ModelFitting Error: Model has critical EDA Issue

[ecalamia101697]

Hello!
I've been running a model in Databricks for a few months now and have just began running into this error when I'm executing mmm.sample_posterior(). I have not changed any of the variables or how the data is structured. I have also done checks to ensure my data does not have any nans or infs. Could I please get assistance on this? This model delivery has a strict timeline we need to stick to. Thanks!

ModelFittingError: Model has critical EDA issues. Please fix before running `sample_posterior`.

Check type: MULTICOLLINEARITY
- An error occurred during running check_vif: MissingDataError('exog contains inf or nans')
For further details, please refer to `Meridian.eda_outcomes`.

[dattatreyam23]

Hi @ecalamia101697,

Thank you for contacting us!

To assess multicollinearity, variance inflation factor (VIF) is computed for all scaled treatment units (including scaled reach ×
frequency for RF and ORF channels) and scaled control variables. A VIF estimates the extent to which the variance of an explanatory variable is inflated due to collinearity with other variables in the model. A VIF of 1 indicates no collinearity, while higher values suggest increasing levels of multicollinearity.

For a geo model, VIF is first computed for each variable across all geos and times. An ERROR is triggered if any variable can be expressed nearly perfectly as a linear combination of other variables (VIF exceeds the default threshold of 1000). For a national model, VIF is computed for each variable across all times. An ERROR is triggered if the VIF of a variable exceeds the default threshold of 1000.

Since your raw data is clean, could you please get back to us with the following:

• For the error message MissingDataError('exog contains inf or nans'), it can still caused by problems like data of non-float type[See example]. We would suggest you to check if the issue still persists after the data has been converted to float.

• Could you please share with us version of Meridian that you are currently using?

You may read about the EDA checks in detail, from our documentation here

Let us know if dropping those variables resolves the issue. Feel free to reach out to us with any further queries regarding the same.

Thank you,

Google Meridian Support Team

[ecalamia101697]

Hello! Thanks for the response.
I tried converting all integers to floats and the error persists. I am also running on version 1.5.2.

As a note this is the full error when I run eda_outcomes. It doesn't seem to be due to a high VIF, but more a lack thereof.

multicollinearity=EDAOutcome(check_type=<EDACheckType.MULTICOLLINEARITY: 3>, findings=[EDAFinding(severity=<EDASeverity.ERROR: 3>, explanation="An error occurred during running check_vif: MissingDataError('exog contains inf or nans')", finding_cause=<FindingCause.RUNTIME_ERROR: 5>, associated_artifact=None)], analysis_artifacts=[]), pairwise_correlation=EDAOutcome(check_type=<EDACheckType.PAIRWISE_CORRELATION: 1>, findings=[EDAFinding(severity=<EDASeverity.ATTENTION: 2>, explanation='Some variables have perfect pairwise correlation in certain geo(s). Consider checking your data, and/or combining these variables if they also have high pairwise correlations in other geos.', finding_cause=<FindingCause.MULTICOLLINEARITY: 2>, associated_artifact=PairwiseCorrArtifact(level=<AnalysisLevel.GEO: 3>, corr_matrix=<xarray.DataArray 'correlation_matrix' (geo: 33, var1: 161, var2: 161)> Size: 3MB

Thanks!

[ethancalamia]

@dattatreyam23 Any updates on the above? This is a time sensitive ask as my team and I are trying to stick to a deadline.

[dattatreyam23]

Hi @ethancalamia ,

Thank you for your patience.

It seems that while your raw input data may be perfectly clean, the transformations that occur internally within the model can sometimes unexpectedly generate inf or NaN values. This internal transformed data is likely the root cause of the issue.

We request that you inspect your model object mmm to check the following attribute for any inf or NaN values:

mmm.eda_engine.treatment_control_scaled_ds

You may let us know if that helps so we can determine the next best step to get this resolved for you.

Thanks,

Google Meridian Support Team

[ethancalamia]

Thanks @dattatreyam23 !
Did some exploration on my end and there were a few dates where the Population column was 0 causing a divide by 0 error that was hidden behind the inf and NaN values. This was ultimately due to lack of data during those two weeks which is common for our data. Interesting that it did not arise earlier during modeling as we didn't make any adjustments to our data.

Fixed this so that the population column always has a value > 0 and model runs fine!

Thanks for your help!