Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

22
A Mixture Rasch Model- Based Computerized Adaptive test for Latent Class identification Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Transcript of Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Page 1: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

A Mixture Rasch Model-Based Computerized Adaptive test for Latent Class identification

Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Page 2: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

The Mixture Rasch Model

)](exp[1

1

icjcjic bP

)](exp[1

1

icjccckic

ccji bPP

ci

xjic

xjicc

i

xjic

xjicc

jc PP

PPP

1

1

)1()(

)1()(

Page 3: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Starting Point (first item)◦ “Best guess”, “Use what you’ve got”, or “Start

easy”.◦ Selecting five items randomly from the calibration

item pool Item Selection Algorithm

◦ Fisher information◦ Kullback-Leibler (KL) information

Termination Rule◦ Fixed-length◦ Fixed-precision

The Implementation of a Mixture Rasch Model-Based CAT

Page 4: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

The latent trait measured within each latent class is unidimensional but the latent traits measured across latent classes are multidimensional.

Estimation of ability parameters◦ One single latent ability parameter◦ Class-specific ability parameters

KL information

)(

)(log)(

)(

)(log)()||(

]);(

);(log[)||(

1

22

1

2221

1

221 2

i

ii

i

iii

i

ii

Q

QQ

P

PPKL

xL

xLEKL

Page 5: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Estimation of a single latent ability parameter, to maximize the KL information between two latent classes at the current ability estimate.◦

◦ Maximizes the information to distinguish between the latent classes conditional on the current ability estimate.

◦ Appropriate for used when the same latent ability is measured across latent classes.

Method 1

)ˆ,(

)ˆ,(log)ˆ,(

)ˆ,(

)ˆ,(log)ˆ,()||(

1

22

1

2221

cQ

cQcQ

cP

cPcPccKL

i

ii

i

iii

Page 6: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Estimation of a single latent ability parameter, to maximize the distinction between latent classes as well as between the current ability estimate and its true value.◦

◦ Maximizes the information to distinguish between both latent classes and the upper and lower bounds of the interval set around the current ability estimate.

◦ Appropriate for used when the same latent ability is measured across latent classes.

Method 2

)ˆ,(

)ˆ,(log)ˆ,(

)ˆ,(

)ˆ,(log)ˆ,()ˆ,||ˆ,(

1

22

1

2221

Li

UiUi

Li

UiUiULi

cQ

cQcQ

cP

cPcPccKL

. where,ˆˆ and ˆˆn

cUL

Page 7: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Estimation of one latent ability for each latent class, to maximize the distinction between latent classes and between current ability estimates for each latent class.◦

◦ No interim latent class membership updating.

Method 3

)ˆ,(

)ˆ,(log)ˆ,(

)ˆ,(

)ˆ,(log)ˆ,()ˆ,||ˆ,(

11

2222

11

22222211

cQ

cQcQ

cP

cPcPccKL

i

ii

i

iii

Page 8: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Combine Method 1 and 3, is a sum of the weighted KL information based on each class-specific ability estimate makes use of all possible sources of information◦

◦ Only appropriate for use when the same latent trait us measured across the two classes.

