Robustness of Latent Profile Analysis to Measurement Noninvariance Between Profiles

Area of Science:

• Psychometrics
• Statistical Modeling
• Behavioral Sciences

Background:

• Latent profile analysis (LPA) is used to identify subgroups based on continuous indicators.
• Researchers often use composite or factor scores as indicators in LPA.
• This approach omits measurement models and potential noninvariance across profiles.

Purpose of the Study:

• To examine the robustness of LPA when measurement noninvariance is unmodeled.
• To assess LPA's accuracy in class enumeration and parameter recovery using composite/factor scores.
• To understand the impact of unmodeled noninvariance on LPA results.

Main Methods:

• A simulation study was conducted.
• Latent profile analysis (LPA) was performed using composite or factor scores as indicators.
• The simulation manipulated the degree of noninvariance across latent profiles.

Main Results:

• LPA demonstrated relatively high correct class enumeration rates under conditions of small noninvariance, large class separation, large sample size, and equal proportions.
• Significant bias in profile indicator mean differences was observed with unmodeled intercept and loading noninvariance.
• The findings highlight the sensitivity of LPA to measurement invariance.

Conclusions:

• LPA can be robust to minor noninvariance, but applied researchers must be cautious.
• Unmodeled noninvariance can lead to biased parameter estimates, potentially misinterpreting subgroup differences.
• Careful consideration of measurement invariance is crucial when using composite or factor scores in LPA.