Hi Lindsay,

If I understand correctly, you have a number of measured variables from the census about neighborhoods (zip code based?), and believe these can be accounted for by a smaller set of (continous? discrete?) latent variables, and you know these variables? Is that right? If you provide the correlation matrix, n, and model that should help readers.

 I would suggest to explore literature which can give you idea about the correlations between these variables. Next important thing is model fit has to be in the acceptable range. Keep one thing in mind we are trying to confirm the theory and SEM is just a tool. In Social science its usually difficult to get high fit indices. RMSEA should be in the range of 0.05-0.07, CFI - 0.8 - moderate fit and 0.9 or greater than 0.9 is a good fit. Take combination of fit indices- RMSEA, SRMR, GFI or AGFI and CFI.   Create different models based on the theory and choose the best.I hope it will help you to resolve your problem.All the best for your research. 

SES is typically conceived of as a formative variable, where all of its indicators form the construct. This is different from a reflective model (typical for CFA), where a construct gives rise to various indicators.

Here is an example of formative modeling: https://goo.gl/CxcAEA - also see work by Bagozzi. Formative models can (rightfully) be criticized for not being a measurement model - you cannot know how appropriate the indicators are, as you can with a reflective model (see https://goo.gl/QQwfqu). But in the case of SES, you have a strong argument for 'going formative.'

And incidentally, there is also a way to model the reliability of formative indicators - not possible/relevant here, perhaps, but in case it's of interest, see http://epub.wu.ac.at/3040/

Good luck!

Thank you all! These are very helpful answers. From the resources posted, an SEM seems like the best to use, and what I've done so far is approaching a good fit, but does not reach the ranges that are recommended for the fit statistics. My latent variables are continuous (neighborhood crime, safety, social support, and SES), the n is 2678, and the correlations are complex because there's about 30 endogenous variables that I'm using, but a picture of the partial correlation table is attached. I don't really understand how to specify the variance or covariance with such a complicated model.

If anyone can help me figure out how to improve the model, I would greatly appreciate it. The code with the best fit (CFI 0.92, SRMR 0.06, RMSEA 0.08) is as follows:

proc calis data=long2 modification;

path nses>med_hh_inc, nses>p_poverty, nses>p_bachplus, nses>p_50kless, nses>p_unemployed, nses>nod_rate, nses>nts_rate,

nsafety>clean, nsafety>build, nsafety>noise, nsafety>otbuild, nsafety>afraid, nsafety>chance, nsafety>trouble, nsafety>ndvi,

nsocial>dofavors, nsocial>visit, nsocial>closeknit, nsocial>helpneigh, nsocial>cantrust,

ncrime>asst, ncrime>burg, ncrime>larc, ncrime>murd, ncrime>mveh, ncrime>perc, ncrime>proc, ncrime>rape, ncrime>robb, ncrime>totc;

pvar nses=1, nsafety=1, ncrime=1, nsocial=1;

pcov afraid build chance trouble cantrust helpneigh otbuild clean noise, ndvi perc asst robb larc proc mveh burh murd rape, robb p_bach_plus;

run;

To enhance the model fit in CFA  we need to ensure that correlation between variables are significant ( Moderate to high). This is the first step. In SEM, to enhance the model fit we retain the most significant factors. I am not familiar with SAS. I did my SEM using AMOS. There we have modification indices which we use to enhance the fit. Your RMSEA is slightly higher i believe if you can remove any factor or variable which is not having much significance ( It means the factor's contribution to the model is less) the model fit may increase. SRMR seems to be fine. 

Reference : Hu and Bentler (1999) empirically examine various cut-offs for many of these measures, and there data suggest that to minimize Type I and Type II errors under various conditions, one should use a combination of one of the above relative fit indexes and the SRMR (good models < .08) or the RMSEA (good models < .06). RMSEA -0-05-0.08 is a reasonable fit. ( Browne and Cudeck,1993)

Refer : Reporting Structural Equation Modeling and Confirmatory Factor Analysis Results: A Review by

James B. Schreiber, Amaury Nora, Frances K. Stage, Elizabeth A. Barlow, and Jamie King

Just an FYI. If you are doing this type of research internationally, extremely difficult to get a coherent SES as the criteria will majorly shift among nations. 

these data you are using are not for SEM purpose. SEM is there to conceptualize things that are not tangible (latent variable) so you can use multiple indicators to measure the variable (along with errors too). Those variables you are using are defined and therefore there is no reason to specify them as indicators to measure ?? (what factor?). economic models would be better for you.