New Method: Causal Inference with Machine Learning (Advanced Topics)
- Maria Alice Maia
- Jun 30, 2025
- 3 min read
Stop Asking "Did It Work?" Start Asking "WHO Did It Work For?"
I see it constantly. Companies spend millions on a new training program, a new marketing campaign, a new sales tool. Then they run a basic analysis, look at the average result, and get a lukewarm, useless answer. "Productivity went up 2%." "Sales saw a tiny bump."
This is kindergarten data analysis. And it's killing your ROI.
The problem? You're treating your employees, your customers, your teams like a monolithic blob. You're celebrating a meaningless average while the real, actionable insights are screaming, unheard, from within the data. You are "Doing Data Wrong" by settling for a single, blurry number when you should be demanding a high-resolution map of impact.
Let's get specific.
The "Wrong Way" in HR: An HR department launches an expensive leadership program for all mid-level managers. After a year, they compare the promotion rates of trained vs. untrained managers and find a small, disappointing lift. The CFO questions the budget. The C-suite is unimpressed. Was it a failure? Who knows. The data tells them nothing.
This is the swamp of naive comparisons. It's lazy, and it’s costly.
The "Right Way" with Causal Machine Learning (CIML): Forget averages. The real question is about heterogeneous treatment effects. It's a concept from the frontier of causal inference and machine learning, and it’s what separates high-impact data teams from the rest. The goal is to move beyond "what is the average effect" to "what is the effect for this specific type of person?"
A sharp data team doesn't just run a simple A/B test. They use powerful methods like Causal Forests—an evolution of the machine learning techniques many of you already use. These models tear through high-dimensional data (performance reviews, team structure, tenure, project history) to uncover the underlying patterns of cause-and-effect.
Back to our HR example: The CIML model doesn't give a single average. It gives you a crystal-clear segmentation of your people. It says:
"The training has a massive positive impact on promotion rates for managers with less than 2 years of experience leading small, fast-growing teams."
"It has zero, or even a negative, effect on senior managers in large, stable departments—they're bored and disengaged by the content."
This is no longer a blurry average. This is a strategic playbook. This is Optimal Policy Learning.
You don’t need a bigger budget; you need a smarter one. You can now target the training to the exact group who will benefit, maximizing its ROI. For the other group, you can design a more appropriate, advanced-level program. You stop wasting money and start making targeted, impactful investments.
What to do NOW:
Managers: Stop accepting average results. Start asking your data teams: "Can you show me the heterogeneity in this result? Which segments are driving the impact? What is the profile of the person for whom this intervention is most effective?" Demand a map, not a single number.
Tech & Data Professionals: Your job isn't just to build predictive models. The real challenge, the real value, is in building models that estimate causal impact. You need to master techniques that can handle the "fundamental problem of causal inference" and deliver insights that guide policy. This is how you go from being a service department to a strategic driver of the business.
This isn't tech-for-tech's-sake. This is about connecting advanced methods directly to business profit and loss. It's about having the courage to admit that "one-size-fits-all" is a recipe for failure.
My knowledge is not mine to keep. It’s for us to use, to build better, smarter, more human-centric businesses. Let's fix our broken data practices, together.
Passionate about this? Join my email list for more no-nonsense, research-backed insights to unlock real data value. Let's build a community dedicated to getting this right. If you have a real-world case you want to crack, schedule a 20-minute, no-nonsense consultation call with me.
