Re: [attribution] Methodological concerns regarding Attribution Level 1 and causal measurement

Awesome -  couple of things come to mind
- is the hex the only spatial unit? For some uses, some people want spatial units to aggregate into cities, states, provinces, CMA and DMA. Using a small spatial unit as the bases enables aggregations into larger units like Lego.
Do Hexes enable that better than census blocks (which align with one of the major gold standard data sources - the Census)
Hexes might be the best to avoid depending on political boundaries that change but they should also support larger poltical boundaries for specific uses.

Next up - two levels of spatial - the person’s resting night time location (aka Home) and their current roaming location are very useful.

The coarse MMM is critical to ensure buying agents have the right information to make buying decision activated with in the spatial cell. This is what I presented in Brussels two weeks ago - we need MMM to replace “attribution” so that agents are working with “attributed conversion events” instead of gross conversions.

c



> On Jul 14, 2026, at 11:28 AM, Nathan Woodman <nate@proofindata.com> wrote:
> 
> Rick, Chris, all,
> 
> The thread has moved quickly and the core proposal is now spread across several replies. I want to put the full version in one place on the list, for the record.
> 
> The suggestion is narrow. Standardize the unit of analysis. Once the unit is standardized, the inputs and the experimental designs that consume it can remain open.
> 
> Geography sidesteps the cross-site diversion problem. Every operator sees the same geographic units: every publisher, every advertiser, every measurement vendor, every channel including CTV, out-of-home, audio, and in-store outside the browser. No identity bridging is required, and the randomization unit is consistent across the ecosystem by construction. But geography alone is not enough. The unit of analysis is spatial × temporal, and both need to be standardized for cross-channel composition to actually work. A CTV ad airing Wednesday 8pm Eastern, an out-of-home impression at 8:15pm, and a web conversion at 8:30pm only become composable across channels if they all resolve to the same (cell, time-bin) buckets.
> Spatial unit: the Uber H3 hexagonal hierarchical grid, with R8 (roughly 460 meter hex edge) as the default resolution.
> Hexagonal cells are uniform in neighbor relationships, with none of the diagonal-versus-orthogonal asymmetry that square grids carry.
> Hierarchical resolution (R0 through R15) supports country scale down to roughly 5 meters, with clean parent and child rollup.
> H3 is open source and already widely adopted by geospatial vendors, ride-share, retail planning, and increasingly ad tech.
> Cells do not depend on political or administrative boundaries. They are stable across countries, across time, and across boundary changes. This matters substantively for longitudinal measurement and for cross-country interoperability.
> 
> Temporal unit: ISO 8601 week plus hour-of-day in UTC, with a local-time annotation for interpretability.
> ISO week is internationally standardized and naturally hierarchical: year to quarter to month to week to day to hour.
> UTC sidesteps the daylight-saving discontinuities that break clean bucketing.
> Hour-of-day UTC paired with a published local-time annotation lets analyses use clock-time semantics without DST corruption.
> Calendar context (religious calendars, school sessions, civic events) can be layered as optional metadata that any bucketing scheme can incorporate without changing the base unit.
> 
> Composite canonical bucket: (H3 cell at declared resolution) × (ISO week) × (hour UTC), with the per-campaign histogramIndex resolving to this composite. Coarser buckets (cell × week, or cell × month) work for slower-cycle analyses like MMM. Finer buckets (cell × week × hour) work for day-part-aware causal experiments.
> 
> Once the unit is standardized, four things become possible that are not today.
> Consistent bucketing across campaigns, operators, and measurement paradigms. Observational attribution, causal geo experiments per Rick's methodology, MMM, and retail-trade-area attribution all use the same units.
> Browser histogram outputs become a shared input for external geo-time-aware models. A geo-RCT outside the browser can consume browser attribution histograms as one of its inputs, with no custom integration per partner.
> Non-browser channels (CTV, out-of-home, audio, in-store) can geo-time-resolve impressions and conversions to the same units. The browser API becomes a contribution to cross-channel measurement rather than a silo within it.
> Bidirectional integration. Browser histograms feed external geo-time experiments, and external experiments inform how histograms are bucketed for new campaigns. To Chris's point, if the hex is in or derived from the bid request, the same unit serves bid-time reach and frequency management, not only post-bid measurement.
> 