Method 4

)ˆ,(

)ˆ,(log)ˆ,(

)ˆ,(

)ˆ,(log)ˆ,()||(

)ˆ,(

)ˆ,(log)ˆ,(

)ˆ,(

)ˆ,(log)ˆ,()||(

21

2222

21

222221ˆ

11

1212

11

121221ˆ

2

1

cQ

cQcQ

cP

cPcPccKL

cQ

cQcQ

cP

cPcPccKL

i

ii

i

iii

i

ii

i

iii

Page 9: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

12 Item selection methods

Method

1

01

2221

)ˆ,|(

)ˆ,|(log)ˆ,|()||(

xi

iii

cxP

cxPcxPccKL

1

01

2221

)ˆ,|(

)ˆ,|(log)ˆ,|()ˆ,||ˆ,(

xLi

UiUiULi

cxP

cxPcxPccKL

1

011

22222211

)ˆ,|(

)ˆ,|(log)ˆ,|()ˆ,||ˆ,(

xi

iii

cxP

cxPcxPccKL

1

021

222221ˆ

1

011

121221ˆ

)ˆ,|(

)ˆ,|(log)ˆ,|()||(

)ˆ,|(

)ˆ,|(log)ˆ,|()||(

2

1

xi

iii

xi

iii

cxP

cxPcxPccKL

cxP

cxPcxPccKL

1

02

1111

)ˆ,|(

)ˆ,|(log)ˆ,|()||(

xi

iii

cxP

cxPcxPccKL

1

02

1112

)ˆ,|(

)ˆ,|(log)ˆ,|()ˆ,||ˆ,(

xLi

UiUiULi

cxP

cxPcxPccKL

1

012

21212112

)ˆ,|(

)ˆ,|(log)ˆ,|()ˆ,||ˆ,(

xi

iii

cxP

cxPcxPccKL

1

022

212112ˆ

1

012

111112ˆ

)ˆ,|(

)ˆ,|(log)ˆ,|()||(

)ˆ,|(

)ˆ,|(log)ˆ,|()||(

2

1

xi

iii

xi

iii

cxP

cxPcxPccKL

cxP

cxPcxPccKL

i i ccKL )ˆ||()||( 21 ccKLi )||( 12 ccKLi

1

0 )ˆ,|(

)ˆ,ˆ|(log)ˆ,ˆ|(

x iicxP

cxPcxP

1

0 )ˆ,|(

)ˆ,ˆ|(log)ˆ,ˆ|(

x iLi

UU

cxP

cxPcxP

1

01

22

)ˆ,|(

)ˆ,ˆ|(log)ˆ,ˆ|(

x iicxP

cxPcxP

1

02

22

1

01

11

)ˆ,|(

)ˆ,ˆ|(log)ˆ,ˆ|(

)ˆ,|(

)ˆ,ˆ|(log)ˆ,ˆ|(

x ii

x ii

cxP

cxPcxP

cxP

cxPcxP

Page 10: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Memberships: 2; 5000 examinees for each class.

Four item pools, each with 500 items.

Mixing proportion: 50% for both latent classes.

Test length: 20-item

Large item separation

Small item separation

Large ability separation Pool 1 Pool 2

Small ability separation Pool 3 Pool 4

Page 11: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Ability estimation◦ For Method 1 & 2: a single ability estimate across

classes. Administration of item estimated a latent class

membership estimated ability parameter. Sequentially administered item and updated latent class

membership and ability parameter.◦ For Method 3 & 4: class-specific ability estimates.

Administration of item estimated class-specific ability parameters.

Sequentially administered item and updated ability parameters.

The latent class membership only estimated when the last item was administered.

Page 12: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

Results

Page 13: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)
Page 14: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)
Page 15: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)
Page 16: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)
Page 17: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)
Page 18: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)
Page 19: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)
Page 20: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

•The distribution of the converged posterior classification decisions as a function if item sequence (5-20) in the CAT administration.•The classification became stabilized or converged for more than 70% of the examinees after administration of the first five items.

Page 21: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

•The number of examinees whose classification converged at Item 5 was smaller than that for Pool 1, due to less KL information provided by Pool 2.•All alternatives under Method 2 required fewer items to produce stable classification decisions for a majority of the examinees.

Page 22: Hong Jiao, George Macredy, Junhui Liu, & Youngmi Cho (2012)

If more than two latent classes involve in the test, are these KL methods still workable?

To consider mixture model in computerized classification test.

Why the random item selection yielded significantly the most accurate estimates of person ability, compared to the proposed four methods.

The speedness behavior is a kind of latent class. To add this condition by setting the only last several items with latent class model.

Questions