Frequently Asked Questions

What are the new features in IRTPRO™ 4.1, 4.2 and 5.0?

Check out the New Features page for more details.

Can IRTPRO spreadsheet files (*ssig) be exported to SPSS, SAS, STATA, etc. formats?

This feature is available in IRTPRO™ 4 and is accomplished as follows:

  • Use the File, Open option to open an .ssig file
  • Use the File, Export option to export this file to a different format

From the drop-down list, select SPSS Data File(*.sav). The Save As dialog enables one to save the file in any folder. The default folder is the one where the .ssig file is located and the default name is that of the .ssig file.

Can IRTPRO be run in batch mode?

There are two ways to run jobs described by existing .irtpro command files; they use the commands “-Run” and “-RunList“:

-Run
To run the command lines in, for example, the file “Spelling.irtpro“, assuming it is in the C:\IRTPRO Examples folder as installed by default, proceed as follows:

“-Run” is used to open IRTPRO™ and run the analyses described in a single .irtpro command file. In the Command Prompt window, enter a line of the form (for Windows 7):

"c:\program files (x86)\irtpro21\irtpro" –Run "c:\irtpro examples\unidimensional\2PL\Spelling.irtpro"

or (for Windows XP):

"c:\program files\irtpro21\irtpro" –Run "c:\irtpro examples\unidimensional\2PL\Spelling.irtpro"

Notes:
The difference between Windows 7 and XP is the presence or absence of ” (x86)” in the folder name for the IRTPRO™ executable.

In commands typed in the Command Prompt window, or in .bat files executed there, double quotes are required around path/file names that contain spaces.
Several lines of the form(s) above can be created one after another in a file with the extension .bat; then that file can be double-clicked, or invoked in the Command Prompt window, and the lines will be executed one after another.

The .irtpro files can either be created by the IRTPRO™ GUI, or appropriately edited versions of such files, or (possibly) even created by some user-written special-purpose software.
It may be useful to select the -irt.txt, or other .txt file, output in the commands in the .irtpro file, to yield ASCII files that can be post-processed after a sequence of “batched” runs.

-RunList
“-RunList” is used to open IRTPRO™ and run the analyses described in a list of .irtpro command files; the paths/names for the .irtpro command files are specified as lines of a .txt file.

In the Command Prompt window, enter a line of the form (for Windows 7):

"c:\program files (x86)\irtpro21\irtpro" –RunList "c:\irtpro examples\unidimensional\2PL_examples.txt"

or (for Windows XP):

"c:\program files\irtpro21\irtpro" –RunList "c:\irtpro examples\unidimensional\2PL_examples.txt"

in which the file “c:\irtpro examples\unidimensional\2PL_examples.txt” is, for example, the following four lines:

c:\irtpro examples\unidimensional\2PL\AACL3_21Items.irtpro
c:\irtpro examples\unidimensional\2PL\lsat6.irtpro
c:\irtpro examples\unidimensional\2PL\SelfMon4items2PL.irtpro
c:\irtpro examples\unidimensional\2PL\Spelling.irtpro

Note that no quotes are used around the filenames in the .txt file, even though the paths contain spaces.

How can one import ASCII files to the .ssig format?

This is accomplished using the utility ascii2ssig.exe. The purpose of this utility is to convert text data to the IRTPRO™.ssig format. By typing in ascii2ssig from the DOS Command prompt

ascii2ssig <file> [<number>] [/delim[=]'<char>'] [/header]

where:

  • <file>                 Text data file name, e.g. c:\mydata.csv
  • <number>          Number of columns to convert. Default is all.
  • ‘<char>’             Delimiter used in text data. For .csv file, /delim = ‘,’ is the default.
  • /header             Variable name in the first line.

As an example, suppose that we want to import the file AACL_21.csv containing 21 items, in the Command Prompt window, enter the line:

ascii2ssig aacl_21.csv 21 /delim=',' /header

Suppose that there are ten similar files called aacl_21_1.csv, aacl_21_2.csv, aacl_21_3.csv, …, aacl_21_10.csv, then all ten can be converted to .ssig files using import10.bat

Ascii2ssig aacl_21_1.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_2.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_3.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_4.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_5.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_6.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_7.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_8.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_9.csv 21 /delim=’,’ /header
Ascii2ssig aacl_21_10.csv 21 /delim=’,’ /header

Does IRTPRO™ impose any priors on item parameters by default?

IRTPRO™ does not impose priors on any item parameters or group means/(co) variances by default. Univariate priors (e.g., beta prior on lower asymptotes) can be activated via the advanced options dialog box. Univariate priors are supported by all estimation algorithms.

