Re: Best model/analytic strategy given data?

Posted by Rich Ulrich on May 31, 2013; 8:08pm
URL: http://spssx-discussion.165.s1.nabble.com/Best-model-analytic-strategy-given-data-tp5720498p5720514.html

Gene (and Deena),

SEM is not simple.  It certainly is not a magical way to make good sense of
data that are not readily modeled in simpler terms.  Rather, it is a sophisticated
presentation of a whole set of inter-relationships. 

I suspect that the advisor is giving excellent and proper advice when he rejects
SEM for *this* set of proposed predictors and mediators. 

Cross-sectional?  How to you differentiate cause versus effect? 
And:  How do you "operationalize" constructs like "social bonds" and "criminal
justice injustices" on a subject-level?

Finally, I wonder if there is going to be enough information -- enough delinquency --
in any "general sample" of 800.  Positively skewed, yes.  Are there more that 50
cases that are not zero?  Off-hand, recoding to (0, 1, 2+) seems like not-a-bad-idea,
but what-to-look-at also depends on what the observed distribution happens to be.

--
Rich Ulrich

---

Date: Fri, 31 May 2013 11:27:00 -0400
From: [hidden email]
Subject: Re: Best model/analytic strategy given data?
To: [hidden email]

Deena,

It seems that your chair has already set limits to what you can be allowed to do. Given those, I doubt that anything that anybody on this listserv or the semnet listserv will change her/his limits. I think your problem is small ‘p’ political rather than statistical. I hope that you have a statistics/methodology person in your department/school or an ‘outside’ such person who might be respected by your chair and who you can work with to persuade your chair to give up his/her preconceived and incorrect ideas. Good luck.  Gene Maguin

 

 

 

 

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Deena Isom
Sent: Friday, May 31, 2013 9:14 AM
To: [hidden email]
Subject: Best model/analytic strategy given data?

 

Hi Everyone!

 

I am writing my dissertation proposal, and my advisor would like me to outline my potential analytic strategy.  I, however, am using a restricted secondary data set which I have not received yet and only have access to the survey itself.  So, I can only guess from past publications using the data and what little information is on the website about missing data and so forth. 

 

The data is cross-sectional with potentially around 800 respondents.

 

In general, I plan to test the following relationships:

 

racial discrimination (RD) and criminal justice injustices (CJI) --> delinquency

 

RD and CJI --> negative emotions and social bonds --> delinquency

 

Racial Socialization X (RD and CJI) --> negative emotions and social bonds --> delinquency

 

So,  mediating effects and interaction effects....

 

Given that data comes from a general sample, the delinquency indicator is likely positively skewed.  Furthermore, many of the IVs (which there are more than listed above) are likely correlated. 

 

I had originally proposed to use SEM, but my advisor didn't like that idea because of the cross-sectional nature of the data.  So, given the limits of the DV, I decided I could recode it from a count of "have you ever done (a given delinquent act)" to an ordered categorical variable of non-offender, one-time offender, and repeat offender.  Doing this I proposed to used ordered logit, which also allows for multicollinearity.  He didn't like me recoding the DV and wants it to remain a count....

 

So, I would appreciate any advice on what analytical strategy to propose.  I don't want to do OLS.  But, what else can I do?  Also, if I do go that route, I would need to log the DV, correct?

 

Again, any advice pointing me towards a more sophisticated strategy than OLS, yet still simplified, would be greatly appreciated.

 

Thanks!!  Have a great weekend!  

--
~Deena