How IMA Reconciles Weekly Channels with Daily Campaigns
This chapter examines Integrated Marketing Attribution, or IMA, as the concrete bridge between strategic channel measurement and tactical campaign decisions. The question is not how to make a weekly number look daily. It is how to change time and resolution while preserving the meaning of the channel contribution that anchors the calculation. By the end, the reader should be able to follow what enters each stage, what crosses each interface, and where the resulting campaign detail remains a model-based allocation rather than independent causal proof.
IMA is built around a simple asymmetry. Marketing Mix Modeling, or MMM, is useful precisely because it looks at aggregated activity over a longer horizon. Campaign optimization needs more detail than that aggregate view naturally supplies. IMA keeps the aggregate model in charge of the quantity to be explained, then uses a separate layer to distribute that quantity among campaigns. The design gains granularity by adding constraints, not by pretending that the original constraints do not exist.
The anchor: a weekly channel contribution
The first stage estimates a weekly MMM from aggregated sales, pricing, promotion, macroeconomic variables, and weekly media series. The source describes two outputs that matter to IMA. One is the incremental sales contribution attributed to each channel for each week. The other is the channel-level coefficient and adstock information that describes how media exposure carries into later outcomes. These outputs define the boundary of the downstream task: IMA is not asked to rediscover total sales or invent a new channel total. It must work with the channel contribution MMM has already estimated.
That distinction gives MMM a strategic role. A channel-level result can support longer-horizon budget allocation because it aggregates enough activity to reduce some of the instability that appears when thousands of campaigns are modeled separately. It also provides a common scale for channels that may use different campaign structures. The price of that stability is that the result is too coarse to answer every tactical question. A weekly contribution does not identify which campaign within the channel deserves a change today.
IMA treats that limitation as an interface rather than a contradiction. The upstream model owns the weekly channel quantity. The downstream process receives that quantity as a target to be represented at a finer time scale. This is an important form of semantic discipline: the downstream model cannot improve its campaign detail by quietly changing the target from incremental channel sales to total sales, conversions, or an unqualified performance score.
The source’s schema makes the resolution change explicit. The modeling window covers three years of historical data. MMM operates in weekly intervals, while IMA produces daily campaign-level attribution. Media can be represented by impressions, clicks, or spend, and the outcome used for the granular layer is campaign-level incremental sales. The evaluation is performed by aggregating IMA back to the weekly channel level and comparing it with the corresponding MMM contribution. The comparison level is part of the model’s meaning, not a formatting detail added at the end.
The second stage performs the temporal handoff. A naive method might split each weekly contribution across days in proportion to daily media volume. That would be easy to explain, but it would discard the carryover behavior learned by MMM. Advertising can influence sales after the original exposure, and the distribution of that effect may not follow the daily pattern of raw impressions or spend.
IMA instead converts the weekly adstock parameters to a daily scale and tunes them where needed to reflect daily dynamics. It applies those parameters to daily channel media, creating adstock-transformed daily series. Those transformed series distribute the weekly channel-level incremental sales into daily channel-level incremental contributions. The invariant is not that each day receives a fixed fraction of the week. The invariant is that the daily allocation retains the channel’s modeled carryover structure and sums back to the weekly channel contribution.
This is why the time conversion must be documented. A “daily channel target” is not a direct daily observation of incremental sales. It is the result of taking a weekly modeled contribution and using a temporal response assumption to locate that contribution across days. If the daily adstock conversion is changed, the campaign inputs and the target change together. The resulting values may still be useful, but they answer a different temporal question.
The arrows show the ownership of the main quantities. MMM creates the weekly anchor and the carryover information. Temporal disaggregation translates that anchor into a daily target. The channel-specific models allocate within that target. Finally, the daily campaign results are aggregated back to the same weekly channel level for a consistency check. The return arrow is a comparison, not a feedback claim: the source describes IMA as downstream of MMM, not as a replacement that feeds new campaign estimates back into the upstream model.
The allocation layer: Bayesian models inside each channel
Once the daily channel target exists, IMA fits an independent attribution model for each channel. For one channel, the target is the daily channel-level incremental sales contribution created by temporal disaggregation. The predictors are daily campaign media series within that channel, such as spend, impressions, or clicks, transformed to reflect carryover. If there are M campaigns, the model estimates an intercept and one effectiveness coefficient for each campaign.
The independence is purposeful. Separating channels prevents the campaign variables from every channel from entering one very large regression at once. It makes the models easier to fit repeatedly and limits the immediate dimensionality of the allocation problem. It also respects the meaning of the target: a channel model distributes the contribution assigned to that channel among its campaigns.
But independence is also a boundary. A campaign in one channel may interact with exposure in another channel, and the current IMA design does not model structured cross-channel interactions. A search campaign and a social campaign can therefore be analyzed in parallel without the model explaining how their effects reinforce or substitute for one another. Parallel channel models are a scalability choice, not evidence that channels are causally isolated.
