Logistik {difR}

Logistic regression DIF statistic

Description

Calculates the "logistic regression" likelihood-ratio statistics and effect sizes for DIF detection.

Usage

Logistik(data, member, member.type = "group", match = "score",
	 anchor = 1:ncol(data), type = "both", criterion = "LRT", all.cov = FALSE)
 

Arguments

data

numeric: the data matrix (one row per subject, one column per item).

member

numeric or factor: the vector of group membership. Can either take two distinct values (zero for the reference group and one for the focal group) or be a continuous vector. See Details.

member.type

character: either "group" (default) to specify that group membership is made of two groups, or "cont" to indicate that group membership is based on a continuous criterion. See Details.

match

specifies the type of matching criterion. Can be either "score" (default) to compute the test score, or any continuous or discrete variable with the same length as the number of rows of data. See Details.

anchor

a vector of integer values specifying which items (all by default) are currently considered as anchor (DIF free) items. Ignored if match is not "score". See Details.

type

a character string specifying which DIF effects must be tested. Possible values are "both" (default), "udif" and "nudif". See Details.

criterion

a character string specifying which DIF statistic is computed. Possible values are "LRT" (default) or "Wald". See Details.

all.cov

logical: should all covariance matrices of model parameter estimates be returned (as lists) for both nested models and all items? (default is FALSE.

Details

This command computes the logistic regression statistic (Swaminathan and Rogers, 1990) in the specific framework of differential item functioning. It forms the basic command of difLogistic and is specifically designed for this call.

If the member.type argument is set to "group", the member argument must be a vector with two distinct (numeric or factor) values, say 0 and 1 (for the reference and focal groups respectively). Those values are internally transformed onto factors to denote group membership. The three possible models to be fitted are then:

M_0: logit (\pi_g) = \alpha + \beta X + \gamma_g + \delta_g X

M_1: logit (\pi_g) = \alpha + \beta X + \gamma_g

M_2: logit (\pi_g) = \alpha + \beta X

where \pi_g is the probability of answering correctly the item in group g and X is the matching variable. Parameters \alpha and \beta are the intercept and the slope of the logistic curves (common to all groups), while \gamma_g and \delta_g are group-specific parameters. For identification reasons the parameters \gamma_0 and \delta_0 for reference group (g=0) are set to zero. The parameter \gamma_1 of the focal group (g=1) represents the uniform DIF effect, and the parameter \delta_1 is used to model nonuniform DIF effect. The models are fitted with the glm function.

If member.type is set to "cont", then "group membership" is replaced by a continuous or discrete variable, given by the member argument, and the models above are written as

M_0: logit (\pi_g) = \alpha + \beta X + \gamma Y+ \delta X Y

M_1: logit (\pi_g) = \alpha + \beta X + \gamma Y

M_2: logit (\pi_g) = \alpha + \beta X

where Y is the group variable. Parameters \gamma and \delta act now as the \gamma_1 and \delta_1 DIF parameters.

The matching criterion can be either the test score or any other continuous or discrete variable to be passed in the Logistik function. This is specified by the match argument. By default, it takes the value "score" and the test score (i.e. raw score) is computed. The second option is to assign to match a vector of continuous or discrete numeric values, which acts as the matching criterion. Note that for consistency this vector should not belong to the data matrix.

Two types of DIF statistics can be computed: the likelihood ratio test statistics, obtained by comparing the fit of two nested models, and the Wald statistics, obtained with an appropriate contrast matrix for testing the model parameters (Johnson and Wichern, 1998). These are specified by the argument criterion, with respective values "LRT" and "Wald". By default, the LRT statistics are computed.

If criterion is "LRT", the argument type determines the models to be compared by means of the LRT statistics. The three possible values of type are: type="both" (default) which tests the hypothesis H_0: \gamma_1 = \delta_1=0 (or H_0: \gamma = \delta=0) by comparing models M_0 and M_2; type="nudif" which tests the hypothesis H_0: \delta_1 = 0 (or H_0: \delta = 0) by comparing models M_0 and M_1; and type="udif" which tests the hypothesis H_0: \gamma_1 = 0 (or H_0: \gamma = 0) by comparing models M_1 and M_2 (assuming that \delta_1 = 0 or \delta = 0). In other words, type="both" tests for DIF (without distinction between uniform and nonuniform effects), while type="udif" and type="nudif" test for uniform and nonuniform DIF, respectively.

If criterion is "Wald", the argument type determines the logistic model to be considered and the appropriate contrast matrix. If type=="both", the considered model is model M_0 and the contrast matrix has two rows, (0,0,1,0) and (0,0,0,1). If type=="nudif", the considered model is also model M_0 but the contrast matrix has only one row, (0,0,0,1). Eventually, if type=="udif", the considered model is model M_1 and the contrast matrix has one row, (0,0,1).

The data are passed through the data argument, with one row per subject and one column per item. Missing values are allowed but must be coded as NA values. They are discarded from the fitting of the logistic models (see glm for further details).

The vector of group membership, specified with member argument, must hold only zeros and ones, a value of zero corresponding to the reference group and a value of one to the focal group.

Option anchor sets the items which are considered as anchor items for computing the test scores and related logistic regression DIF statistics. Items other than the anchor items and the tested item are discarded. anchor must hold integer values specifying the column numbers of the corresponding anchor items. It is mainly designed to perform item purification. Note that this option is discarded when match is not "score".

The output contains: the selected DIF statistics (either the LRT or the Wald statistic) computed for each item, two matrices with the parameter estimates of both models (for each item) and two matrices of related standard error values. In addition, Nagelkerke's R^2 coefficients (Nagelkerke, 1991) are computed for each model and the output returns both, the vectors of R^2 coefficients for each model and the differences in these coefficients. Such differences are used as measures of effect size by the difLogistic command; see Gomez-Benito, Dolores Hidalgo and Padilla (2009), Jodoin and Gierl (2001) and Zumbo and Thomas (1997). The criterion and member.type arguments are also returned, as well as a character argument named match that specifies the type of matching criterion that was used.

Value

A list with nine components:

stat

the values of the logistic regression DIF statistics.

R2M0

the values of Nagelkerke's R^2 coefficients for the "full" model.

R2M1

the values of Nagelkerke's R^2 coefficients for the "simpler" model.

deltaR2

the differences between Nagelkerke's R^2 coefficients of the tested models. See Details.

parM0

a matrix with one row per item and four columns, holding successively the fitted parameters \hat{\alpha}, \hat{\beta}, \hat{\gamma}_1 and \hat{\delta}_1 of the "full" model (M_0 if type="both" or type="nudif", M_1 if type="udif").

parM1

the same matrix as parM0 but with fitted parameters for the "simpler" model (M_1 if type="nudif", M_2 if type="both" or type="udif").

seM0

a matrix with the standard error values of the parameter estimates in matrix parM0.

seM1

a matrix with the standard error values of the parameter estimates in matrix parM1.

cov.M0

either NULL (if all.cov argument is FALSE) or a list of covariance matrices of parameter estimates of the "full" model (M_0) for each item (if all.cov argument is TRUE).

cov.M1

either NULL (if all.cov argument is FALSE) or a list of covariance matrices of parameter estimates of the "reduced" model (M_1) for each item (if all.cov argument is TRUE).

criterion

the value of the criterion argument.

member.type

the value of the member.type argument.

match

a character string, either "score" or "matching variable" depending on the match argument.

Author(s)

Sebastien Beland
Collectif pour le Developpement et les Applications en Mesure et Evaluation (Cdame)
Universite du Quebec a Montreal
[e-mail], http://www.cdame.uqam.ca/
David Magis
Department of Psychology, University of Liege
Research Group of Quantitative Psychology and Individual Differences, KU Leuven
[e-mail], http://ppw.kuleuven.be/okp/home/
Gilles Raiche
Collectif pour le Developpement et les Applications en Mesure et Evaluation (Cdame)
Universite du Quebec a Montreal
[e-mail], http://www.cdame.uqam.ca/

References

Gomez-Benito, J., Dolores Hidalgo, M. and Padilla, J.-L. (2009). Efficacy of effect size measures in logistic regression: an application for detecting DIF. Methodology, 5, 18-25. doi: 10.1027/1614-2241.5.1.18

Jodoin, M. G. and Gierl, M. J. (2001). Evaluating Type I error and power rates using an effect size measure with logistic regression procedure for DIF detection. Applied Measurement in Education, 14, 329-349. doi: 10.1207/S15324818AME1404_2

Johnson, R. A. and Wichern, D. W. (1998). Applied multivariate statistical analysis (fourth edition). Upper Saddle River, NJ: Prentice-Hall.

Magis, D., Beland, S., Tuerlinckx, F. and De Boeck, P. (2010). A general framework and an R package for the detection of dichotomous differential item functioning. Behavior Research Methods, 42, 847-862. doi: 10.3758/BRM.42.3.847

Nagelkerke, N. J. D. (1991). A note on a general definition of the coefficient of determination. Biometrika, 78, 691-692. doi: 10.1093/biomet/78.3.691

Swaminathan, H. and Rogers, H. J. (1990). Detecting differential item functioning using logistic regression procedures. Journal of Educational Measurement, 27, 361-370. doi: 10.1111/j.1745-3984.1990.tb00754.x

Zumbo, B. D. and Thomas, D. R. (1997). A measure of effect size for a model-based approach for studying DIF. Prince George, Canada: University of Northern British Columbia, Edgeworth Laboratory for Quantitative Behavioral Science.

See Also

difLogistic, dichoDif

Examples

## Not run: 

 # Loading of the verbal data
 data(verbal)

 # Testing both types of DIF simultaneously
 # With all items, test score as matching criterion
 Logistik(verbal[,1:24], verbal[,26])

 # Returning all covariance matrices of model parameters
 Logistik(verbal[,1:24], verbal[,26], all.cov = TRUE)

 # Testing both types of DIF simultaneously
 # With all items and Wald test
 Logistik(verbal[,1:24], verbal[,26], criterion = "Wald")

 # Removing item 6 from the set of anchor items
 Logistik(verbal[,1:24], verbal[,26], anchor = c(1:5, 7:24))

 # Testing for nonuniform DIF
 Logistik(verbal[,1:24], verbal[,26], type = "nudif")

 # Testing for uniform DIF
 Logistik(verbal[,1:24], verbal[,26], type = "udif")

 # Using the "anger" trait variable as matching criterion
 Logistik(verbal[,1:24],verbal[,26], match = verbal[,25])

 # Using the "anger" trait variable as group membership
 Logistik(verbal[,1:24],verbal[,25], member.type = "cont")
 
## End(Not run)