Re: st: reproducing published CFA results using -ssd- / -sem-

[Stas Kolenikov <skolenik@gmail.com>]

Danke shoen, Dirk.

I tried -iter(0)- originally, too. However, since Stata was failing
before the first iteration, it did not produce -e(b)-, as I was hoping
it would. -trace- did help, to some extent, but the parameter vector
only has non-descript -r1- and -c25- rather than nicely formatted
labels.

The default starting values are given by instrumental variables, and
can be turned off with -noivstart-. I described this approach in my
-confa- paper, and Stata used it for the general SEMs, as well
(although arguably without much documented trail). Unfortunately,
MTMMs are tricky, and getting the instruments right may be
complicated, and for whatever reasons, IV fails with this model+data
combination. With -noivstart-,  the covariances were padded with
zeroes, so the model estimation could proceed. Without -noivstart-,
i.e., with instrumental variables initialization, Stata may have
overwritten the covariances I constrained to be zero, which came into
contradiction and/or made the model underidentified.

-- Stas Kolenikov, PhD, PStat (SSC)
-- Senior Survey Statistician, Abt SRBI
-- Opinions stated in this email are mine only, and do not reflect the
position of my employer
-- http://stas.kolenikov.name



On Wed, Mar 13, 2013 at 7:12 AM, Dirk Enzmann
<dirk.enzmann@uni-hamburg.de> wrote:
> The algorithm Stata uses to set initial values I don't know. Perhaps you
> know already, but in any case: You can inspect them at least by using the
> combination of
>
> trace iterate(0) coefl
>
> as options of the -sem- command (note that -coefl- and -stand- can't be used
> simultaneously.
>
> Dirk
>
> Am 13.03.2013 04:26, schrieb Dirk Enzmann:
>
>> Stas,
>>
>> You should specify starting values of the covariances within the
>> methods- and traits factors, as well. The following commands converge.
>>
>> Unfortunately, I found no way to tell Stata to free the constraints of
>> the first indicator per latent variable (in order to fix the variances
>> of the latent variables to 1, instead).
>>
>> The results correspond to results obtainable using MPlus. They do not
>> reproduce all coefficients in Lorenz et al. (2007), however. There seems
>> to be an error in Table 6 of Lorenz et al. (2007).
>>
>> Other modifications are not substantial, they only help to improve the
>> legibility / comprehensibility of the model commands:
>>
>> * --- Start Stata Syntax: ----------------------------
>>
>> clear
>>
>> /* Reproducing CFA of Lorenz et al. (2007)
>>
>> Lorenz, F.O., Melby, J.N., Conger, R. D., & Xu, X. (2007). The effect
>> of context on the correspondence between observational ratings and
>> questionnaire reports of hostile behavior: A multitrait, multi-
>> method approach. Journal of Family Psychology, 21, 498-509.
>> */
>>
>> #delimit ;
>>
>> ssd init host_or angc_or antb_or
>> crit_so angr_so argu_so shou_so
>> crit_sp angr_sp argu_sp shou_sp
>> crit_po angr_po argu_po shou_po
>> crit_pp angr_pp argu_pp shou_pp;
>>
>> ssd set obs 292;
>>
>> /*
>> ssd set means 4.13 2.14 5.33
>> 2.27 1.70 1.87 1.21
>> 1.26 1.61 1.86 1.18
>> 2.41 1.76 2.11 1.30
>> 1.42 1.64 1.95 1.22;
>> */
>>
>> ssd set sd 2.36 1.70 1.87
>> 1.37 1.05 1.23 0.52
>> 0.98 0.96 1.38 1.06
>> 1.43 1.07 1.27 0.68
>> 1.17 1.07 1.36 1.23;
>>
>> ssd set correlations
>> 1.00 \
>> .73 1.00 \
>> .79 .58 1.00 \
>> .37 .28 .38 1.00 \
>> .39 .34 .42 .57 1.00 \
>> .39 .29 .42 .61 .78 1.00 \
>> .19 .24 .30 .39 .59 .58 1.00 \
>> .11 .02 .11 .35 .25 .35 .21 1.00 \
>> .26 .20 .21 .34 .31 .35 .21 .53 1.00 \
>> .30 .21 .23 .32 .30 .35 .12 .52 .65 1.00 \
>> .26 .19 .20 .35 .36 .34 .32 .60 .70 .64 1.00 \
>> .48 .31 .39 .37 .35 .37 .15 .16 .24 .20 .18 1.00 \
>> .44 .35 .38 .35 .49 .46 .25 .25 .29 .24 .30 .62 1.00 \
