Help with Censored Regression

Questions
21

Censored model solved :slight_smile:

Correct model is this…

def censored_regression(x, y, bounds):
    
    # data pre-processing
    left_censored = (y <= bounds[0])
    x_lc = x[left_censored]
    y_lc = y[left_censored]
    
    right_censored = (y >= bounds[1])
    x_rc = x[right_censored]
    y_rc = y[right_censored]
    
    uncensored = (y>bounds[0]) & (y<bounds[1])
    x = x[uncensored]
    y = y[uncensored]

    with pm.Model() as model:
        m = pm.Normal("m", mu=0, sd=1)
        c = pm.Normal("c", mu=0, sd=1)
        σ = pm.HalfNormal("σ", sd=1)
        y_likelihood = pm.Normal("y_likelihood", mu=m * x + c, sd=σ, observed=y)
        left_censored = pm.Potential("left_censored", normal_lcdf(m * x_lc + c, σ, y_lc))
        right_censored = pm.Potential("right_censored", normal_lccdf(m * x_rc + c, σ, y_rc))

    with model:
        trace = pm.sample()

    return model, trace

Screenshot 2021-01-24 at 14.01.01
Screenshot 2021-01-24 at 14.01.011110×746 69.6 KB

Screenshot 2021-01-24 at 14.01.04
Screenshot 2021-01-24 at 14.01.04950×634 105 KB

If people can confirm that this looks kosher, then I’m happy to put the truncated and censored regression code together in an example notebook.

22

Looks good to me!

23

@AlexAndorra I was suggesting to @drbenvincent to develop an example of Censored and Truncated regression to the pymc-examples, since he has spent quite some time tinkering and debugging about these in pymc3. Since you are more familiar with the pymc-examples, could you share your opinion as to whether this makes sense (I don’t want to waste anyone’s time :slight_smile: )

24

I’ve made a pull request for a notebook on truncated regression. I’m “ok” with GitHub, so I have little confidence that I’ve done it correctly.