So, what you’re suggesting is to use either the impression variable or the spend variable, not both together, right?

But the official guideline mentions analyzing both variables together — so I just want to make sure I understand correctly.
In my case, since some media channels have extremely low CPM, you’re saying it’s not ideal to include both spend and impressions at the same time, correct?

I’ll follow your advice and test them separately — first focusing on media_spend, and then on media_impression.
Previously, I applied a log transformation because the range of impressions varied so widely, but as you mentioned, I’ll skip the log transformation this time.

Also, I’m curious about whether multicollinearity happens only in my case or if it’s common across other media as well.
It seems natural that spend and impressions would be highly correlated, since higher spend typically drives more impressions.
So in such cases, are you generally recommending not to include both variables (spend and impressions) in the model simultaneously?
Because it feels like multicollinearity between the two is inevitable given their relationship.

But I’m a bit curious about something —
the impression variable is mainly used for causal inference, right?
Then why would multicollinearity occur between the impression and spend variables?

Also, I think there might be a small mistake in the sentence “Spend data is not strictly mandatory if your goal isn’t accurate ROI estimates.” —
did you actually mean to say “impression data” instead of “spend data”?

This is a snippet from my code.
So, according to what you said, you mean I should test the model by including either the media_spend variable or the media_impression variable, but not both together, right?

coord_to_columns = load.CoordToColumns(
    time='time',
    controls=['GQV'],
    population='population',
    kpi='TotalUU',
    media=[
        'media1_impression_log', 'media2_impression_log',
        'media3_impression_log', 'media4_impression_log',
        'media5_impression_log', 'media6_impression_log',
        'media7_impression_log', 'media8_impression_log'
    ],
    media_spend=[
        'media1', 'media2', 'media3', 'media4', 'media5', 'media6',
        'media7', 'media8'
    ],
    non_media_treatments=[]
)

correct_media_to_channel = {
    'media1_impression_log': 'media1',
    'media2_impression_log': 'media2',
    'media3_impression_log': 'media3',
    'media4_impression_log': 'media4',
    'media5_impression_log': 'media5',
    'media6_impression_log': 'media6',
    'media7_impression_log': 'media7',
    'media8_impression_log': 'media8'
}

correct_media_spend_to_channel = {
    'media1': 'media1',
    'media2': 'media2',
    'media3': 'media3',
    'media4': 'media4',
    'media5': 'media5',
    'media6': 'media6',
    'media7': 'media7',
    'media8': 'media8'
}

But let’s say after running the models separately, I get the following results:
when using only the spend variable, media A shows a high ROI (contribution relative to spend),
but when using only the impression variable, media A shows the lowest ROI (contribution relative to impressions).

How should I interpret that?
I’m asking because I expect that kind of result might happen when I run the models separately like this.

Also, when using only the impression variable, what kind of prior should I set?
Previously, I just used the default priors defined by your team —
but in this case, I’m wondering whether I should change them or keep them as is.

Hello @mona6332 ,

Thank you for contacting us again!

So, what you’re suggesting is to use either the impression variable or the spend variable, not both together, right?

But the official guideline mentions analyzing both variables together — so I just want to make sure I understand correctly. In my case, since some media channels have extremely low CPM, you’re saying it’s not ideal to include both spend and impressions at the same time, correct?"

Sorry for the confusion. Both variables (spend and impressions/conversions) can be included in the model (and it is generally recommended to do so). Impressions (media column) are used in the Adstock Hill transformations and spend (media_spend column) is used in ROI calculation. If you only include spend, then it will be used in both Adstock Hill transformations and ROI calculation.
Impressions and spend for a channel can be (and usually are) correlated, that's not an issue for Meridian. In your case, multicollinearity might not be an issue but the log-transformed impressions might be the issue. Meridian already incorporates non-linear effects, specifically diminishing returns, through the Hill function internally. Unless you have a specific reason to include the log-transformed impressions, it is generally not recommended.

Also, I think there might be a small mistake in the sentence “Spend data is not strictly mandatory if your goal isn’t accurate ROI estimates.” — did you actually mean to say “impression data” instead of “spend data”?

