different results from plot_posterior_predictive when run twice

[Alcampopiano]

Hello and thanks for the great library.

When I run mmm.sample_posterior_predictive(mmm.preprocessed_data['X'], extend_idata=True, combined=True) followed by mmm.plot_posterior_predictive(original_scale=True);, I get the following plot:

Then, if I run the same code again, I get this plot:

Is this expected?

I am using v0.3.3

[wd60622]

Can you share your setup code? Are there any fits as well?

[Alcampopiano]

I'm be happy to! Here are a few of the plots/tables used in the example guide as well as my model config (just uniform priors mostly); however, even when I use more informed priors I find this same effect.

All r_hats are 1.01 or 1.

Please let me know if this helps! Glad to share other metrics.

my_model_config = {
    "intercept": {"dist": "Uniform", "kwargs": {"lower": 0, "upper": 1}},

    "beta_channel": {
        "dist": "Uniform",
        "kwargs": {"lower": 0, "upper": 1},
    },
    "alpha": {"dist": "Uniform", "kwargs": {"lower": 0, "upper": 1}},
    "lam": {"dist": "Uniform", "kwargs": {"lower": 0, "upper": 1}},
    "likelihood": {
        "dist": "Normal",
        "kwargs": {"sigma": {"dist": "HalfNormal", "kwargs": {"sigma": 2}}},
    },
    "gamma_control": {"dist": "Uniform", "kwargs": {"lower": 0, "upper": 1}},
    "gamma_fourier": {"dist": "Uniform", "kwargs": {"lower": 0, "upper": 1}},
}

[wd60622]

Hi @Alcampopiano

Was there additional setup prior to calling sample_posterior_predictive? I have not encountered this and the example notebook don't experience this as well.

Would you happen to know if this has been resolved with the later versions?

[Alcampopiano]

Hi @wd60622

Thanks very much for checking in.

Unfortunately, I dont have that project anymore—otherwise I'd be happy to try and reproduce.

I suspect that newer versions of pymc-marketing are not affected as you mention but cannot confirm as I switched to using pymc.

Thanks for all the great work!