Why are MaskedArrays not supported anymore and is there suggested circumvention?

Questions v5
10

Hey @ricardoV94 I am facing a similar issue. I have a model where my observed outcome variable has missing data. I would like to automatically impute the data using the sampling distribution. I have a pymc model that looks like this:

with pm.Model(coords=coords) as STS:

    time = pm.Data("time", train_time_scaled)
    
    ft = pm.Data("fourier_series", train_fourier_series.to_numpy())

    k0 = pm.Normal("k0", 0, 1)
    m = pm.Normal("m", 0, 1)

    b = pm.Normal("b", 0, 1, dims="fourier_coef")
    seasonal = pm.Deterministic("seasonal", pm.math.dot(ft, b))

    mu = np.linspace(0.1, 0.8, len(coords['changepoints']))
    s = pm.Data("s", mu, dims="changepoints")

    delta = pm.Laplace("delta", mu=0, b=0.2, dims="changepoints")
    a = pm.Deterministic("a", pm.math.where(time[:, None] < s[None, :], np.array(0.), np.array(1.)))

    gamma = pm.Deterministic("gamma", -s*delta, dims="changepoints")
    func = (k0 + pm.math.dot(a, delta))*time + (m + pm.math.dot(a, gamma))

    trend = pm.Deterministic("trend", func)

    error = pm.HalfNormal("error", sigma=1)
    pm.Normal(
        "likelihood",
        mu=trend + seasonal,
        sigma=error,
        observed=train_xs_scaled
    )

I am passing train_xs_scaled directly because it has missing values so I am not able to use pm.Data()

Now I want to produce OOS forecasts using this model. For that I have:

with STS:
    pm.set_data(
        new_data=dict(
            time=test_time_scaled,
            fourier_series=test_fourier_series
        )
    )
    predictions = pm.sample_posterior_predictive(idata)

But this results in an error:

IndexError: boolean index did not match indexed array along dimension 0; dimension is 49 but corresponding boolean 
dimension is 432

Where the training examples have a length of 432 and the testing examples have a dimension of 49. Is there any way I can modify the model so that it does not use the old observed data for the ‘likelihood’ since they should not be needed anyway to produce the forecast?