Control Channels - Contribution

[shubhambooking2]

i have a model where i have multiple channels as part of control. Is there a way to get the summary indicating the contribution(%) associated with each of the control channel. Currently i see that we get the overall baseline contribution, however understanding i wanted to understand how much of contribution is associated with true baseline (comprising of trend, seasonality, knots etc.) and seperately the contribution of each of other control variables

[brettackers]

+1 on this.
we have a control where we had a major site update that saw a large increase in our KPI, however, the contribution output doesn't reflect this

[gauravsab1]

The contribution of control variables can be computed using the analyzer.expected_impact method. You can iteratively set each control to 0 (or any other reasonable baseline value) and compare to the case where that same control is set to its usual value in the dataset.

Note that the "overall baseline" (the orange bar in this chart) is computed as expected revenue in the absence of all media channels (all media set to 0; all other variables held at the provided values in the dataset).

[sonriks6]

Hello,

This is what I use so far:

outcome_tf = model_analysis.expected_outcome(
aggregate_times=False,
use_kpi=True
)
outcome = pd.DataFrame(tf.reduce_mean(outcome_tf, axis=[0, 1]).numpy())
outcome.to_csv(f'./{brand}_{market}_incremental_kpi.csv')

But have no clue on how to get the %contribution into the baseline of the control factors I'm using, any suggestions??

[gmoosman]

Thanks, @sonriks6 !
That is useful.

I am not used to the process. Do you have any code on how to set specific controls to 0 in order to compare to the case where that same control is set to its usual value in the dataset? (as @gauravsab1 writes)

[Amanda-MCN]

Having read this in the documentation I just want to check that this solution is valid

[dattatreyam23]

Hi @Amanda-MCN,

Thank you for reaching out to the Google Meridian support team.

The proposed solution, using the analyzer.expected_outcome method provides an approximate workaround for the estimation of baseline contributions of control variables by setting each control to zero. However, the theory as mentioned in the documentation on Baseline Breakdown still stands valid. Control variables in Meridian do not produce causal effects or baseline contribution percentages, as they are primarily utilised to improve prediction accuracy, rather than acting as additive components in expected outcome. Hence, although the workaround is useful for exploratory purposes, it should not be relied upon for precise analysis of baseline contribution.

Additionally, we have recently implemented the support for organic media and non-media treatment variables. The contribution breakdown now will show the contribution of this variable as well. Please check our documentation to get an in-depth view of the same.

Feel free to reach out if you have any further questions regarding this.

Thank you,

Google Meridian Support Team

[Amanda-MCN]

Thanks for the clarification. As another clarification, I assume I can use expected_outcome to get the knots contribution?

[sonriks6]

Hello,

There's also the option to compute the gamma_c to get the control variable contribution:

gamma_C = meridian.inference_data.posterior.gamma_c
mean_chains = gamma_C.mean(dim=('chain', 'draw'))
print(f"{mean_chains}")

You get and array with those insights but not sure if this is percentage attributed or how to read the results(?)