When should priors be used in IRTPRO™?

Like Multilog, IRTPRO™ uses maximum marginal likelihood estimation, which involves no priors, when doing item parameter calibration, unless the user instructs otherwise. The user may select item parameter priors through the Priors tab in Advanced Options dialog box. Or, alternatively, write in the Priors syntax section. Three kinds of prior densities are supported: Beta, Log-normal, and normal. These univariate priors can be imposed on any item parameters. However, for the guessing parameter, the normal prior actually means logit-normal on the logit of guessing probability. The log-normal prior is mostly meant for the slope parameter in unidimensional models. The beta prior is supported since people are familiar with the Beta priors in Bilog and Parscale. Due to the general multidimensional and hierarchical model on top of which IRTPRO™ is built, we chose not to assume default priors as does Bilog or Parscale (which have more limited focuses to unidimensional analysis and can therefore make more assumptions about the user’s intent). Hence, by default, no priors are imposed on any parameter in IRTPRO™. This can cause problems if the 3PL model is used. Unless sample size is huge, a logit-normal or beta prior on the lower asymptote parameter should be imposed at the very least.

Can IRTPRO™ handle the testlet response theory model?

Wainer and colleagues’ testlet response theory model can be understood as either a second-order factor analysis model or as a restricted bifactor model. There is a paper on this topic: Rijmen, F (2010). Formal relations and an empirical comparison between the bi-factor, the testlet, and a second order multidimensional IRT model. Journal of Educational Measurement.

This model can be specified directly in IRTPRO™ as a restricted bifactor model with equality restrictions imposed on the within-item slopes (general and group-specific) while at the same time the variances of the group-specific dimensions are set free. If Bock-Atikin EM is used and dimension reduction box is checked (in advanced options), IRTPRO™ automatically reduces the dimension of integration to two-fold. Furthermore, multiple-group versions of the model are also supported in IRTPRO™.

Can IRTPRO™ do IRTLRDIF?

In principle, the answer is yes, provided that the correct compact and augmented models are being compared but IRTPRO™ does not offer the automated all-other anchor sweep testing facilities offered by the program IRTLRDIF. IRTPRO™ uses an updated version of the Wald DIF test, however, which is more generalizable to the situation of more than 2 groups and there is a similar anchor-all test-all approach for DIF sweeps.

Can maximum likelihood scale scores be computed?

IRTPRO™ does not compute ML scale scores. It computes EAP and MAP scores. However, one can approximate ML scoring in MAP by manually editing the IRTPRO™ scoring syntax file such that the groups have a very large variance (implying a vague prior on theta), e.g., 100.

Can IRTPRO™ score persons in a dataset correctly if some items have one or more categories that none of the individuals responded to?

When using IRTPRO™ to score individuals in a dataset using a set of item parameters from an external file, one of the items might not have any individuals in the dataset that responded to an item in, for example, the first category. IRTPRO™ will not score correctly in this case. To solve this, one could add a fake person to the dataset with a response of 1 for that variable so that it would recode people correctly, and then drop the fake person after scoring the dataset. This is an annoyance, but there is at present (and for the foreseeable future) no easier solution. When this happens to us, or when we fear it might happen, we just add as many “fake persons” to the dataset as there are response categories. For five-category items that’s five: If the codes are 1-5, then there’s one that’s all 1s, the next is all 2s, etc. up to all 5s.

What is the format of the -prm.txt parameter output / input file?

The –prm.txt output file is a tab-delimited ASCII .txt file that contains one line for each item and/or group. It is an output file from item calibration runs, and may be used as a parameter input file for stand-along scoring runs.
The tab-delimited format makes it easy for users to edit, or store, by opening it with, say, Excel.
The one-line-per-item format makes it easy to “mix and match” lines from multiple runs of the program to “assemble” new tests—new combinations of items one may want to score.
Each line contains a variable number of fields, depending on the number of dimensions and the item response model:

  1. For all items / groups, the first field is the label.
  2. For all items / groups, the second field is the number of dimensions (“nfac”; an int).
  3. For all items / groups, the third field is itemtype: 1 = 3PL, 2 = Graded, 3 = Nominal; 0 means population distribution parameters [mean and (co)variances].

For itemtype = 1, 2, or 3 (these are really items):

4. The fourth field is the number of categories (ncat; an int).
An additional column is added here, if itemtype = 3 (nominal). This is a single integer signaling whether the item is actually to be treated as a GPC item. Recall that GPC is mathematically a special case of the nominal model. If this number is 1, the item is GPC. If zero, the item is not GPC, hence it is nominal (see Appendix below).
5) For all models, the fifth (set of) field(s) contains nfac real values, the slope(s); these are the parameters labeled “a” in the “Item Parameter Constraints” window.

