🔎 Why measurement bias matters
Many social-science concepts are latent and observed only through indicators. When an external variable of interest directly affects those indicators, estimates of its relationship with the latent trait become biased. These violations of measurement invariance can, for example, confound true cross-country differences in postmaterialism with differences in how questions are interpreted.
🧭 What this paper does
Extends the recently introduced EPC-interest (an expected-parameter-change measure for sensitivity analysis) from continuous-data latent variable models to models that use categorical indicators. The extension makes it possible to quantify how much direct effects of an external variable on categorical indicators would alter the estimated relationship between that variable and the latent construct.
🛠️ How the extension works
- Adapts the EPC-interest framework to latent variable models with categorical indicators.
- Provides a sensitivity-analysis tool that evaluates the importance of potential measurement noninvariance and informs decisions about partial measurement invariance.
- Frames the change in the substantive estimate as the parameter of interest, allowing direct assessment of bias from measurement effects.
📋 Demonstrations with real data
- U.S. Senate roll-call votes: applies the extended EPC-interest to categorical voting data to illustrate detection of measurement-related distortions in latent trait estimates.
- World Values Survey postmaterialism rankings: assesses whether respondent rankings reflect true latent differences or measurement variation across groups.
💡 Why it matters
The extension equips researchers who use categorical indicators with a principled sensitivity-analysis method to:
- Detect and quantify bias from measurement noninvariance
- Inform decisions about imposing partial measurement invariance
- Improve confidence in substantive inferences from latent variable models in comparative and applied research
