marketing-attribution
Attribution and incrementality measurement: attribution models, marketing mix modelling (MMM), multi-touch attribution (MTA), incrementality testing, conversion lift, holdout tests, Shapley values, Ro
SKILL.mdskills/marketing-attribution/SKILL.md
GitHubmarketing-attribution
When to use this skill
Activate when the user asks about: attribution models, marketing mix modeling (MMM), multi-touch attribution (MTA), incrementality testing, conversion lift, holdout tests, Shapley values, Robyn, LightweightMMM, iROAS, UTM governance, walled garden measurement, view-through vs click-through attribution, budget optimization by channel, causal inference for marketing, difference-in-differences.
Core behavior
- Always clarify which attribution question is being asked: "which channel gets credit?" vs "what is the true incremental impact?"
- Explain the limitations of each model in context (last-click undervalues upper funnel; MMM requires 2+ years of data)
- Recommend the right tool for the objective: MTA for tactical optimization, MMM for strategic budget allocation, incrementality for ground truth
- Flag walled garden data discrepancies (Meta vs Google vs internal)
- Express results as iROAS (incremental ROAS) not just attributed ROAS when possible
Model selection guide
| Question | Recommended model |
|---|---|
| Which channel to cut budget? | MMM |
| Which keyword drives conversions? | Data-driven MTA |
| Did this campaign actually work? | Geo holdout / Conversion Lift |
| How to split credit across touchpoints? | Shapley MTA |
| Is my Google Ads ROAS real? | Incrementality test |
Key patterns
Shapley value MTA (Python)
from itertools import combinations
def shapley_values(channels, conversion_fn):
n = len(channels)
shapley = {c: 0 for c in channels}
for channel in channels:
others = [c for c in channels if c != channel]
for r in range(len(others) + 1):
for subset in combinations(others, r):
weight = 1 / (n * len(list(combinations(others, r))) )
marginal = conversion_fn(list(subset) + [channel]) - conversion_fn(list(subset))
shapley[channel] += weight * marginal
return shapley
Difference-in-Differences (DiD)
import statsmodels.formula.api as smf
# did = (treated_post - treated_pre) - (control_post - control_pre)
model = smf.ols('conversions ~ treated * post_period + C(region)', data=df).fit()
did_estimate = model.params['treated:post_period']
UTM taxonomy for WPPMedia clients
utm_source: google / meta / dv360 / ttd / email / direct
utm_medium: cpc / paid_social / display / video / organic
utm_campaign: {client}_{objective}_{year_month}
utm_content: {creative_id}_{format}
utm_term: {keyword} (Google Ads only)
Output formats
- Model comparison table with pros/cons for the specific use case
- Python or R code for statistical models
- Budget reallocation recommendation based on iROAS by channel
Tools
Python (Robyn, LightweightMMM, PyMC, pandas, sklearn), R, BigQuery, Google Analytics, Meta Business Suite