Joint maximum likelihood estimation of item parameters and examinee attribute profiles

Description

This function returns joint maximum likelihood estimates of item parameters and examinee attribute profiles in cognitive diagnostic models. The algorithm starts from the nonparametric estimation of attribute profiles, implemented by the AlphaNP function, and then iteratively estimates item parameters and attribute profiles using conditional maximum likelihood estimation until the algorithm converges. Currently supported models include the DINA model, the DINO model, he NIDA model, the G-NIDA model, and the R-RUM model.

Usage

1
2
3
JMLE(Y, Q, model = c("DINA", "DINO", "NIDA", "GNIDA", "RRUM"), 
NP.method = c("Weighted", "Hamming", "Penalized"), wg=1, ws=1, 
conv.crit.par = 0.001, conv.crit.att = 0.01, max.ite = 100)

Arguments

Y

A matrix of binary responses. Rows represent persons and columns represent items. 1=correct, 0=incorrect.

Q

The Q-matrix of the test. Rows represent items and columns represent attributes. 1=attribute required by the item, 0=attribute not required by the item.

model

Currently support five models: "DINA", "DINO", "NIDA", "GNIDA", and "RRUM". The default is "DINA".

NP.method

The method of the nonparametric estimation in the initial stage. "Hamming": the plain Hamming distance method; "Weighted": the Hamming distance weighted by inversed item variance; "Penalized": the Hamming distance weighted by inversed item variance and specified penalizing weights for guess and slip. The default is "Weighted".

wg

Additional argument for the "penalized" NP.method. wg is the weight assigned to guessing in the DINA or DINO models. A large value of wg results in a stronger impact on Hamming distance (larger loss function values) caused by guessing.

ws

Additional input for the "penalized" NP.method. ws is the weight assigned to slipping in the DINA or DINO models. A large value of ws results in la stronger impact on Hamming distance (larger loss function values) caused by slipping.

conv.crit.par

The critical value for the maximum absolute change in all item parameters values to determine convergence.

conv.crit.att

The critical value for the percentage of examinee attribute profiles that are changed to determine convergence.

max.ite

The maximum number of iterations allowed.

Value

alpha.est

JMLE estimates of examinee attribute profiles. Rows represent persons and columns represent attributes. 1=examinee masters the attribute, 0=examinee does not master the attribute.

par.est

JMLE estimates of item parameters, including par.est$slip, par.est$guess, par.est$se.slip, and par.est$se.guess for the DINA, DINO, NIDA, and GNIDA models, and par.est$pi, par.est$r, par.est$se.pi and par.est$se.r for the R-RUM model. Note that for the G-NIDA model and the R-RUM model, the item parameter estimates and standard errors are not available for the entries where the Q-matrix is 0.

n.tie

Number of ties in the final log-likelihood among the candidate attribute profiles for each person. When we encounter ties, one of the tied attribute profiles is randomly chosen.

undefined.flag

A binary vector indicating whether the parameters of each item are undefined. 1=undefined, 0=defined.

loglike

The final overall log-likelihood value from the estimated item parameters and attribute profiles based on the specified model.

convergence

A message on whether the algorithm converged.

n.ite

Number of iterations performed.

loglike.matrix

The values for the log-likelihood function in the last iteration for each candidate attribute profile by each person. Rows represent candidate attribute profiles in the same order with the pattern matrix; columns represent different examinees.

est.class

The final class number (row index in pattern) for each person's attribute profile. It can also be used for locating the log-likelihood value in loglike.matrix for the estimated attribute profile for each person.

NP.loss.matrix

The values for the loss function of the nonparametric estimation of Alpha. Rows represent candidate attribute profiles in the same order with the pattern matrix; columns represent different examinees.

NP.alpha.est

The estimates of examinee attribute profiles from the initial nonparameteric estimation.

NP.est.class

The class number (row index in pattern) for each person's attribute profile from the initial nonparametric classification. It can also be used for locating the loss function value in NP.loss.matrix for the estimated attribute profile for each person.

pattern

All possible attribute profiles in the search space.

model

The chosen model.

Q

The Q-matrix of the test.

References

Chiu, C. (2011). Flexible approaches to cognitive diagnosis: nonparametric methods and small sample techniques. Invited session of cognitive diagnosis and item response theory at 2011 Joint Statistical Meeting.

Chiu, C. Y., & Douglas, J. A. (2013). A nonparametric approach to cognitive diagnosis by proximity to ideal response patterns. Journal of Classification 30(2), 225-250.

See Also

AlphaMLE, AlphaNP, ParMLE, print.JMLE, plot.JMLE

Examples

1
2
3
4
5
6
7
8
9
data("Data.DINA")
JMLE.result <- JMLE(Data.DINA$response, Data.DINA$Q, model="DINA", conv.crit.par=0.001, 
conv.crit.att=0.001, max.ite=100)
print(JMLE.result)  # Print the estimated item parameters, standard errors, 
#and examinee attribute profiles
plot(JMLE.result, nperson=1)   # Plot the sorted loss function of different 
#attribute profiles for this examinee
ItemFit(JMLE.result)
ModelFit(JMLE.result)