If SCOREFILE=filename is specified, a file is output which contains the measure and model standard error corresponding to every possible score on a test consisting of all the non-extreme items.
The SCOREFILE= and Table 20 person ability estimates are estimated on the basis that the current item difficulty estimates are the "true" estimates. These are the person estimates if you anchored (fixed) the items at their reported estimates. PFILE=, Table 17 and the other Person Measure Tables show the person abilities that are the maximum likelihood estimates at the current stage of estimation. To make these two sets of estimates coincide, please tighten the convergence criteria in your Winsteps control file:
CONVERGE=L
LCONV=*.001 ; or tighter
If you want the score file for person measures including the extreme (zero, perfect) items, then
1. Run a standard analysis.
2. Output: IFILE=if.txt SFILE=sf.txt
3. Run the analysis again, with Extra specifications: IAFILE=if.txt SAFILE=sf.txt
4. The person measures will have altered somewhat to adjust for the imputed difficulties of the extreme items.
5. Output Table 20 and SCOREFILE=
SCOREFILE=? opens a Browse window
This is also shown in Table 20. It has 2 heading lines (unless HLINES=N or ROW1HEADING=N), and has the format:
KID SCORE FILE FOR KNOX CUBE TEST Dec 12 2020 10: 8 USCALE=1.00
SCORE MEASURE S.E. INFO NORMED S.E. FREQUENCY % CUM.FREQ. % PERCENTILE 1 ... 4 ... 18
3 -6.66 1.88 .28 217 85 1 2.9 1 2.9 1.4 1.00 .00 .00
4 -5.30 1.11 .81 278 50 0 .0 1 2.9 2.9 1.00 .29 .00
5 -4.35 .88 1.29 321 40 1 2.9 2 5.7 4.3 1.00 .51 .00
6 -3.64 .82 1.49 353 37 2 5.7 4 11.4 8.6 1.00 .68 .00
7 -2.97 .82 1.48 383 37 2 5.7 6 17.1 14.3 1.00 .81 .00
8 -2.26 .88 1.29 415 40 2 5.7 8 22.9 20.0 1.00 .90 .00
9 -1.39 .99 1.01 454 45 3 8.6 11 31.4 27.1 1.00 .95 .00
10 -.26 1.11 .81 505 50 12 34.3 23 65.7 48.6 1.00 .98 .00
11 .94 1.05 .90 559 48 5 14.3 28 80.0 72.9 1.00 1.00 .00
12 1.96 .98 1.05 605 44 4 11.4 32 91.4 85.7 1.00 1.00 .00
13 2.88 .95 1.12 646 43 1 2.9 33 94.3 92.9 1.00 1.00 .00
14 3.76 .94 1.14 686 42 2 5.7 35 100.0 97.1 1.00 1.00 .00
15 4.65 .96 1.07 726 43 0 .0 35 100.0 100.0 1.00 1.00 .00
16 5.73 1.16 .75 775 52 0 .0 35 100.0 100.0 1.00 1.00 .00
17 7.15 1.90 .28 839 86 0 .0 35 100.0 100.0 1.00 1.00 .00
1. SCORE: Score on test of all items. TOTALSCORE=Yes includes extreme items. TOTALSCORE=No excludes extreme items (if any).
The score file shows integer raw scores, unless there are decimal weights for IWEIGHT=. In which case, scores to 1 decimal place are shown. To obtain other decimal raw scores for short tests, go to the Graphs pull-down menu. Select "Test Characteristic Curve". This displays the score-to-measure ogive. Click on "Copy data to clipboard". Open Excel. Paste. There will be to three columns. The second column is the measure, the third column is the raw score.
2. MEASURE: Measure (user-scaled by USCALE=)
If this value is not the same as in PFILE=
1. Missing data. The measure in the PFILE= is based on the responses observed. The measure in the SCOREFILE= is based on all the items.
2. Convergence criteria. The convergence criteria: CONVERGE=, LCONV=, RCONV= are not tight enough for your data. Please try setting LCONV= and RCONV= to smaller values. SCFILE= measures are estimated from the item estimates reported in Table 14 using the technique at https://www.rasch.org/rmt/rmt102t.htm
a. PFILE= values are obtained from the main estimation procedure, controlled by LCONV=, RCONV= and CONVERGE=
b. SCOREFILE= person values are obtained from the stored item estimates from the main analysis, controlled by LCONV/1000 and RCONV/1000 which is higher numerical precision than a.
If you need 1. and 2. to report the same values please set LCONV= and/or RCONV= smaller. For example, LCONV=.001 and RCONV=.005. This will cause more estimation iterations.