For itemtype = 1 (3PL):
6) The sixth field contains a real value, the intercept parameter; these are the parameters labeled “c” in the “Item Parameter Constraints” window.
7) The seventh field contains a real value, the logit of the guessing parameter; these are the parameters labeled “g” in the “Item Parameter Constraints” window.
8) if (nfac == 1) The eighth field contains a real value, the threshold parameter.*
8/9) if (nfac > 1) The eighth field contains a real value, the guessing parameter; this value is the ninth field if (nfac == 1).*
*Note: fields marked with * are for the user’s convenience, and will not be used when this file is read as either stating values or as parameter values for scoring.

For itemtype = 2 (Graded):
6) The sixth (set of) field(s) contains (ncat – 1) real values, the intercepts(s) these are the parameters labeled “c” in the “Item Parameter Constraints” window.
7) if (nfac == 1) The seventh (set of) field(s) contains (ncat – 1) real values, the thresholds(s).*
*Note: fields marked with * are for the user’s convenience, and will not be used when this file is read as either stating values or as parameter values for scoring.

For itemtype = 3 (Nominal):
6) The sixth field contains an integer, T-matrix type, 0 = Trend, 1 = Identity, 2 = User-supplied.
7) The seventh (set of) field(s) contains (ncat – 1) real values, the contrast(s) for the scoring functions. The first of these values is the “1.0” that is always the value of α1; the rest are subsequent values of the αs. [These are the parameters labeled “α” in the “Item Parameter Constraints” window.]
8) The eighth field contains an integer, L-matrix type, 0 = Trend, 1 = Identity, 2 = User-supplied.
9) The ninth (set of) field(s) contains (ncat – 1) real values, the contrast(s) labeled (γs) in the “Item Parameter Constraints” window, for the intercepts.
10-11) if (nfac == 1) The tenth set of fields contain the ncat values of the “ak” parameters, and the eleventh set of fields contain by the ncat values of the “ck” parameters of the old (Bock, 1972) parameterization of the nominal model.*
12) if (nfac == 1) and the model is actually GPC, The twelfth set of fields contains the “b” parameter from (one of) Muraki’s paramterization(s), and then the next ncat fields contain the corresponding “d” parameters—the first “d” is always 0.0.*
*Note: fields marked with * are for the user’s convenience, and will not be used when this file is read as either stating values or as parameter values for scoring.

13) Note: At a later time, the 13th set of fields will contain the elements of any user-supplied T matrix, followed by the elements of any user-supplied L matrix; that is not yet implemented.
For itemtype = 0 (these are really group / population parameter values):
4) The fourth (set of) field(s) contains nfac real values, the mean(s) (μ).
5) The fifth (set of) field(s) contains nfac*(nfac+1)/2 real values, the covariance(s), in the order σ1 1, σ2 1, σ2 2, σ3 1, σ3 2, σ3 3, σ4 1, …
6) if (nfac == 1) The sixth field contains a real value, the standard deviation.*
*Note: fields marked with * are for the user’s convenience, and will not be used when this file is read as either stating values or as parameter values for scoring.

Appendix: The GPC Model
When the model does not contain free scoring function contrasts, the nominal model (itemtype = 3) is actually a GPC. The way it is done is best illustrated with an example. In the zip archive, one can find the html output for a 3-item nominal+GPC run. The first two items are nominal, with the 3rd being GPC. Basically, if the itemtype is 3 (nominal), the procedure reading the PRM file should expect an additional integer. If it is 0, the model is not GPC, and hence should be treated as nominal; if 1, the model is GPC.
Here is an excerpt from the -PRM.txt (in the examples) output (the first couple of columns) for items 2 and 3:
Nominal4 1 3 4 0 1.01469 0 1.00000 0.85037
0.80860 0 0.70135 0.46954 0.38907
GPC4 1 3 4 1 1.13518 0 1.00000 0.00000
0.00000 0 0.92548 0.01721 -0.05477
As one can see, the order of PRM listing is the label (indicating one is nominal and the other is GPC), the number of factors (1-factor in this case), item type (both equal to 3), the number of categories (both ncat=4), indicator for GPC (item 2 is not GPC and item 3 is GPC), the slope estimate (there is one slope because this is a 1-factor model), the type of T matrix (scoring function contrast matrix — this case equal to 0 = Trend), (ncat-1) scoring function contrast estimates, the type of L matrix (the intercept contrast matrix — again 0 = Trend). One thing to note is that for a GPC model, the T matrix type is always 0, and the first scoring function contrast is always 1.0, and the remaining scoring function values are all 0. On the other hand, a true nominal model does not have such restrictions.

