# random.polychor.pa: A Parallel Analysis with Randomly Generated Polychoric... In random.polychor.pa: A Parallel Analysis with Polychoric Correlation Matrices

## Description

The function performs a parallel analysis using simulated polychoric correlation matrices. The eigenvalues (extracted following both FA and PCA methods) from each random generated polychoric correlation matrix and from the polychoric correlation matrix of real solutions from Polychorich vs Pearson correlations, FA vs PCA and PA vs MAP are presented. Random data sets are simulated assuming or a uniform or a multinomial distribution or via the bootstrap method of resampling (i.e., random permutations of cases). Also Multigroup Parallel analysis is made available for random (uniform and multinomial distribution and with or without difficulty factor) and bootstrap methods. An option to choose between default or full output is also available as well as a parameter to print Fit Statistics (Chi-squared, TLI, RMSEA, RMR and BIC) for the factor solutions indicated by the Parallel Analysis.

## Usage

 1 2 3 4 random.polychor.pa(nvar="NULL", n.ss="NULL", nrep, nstep="NULL", data.matrix, q.eigen, r.seed = "NULL", diff.fact=FALSE, distr="NULL", comparison = "random", fit.pa=FALSE, print.all=FALSE, continuity=0.0, wght="NULL")

## Arguments

 nvar Number of variables (items) in the raw data matrix. From version 1.1 of the function, it is no more needed to specificy nvar as this information is derived from the number of columns of the data.matrix. Default value is set to "NULL" for compatibility with past version of the function n.ss Number of participants of the raw data matrix. From version 1.1 of the function, it is no more needed to specify n.ss as this information is derived from the number of rows of the data.matrix. Default value is set to "NULL" for compatibility with past version of the function. nrep Number of random samples that should be simulated nstep Number of ordered categories of the item (e.g., Likert-like 3 ordered category item). This information is no longer needed as the new version of the function (1.1) allows also for items with varying number of categories. The number of categories from each item is derived directly from the data.matrix. A table summarizing the different groups of item with different number of categories will be showed. Default value is set to "NULL" for compatibility with past version of the function. data.matrix the name of raw data matrix. The raw data.matrix should be numeric and none of the ordered category should be coded as 0 (zero). No automatic recode routine is provided within the function to deal with alphanumeric content of the ordered categories of manifest variables. So the user performs all these recodings before running the function. q.eigen a number comprised within the interval of 0 and 1 and indicating the quantile that is used to choose the number of non-random factors (e.g., .50 or .95 or .99) r.seed eventually, a preferred number that will be used to initialize the random generator. Default value: 1335031435. diff.fact default value is FALSE and in this case the function will estimate random datasets without trying to reproduce each observed category with the same probability as that observed in the empirical dataset provided. If the paramether is set on TRUE, the function will simulate random samples with the same proportion of each category for each item as that of empirical dataset. This parameter allows to simulate data sets that have the same item difficulties distribution as well as the same difficulty factors of the real data (Ledesma and Valero-Mora, 2007) distr this parameter allows to choose between uniform (that is also the default) or multinomial distribution for the simulation of random data sets. comparison four methods are now available: random, bootstrap, multigroup-random, multigroup-bootstrap. Allowed values are: random for random simulated dataset; bootstrap for boostrap simulations with permutation as method of resampling; random-mg for random multi-group simulated datasets; bootstrap-mg for bootstrap multigroup simulated datasets. When bootstrap method is choosen, the bootstraped eigenvalue distribution is compared with that obtained from empirical data and the PA algorithm will indicate the number of empirical eigenvalues that are greater than the corresponding bootstraped eigenvalues. fit.pa when set to TRUE, the fit statistics (Chi-squared, TLI, RMSEA, RMR, BIC) for all factor solutions indicated by Parallel Analysis will be printed. This parameter allows a call to the fa() function of the psych package, and passes just the number of factor to retain with all remaining values set to the default (i.e., fm="minres", rotate="oblimin", etc.) print.all when set to TRUE, all the simulated (random or bootstrap) and empirical eigenvalue distributions will be printed for each samples. When comparison is equal to bootstrap or bootstrap-mg also the bootstrap bias and standard error estimates will be printed. continuity by default is set to 0.0 meaning no correction for continuity will be applied. However if needed users may add the correction for continuity by setting the parameter to 0.5 to handle zero count cells. The parameter is passed to the polychoric function of the psych package. wght a vector of positive and non-zero weights of the same length of the dataset and with no missing values can be passed to the function in order to compute a weighted polychoric (or pearson) correlation matrix. The default is set to "NULL" meaning that weights are all equal to 1/nrow(data.matrix)

## Details

The function perform a parallel analysis (Horn, 1976) using randomly simulated polychoric correlations and generates nrep random samples of simulated data with the same number of participants and of variables of the provided data.matrix. The function will read the entered data.matrix and extracts: the number of units (i.e., number of rows); the number of variables (i.e., number of columns); and the number of categories of each item. From version 1.1, the function accepts also variables with varying number of categories (e.g., three items with only two categories and two items with three categories, etc.). In version 1.1.1, the function is also able to manage supplied data.matrix in which variables represent factors (i.e., variables with ordered categories) may cause an error when the Pearson correlation matrix is calculated. The information in the supplied data.matrix are used to generate the nrep random raw datasets with the same characteristics of the original real data set. So only three information are needed for the problem to run: the number of replications (nrep), the data matrix (data.matrix) and the percentile to be used (q.eigen).

