This selects which of LCONV= and RCONV= set the convergence criterion. See convergence considerations.
CONVERGE=L
|
LCONV= for "Logit change size" controls convergence.
Iteration stops when the biggest logit change is less or equal to LCONV=, or when the biggest logit change size increases (divergence).
|
CONVERGE=R
|
RCONV= for "Residual size" controls convergence.
Iteration stops when the biggest residual score is less or equal to RCONV=, or when the biggest residual size increases (divergence).
|
CONVERGE=E
|
Either LCONV= for "Logit change size" or RCONV= for "Residual size" controls convergence.
Iteration stops when the biggest logit change is less or equal to LCONV=, or when the biggest residual score is less or equal to RCONV=, or when both the biggest logit change size increases and the biggest residual size increases (divergence).
|
CONVERGE=B
|
Both LCONV= for "Logit change size" and RCONV= for "Residual size" controls convergence.
Iteration stops when both the biggest logit change is less or equal to LCONV= and the biggest residual score is less or equal to RCONV=, or when both the biggest logit change size increases and the biggest residual size increases (divergence).
|
CONVERGE=F
|
Force both LCONV= for "Logit change size" and RCONV= for "Residual size" to control convergence.
Iteration stops when both the biggest logit change is less or equal to LCONV= and the biggest residual score is less or equal to RCONV=.
|
Example 1: We want to be take a conservative position about convergence, requiring both small logit changes and small residual sizes when iteration ceases.
CONVERGE=Both
Example 2: We need very high precision, then specify:
CONVERGE=BOTH ; both score-residual and logit-change criteria
RCONV=.001 ; at most, one tenth of the smallest score residual needed.
LCONV=.00001 ; at most, one tenth of the highest precision to be reported.
These values are much, much smaller than the natural precision of the data, which is .5 raw score points, or the logit precision (S.E.) of the ability measures.
Example 3: We want to set the convergence criteria to match BIGSTEPS version 2.59
CONVERGE=B ; the criteria were LCONV= and RCONV=
RCONV= 0.5 ; the BIGSTEPS standards or whatever value you used
LCONV= .01
Example 4: We want to set the convergence criteria to match Winsteps version 3.20
CONVERGE=E ; the criterion was LCONV or RCONV
RCONV= 0.5 ; the 3.20 standards or whatever value you used
LCONV= .01
Example 5: We want the convergence criteria to match Winsteps version 2.85
CONVERGE= F ; force both LCONV and RCONV to be met
RCONV= 0.5 ; the 2.85 standards or whatever value you used
LCONV= .01
You may also want:
WHEXACT=NO ; centralized Wilson-Hilferty was the default
Example 6: Question: With anchored analyses, iterations never stop!
|----------------------------------------------------------------------------|
| JMLE MAX SCORE MAX LOGIT LEAST CONVERGED CATEGORY STEP |
| ITERATION RESIDUAL* CHANGE EXID BYCASE CAT RESIDUAL CHANGE |
|----------------------------------------------------------------------------|
| 1 -239.04 .5562 1993 392* 6 85.70 -.5960|
| 2 -105.65 -.1513 1993 392* 4 -28.92 .2745|
.....
| 18 -5.35 -.0027 2228 352* 3 2.35 .0146|
| 19 -5.16 .0029 2228 352* 3 2.31 .0106|
| 20 -5.05 .0025 2228 352* 3 2.28 .0055|
| 21 -5.00 .0010 2228 352* 3 2.26 .0075|
| 22 -4.99 -.0008 2228 352* 3 2.25 .0025|
.....
| 170 -5.00 -.0011 1377 352* 3 1.96 .0109|
| 171 -5.00 .0018 187 352* 3 1.96 .0019|
.....
The standard convergence criteria in Winsteps are preset for "free" analyses. With anchored analyses, convergence is effectively reached when the logit estimates stop changing in a substantively meaningful way. This has effectively happened by iteration 20. Note that the logit changes are less than .01 logits - i.e., even the biggest change would make no difference to the printed output (which is usually reported to 2 decimal places)
To have the current Winsteps do this automatically, set
CONVERGE=L
LCONV=.005 ; set to stop at iteration 22 - to be on the safe side.
