others can provide additional feedback.

Ryan

On Thu, Feb 23, 2012 at 8:13 AM, Benjamin Spivak <[hidden email]> wrote:

Hi Ryan,

Thank you very much, I will look at Rasch models and attempt to apply it to my work. Could I trouble you with another question. 

Say, I do find that there is a uni-dimensional construct underlying comprehension. How could this assist me to look for differences between conditions?

Regards,

Ben

On 23 February 2012 23:41, R B <[hidden email]> wrote:

Benjamin,

The first thing I would consider would be whether there is an underlying unidimensional construct underlying comprehension. You could fit a Rasch model via Winsteps, which allows for items with varying response options.

Ryan

On Thu, Feb 23, 2012 at 1:14 AM, Benjamin Spivak <[hidden email]> wrote:

Dear List,

 

Please help I am stuck. I am performing some jury research with some very strange result and I want to find the ideal statistical model to fit my data.

 

My study. Essentially a 2x3 jury simulation study measuring comprehension of law. I also have over 18 DV's (questions) for comprehension. I have considered condensing the questions into one variable. But Cronbach's alpha is quite low for any combination of questions that I can relate to one construct.

 

Distribution is extremely non-normal and there is a heterogeneity of variance between groups. Also, because the jury study uses deliberating groups I have violated the assumption of independence.

 

Considering all this, I am having real trouble determining what the best approach for analysis would be. I have tried multi-level linear modelling, but I get results that I cannot make sense of.

 

I am at my wits end. If somebody could help, I would be in their debt.

 

Regards,

 

Ben.

 

 

 

 I would like to analyze individual jury data, but clearly my model violates the assumption of independence. So I have to account for this.

[Jon K Peck]

Rasch models are available in Statistics via several extension commands.

STATS IRM fits a 3-parameter item response model
STATS GRM fits a graded response model to ordinal data
SPSSINC RASCH estimates an item response model

All of these can be downloaded from the SPSS Community website (www.ibm.com/developerworks/spssdevcentral)  The R Essentials, also available from that site, is a prerequisite.


Jon Peck (no "h") aka Kim
Senior Software Engineer, IBM
[hidden email]
new phone: [phone]




From:        Benjamin Spivak <[hidden email]>
To:        [hidden email]
Date:        02/23/2012 06:20 AM
Subject:        [SPSSX-L] Fwd: What is the best approach for my research?
Sent by:        "SPSSX(r) Discussion" <[hidden email]>

---

Hi Ryan,

Thank you very much, I will look at Rasch models and attempt to apply it to my work. Could I trouble you with another question. 

Say, I do find that there is a uni-dimensional construct underlying comprehension. How could this assist me to look for differences between conditions?

Regards,

Ben


On 23 February 2012 23:41, R B <ryan.andrew.black@...> wrote:
Benjamin,

The first thing I would consider would be whether there is an underlying unidimensional construct underlying comprehension. You could fit a Rasch model via Winsteps, which allows for items with varying response options.

Ryan


On Thu, Feb 23, 2012 at 1:14 AM, Benjamin Spivak <benjamin.spivak@...> wrote:
Dear List,
 
Please help I am stuck. I am performing some jury research with some very strange result and I want to find the ideal statistical model to fit my data.
 
My study. Essentially a 2x3 jury simulation study measuring comprehension of law. I also have over 18 DV's (questions) for comprehension. I have considered condensing the questions into one variable. But Cronbach's alpha is quite low for any combination of questions that I can relate to one construct.
 
Distribution is extremely non-normal and there is a heterogeneity of variance between groups. Also, because the jury study uses deliberating groups I have violated the assumption of independence.
 
Considering all this, I am having real trouble determining what the best approach for analysis would be. I have tried multi-level linear modelling, but I get results that I cannot make sense of.
 
I am at my wits end. If somebody could help, I would be in their debt.
 
Regards,
 
Ben.
 
 
 
 I would like to analyze individual jury data, but clearly my model violates the assumption of independence. So I have to account for this.

[Zuluaga, Juan]

This kind of multivariate data, non-normal, non independent, seems to be of the type that Ecologists are used to. 
_Numerical Ecology_ by Legendre & Legendre (1998) may be helpful.  
While SPSS will give you the basic classical and nonmetric MDS and PCA, you may have to use other tools, like the package vegan in R. (Check out the companion volume _Numerical Ecology in R_).

---

From: Benjamin Spivak [[hidden email]]