PHD Discussions Logo

Ask, Learn and Accelerate in your PhD Research

Question Icon Post Your Answer

Question Icon

Which family of generalized linear models (e.g., logistic regression, probit models) is optimal for estimating the probability of a binary outcome as a function of a continuous predictor, and how should model fit and assumptions be evaluated?

I'm modeling the probability of cell viability (live/dead) as a function of increasing drug concentration. While logistic regression is the default, I want to ensure I'm using the most appropriate link function and rigorously checking the model's assumptions beyond a simple significance test. The probit model is also mentioned in the literature for such dose-response relationships.

 

All Answers (1 Answers In All)

By Sohail Answered 4 months ago

For your dose-response viability data, both logistic and probit models are excellent and typically yield nearly identical substantive conclusions. In my practice, I usually start with logistic regression due to its more intuitive odds-ratio interpretation. The optimal choice can hinge on the underlying error distribution you assume; probit implies a normal latent variable. More critical is the evaluation. I would recommend examining the Hosmer-Lemeshow test for global fit, scrutinizing residual plots (like deviance residuals vs. predicted values), and checking for influential points. Ensure your continuous predictor has a reasonably linear relationship with the log-odds.

Your Answer