Causality and the Counterfactual

This unbridgeable gap is the Fundamental Problem of Causal Inference. This chapter introduces the standard mathematical language used to reason about it — the Potential Outcomes Framework — and shows precisely why a naive comparison of treated and untreated units overstates causal effects.

The Executive Question: What Is the True Causal Lift?

Suppose a marketing dashboard reports that customers who received a discount coupon spent more on average than customers who did not. It is tempting to subtract the two averages and call the difference "the coupon's lift." The framing is wrong.

The right question is not do recipients spend more than non-recipients? — that is descriptive arithmetic. The right question is would the recipients have spent less if they had not received the coupon? Figure 1 illustrates the branching structure of this dilemma: for any single unit at any single moment, we observe one of two possible paths and never both.

The Potential Outcomes Framework

Let $i$ index a unit (a customer, a store-week, a state-month) and let $D_i \in \{0,1\}$ be a binary treatment indicator: $D_i = 1$ if unit $i$ received the treatment, $D_i = 0$ otherwise.

For every unit, define two potential outcomes:

• $Y_i(1)$ — the outcome unit $i$ would experience under treatment.
• $Y_i(0)$ — the outcome unit $i$ would experience under no treatment.

The individual treatment effect is the difference between these two states:

The Fundamental Problem of Causal Inference is that we only observe the realized outcome, which depends on whether unit $i$ was actually treated:

If unit $i$ was treated, we see $Y_i(1)$ and $Y_i(0)$ is missing. If untreated, we see $Y_i(0)$ and $Y_i(1)$ is missing. The individual effect $\tau_i$ is therefore never observable for any single unit.

Because we cannot recover individual effects, decision-makers focus on averages over units. The Average Treatment Effect is

and the Average Treatment Effect on the Treated narrows the average to units that actually received the treatment:

ATE and ATT generally differ. If your treatment is opt-in — coupons that loyal customers self-select into, software features that power users enable — the treated population is not a random draw from the full population, and the ATT will reflect the response of the kind of unit that takes the treatment.

Why Naive Comparisons Fail

If we cannot see individual counterfactuals, can we just compare group averages? The naive difference is

Substitute the observed outcomes with potential outcomes (using $Y_i = Y_i(1)$ when $D_i = 1$ and $Y_i = Y_i(0)$ when $D_i = 0$) and the difference splits cleanly into two terms:

This is one of the most important equations in applied business analytics. It says: the naive treated-minus-untreated difference equals the true ATT plus selection bias, where selection bias is the difference in untreated-state potential outcomes between the two groups.

Selection bias is zero only when the two groups would have looked identical in the absence of treatment. That is rarely true in observational data, because units choose, or are chosen for, treatment based on factors that also drive the outcome. Loyal customers opt into coupons; high-margin stores adopt new features first; states that legalize a policy differ in unobserved ways from states that do not.

A numerical illustration

Suppose loyal customers self-select into a coupon program. We observe an average spend of $\$15.50$ in the coupon group and $\$9.20$ in the non-coupon group, a naive lift of $\$6.30$. The decomposition above warns us that this $\$6.30$ is

The bias term is plausibly large and positive: the customers who took the coupon were the kind of customer who would have spent more even without it. Without a design that controls selection, we have no way to split the $\$6.30$ into its two components.

The Four Standard Counterfactual Designs

The rest of Part III is, essentially, four families of methods for constructing a credible stand-in for $\mathbb{E}[Y_i(0) \mid D_i = 1]$ — the missing untreated potential outcome of the treated group.

Each method buys credibility at the cost of an assumption you can name and defend. The Decision Question Card from the previous article forces you to choose which counterfactual you are willing to defend before you choose a method.

Data Case: Colorado Housing and the Synthetic Counterfactual

To see counterfactual construction in action with real data, consider Colorado's January 2014 legalization of recreational cannabis. A real-estate investor wants to know whether the policy moved housing values. The unit of analysis is the state-month and the outcome is the Zillow Home Value Index.

The Fundamental Problem applies at the state level: we observe Colorado after legalization, but we never observe the Colorado that would have existed had legalization not happened. A naive before-after comparison conflates the policy with the broader US housing recovery that was already underway in 2014.

The synthetic control method constructs a credible counterfactual by searching for a weighted combination of donor states — states where the policy did not change — whose weighted housing trajectory closely matches Colorado's pre-treatment path. Figure 2 shows the actual and synthetic paths; Figure 3 shows the donor weights.

The post-2014 gap between Colorado and the synthetic counterfactual averages roughly twenty percent. That is the design's estimate of the causal effect of legalization on Colorado housing values, under the assumption that no other Colorado-specific shock after 2014 could have produced the gap. We will return to this design in Chapter 7 and stress-test that assumption with placebo donor states.

For now, treat this case as a worked example of the framework: an action, a unit, an outcome, and a counterfactual that was built — not assumed — from the data.