19
Item response modeling of paired comparison and ranking data

Item response modeling of paired comparison and ranking data

Embed Size (px)

Citation preview

Item response modeling of paired comparison and ranking data

paired comparison and ranking data

• paired comparison– n(n-1)/2 pairs

• Ranking– n!– A special case of paired comparison when no

intransitive pattern• Thurstone’s model (1927)– Utility (property of item)

Ranking

• Ypair = 1 if ti – tk > 0• Design matrix:• Utility distribution:– Case 3: – Case 5:

• Separate person parameter out of random error:– Note they think of loading parameter as item

attributes (e.g., male actors versus female athletes)

Pairwise comparison

• Add intransitive error:• Then

• Identification:– Origin and scale: N(0,I) for η– Rotation: -------------------------– Additions due to pairwise design:• Fix loading parameters of a statement to 0.• Fix one mean parameter of a statement to 0.• Fix one unique variance of a statement to 1.

• Identification:– At least n=5, 6, 7 for m = 1, 2, 3 (number of

dimension). Why?– If not, require more constraints! What it is?• Constrain All the covariance matrix (I have tried this!)

Thurstonian IRT model• Recall• Do substitution

• Take n=3 as example:

Parameter estimation

• MML may be infeasible because ICCs are conditionally dependent for Thurstone IRT model.

• Limited information method is applicable by using Mplus.

• But d.f. should be modified for ranking data:

Item characteristic function

• Recall Y* =

• Re-expressed as

• Note

Latent trait estimation, information functions, and reliability estimation

• Locally independence is violated!• MAP• Information function

• Reliability – 1.– 2.

Simulation studies

• To estimate• Sample size: 200, 500, 1000 • Item size: 6, 12• Equal or unequal variance:

Results

Results

How "close" are the observed values to those which would be expected under the fitted

model?

MAP

Vocational interest (pairwise comparison)

• Unrestricted thresholds: p=.046, RMSEA=.016• Equal w: p=.000, RMSEA=.025• Constrained thresholds: p=.000, RMSEA=.025• Reliability: .62 (theoretical); .43 (empirical) due to shrunken MAP

RealisticInvestigativeArtisticConventionalSocialEnterprising

Vocational interest (pairwise comparison)

Work motivation (ranking data)

• Chi-square fit index: p=.000, RMSEA=.062

Work motivation (ranking data)

• Reliability: .74 (theoretical); .76 (empirical)

Discussion

• Locally dependence– Using MCMC

• Discrimination: positively vs negatively worded• Multiple traits• Forced-choice design