>> .50 .42 .47 .40 .51 .53 .30 .24 .32 .28 .32 .62 .74 1.00 \
>> .26 .28 .24 .25 .30 .26 .27 .16 .19 .18 .19 .26 .44 .44 1.00 \
>> .31 .21 .28 .34 .29 .34 .24 .33 .42 .34 .42 .46 .45 .44 .22 1.00 \
>> .34 .24 .25 .34 .31 .28 .20 .22 .44 .31 .46 .38 .43 .39 .19 .65 1.00 \
>> .28 .17 .24 .32 .27 .31 .17 .20 .45 .32 .43 .34 .38 .44 .23 .54 .63 1.00 \
>> .32 .17 .28 .33 .30 .29 .20 .26 .46 .33 .50 .44 .46 .42 .25 .68 .77 .64
>> 1.00;
>>
>> ssd status;
>>
>> foreach method in OR_Fac SO_Fac SP_Fac PO_Fac PP_Fac {;
>> foreach trait in Crit_Fac Angr_Fac Argu_Fac Shou_Fac {;
>> local c_cov `c_cov' `method'*`trait'@0;
>> };
>> };
>> di "cov( `c_cov' )";
>>
>> foreach factor in OR_Fac SO_Fac SP_Fac PO_Fac PP_Fac
>> Crit_Fac Angr_Fac Argu_Fac Shou_Fac {;
>> local v_opt `v_opt' (`factor', init(1));
>> };
>> di "var( `v_opt' )";
>>
>> foreach factor in SO_Fac SP_Fac PO_Fac PP_Fac {;
>> local c_cov `c_cov' (OR_Fac*`factor', init(0.2));
>> };
>> foreach factor in SP_Fac PO_Fac PP_Fac {;
>> local c_cov `c_cov' (SO_Fac*`factor', init(0.2));
>> };
>> foreach factor in PO_Fac PP_Fac {;
>> local c_cov `c_cov' (SP_Fac*`factor', init(0.2));
>> };
>> local c_cov `c_cov' (PO_Fac*PP_Fac, init(0.2));
>> foreach factor in Angr_Fac Argu_Fac Shou_Fac {;
>> local c_cov `c_cov' (Crit_Fac*`factor', init(0.2));
>> };
>> foreach factor in Argu_Fac Shou_Fac {;
>> local c_cov `c_cov' (Angr_Fac*`factor', init(0.2));
>> };
>> local c_cov `c_cov' (Argu_Fac*Shou_Fac, init(0.2));
>> di "cov( `c_cov' )";
>>
>> sem (OR_Fac@1 -> host_or@1)
>> (OR_Fac -> (angc_or, init(1))
>> (antb_or, init(1))
>> )
>> (SO_Fac@1 -> crit_so)
>> (SO_Fac -> (angr_so, init(1))
>> (argu_so, init(1))
>> (shou_so, init(1))
>> )
>> (SP_Fac@1 -> crit_sp)
>> (SP_Fac -> (angr_sp, init(1))
>> (argu_sp, init(1))
>> (shou_sp, init(1))
>> )
>> (PO_Fac@1 -> crit_po)
>> (PO_Fac -> (angr_po, init(1))
>> (argu_po, init(1))
>> (shou_po, init(1))
>> )
>> (PP_Fac@1 -> crit_pp)
>> (PP_Fac -> (angr_pp, init(1))
>> (argu_pp, init(1))
>> (shou_pp, init(1))
>> )
>> (Crit_Fac@1 -> crit_sp)
>> (Crit_Fac -> (crit_so, init(1))
>> (crit_po, init(1))
>> (crit_pp, init(1))
>> )
>> (Angr_Fac@1 -> angr_sp)
>> (Angr_Fac -> (angr_so, init(1))
>> (angr_po, init(1))
>> (angr_pp, init(1))
>> )
>> (Argu_Fac@1 -> argu_sp)
>> (Argu_Fac -> (argu_so, init(1))
>> (argu_po, init(1))
>> (argu_pp, init(1))
>> )
>> (Shou_Fac@1 -> shou_sp)
>> (Shou_Fac -> (shou_so, init(1))
>> (shou_po, init(1))
>> (shou_pp, init(1))
>> ), cov( `c_cov' ) var ( `v_opt' ) noconstant stand;
>>
>> #delimit cr
>>
>>
>> * --- End Stata Syntax. ------------------------------
>>
>> Dirk
>>
>> Tue, 12 Mar 2013 09:38:04 -0500, Stas Kolenikov <skolenik@gmail.com>
>> wrote:
>>
>>
>> http://hsphsun3.harvard.edu/cgi-bin/lwgate/STATALIST/archives/statalist.1303/date/article-544.html
>>
>>
>>
>> ========================================
>> Dr. Dirk Enzmann
>> Institute of Criminal Sciences
>> Dept. of Criminology
>> Rothenbaumchaussee 33
>> D-20148 Hamburg
>> Germany
>>
>> phone: +49-(0)40-42838.7498 (office)
>> +49-(0)40-42838.4591 (Mrs Billon)
>> fax: +49-(0)40-42838.2344
>> email: dirk.enzmann@uni-hamburg.de
>>
>> http://www2.jura.uni-hamburg.de/instkrim/kriminologie/Mitarbeiter/Enzmann/Enzmann.html
>>
>> ========================================
>
>
> --
> ========================================
> Dr. Dirk Enzmann
> Institute of Criminal Sciences
> Dept. of Criminology
> Rothenbaumchaussee 33
> D-20148 Hamburg
> Germany
>
> phone: +49-(0)40-42838.7498 (office)
>        +49-(0)40-42838.4591 (Mrs Billon)
> fax:   +49-(0)40-42838.2344
> email: dirk.enzmann@uni-hamburg.de
> http://www2.jura.uni-hamburg.de/instkrim/kriminologie/Mitarbeiter/Enzmann/Enzmann.html
> ========================================
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/