Connecting TRIBE v2 to Triplewhale: an end-to-end research module
Can predicted brain activation tell you anything about real ad performance? An honest research-only walkthrough — six pipeline steps, eight testable hypotheses, and the statistical and licensing landmines to dodge.
Research only. Not commercially deployable.
TRIBE v2 is CC-BY-NC 4.0. You can run this pipeline against your own ad library and attribution data for academic exploration, learning, or publication. You cannot use the output to inform paid creative, client work, or any commercial decision — even your own brand's. Triplewhale's data also has terms of service; consult them before redistributing or republishing any extracted metrics.
CC-BY-NC 4.0 license textWhat this module is asking
Imagine you have two pieces of evidence about an ad. One is a guess of what someone's brain did while watching it (TRIBE). The other is what people actually did — clicks, purchases, refunds (Triplewhale).
A natural question: do they line up? When the brain model says 'this ad recruits the part of the brain that processes faces a lot,' do those ads actually convert better? The honest answer is 'probably sometimes, in narrow ways, with caveats.'
This module walks through how you'd answer that question carefully — not to deploy something commercial, but to understand what these tools can and can't tell us.
In one line: a research notebook, end-to-end. Six steps from raw ads to a publishable finding — or a publishable null result, which is often more useful.
min ads to detect r=0.3 at 80% power
min for r=0.2 at 80% power
ROI × metric tests — FDR is mandatory
commercial deployments permitted
Six steps. From ads in your library to a publishable finding.
Click through the pipeline. Each step describes what's happening, the data shape it produces, and the gotcha that most often torpedoes this kind of cross-domain analysis. Implementation details are intentionally left to the researcher.
You need every ad you'll analyze in two places at once: the actual video file (for TRIBE) and a stable identifier that maps to Triplewhale's ad-level metrics. Meta's ad_id is the cleanest join key — it survives across both sides.
Per-ad row
- ad_id (Meta) — primary key
- video_path (local mp4 for TRIBE)
- campaign_id, adset_id (for grouping)
- first_run_date (filters out fresh / fatigued ads later)
Don't analyze ads with under ~14 days of run-time. Triplewhale metrics for fresh ads are noisy; TRIBE will give you predictions either way, but you'll be correlating noise.
Eight testable research questions.
Each card is a real research question you could answer with this pipeline. Tier reflects how likely the question is to produce a clean finding given sample-size realities and confound landscape — start tractable.
Statistical caveats
Five places this analysis goes wrong
Cross-domain correlational work sounds simple and is hard. These five caveats account for most of the gap between 'we found something' and 'we found something real'.
Where the line is.
- Running the pipeline against your account for personal learning
- Publishing methods + findings (or null findings) openly
- Open-sourcing notebooks under CC-BY-NC
- Citing in academic / conference contexts
- Feeding insights into your in-house creative process
- Selling reports / dashboards based on TRIBE predictions
- Wrapping the pipeline in a SaaS for clients
- Quietly using findings to direct creative at your day job
The playbook
Eight rules for honest research
your team's coverage
Sources
What we read to build this
The science is fascinating. The license is hard.
If you're researching this academically, the pipeline above is yours. If you're making creative decisions for commercial brands, Shuttergen runs the structural-variation playbook on a behavioral-signal loop — without the license risk and without the statistical landmines.
Get started free