> This does not require changes to the API surface. saveImpression() and measureConversion() can carry composite H3 × ISO-week × hour identifiers as histogramIndex values where the campaign chooses, and campaigns that prefer other bucketing dimensions can use them. The spec would just publish the convention so cross-operator and cross-channel composition becomes routine rather than a per-partner custom integration.
> The benefit is structural. Attribution Level 1 stops being a browser-only attribution layer and becomes the browser-edge contribution to a broader cross-channel measurement system. The privacy model, DP, MPC aggregation, and per-site budgets stays exactly as it is. This also addresses the cross-site diversion gap, because the randomization unit is geographic rather than user-identifier, and it does so without requiring the API or the privacy model to change.
> 
> Best,
> Nathan Woodman
> 
> On Thu, Jun 4, 2026 at 9:58 AM Chris Williams <chris_williams65@icloud.com <mailto:chris_williams65@icloud.com>> wrote:
>> Imo - there is no such thing as browser attribution.
>> 
>> Attribution cannot be executed without consideration of non-digital media and non-media factors eg price changes. 
>> Attribution can ONLY occur in the geo-temporal space on a probabilistic modelled method.
>> Hence the importance of standardized Geo - Canada uses postal codes, much smaller than DMAs.
>> 
>> MTA needs to be re-defined as Multi-Touch Analysis - an input into MMM that helps to illuminate cross media touchpoints and multi-channel dependancies. Look at how Google Meridian uses Search as a confounding factor in MMM. No medium works alone.
>> 
>> I’m doing a webinar with CIMM on MMM June 8th where I stress the importance of geo (thanks Rick) in the outcome measurement process in addition to linking it back to the planning, buying and campaign set up.
>> 
>> c
>> 
>> 
>> <1780415793990.jpeg>
>> CIMM is hosting its next Research Seminar, Sharpening the Competitive Edge of TV's Ad Proposition, on Monday, June 8 from 2:00–3:30 PM ET. Open to the entire industry, the virtual session will… | Coalition for Innovative Media Measurement (CIMM)
>> linkedin.com
>>  <https://www.linkedin.com/posts/cimm-is-hosting-its-next-research-seminar-share-7467605088182861824-1Abx/?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAAA6vMBieNPQNTL87LwTLuoLySm0oqEr7M>CIMM is hosting its next Research Seminar, Sharpening the Competitive Edge of TV's Ad Proposition, on Monday, June 8 from 2:00–3:30 PM ET. Open to the entire industry, the virtual session will… | Coalition for Innovative Media Measurement (CIMM) <https://www.linkedin.com/posts/cimm-is-hosting-its-next-research-seminar-share-7467605088182861824-1Abx/?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAAA6vMBieNPQNTL87LwTLuoLySm0oqEr7M>
>> linkedin.com <https://www.linkedin.com/posts/cimm-is-hosting-its-next-research-seminar-share-7467605088182861824-1Abx/?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAAA6vMBieNPQNTL87LwTLuoLySm0oqEr7M>
>> 
>>> On Jun 4, 2026, at 8:58 AM, Nathan Woodman <nate@proofindata.com <mailto:nate@proofindata.com>> wrote:
>>> 
>>> Thanks Chris - your point extends the case. If H3 is in or derived from the bid request, the same canonical bucket serves both bid-time reach/frequency management and post-bid attribution histograms.
>>> 
>>> Two requirements this puts on browser attribution:
>>> 
>>> Browser attribution needs to be extensible so it fits into external measurement frameworks (geo RCTs, MMM, panel).
>>> The attribution model's parameters need to be tunable so outputs are cross-validatable against external measurement.
>>> Standardized H3 × ISO-week × hour parameters align with both - it is the same dimension and on both sides of the transaction. 
>>> 
>>> Best, Nate
>>> 
>>> 
>>> 
>>> 
>>> On Mon, Jun 1, 2026 at 9:49 PM Chris Williams <chris_williams65@icloud.com <mailto:chris_williams65@icloud.com>> wrote:
>>>> Geo-temporal also takes care of reach and frequency if the Hex is included in the bid request.
>>>> The buying agent makes decisions on how many impressions (and which format the impressions are in) are required in each Hex to achieve Reach goals and manage frequency DURING the buy not measuring it after.
>>>> 
>>>> 
>>>> <1777905482258.png>
>>>> Beyond Programmatic: A Holistic Architecture for Agentic Media
>>>> linkedin.com
>>>>  <https://www.linkedin.com/pulse/beyond-programmatic-holistic-architecture-agentic-media-williams-3eb9f/>Beyond Programmatic: A Holistic Architecture for Agentic Media <https://www.linkedin.com/pulse/beyond-programmatic-holistic-architecture-agentic-media-williams-3eb9f/>
>>>> linkedin.com <https://www.linkedin.com/pulse/beyond-programmatic-holistic-architecture-agentic-media-williams-3eb9f/>
>>>> 
>>>> 
>>>>> On Jun 1, 2026, at 5:39 PM, Nathan Woodman <nate@proofindata.com <mailto:nate@proofindata.com>> wrote:
>>>>> 
>>>>> Hi Rick, Charlie, all,
>>>>> 
>>>>> I want to support Rick's argument for geo-based methods and engage with Charlie's acknowledgment that cross-site diversion is the unsolved gap, by suggesting a specific path forward: standardize the unit of analysis. Once the unit is standardized, the inputs and the experimental designs that consume it can remain open.
>>>>> 
