The Lineup Protocol: Using Simulation to Improve Model Diagnostics in Binary Logistic Regression
What does a good diagnostic plot look like? When we look at a residual or QQ-plot, we are implicitly comparing it to what we view as a good plot in our mind. It turns out that we as humans are exceptional at picking patterns out of randomness and hence tend to over-interpret visual diagnostics. We attempt to remedy this problem by simulating good diagnostic plots for a viewer to use as references when they evaluate a graphical display. In this talk, we will look at a variety of simulation-based diagnostics in the context of binary logistic regression (regression where the response can only be 0 or 1). We explore how to simulate data according to a fitted model and discuss how simulation can help us detect model violations. In particular, we look at the results of a simulation study for a specific type of simulation-based diagnostic, the lineup protocol.