MMMs vs causal inference for marketing attribution

Marketing mix models are an old way to figure out how marketing stuff like ads and promos affect sales. They use statistics to look at sales data and marketing data over a long time. Then they try to find patterns that show how the marketing caused the sales.

These models have been used by big companies for a real long time, like 50 years or more. It’s a pretty standard thing in marketing. A bunch of smart folks with PhDs make these models with fancy math.

But there are some problems with marketing mix models:

• They assume the patterns in the past will keep happening in the future
• They have a hard time proving the marketing actually caused the sales
• They can get messed up by other stuff changing in the market

So yeah, they’re not perfect. But they were the best we had for a while. Companies spent a ton of cash on them.

New hotness: Causal inference techniques

Okay, so in the past 20 years or so, some new ideas came out for figuring out causation from data. The big ones are:

Difference-in-differences (DiD)

DiD looks at how something changes in two groups over time – a treatment group that got marketed to, and a control group that didn’t. By comparing the diff in the diffs between the groups, it backs out the impact of just the marketing.

Instrumental variables (IV)

IV uses a special “instrument” variable that affects sales only through the marketing variable. This isolates the causal effect of the marketing. It filters out other noise that messes up the analysis.

These new techniques are dope because:

1. They directly estimate causal effects, not just correlations
2. They adjust for time trends and market changes
3. They’re easier to explain than some crazy regression model
4. You can use them on just a couple marketing campaigns – don’t need years of data

In head-to-head tests, the causal inference joints beat the pants off marketing mix models for figuring out the real ROI of marketing. Mix models tend to give all the credit to the biggest, longest-running campaigns. DiD and IV can show the punch of scrappy upstart campaigns.

Why are companies still using marketing mix models?

Aight, so if the new causal stuff is so great, why are companies still paying for crusty marketing mix models? Few reasons:

1. Switching costs. They’ve built whole teams and processes around mix.
2. Familiarity. Execs trust what they know. Mix models feel safe.
3. Sunk costs. Hard to admit you wasted a decade budget on an inferior approach.
4. Inertia. Change is hard. Gotta get lots of folks on board to pivot.

But the change is coming. More and more firms are checking out DiD and IV. The results speak for themselves. You can’t argue with cold, hard, causal effects.

Recommendations

Marketers need to get with the times. If you want real talk on what’s driving sales, you gotta get down with causal inference. My suggestions:

• Start with a small pilot. Pick a couple campaigns and analyze them with DiD or IV. Compare that to your usual mix model. See which has more insight.
• Get your data in shape. You need clean data on who got marketed to. Put in place a dope data pipeline.
• Train the team. Send your smart folks to causal inference bootcamps. Have them read papers and run tests.
• Educate the bosses. Show them the results. Explain it in plain talk. Get them hyped on causality.
• Phase out mix models. Rip off the bandaid. Might sting at first but it’s better in the long run.

Following these steps will get you an attribution system that actually knows what’s up. No more mix model black boxes. Just real, causal, money-making insights.