McKelvey and Zavoina (1975) [269] Proposes a pseudo-R2 for (ordinal) probit models, based on estimating R2 for an underlying latent linear regression.
Kent and O'Quigley (1988) [221] Survival models, e.g. proportional hazards. Considers a transformation T* of the survival time T such that T given X has a Weibull distribution and log T is linear in X. Then proposes a pseudo R2 which is an estimate of the R2 for this linear model. Also proposes an R2 based on proportional reduction in Kullback-Leibler divergence between models. Estimates for this, one of which is R2LR.
Hage and Mitchell (1992) [173] Compares pseudo-R2 measures of McKelvey and Zavoina (1975) [269], Aldrich and Nelson (1984) [16] (rescaled to have 0-1 values) and R2L for binary logit and probit models. Simulations to evaluate performance as estimators of an underlying latent R2. Comments on adjusting for number of parameters, sensitivity to distributional assumptions and the existence of a latent outcome.
Laitila (1993) [231]
Proposes a pseudo-R2 for logit, probit and tobit models. Based on the
idea of a continuous latent variable underlying the binary or censored
obaserve response. The proposed statistic estimates the standard
R2=var($\hat{Y}$)/var(Y) for the latent linear regression
(the same as R2 of McKelvey and Zavoina (1975) [269])
Veall and Zimmermann (1994a) [374] Review of main pseudo-R2 measures for binary regression models. Simulations to assess which best approximates R2 for an underlying latent linear model (the McKelvey and Zavoina measure [269] does best).
Veall and Zimmermann (1994b) [375] Veall and Zimmermann (1994a) [374] for tobit models.