Simulated data file = ""

A file of simulated data can be constructed from the measures estimated (or anchored) for the main analysis. It will have one simulated observation for each observation in the original data file. Each simulation is unique, so that multiple different simulations can be obtained with the Output Files menu clicking on Simulated Data file.

 

Simulated data file = filename

The simulated data can be analyzed using Data=filename in the original specification file (or enter at Extra specifications? prompt). Comment out any Dvalues= specifications in the original specification file.

 

The simulated data file has the basic Facets data format:

 

; Simulated data matching the empirical data structure

; Ratings of Scientists (edited to illustrate ambiguity in measurement)

; matching: C:\FACETS\examples\subsets.txt

1,2,1,7 ; 9 ; 1,2,1 are facet elements. 7 is simulated. 9 is the original data value.

1,2,2,7 ; 7

1,2,3,4 ; 5

1,2,4,9 ; 8

1,2,5,3 ; 5

 

 

Example 1: We need to compute the S.E. for every element including sampling error. This is "parametric bootstrapping". Also

Simulate and analyze 1000 Facets data sets from lfs.txt in one folder.

Save the following as x.bat in c:\Facets\examples, and then double-click on x.bat.

If you are using Minifac, then change \Facets to \Minifac -

The S.D. of the estimates for each element is its total S.E.

 

SET /A COUNT=1

:LOOP

echo Loop number %COUNT%

rem do this 1000 times

IF %COUNT% == 1001 GOTO END

rem generate simulate data file from lfs.txt

START /WAIT ..\Facets BATCH=YES lfs.txt specfile.out.txt simul=s%COUNT%.txt

rem analyze simulated data using in original specification file

START /WAIT ..\Facets BATCH=YES lfs.txt s%COUNT%.out.txt data=s%COUNT%.txt

SET /A COUNT=COUNT+1

GOTO LOOP

:END

PAUSE

 

Example 2: Simulate data corresponding to various types of rater behavior.

1. We conceptualize the rater effects we want to investigate, for instance "halo effect".

2. We formulate statistical models corresponding to each of the rater effects, for instance "halo effect" = all observations by a rater of a person are the same as the first observation.

3. We propose the parameter values which would correspond to each of those rater effects.

4. We use the statistical models of 2. and the parameter values of 3. to generate the data.

 

If the models in 2. are Rasch models that can be simulated by Facets, then we can use the parameter values in 3. as anchor values in Facets analyses. Then generate data in 4. using the "simulate data file" option in Facets. In order to make the Facets program run, we give it some data, but it does not matter what the data are, because Facets will use the anchor values, not values estimated from the data, to generate the new data.

 

If the models in 2. are not models that can be simulated by Facets, then we can formulate the data directly that match what we intend, for instance 3 3 3 3 3 3 could be one rater-person data string for "halo effect", and 4 4 4 4 4 4 could be another data string. Or we can use general-purpose simulation software, such as Simfit, simfit.usal.es/english/default.htm

 

Example 3: Discover the estimation bias in a set of Facets estimates.

Use the batch file in Example 1, to simulate data matching your dataset. Then compare the standard deviations of the estimates. Here are the numbers from an dataset with 3 facets and 5 simulations:

 

 

Example 4: The simulated data file is to have a different data pattern than the original data file, but be based on the same element measures.

1. Analyze your original data.

2. Output an anchor file (Anchorfile=) with no data but all elements anchored.

3. Construct a dummy data file of all the same response values, such as "1", with the data pattern you want. You can use Excel to do this.

4. Analyze the anchorfile as your specification file and the dummy data file as your data file.

5. Output the simulated data file. This will now have the data pattern that you want, and match the measures of the original dataset.

 

Example 5: Simulate a data file with more persons: One approach:

1. From your current Facets analysis, output an Anchorfile=

2. In the anchor file, add to the person facet in Labels=, the new person elements and their anchor values. These can be generated using Excel and your desired mean and S.D. of the additional person measures.

3. Add to the data new dummy observations, such as "1", with the new persons, the raters, the items. You can use Excel to do this

4. Analyze the anchorfile with Facets, using the new data file. Everything should be anchored.

5. Ignore the output of the analysis.

6. Output file. Simulate a new simulated data file.

7. Unanchor everything in the anchorfile. Save it as your new specification file.

8. Analyze the unanchored specification file with the simulated data file.

 

Example 6: Simulate a data file with more persons: Another approach:

You want to go from 25 candidates to 100 candidates.

Say your original candidate numbers are 1-25

Copy that dataset and change the candidate numbers to 101-125

Copy the dataset again and change the candidate numbers to 201-225

Copy the dataset again and change the candidate numbers to 301-325

Put all four datasets together.

Change your Facets Labels= from 1-25 to 1-325 (missing element numbers do not matter)

Analyze the combined dataset of 4x25 = 100 candidates, and simulate data from it.

 

Example 7: Estimate the person "test" reliability for an incomplete design:

1. Analyze the original data with Facets. A generic reliability will be reported for the persons, but without considering the specifics of the design

2. Simulate multiple datasets to match the original design - see Example 1. No anchoring

3. Analyze each of the dataset as though it is the original data.

4. Compute the mean and variance of the estimates for each element.

5. Better reliability estimate = (observed variance of the means - average of the element variances) / observed variance of the means