The primary purpose of providing Spend Data (media_spend or rf_spend) is strictly for the denominator in Return on Investment (ROI) and Marginal ROI (mROI) calculation. If your goal is not to calculate ROI or mROI (fore.g., if you are only interested in measuring incremental outcome (contribution) or estimating the response curve shape), then the spend data is not needed for the core function of the regression model itself.

Feel free to reach out for any further queries or feedback regarding Meridian.

Google Meridian Support Team

Issue Summary: NameError after model training — likely due to variable scaling instability

Context

While training a Meridian MMM model, the model executed successfully through prior sampling and posterior sampling phases.
However, when calling the summary output function, the following error occurred:

NameError: name 'mmm_summarizer' is not defined

This error initially appeared to be an import issue, but upon deeper inspection, it seems related to numerical instability during the model’s sampling phase, caused by unscaled (large-magnitude) input variables.

Steps to Reproduce

Load CSV data using the Meridian CsvDataLoader Define priors and model spec:

coord_to_columns = load.CoordToColumns(
    time='time',
    controls=['GQV'],
    population='population',
    kpi='TotalUU',
    media=[
        'media1_impression_log', 'media2_impression_log',
        'media3_impression_log', 'media4_impression_log',
        'media5_impression_log', 'media6_impression_log',
        'media7_impression_log', 'media8_impression_log'
    ],
    media_spend=[
        'media1', 'media2', 'media3', 'media4', 'media5', 'media6',
        'media7', 'media8'
    ],
    non_media_treatments=[]
)

correct_media_to_channel = {
    'media1_impression_log': 'media1',
    'media2_impression_log': 'media2',
    'media3_impression_log': 'media3',
    'media4_impression_log': 'media4',
    'media5_impression_log': 'media5',
    'media6_impression_log': 'media6',
    'media7_impression_log': 'media7',
    'media8_impression_log': 'media8'
}

correct_media_spend_to_channel = {
    'media1': 'media1',
    'media2': 'media2',
    'media3': 'media3',
    'media4': 'media4',
    'media5': 'media5',
    'media6': 'media6',
    'media7': 'media7',
    'media8': 'media8'
}

loader = load.CsvDataLoader(
    csv_path=r"C:\Users\Desktop\meridian_data.csv",
    kpi_type='non_revenue',
    coord_to_columns=coord_to_columns,
    media_to_channel=correct_media_to_channel,
    media_spend_to_channel=correct_media_spend_to_channel,
)
data = loader.load()

roi_mu = 0.2
roi_sigma = 0.9
prior = prior_distribution.PriorDistribution(
    roi_m=tfp.distributions.LogNormal(roi_mu, roi_sigma, name=constants.ROI_M)
)
model_spec = spec.ModelSpec(prior=prior)
mmm = model.Meridian(input_data=data, model_spec=model_spec)
mmm.sample_prior(500)
mmm.sample_posterior(
    n_chains=5,
    n_adapt=500,
    n_burnin=500,
    n_keep=1000,
)
mmm_summarizer.output_model_results_summary(
    'summary_output_MERIDIAN_.html',
    save_dir,
    '2024-01-07',
    '2025-08-31'
)

Then, the following error occurs

NameError: name 'mmm_summarizer' is not defined

Observations

When the model is trained without scaling (e.g., large values in media spend variables, up to millions),
the error occurs after posterior sampling completes — during summarization.

However, when I divide the media spend variables by 1000,
the entire workflow runs smoothly and the HTML summary is successfully generated.

This suggests that the error is not caused by a missing import, but by numeric overflow / instability during posterior sampling that disrupts the internal state before mmm_summarizer runs.
As a result, the summarizer is never properly initialized, or the internal references are lost.

Hypothesis

Root cause: unscaled input features (e.g., spend in large magnitudes) caused overflow or NaN values during posterior sampling steps.

This likely interrupted the internal object graph, resulting in a misleading NameError instead of a numerical warning.

When the same data was scaled down (e.g., spend ÷ 1000), the model trained and summarized correctly.

Hello @mona6332 ,

Thank you for contacting us again!

Please include the following lines in your code and try again, then let us know if you encounter any further issues.

from meridian.analysis import summarizer
mmm_summarizer = summarizer.Summarizer(mmm)

Feel free to reach out for any further queries or feedback regarding Meridian.

Google Meridian Support Team