Core Concept: Difference-in-Differences (DiD) - Basic Setup & Parallel Trends Assumption
- Maria Alice Maia
- Jan 7
- 3 min read
Your firm rolled out a new project management tool in the Strategy department. You compared their change in productivity to the Operations department and found a 10% lift. A clear win for the new tool, right?
Not so fast. What if the Strategy unit was already on a faster improvement trajectory?
Welcome to the world of Difference-in-Differences (DiD), one of the most powerful and commonly used methods for causal inference—and one of the most easily botched.
The logic of a simple DiD is beautiful. You have two groups (Treated, Control) and two periods (Before, After). To find the treatment's effect, you calculate:
(Treated After - Treated Before) - (Control After - Control Before)
You use the change in the control group to proxy for the trend that would have happened anyway, and you subtract it out. It's an elegant way to account for time-varying confounders like market shifts or seasonality that we discussed last week.
But this entire method rests on one critical, untestable assumption: The Parallel Trends Assumption.
You must assume that, in the absence of the treatment, your treated group would have followed the exact same trend as your control group.
The "Doing Data Wrong" Scenario: Assuming Parallel Trends
Let’s go back to our Consulting Firm example.
The Analysis: The firm's analyst calculates the DiD and finds the new software led to a 10% increase in project profitability for the Strategy unit compared to the Operations unit.
The Hidden Flaw: They never checked the historical data. What if, for the three years before the new software, the Strategy unit's profitability was already growing 5% faster than the Operations unit's each year? If so, at least half of the "effect" they measured was just a pre-existing difference in trends that had nothing to do with the software. The parallel trends assumption was violated, and the result is biased.
The Right Way: You Must Defend the Assumption
You can never prove parallel trends hold in the post-treatment period (that's the counterfactual), but you absolutely must provide evidence that they were holding before the treatment. This is non-negotiable.
The Action: Plot the outcome for both groups for multiple pre-treatment periods. Do the lines move in parallel? If they are diverging or converging before the treatment begins, your core assumption is weak, and a simple DiD is not the right tool.
The Litmus Test: As the recent literature emphasizes, this "pre-trends test" is a crucial diagnostic for the credibility of any DiD study.
My time as both a consultant at FALCONI and a leader driving strategy at companies like Itaú and Alura taught me this: no two business units are ever identical. You cannot simply assume their trends are parallel. You must show the data. A DiD analysis without a pre-trends plot is, in my view, an incomplete and untrustworthy analysis.
My mission is to help you move beyond just running the model to deeply understanding and stress-testing the assumptions that make it work. This knowledge isn't mine to keep.
If you’re ready to build more credible, defensible causal insights, join my movement. Subscribe to my email list.
And if you're working on a DiD analysis and have that nagging feeling about your control group, book a 20-minute, no-nonsense consultation with me. Let's look at the trends together.