The Bayesian part addresses a second interface problem: campaign media series are often highly correlated. Campaigns may run during the same flight, share a creative family, or follow a common calendar. In the source’s representative channel, the adstock-transformed campaign variables show substantial off-diagonal correlation and visible blocks associated with campaign families and overlapping flighting. In an ordinary regression, that structure can make coefficients unstable. Small changes in the data may produce large changes in individual estimates, inflated variance, or even implausible sign reversals.
IMA uses information from the aggregate anchor to regularize this underdetermined allocation. The MMM channel-level media effectiveness coefficient becomes the prior mean shared across campaign-level coefficients, and the corresponding MMM offset informs the intercept. The campaign coefficients and intercept use truncated-normal priors constrained to be nonnegative. The constraint encodes the paper’s economic assumption that baseline contribution and media exposure do not reduce sales. Weakly informative prior scales are intended to leave room for the data to move the posterior away from the anchor when the campaign-level signal is strong enough.
The important point is what a shared prior does—and does not do. It gives correlated, sparse campaign estimates a stable reference. It does not force every campaign to receive the same contribution or a proportional share. The source reports that posterior estimates diverge meaningfully across low performers, moderate contributors, and high-impact outliers even though they begin around a common channel-level prior. Bayesian shrinkage stabilizes differentiation; it does not eliminate it.
The model can be written conceptually as a daily balance. The channel’s daily incremental target is represented by an intercept plus the sum of each campaign’s coefficient multiplied by its adstock-transformed media. The intercept captures residual channel contribution not explained by the campaign media inputs. Each campaign’s contribution is calculated from its fitted coefficient and transformed series. The allocation is useful because these pieces are connected to the channel target rather than estimated in isolation.
That balance also explains why the choice of media series belongs in the semantic contract. A coefficient attached to spend does not mean the same thing as a coefficient attached to impressions. An adstocked impression series is not the same exposure as a raw impression series. If a downstream user changes the predictor while retaining the label “campaign contribution,” the model may still run, but the interpretation has changed. The input definition must travel with the coefficient and with the resulting contribution.
The nonnegative prior is another visible assumption. It may be reasonable for a setting in which the model is intended to represent positive media contribution and residual baseline, but it does not establish that every real campaign or interaction can only have a nonnegative effect. Creative fatigue, cannibalization, targeting changes, or cross-channel substitution could make the economic story more complicated. The source uses the constraint as part of its modeling specification; it does not validate the assumption for every channel or campaign.
What reconciliation proves
The final interface is the evaluation rule. IMA’s daily campaign predictions are aggregated to weekly totals and compared with the corresponding weekly MMM-derived channel contributions. In the paper’s representative example, the aggregated predictions closely track the channel target, with a reported R-squared of 0.98. That result supports a specific conclusion: the downstream layer can preserve the aggregate channel signal while producing differentiated campaign-level estimates despite campaign correlation.
It does not support every stronger conclusion a dashboard might suggest. A high reconstruction score does not prove that the allocation is the unique decomposition of the channel effect. It does not independently validate each campaign coefficient through a randomized campaign experiment. It does not show that the daily timing is correct for every channel, because the paper reports a representative channel and does not provide a complete production parameter table. And it does not establish cross-channel interaction effects, because the models are fit independently by channel.
This boundary is not a weakness to hide; it is the reason for the architecture. The value of IMA is operational: a marketing organization can obtain a stable, differentiated view of campaigns without requiring complete user-level tracking, while retaining a check against a strategic channel estimate. The evidence is strongest for consistency with the MMM anchor and for the practical usefulness of regularized decomposition. The evidence is weaker for independent causal attribution at the campaign level.
There is also a one-way flow of trust in the described design. MMM informs the temporal and prior structure of IMA. IMA then produces a finer representation that can be checked by aggregating back to MMM’s level. If that check fails, the team should inspect the semantic interfaces: campaign-to-channel membership, daily media definitions, adstock conversion, target construction, or aggregation. A mismatch is a diagnostic signal. It is not automatically evidence that the campaigns are ineffective or that the upstream model is wrong.
For a cross-channel measurement system, this gives a practical rule. Preserve the meaning of the target while changing resolution; preserve the carryover rule while changing time units; preserve the channel hierarchy while allocating campaigns; and label the output as a model-constrained contribution. Those rules let tactical users act on additional detail without mistaking detail for certainty.
The next question is broader than IMA’s internal pipeline. Even when the upstream and downstream model interfaces are aligned, the evidence can be altered before it reaches the analyst. Privacy reporting may remove matches, break identity continuity, narrow attribution windows, suppress small cells, or add noise. The measurement contract must therefore be tested not only across model stages, but also across the reporting system that determines what signal survives.