(controlled by LCONV=, RCONV=, CONVERGE=, MPROX=, MJMLE=, CUTLO=, CUTHI=)

 

 TABLE 0.2 LIKING FOR SCIENCE (Wright & Masters p. ZOU042ws.txt Oct  9  9:00 2002

INPUT: 76 PUPILS, 25 ACTS                                          WINSTEPS 3.36

--------------------------------------------------------------------------------

 

                             CONVERGENCE TABLE

 

+----------------------------------------------------------------------------+

|    PROX          ACTIVE COUNT       EXTREME 5 RANGE      MAX LOGIT CHANGE  |

| ITERATION   PUPILS  ACTS    CATS     PUPILS  ACTS       MEASURES  STRUCTURE|

|----------------------------------------------------------------------------|

|        1       76      25     3       3.59    1.62        3.1355    -.1229 |

|        2       74      12     3       4.03    1.90         .3862    -.5328 |

|        3       74      12     3       4.19    1.96         .1356    -.0783 |

WARNING: DATA ARE AMBIGUOUSLY CONNECTED INTO 6 SUBSETS. MEASURES ACROSS SUBSETS ARE NOT COMPARABLE

 see Connection Ambiguities

----------------------------------------------------------------------------|

|    JMLE     MAX SCORE   MAX LOGIT    LEAST CONVERGED     CATEGORY STRUCTURE|

| ITERATION   RESIDUAL*    CHANGE    PUPIL  ACT      CAT   RESIDUAL   CHANGE |

|----------------------------------------------------------------------------|

|       1        -2.04       .2562       7      5*      2       -.72    .0003|

+----------------------------------------------------------------------------+

Standardized Residuals N(0,1)  Mean: .03 S.D.: 1.24

 

Look for scores and residuals in last line to be close to 0, and standardized residuals to be close to mean 0.0, S.D. 1.0.

 

The meanings of the columns are:

PROX   normal approximation algorithm - for quick initial estimates

ITERATION  number of times through your data to calculate estimates

 

ACTIVE COUNT number of parameters participating in the estimation process after elimination of deletions, and perfect (maximum possible) and zero (minimum possible) scores:

PERSONS person parameters

ITEMS  item parameters

CATS  rating scale categories - shows 2 for dichotomies

 

These counts may reduce because persons, items or categories may have been

1.deleted

2.dropped because they have no responses

3.dropped from standard estimation because they are unanchored and have extreme scores. These will be reported with Bayesian estimates.

 

EXTREME 5 RANGE

PERSONS The current estimate of the spread between the average measure of the top 5 persons and the average measure of the bottom 5 persons.

ITEMS The current estimate of the spread between the average measure of the top 5 items and the average measure of the bottom 5 items.

MAX LOGIT CHANGE

MEASURES maximum logit change in any person or item estimate. This i expected to decrease gradually until convergence, i.e., less than LCONV=.

STRUCTURE maximum logit change in any structure measure estimate - for your information - need not be as small as MEASURES.

 

JMLE   JMLE joint maximum likelihood estimation - for precise estimates

ITERATION  number of times through your data to calculate estimates

It is unusual for more than 100 iterations to be required

MAX SCORE RESIDUAL maximum score residual (difference between integral observed core and decimal expected score) for any person or item estimate - used to compare with RCONV=. This number is expected to decrease gradually until convergence acceptable.

*   indicates to which person or item the residual applies.

 

MAX LOGIT CHANGE maximum logit change in any person or item estimate - used to compare with LCONV=. This number is expected to decrease gradually until convergence is acceptable.

 

LEAST CONVERGED element numbers are reported for the person, item and category farthest from meeting the convergence criteria.

*   indicates whether the person or the item is farthest from convergence.

  the CAT (category) may not be related to the ITEM to its left. See Table 3.2 for details of unconverged categories.

 

CATEGORY RESIDUAL maximum count residual (difference between integral observed count and decimal expected count) for any response structure category - for your information. This number is expected to decrease gradually. Values less than 0.5 have no substantive meaning.

 

STRUCTURE CHANGE maximum logit change in any structure calibration. Not used to decide convergence, but only for your information. This number is expected to decrease gradually.

 

Standardized Residuals These are modeled to have a unit normal distribution. Gross departures from mean of 0.0 and standard deviation of 1.0 indicate that the data do not conform to the basic Rasch model specification that randomness in the data be normally distributed.