Measurement Bias Detection Through Factor Analysis Barendse, M. T., Oort, F. J. Werner, C. S.,...

20
Measurement Bias Detection Through Factor Analysis Barendse, M. T., Oort, F. J. Werner, C. S., Ligtvoet, R., Schermelleh-Engel, K.

Transcript of Measurement Bias Detection Through Factor Analysis Barendse, M. T., Oort, F. J. Werner, C. S.,...

Measurement Bias Detection Through Factor Analysis

Barendse, M. T., Oort, F. J. Werner, C. S., Ligtvoet, R., Schermelleh-Engel, K.

Defining measurement bias

• Violation of measurement invariance

Where V is violator• If V is grouping variable, then MGFA is suitableIntercepts – uniform biasFactor loadings – non-uniform bias (vary with

t)

Restricted Factor Analysis (RFA)

• Advantages of RFA over MGFA:V can be continuous or discrete, observed or

latentInvestigate measurement bias with multiple Vs.More precise parameter estimates and larger

power• Disadvantage of RFA:Not suited for nonuniform bias (interaction term)

Approaches for non-uniform bias

• RFA with latent moderated structural equations (LMS)

---- Simulation (categorical V) showed at least as good as MGFA

• RFA with random regression coefficients in structural equation modeling (RSP)

---- performance unknown

This paper…• Compared methods:MGFA RFA with LMSRFA with RSP• Measurement biasUniformNonuniform• ViolatorDichotomousContinous

Data generation (RFA)

• True model:

• Uniform bias: . Nonuniform bias: • T and v are bivariate standard normal

distributed with correlation r• e is standard normal distributed• u is null vector

0b 0c

Simulation Design

For continuous V:• Type of bias (only on item 1): No bias (b=c=0), uniform bias(b=0.3,c=0), nonuniform bias (b=0,c=0.3), mixed bias (b=c=0.3)• Relationship between T and V Independent (r=0), dependent (r=0.5)

Simulation Design

For dichotomous V:• V=-1 for group 1 and v=1 for group 2• Model can be rewritten into

• Relationship between T and V: Correlation varies!

.)()(

,)()()2(

)1(

detcabux

detcabux

)1,4.0(~

)1,4.0(~)2(

)1(

NT

NT

The MGFA method

• When v is dichotomous, regular MGFA• When v is continuous, dichotomize x by V• Using chi-square difference test with df=2Uniform : interceptsNonuniform: loadings

The RFA/LMS method

• V is modeled as latent variable:Single indicatorFix residual variance (0.01)Fix factor loading• Three-factor model: T, V, T*V• Robust ML estimation• Chi-square test with S-B correction: : uniform bias : nonuniform bias

0b0c

RFA/RSP method

• Replacing with , where is a random slope.

• Robust ML estimation• Chi-square test with S-B correction: : uniform bias : nonuniform bias

0b

0c

Single & iterative procedures• Single run procedure: test once for each item• Iterative procedure: 1)Locate the item with the largest chi-square

difference2)Free constrains on intercepts and factor

loadings for this item and test others3)Locate the item with the largest chi-sqaure

difference 4)…5)Stops when no significant results exist or half

are detected as biased

Results of MGFA – single run

• Shown in Table 2.• Conclusion:1.better with dichotomous than with

continuous V; 2.non-uniform bias is more difficult to detect

than uniform bias; 3.Type I error inflated.

Results of MGFA – iterative run

• Shown in Table 3.• Conclusion:1.Iterative procedure produces close power as

single run does.2.Iterative procedure produces better

controlled Type I error rate.

Results of RFA/LMS & RFA/RSP - single run

• Shown in Table 4 and Table 5.• Conclusion:1.LMS and RSP produce almost equivalent

results. 2. larger power than MGFA with continuous V.3.More severely inflated Type I error rates

Results of RFA/LMS & RFA/RSP - iterative run

• Shown in Table 6.• Conclusion:1.Power is close to the single run2.Type I error rates are improved

Results of estimation bias - MGFA

• Shown in Table 7.• Conclusion:1.Bias in estimates is small2.Bias in SD is non-ignorable3.Smaller bias in estimates for dichotomous V

(dependent T&V)

Results of estimation bias - RFA

• Shown in Table 8 & 9• Conclusion:1.Similar results for LMS and RSP2.Small bias in estimates3.Non-ignorable bias in SD4.Smaller SE than MGFA5.Smaller bias in estimates than MGFA with

dependent T&V, continuous V.

Discussion

• Nonconvergence occurs with RFA/LMS

Non-convergence• Summary: