| Message/Author |
|
|
|
|
I have to do a polychoric CFA (WLSMV estimation) because I want to assess a presumed 5-factor structure of 31 items(4 answer options, ordinal).
Theoretically a bi-factor model would make most sense, such that all items load on a general factor whilst each item also loads on one subfactor (see input below).
However, I tried it multiple times and always get multiple errors like this:
the sample correlation between u1_1 and u2_5 is -0.986 due to one or more zero cells in their bivariate table. Information from these variables can be used to create a new variable.’’
I got 33 errors like this with different variable combinations. I checked the correlations between these variables in other statistics software but these tend to be low to moderate only.
I think because of this I also cannot retrieve any model fit indices, right?
Do you have any suggestions how to do is? I cannot just combine the variables into a new variable as suggested in the error message because theoretically this mostly doesn't make sense (and the correlations are not really high, in other programs). Do you have any idea how this is possible? |
|
|
|
|
VARIABLE:
NAMES ARE u1_1 u1_2 u1_3 u1_4 u1_5 u2_1 u2_2 u2_3 u2_4 u2_5 u2_6;
NAMES ARE u3_1 u3_2 u3_3 u3_4 u3_5 u3_6 u4_1 u4_2 u4_3 u4_4 u4_5 u4_6 u4_7;
NAMES ARE u5_1 u5_2 u5_3 u5_4 u5_5 u5_6 u6_1 u6_2 u6_3 u6_4 u6_5 u6_6 u6_7;
NAMES ARE u7_1 u7_2 u7_3 u7_4 u7_5 u7_6 u8_1 u8_2 u8_3 u8_4 u8_5 u8_6;
NAMES ARE u9_1 u9_2 u9_3 u9_4 u9_5 u9_6;
USEVAR ARE u1_1-u5_6;
MISSING ARE ALL (999);
CATEGORICAL ALL;
ANALYSIS:
type IS general;
estimator IS WLSMV
MODEL:
f1 BY u1_1* u1_2 u1_3 u1_4 u1_5;
f2 BY u2_1* u2_2 u2_3 u2_4 u2_5 u2_6;
f3 BY u3_1* u3_2 u3_3 u3_4 u3_5 u3_6;
f4 BY u4_1* u4_2 u4_3 u4_4 u4_5 u4_6 u4_7 u4_8;
f5 BY u5_1* u5_2 u5_3 u5_4 u5_5 u5_6;
fg BY u1_1* u1_2 u1_3 u1_4 u1_5 u2_1 u2_2 u2_3;
fg BY u2_4* u2_5 u2_6 u3_1 u3_2 u3_3 u3_4;
fg BY u3_5* u3_6 u4_1 u4_2 u4_3 u4_4 u4_5 u4_6;
fg BY u4_7* u5_1 u5_2 u5_3 u5_4 u5_5 u5_6;
fg WITH f1-f5 @0;
f1 WITH f2-f5 @0;
f2 WITH f3-f5 @0;
f3 WITH f4-f5 @0;
f4 WITH f5 @0;
fg @1;
f1 @1;
f2 @1;
f3 @1;
f4 @1;
f5 @1;
OUTPUT:
STANDARDIZED MODINDICES (ALL); |
|
|
|
|
| Perhaps the other programs don't declare the variables as categorical. That gives attenuated correlations. When you have many zero cells in your bivariate tables, polychoric correlations aren't really appropriate. You could use ML or Bayes estimators while still declaring the variables as categorical - those estimators do not depend on polychoric correlations. ML will be hard with as many as 5 factors but Bayes should be fine. |
|
|
|
|
| Also, we ask not to double post. Longer messages should be sent to Support. |
|
|
|
|
Thank you a lot for the information, and sorry for double posting!
I'm still confused about the correlations, because I performed spearman rho correlations in SPSS with the same data, and this also treats variables as categorical right? I'm still trying to figure out how this can be explained. |
|
|
|
|
| Spearman rho is a different correlation coefficient which considers ranking information. So it is not comparable to polychorics and factor analysis results will be different. |
|
|
| Back to top |