In using the MH-RM Algorithm, when should one check the AR value?

Check the AR value in the iteration output whenever the MH-RM algorithm is selected. Ideally, we’d like to keep the AR to something around .25 for models that are high dimensional (i.e., more than 3 or 4 factors). AR values larger than .25, say, .4 or .5 may be acceptable if the dimensionality is not too high. If the AR value is too high, the proposal dispersion constant can be increased. If it is too low, the dispersion constant should be decreased. If one starts at 1.0 and observe the AR is larger than the target value, do step-halving. One would stop the iterations, cutting the constant to .5 and restart the iterations, watch how well the AR does with the new dispersion. If it looks good, stop, if not, either cut or increase the constant in steps of .1 or .2.

AR is ultimately going to oscillate around a target value due to randomness. An AR of .28 is probably as good as an AR of .25. The MH chain is going to be just about as efficient.

See also Roberts, G.O., & Rosenthal, J.S. (2001). Optimal scaling for various Metropolis-Hastings algorithms. Statistical Science, 16, 351–367.

When is the M2 Fit statistic printed?

IRTPRO™ will only print M2 and related statistics when BAEM is used. This is because we have not been able (yet) to implement a good way to compute the posterior probabilities and Jacobians and weights needed for M2 with adaptive quadrature points. So for now, BAEM has all the fit statistics but ADQ and MH-RM has few.

Can you provide more details regarding the Logit(- g) or the g parameter used in 3-PL models?

The g parameter is the so-called pseudo-guessing (lower asymptote parameter, c parameter, etc.) in the 3PL model. The logit of g is the parameter that is actually estimated in IRTPRO™ because the re-parameterization produces a more nearly quadratic log-likelihood that is easier for optimization. In other words, we do not estimate the item guessing probabilities, but rather their log-odds.
Note:
If logit(g) is “z”, then g = 1/(1+exp(-z)), which is the usual guessing parameter.

What does the “c parameter” represent in the IRTPRO™ output (if not the guessing parameter)?

The c parameter is the intercept, which is related to the threshold parameter “b”, where:
b = -c/a, and a is the slope.
So the c’s are really just -a times the corresponding b (threshold). The intercepts are preferred because they generalize well to multidimensional IRT models.
The critical thing is that we are using the intercept/slope parameterization as opposed to the more typical (but only useful for unidimensional IRT) threshold/slope parameterization. In multidimensional IRT, because there are more than one latent variable in the linear predictor, the closest thing to a threshold would be to transform the c values into thresholds as per Eq (4.9) of M.D. Reckase (2009): Multidimensional Item Response Theory, Statistics for Social and Behavioral Sciences, Springer Science + Business Media. Wirth & Edwards’s (2007) paper in Psych Methods also contains a number of conversion formulae that would work.

Can one specify the scaling factor 1.7 in IRTPRO™?

IRTPRO™ does not include the scaling factor D = 1.7 in any equation for any purpose. All trace line equations are written without the 1.7.

Is there any technical information available regarding the GPC model other than the pdf guide?

There are many parameterizations of the GPC model. Some of them are on p. 55 of Thissen, D., Cai, L., & Bock, R.D. (2010). The nominal item response model. In M. Nering & R. Ostini (Eds.), Handbook of polytomous item response theory models: Developments and applications.

The model used in IRTPRO™ is on p. 59.

Conversion to (some of the) GPC parameters is on pp. 60-62.

However, because the GPC model has been expressed in many ways, and implemented in many ways in software, the only way to truly figure out translation between IRTPRO™ and any of the other programs is by cross-checking results.

How can one save edited graphs in IRTPRO?

To illustrate the way that one can save an edited graph, open the file Efficacy.ssig from the IRTPRO Examples\multidimensional\CFA folder and select the Graphics, Univariate… option then select NOSAY and VOTING. Click the OK button to produce the bar charts.

To save the changes to this graph for future use, right-click on the NOSAY graph to activate the 2D Chart Control Properties dialog and click the Save button to invoke the Save As dialog. Enter a filename, for example, efficacy_nosay.oc2 and then click the Save button.

During a new session, when opening the file efficacy.ssig, the option Graphs, Univariate will produce the original, un-edited bar chart for NOSAY. To restore the edited version of this graph, right-click on the graph and click the Load button on the 2D Chart Control Properties dialog. Select the file efficacy_nosay.oc2. 

Click the Open button to return to the 2D Chart Properties dialog window, and then click OK to display the edited version of the bar chart.

