MIXED Multilevel Cross-Level Correlation

[jsberger]

Hi All,

I've got multi-level data (students nested within teachers), and i have some construct scale scores at the student level (level-1) and teacher level (level-2). For reporting purposes, i have to provide correlations among these scale scores, and i'm wondering how to go about this.  I can aggregate the student data (i.e., mean scores) and generate correlations, or i can use the full dataset, where the teacher scores will be replicated c times, where c = the number of students taught by a teacher. 

I came across a PDF on the HLM software website that discusses cross-level correlation calculation, but i don't find it to be particularly clear: http://www.ssicentral.com/hlm/help7/howto/How_to_Calculate_cross-level_correlations.pdf

Does anyone here have any advice, or experience with calculating these correlations as in the hyperlink above?

Thanks in advance,

Jason

[jsberger]

Hi All,

So here's what i have developed thus far in response to my question below, using the sample file created using:

data _null_;
set sample;
file "C:\Users\Jason\Desktop\sample.txt";
put teacherid student_id schoolid post_classman_ability post_pck_ability;
file print;
put teacherid student_id schoolid post_classman_ability post_pck_ability;
run;  

If i correlate two variables using all available data (including duplicate level-2 values), i get a correlation of -.12783.

**correlation between two retaining all data;
proc corr data=sample;
var post_classman_ability post_pck_ability;
run;

If i use a mixed model with the student-level variable as the DV and the teacher as the IV, obtaining the var-covariance matrix of fixed effects (which includes an intercept and the post_pck_ability predictor fixed effects), i can then use that covariance matrix to calculate the cross-level correlation as specified in the link in the original post.

**model with student-level measure as dv and teacher-level measure as iv;
proc mixed data=tssc_y1 method=ml;
class schoolid;
model post_classman_ability = post_pck_ability/ solution corrb covb;
random intercept/ subject=teacherid type=un;
ods output covb=covb;
run;

The IML syntax below attempts to calculate the cross-level correlation, obtaining a value of -.057711.  I'm glad to see the sign in the same direction, but i'm wondering if the decrease from -.12 down to .-05 seems reasonable, of if i'm missing something.

**correlation from covariance;
proc iml;
use covb;
read all var {Col1 Col2} into cov_mat[colname=varNames];
print cov_mat;
start Cov2Corr(S);
**grab the variances of each variable;
var_mat=(vecdiag(S))`;
*print var_mat;
**caculate sd by taking square root of product;
prod_sd=sqrt(var_mat[1,1]#var_mat[1,2]);
*print prod_sd;
**divide each element of covariance matrix by the prod_sd, focus on covariance cells;
r_mat=S/prod_sd;
r=r_mat[1,ncol(r_mat)];
return R;
finish;
r=Cov2Corr(cov_mat);

n=861;
t=sqrt(n-2)*sqrt((r*r)/(1-r*r));
p_onetail=1-probt(t,n-2);
p_twotail=(1-probt(abs(t),n-2))*2;
print r p_onetail p_twotail;
quit;