>>>>> The structural insight from Rick's geo-RCT methodology (https://github.com/rickcentralcontrolcom/geo-rct-methodology) is that geography sidesteps the cross-site diversion problem entirely. Every operator - every publisher, every advertiser, every measurement vendor, every channel including CTV, out-of-home, audio, and in-store outside the browser, sees the same geographic units. No identity bridging required. The randomization unit is consistent across the ecosystem by construction.
>>>>> 
>>>>> But geography alone isn't enough. The unit of analysis is spatial × temporal, and both need to be standardized for cross-channel composition to actually work. A CTV ad airing Wednesday 8pm Eastern, an OOH impression at a billboard at 8:15pm, and a web conversion at 8:30pm only become composable across channels if they all resolve to the same canonical (cell, time-bin) buckets.
>>>>> 
>>>>> Specifically I'd recommend:
>>>>> 
>>>>> Spatial unit - Uber H3 hexagonal hierarchical grid, with R8 (~460m hex edge) as the default resolution. H3 has properties well-suited to advertising measurement:
>>>>> 
>>>>> Hexagonal cells are uniform in neighbor relationships (no diagonal-vs-orthogonal asymmetry that grids have).
>>>>> Hierarchical resolution (R0 through R15) supports country-scale down to ~5-meter-scale with clean parent/child rollup.
>>>>> Open-source, already widely adopted by geospatial vendors, ride-share, retail planning, and increasingly ad-tech.
>>>>> Does not depend on political or administrative boundaries — stable across countries, across time, across boundary changes. This matters substantively for longitudinal measurement and for cross-country interoperability.
>>>>> 
>>>>> Temporal unit — ISO 8601 week + hour-of-day in UTC, with local-time annotation for interpretability. The base time unit being deterministic and timezone-unambiguous is critical:
>>>>> 
>>>>> ISO week is internationally standardized and naturally hierarchical (year → quarter → month → week → day → hour).
>>>>> UTC sidesteps daylight-saving discontinuities that break clean bucketing.
>>>>> Hour-of-day UTC paired with a published local-time annotation lets analyses use clock-time semantics without DST corruption.
>>>>> Calendar context (religious calendars, school sessions, civic events) can be layered as optional metadata that any bucketing scheme can incorporate without changing the base unit.
>>>>> 
>>>>> Composite canonical bucket: (H3 cell at declared resolution) × (ISO week) × (hour UTC) — with the per-campaign histogramIndex resolving to this composite. Coarser buckets (cell × week, or cell × month) work for slower-cycle analyses like MMM; finer buckets (cell × week × hour) work for day-part-aware causal experiments.
>>>>> 
>>>>> Once the unit is standardized, four things become possible that aren't today:
>>>>> 
>>>>> Consistent bucketing across campaigns, operators, and measurement paradigms , observational attribution, causal geo experiments per Rick's methodology, MMM, retail-trade-area attribution all use the same units.
>>>>> 
>>>>> Browser histogram outputs become a shared input for external geo-time-aware models. A geo-RCT outside the browser can consume browser attribution histograms as one of its inputs, with no custom integration per partner.
>>>>> 
>>>>> Non-browser channels (CTV, OOH, audio, in-store) can geo-time-resolve impressions and conversions to the same units. The browser API becomes a contribution to cross-channel measurement rather than a silo within it.
>>>>> 
>>>>> Bidirectional integration: browser histograms feed external geo-time experiments; external experiments inform how histograms are bucketed for new campaigns. Browser attribution stops being a one-way endpoint.
>>>>> 
>>>>> Importantly, this proposal doesn't require changes to the API surface. saveImpression() and measureConversion() can carry composite H3 × ISO-week × hour identifiers as histogramIndex values where the campaign chooses; campaigns that prefer other bucketing dimensions can use them. The spec would just publish the convention so cross-operator and cross-channel composition becomes routine rather than a per-partner custom integration.
>>>>> 
>>>>> The benefit is structural: Attribution Level 1 stops being a siloed browser-only attribution layer and becomes the browser-edge contribution to a broader cross-channel measurement system. The privacy model - DP, MPC aggregation, per-site budgets - stays exactly as it is. The compositionality with geo-RCT methodologies, MMM, CTV measurement, OOH audience measurement, and retail-trade-area attribution changes substantially.
>>>>> 
>>>>> This also addresses the gap Charlie identified , cross-site diversion is solved when the randomization unit is geographic rather than user-identifier, and it does so without requiring the API or the privacy model to change.
>>>>> 
>>>>> I'd be happy to provide more detail on H3 hierarchical structure, ISO week conventions, or specific worked examples of how cross-channel composition would work mechanically. Happy to discuss further on or off list.
>>>>> 
>>>>> Best,
>>>>> 
>>>>> Nathan Woodman Founder & Principal Advisor, Proof In Data
>>>>> 
>>>> 
>>