Uncovering a De-Biased Latent Proxy of Teaching Quality From Heavily Skewed Evaluation Data: A Latent Space Reconstruction Framework via ED-VAE

The psychometric validity of Student Evaluations of Teaching (SETs) is severely compromised by the ‘‘ceiling effect’’, where negative skewness compresses variance and attenuates the sensitivity of conventional linear models. To address this, we propose the Educational De-biasing Variational Autoencoder (ED-VAE), a deep generative framework with a skewness-aware loss function designed to estimate a debiased latent proxy of teaching effectiveness. Leveraging a rare quasi-experimental dataset (N= 22, 429) collected over the 2019–2024 period from 26 application-oriented institutions in China (65% of the 66 instructors taught largely identical curricula across more than one institution), we were able to mitigate confounding from instructor heterogeneity in a way that is seldom feasible in standard SET studies. Experimental results show that the extracted latent quality (Z1) achieves high internal consistency (r≈ 0.96) and consistent cross-context behaviour (r≈ 0.16). To ensure external validity, a 5-fold grouped cross-validation on unseen instructors confirmed out-of-sample generalizability (r≈ 0.156). Furthermore, mixed-effects modelling (HLM) indicates that the latent construct primarily captures individual instructor variance rather than mere institutional response styles. While the OLS fixed-effects baseline failed to detect institutional impact (p= 0.173), ED-VAE reveals a modest but statistically stable correlation (r= 0.126, 95% CI [0.111, 0.140]) between the latent proxy and institutional resources. A Monte-Carlo simulation calibrated to our empirical response distribution (P(score=5) ≈ 92%) shows that ceiling-induced range restriction attenuates Pearson correlations by roughly 44%; under this attenuation, an observedr= 0.126 is consistent with a small-to-moderate construct-level effect of ρ ≈ 0.22. We therefore interpret ED-VAE as a proof-of-concept computational auxiliary, suitable for macro-level research and diagnostic institutional analysis, rather than as a deployable tool for high-stakes individual decisions.