A check for missing values within the real dataset is present and if present will be treated LISTWISE. In this case a warning message will prompt the user signalling how NA were treated (LISTWISE is by now the only treatment considered) and the new sample size. No further checks are made on the raw data, so out-of-range values are not detected and it is on the behalf of the user to make a preliminary check on the reliability of data. A table summarizing the groups of items with different number of categories will be shown along with the main results of the PA.

The function will extract the eigenvalues from each randomly generated polychoric matrices and the requested percentile is returned. Eigenvalues from polychoric correlation matrix obtained from real data is also computed and compared with the eigenvalues extracted from the simulation (Polychoric matrices) in a (scree) plot. Results from classical Factor Analysis and Principal Component Analysis are made available. Separated random data sets are simulated for the two analysis.

Recently, Cho, Li & Bandalos (2009) showed that, in using PA method, it is important to match the type of the correlation matrix used to recover the eigenvalues from real data with the type of correlation matrix used to estimate random eigenvalues. Crossing the type of correlations (using Polychoric correlation matrix to estimate real eigenvalues and random simulated Pearson correlation matrices) may result in a wrong decision (i.e., retaining more non-random factors than the needed). A comparison with eigenvalues extracted from both randomly simulated Pearson correlation matrices and real data is also included.

Finally, for both type of correlation matrix (Polychoric vs Pearson), the two versions (the classic squared coefficient and the 4th power coefficient) of Velicer's MAP criterion are calculated (Velicer, 1976; Velicer, Eaton, & Fava, 2000) by implementing under R the code released by O'Connor (2000) for SPSS, SAS and MATLAB.

As the poly.mat() function used to calculate the polychoric correlation matrix is going to be deprecated in favour of polychoric() function, the random.polychor.pa was consequently updated (version 1.1.2) to account for changes in psych() package.

Version 1.1.3 tackles two problems signalled by users: 1) the possibility to make available the results of simulation for plotting them in other softwares. Now the random.polychor.pa will show, upon request, all the data used in the scree-plot. 2) The function polichoric() of the psych() package does not handle data matrices that include 0 as possible category and will cause the function to stop with error. So a check for the detection of the 0 code within the provided data.matrix is now added and will cause the random.polychor.pa function to stop with a warning message.

In version 1.1.3.5 a paramether was added, diff.fact, in order to simulate random dataset with the same probability of observing each category for each variable as that observed in the provided (empirical) dataset. This paramether was added for those reaserchers that want to replicate random datasets with the same distribution of item difficulties as the real data (Ledesma, Valero-Mora, 2007; Reckase, 2009, pp.216). Finally the search for zeroes within the provided datafile was removed, so data with zeroes are now accepted.

In version 1.1.3.6 a check for the range of quantile (beteen 0 and 1) was added.

In version 1.1.4 was added the possibility to choose between uniform and multinomial distribution, between random and boostrap distribution and between single or multiple sample parallel analysis. Also two switches were added: one allows to print or a default output (including only the number of factors to retain following Parallel Analysis and the scree-plot) or a full output (including the eigenvalues distributions of simulated and empirical data sets); the other switch allows to print a further Fit Statistics for the factor solutions indicated by the Parallel Analysis.

The default value of the comparison is set to random and this means that random uniform samples (runif()) will be computed. Two types of distribution are available: the uniform distribution and the multinomial distribution (Liu and Rijmen, 2008). To switch between them just set accordingly the parameter distr. As method of resampling is now also available the bootstrap method that is based on random permutations of cases obtained via the boot() function.

Multigroup version of random uniform samples may be obtained by setting the parameter comparison to random-mg. To perform multigroup Parallel Analysis simulating random samples with multinomial distributions the parameter distr should be set to multinomial. Multigroup Parallel Analysis is also available for the bootstrap method by setting the comparison parameter to bootstrap-mg. When Multigroup Parallel Analysis is used, the first variable (column) of the dataset should be reserved to the factor used to identify the different samples (i.e., gender, agre groups, nationalities, etc.).

Fit statistics (Chi-squared, TLI, RMSEA, RMR, BIC) for all factor solutions indicated by Parallel Analysis are available when the parameter fit.pa is set to TRUE. Consider that in estimating the these Fit Statistic indexes not allways converge and this may break the computations for the parallel analysis. In this case switch the parameter off. Moreover the values of these indexes are not allways within the expected range.