Is there an easy way to change the appearance of multiple graphs

The answer to this question is yes. Right-click on the first graph to activate the 2D Chart Control Properties dialog. Then select the ChartStyles, Style 1 option and choose the color to display the bar associated with the first category. Next select Style 2 and choose another color for the second category.

To replicate this (and any other changes), left-click on the first graph and select the Copy Style option from the menu that is displayed as a result of this action.

By left-clicking on each of the remaining graphs, the Save, Load, Copy and Paste style options are activated. Choose the Paste Style option to produce the same appearance for each item.

Saving and loading Styles

Once the appearance of a graph has been changed, one could save these changes in a file with extension .csf. This is accomplished by left-clicking on the graph in question and selecting the Save Style option. By doing so, a Save As dialog is displayed. The styles are saved in a file named AACL_21_Style.csf. 

When the same .ssig file is opened and graphs are displayed, one can select the Load Style option (by left-clicking on a graph) and browse for the saved style file. Once this action is completed, use the Copy and Paste options described above to change any one of the remaining graphs.

Can IRTPRO™ conduct an IRT analysis on repeated measures data?

Yes, for some situations and corresponding models IRTPRO™ can analyze data on repeated measurements or longitudinal data. See example 2.2 in Cai, L. (2010-a). A two-tier full-information item factor analysis model with applications. Psychometrika, 75, 581-612.

Does IRTPRO™ have the option of producing the estimated observed scores based on the test characteristic curve?

When the test characteristic curve is plotted, it plots the “estimated observed score.” By clicking the Table button the underlying data being plotted is displayed.

Does IRTPRO™ have the option of superimposing the empirical item response curve to the estimated item response curve?

No. There are no “empirical item response curves” unless one uses point-estimates of the latent variable to create them. Historically, some have done that, but it’s not endorsed here.

IRTPRO™ _does_ provide the S-X2 statistics that are based on the difference between observed and expected frequencies in response categories by summed scores. If one selects the option to print the tables, one can see that information in tabular form, which is interpreted in much the same way as the nonexistent plots would be.

Can IRTPRO™ compute standardized discrimination parameters?

IRTPRO™ has an option to print factor loadings (in the options-miscellaneous box), which are standardized discrimination parameters.

How does one interpret the columns of the score file?

The columns of the score text output file are as follows (for K-factors):

Group Number (1, 2, 3…)
Case IDs (an internal id is always printed, and if the user selects an ID variable, it’s also printed)
K (number of dimensions) columns of score point estimates
K columns of SEs
K(K+1)/2 columns of the lower-triangular portion of the posterior covariance matrix (so only the unique elements are printed as a list)

The last part is omitted for unidimensional IRT (K==1).

When IRTPRO™ is installed on a new laptop computer, some of the dialogs do not display properly

An example, that illustrates this problem is as follows: when the Analysis, Models tab is displayed, it does not show the Constraints…or DIF… buttons.
To fix this, experiment with your display settings and specifically the scale (which should be made smaller). Proceed as follows: Right click on your computer monitor screen (not on an icon) and select the Display settings option. Select the smallest scale value available. If this does not solve the problem, try various sizes for your display resolution

To learn more about display settings one could Google: “Windows 10 How to change screen resolution and size”, or watch this video.

When attempting to run an IRTPRO™ example with all parameters fixed, the program crashes

A frequently asked question is how to use the estimated parameter estimates based on a previous dataset to fix all the parameters based on the same items and models but using a new dataset. To accomplish this, proceed as follows:

  • Read in the previous estimates (-prm.txt file) as starting values (options button) and then save the irtpro syntax file.
  • Open the syntax file with Notepad and replace the command Startvalues: with Constraints:

If all the parameters are fixed, one cannot run the calibration phase, since IRTPRO™ requires at least one free parameter to start the estimation procedure. With no free parameters IRTPRO™ will stop with an error message. If one needs to do calibration to obtain fit statistics and tables based on the new data, proceed as follows:

  • Click on the Models tab, Constraints… button and free one parameter. To ensure that this parameter estimate is close to the one obtained from the previous run, one could choose a parameter from the previous run with a non-significant estimate and could additionally constrain this parameter by using a strict prior (for example N(0,0.01) instead of N(0,3).

 

I have a problem that isn’t addressed here. What now?

We are always willing to help with any IRTPRO™ issues you are having. Send a description of the problem you are having along with your active IRTPRO™ license code, the version of IRTPRO™ you are using, the code you have written, and the dataset (if possible) to IRTPRO@VPGCentral.com. We will respond as soon as possible.