3. S.E.: Standard error (user scaled by USCALE=) - model, because empirical future misfit is unknown.
4. INFO: Statistical information in measure (=1/Logit S.E.²) = observable points of the Test Information Function (TIF).
Measures locally-rescaled, so that sample mean=500, standard deviation=100
5. NORMED: Measure (rescaled)
6. S.E.: Standard error (rescaled)
Sample distribution:
7. FREQUENCY: Count of sample at this measure
8. %: Percent of sample at this measure
9. CUM.FREQ.: Count of sample at or below this measure
10. %: Percent of sample at or below this measure
11. PERCENTILE: The percent of the sample below the current measure.There are several definitions of percentile. Winsteps uses this one: Percentile Rank: The percentile is the cumulative frequency percent for the score below + half the frequency percent for the current score, half-rounded, and constrained to the range 1-99 for non-zero frequencies: . Percentiles for other definitions can be computed from the CUM. FREQ column. See, for instance, how-to-calculate-percentile. If the data are complete, then the raw data and the Rasch measures produce the same percentiles. If the data are incomplete, then the Rasch measures are used because they are robust against missing data.
Expected scores on items:
12. - ... one score for each item. This can be used for inferring raw scores from Angoff-type standard-setting scores.
If CSV=Y, these values are separated by commas. When CSV=T, the commas are replaced by tab characters.
Example 1: You wish to write a file on disk called "MYDATA.SCF.txt" containing a score-to-measure table for the complete test.
SCOREFILE=MYDATA.SCF.txt
Example 2: You want a score-to-measure file for items with known difficulties.
ITEM1=1 ; start of response string
NAME1=1
NI=10 ; number of items
CODES=01 ; valid item codes
IAFILE=*
; known item difficulties here
1 0.53
.......
10 -0.34
*
SAFILE=*
; known structure "step" calibrations here, if rating scale or partial credit items
*
SCOREFILE=sm.txt ; the score-to-measure file - also see Table 20
&END
END LABELS
0101010101 ; two dummy data records
1010101010 ; give every item a non-extreme score
Example 3: Produce a cumulative percentage plot of raw scores:
1) Output the Scorefile= to Excel
2) Scatterplot with lines between points
3) x-axis: SCORES
4) y-axis: Cumulative Frequency %
For the inverse cumulative percentage plot (percentage point plot), switch the x-axis and the y-axis.
Example 4: more exact relationship between raw scores and Rasch measures.
1) "Output Files" menu, SCOREFILE= to Excel.
2) Excel scatterplot the measures against the scores
3) On the Excel scatterplot, a cubic trendline: display formula.
Example 5. Rasch measures and raw scores corresponding to Angoff Ratings.
For each item Angoff value this can give you the nearest person measure and raw score on all the items. The item conditional p-values are the right-most columns in the SCOREFILE=.
Judges' Use of Examinee Performance Data in an Angoff Standard-Setting Exercise for a Medical Licensing Examination: An Experimental Study. Brian E. Clauser, Janet Mee, Su G. Baldwin, Melissa J. Margolis and Gerard F. Dillon. Journal of Educational Measurement. Vol. 46, No. 4 (Winter 2009), pp. 390-407
A Latent Trait Method for Determining Intrajudge Inconsistency in the Angoff and Nedelsky Techniques of Standard Setting
Wim J. Van Der Linden. Journal of Educational Measurement. Vol. 19, No. 4 (Winter, 1982), pp. 295-308
Example. 6: You want PFILE=, Table 20 and SCOREFILE= to report exactly the same person measures (thetas) to 4 decimal places.
The PFILE= output uses the thetas estimated during the Winsteps main data analysis. Arithmetical precision and output are controlled by CONVERGE=, LCONV=, RCONV=, UDECIMALS= . For precision to 4 decimal places, we need LCONV= 0.00003 and maybe even smaller. Suggestion: set USCALE=10000 so that you can see what is happening beyond the 4th decimal place.
Winsteps Table 20 and SCOREFILE= compute the theta values based on the values of the item difficulties (deltas) and Andrich thresholds (taus, for polytomies) saved from the data analysis (IFILE=, SFILE=). The arithmetic precision of the theta computation is much higher LCONV*0.001 and RCONV*0.001 - this is at the request of users who build item banks with the IFILE item difficulties and wish to compare their item banking software to Winsteps. To perform a Winsteps main data analysis to this higher precision would require many iterations, during which the item difficulties would also change slightly.
You could try this:
Do the main analysis, output IFILE=if.txt and SFILE=sf.txt using your settings of CONVERGE=B, LCONV=lvalue, RCONV=rvalue, UDECIMALS=5
Reanalyze the data with anchored values IAFILE=if.txt and SAFILE=sf.txt with LCONV=lvalue/100, RCONV=rvalue/100, UDECIMALS=5, then produce Table 20 and SCOREFILE=