The print.all parameter when set to TRUE will show the tables of eigenvalue distributions for simulated (random or boostrap) and empirical data. For boostrap simulation, also the bias and standard error will be printed. If the diff.fact is set to TRUE when random simulations are requested, then also the weighting matrix used to reproduce the item frequencies in random data set will be printed.

The continuity parameter is passed to polychoric function to handle the correction for continuity. In polychoric function this correction for continuity is set by default to .5 (i.e. correct=TRUE). However in some cases this correction for continuity causes the polychorich function to stop unexpectedly and consequently also random.polychor.pa function stops unexpectedly. So we added this parameter to allow users to bypass this problem. This parameter is set by default to 0.0 (i.e., no correction for continuity applied) and user may add the correction for continuity by setting the value to 0.5.

The wght parameter allows to compute a weighted correlation matrix. If a vector of weights is passed to the function, then a weighted polychoric (pearson) correlation matrix will be computed. Also simulated data will be weighted as well as bootstrap samples if select. In case of multisample option, user have to provide a vectors of weights opportunely set accordingly to grouping factor. Weights will affects all indices computed from Parallel Analysis including MAP

## Value

The default output (print.all=FALSE) prints the number of factors for Polychoric and Pearson Correlation PA methods for Factor Analysis and Principal Components Analysis (PCA) methods along with the number of factors chosen by the two Velicer's MAP criteria (original and 4th power) for both Polychoric and Pearson correlation matrices. Furthermore, the function will return the (scree) plot of the eigenvalues for real (Polychoric vs Pearson correlation matrices) and simulated data (Polychoric vs Pearson correlation matrices). If the parameter print.all is set to TRUE then other than the defaul output, it will also be printed the tables of eigenvalue distributions for simulated (random (uniform or multinomial) or boostrap) and empirical data. For boostrap simulation, also the bias and standard error will be printed. If the diff.fact is set to TRUE when random simulations are requested, then also the weighting matrix used to reproduce the item frequencies in random data set will be printed.

## Note

In running the random.polychor.pa function it should be reminded that it may take a lot of time to complete the simulation. This is due in part to the fact that the estimation of the polychoric correlation matrix is cumbersome and in part to the fact that the code is not optimized, but simply it does the work.

Occasionally, in calculating the polychoric correlation matrix it may occur an error when the matrix is non-positive definite. In this case you have to re-run the simulation.

A note should be made concerning the method used (from version 1.1) to read the raw data.matrix supplied by the user and used to retrieve the three basic information needed to build the random matrices (number of rows, number of columns and the number of categories for each manifest variable). The number of categories for each variable is derived from the raw data.matrix, so if the possible number of categories for a specific item is for example 5, but subjects endorse only three out of the five categories then the random.polychor.pa function will simulate a variable with only three categories. This means that the function guarantees that the empirical and the simulated data matrix are similar, but this also means that by changing the sample of participants the simulated data will change (even if slightly).

In computing the polychoric correlation matrix for both simulated and empirical data sets, the parameter global of the polychoric() function is set to FALSE and the text message "The items do not have an equal number of response alternatives, global set to FALSE" is suppressed with the suppressMessages() function.

## Author(s)

Fabio Presaghi fabio.presaghi@uniroma1.it and Marta Desimoni marta.desimoni@uniroma1.it

## References

Cho, S.J., Li, F., & Bandalos, D., (2009). Accuracy of the Parallel Analysis Procedure With Polychoric Correlations. Educational and Psychological Measurement, 69, 748-759.

Horn, J. L. (1965). A rationale and test for the number of factors in factor analysis. Psychometrika, 32, 179-185.

Ledesma RD, Valero-Mora P (2007) Determining the number of factors to retain in EFA: an easy to use computer program for carrying out parallel analysis. Practical Assessment, Research & Evaluation 12:1–11

Liu, O.L., and Rijmen, F., (2008) A modified procedure for parallel analysis of ordered categorical data. Behavior Research Methods, 40 (2), 556-562. doi: 10.3758/BRM.40.2.556

O'Connor, B. P. (2000). SPSS and SAS programs for determining the number of components using parallel analysis and Velicer's MAP test. Behavior Research Methods, Instrumentation, and Computers, 32, 396-402.

Reckase, M.D. (2009). Multidimensional Item Response Theory. Springer.

Velicer, W. F. (1976). Determining the number of factors from the matrix of partial correlations. Psychometrika, 41, 321-327.

Velicer, W. F., Eaton, C. A., & Fava, J. L. (2000). Construct explication through factor or component analysis: A review and evaluation of alternative procedures for determining the number of factors or components. In R. D. Goffin & E. Helmes (Eds.), Problems and solutions in human assessment: Honoring Douglas N. Jackson at seventy (pp. 41-72). Norwell, MA: Kluwer Academic.