Hi Aki,
I’m sure it’s wrong, but I don’t understand your argument: These tail probabilities are all very close to 1: the predicted value ax+b is much smaller than the limit of detection of 0.5. The censored values close to x = -2 should not be influential observations. The fact that you mention thicker tails tells me that we’re talking past each other.
As a demonstration, I’ve added a non-censored outlier (data point 100). This has a high Pareto k in both the jittered and non-jittered case.
I’ve also changed the scale a bit to get more extreme k values. Interestingly, the Pareto ks for the fest few data points (smallest x) are close to zero.
this is done with arviz 0.22.0 by the way.
best regards,
Chris