Quasi-Experiments & Difference-in-Differences: Interview-Ready Causal Thinking When Randomisation Is Impossible

A food-delivery app adds a small platform fee in some cities, and order volumes dip the next week. Was it the fee, monsoon disruption, a competitor discount, or just payday timing? This is where quasi-experiments matter - they help managers make causal decisions when the business cannot pause reality to run a clean randomised test.

The Big Picture: Causality Is a Credibility Ladder

Randomised A/B tests sit near the top because random assignment creates comparable groups. But many real business questions - store openings, policy changes, price increases, credit-rule changes, city launches - happen without randomisation. Quasi-experiments are the middle layer: not perfect control, but far stronger than “sales changed after we acted.”

Core Explanation: What a Quasi-Experiment Is Really Doing

A causal question has two worlds: what happened after the action, and what would have happened without the action. The second world is the counterfactual - and because we cannot observe it directly, we design a comparison that approximates it.

Think of quasi-experiments as a family of “second-best but disciplined” causal designs. The business action was not randomised, but something about timing, geography, eligibility, rules, or cutoffs creates usable variation.

Difference-in-Differences: The One Formula You Must Be Able to Say

Difference-in-Differences, or DiD, is built for a common business situation: some units receive a treatment, others do not, and you have data before and after the change.

The formula is:

DiD effect = (Treated after - Treated before) - (Control after - Control before)

The first difference removes each group's own baseline. The second difference removes the broader market trend that affected both groups.

A Small Worked Example: Platform Fee and Order Conversion

Suppose a food-delivery platform introduces a new fee in City A but not yet in City B. We track weekly conversion rate: completed orders divided by app sessions.

A weak answer says, “conversion fell by 1.5 percentage points, so the fee caused a 1.5 point drop.” A stronger DiD answer says, “the market was already softening by 0.4 points in the control city, so the estimated incremental effect of the fee is a 1.1 percentage-point reduction.”

The DiD Assumption Stack: What Must Be True

DiD is powerful because it is simple. It is dangerous for the same reason. The estimate is only credible if the treated and control groups would have evolved similarly without the intervention.

Diagnostics: How to Check Whether Your DiD Is Credible

Interviewers like DiD because it tests whether you can think beyond formulas. A good analyst asks: “What would make my control group believable?” Use these checks.

Real Indian Example: RBI Rule Change and Digital Lending

Quasi-experiments are common in regulated Indian markets because firms cannot randomise regulation. For example, when the Reserve Bank of India restricts or permits a specific lending process, affected lenders experience a shock that competitors may not experience in the same way.

An analyst studying a digital lending restriction could compare the affected product line before and after the rule change against a similar unaffected product or peer set. The primary driver of identification is the regulatory timing; supporting drivers are comparable pre-period trends, similar customer segments, and careful controls for macro credit demand. The strategic “so what”: in finance interviews, DiD helps separate policy impact from broader market cycles.

Case Study: Zomato and the Platform Fee Question

Situation: Food-delivery platforms operate on thin per-order economics. A small platform fee can improve contribution per order, but it can also reduce conversion, frequency, or customer trust. Zomato's platform-fee changes, reported publicly from 2023 onward, were closely watched because they affected unit economics in a highly competitive Indian market.

The move: Instead of treating the fee as a simple before-after business event, an analyst can use a quasi-experimental lens. If the fee was exposed to different users, cities, or periods at different times, not-yet-treated markets can act as controls. The DiD question becomes: did treated markets change more than comparable control markets after the fee, over and above normal seasonality and demand movements?

The outcome or lesson: The value is not merely calculating one number. The primary driver of credible learning is staggered exposure that creates comparison. Supporting drivers include high-frequency transaction data, comparable city cohorts, pre-period trend checks, and controls for festivals, weather, discounts, and competitor promotions. The lesson: pricing experiments in consumer internet businesses must estimate both revenue lift and behavioural drag.

A shallow answer says, “fees increased revenue.” A complete answer says, “fees may increase per-order revenue, but DiD helps estimate whether the incremental revenue is offset by lower conversion or frequency after removing market-wide changes.”

How AI Changes Quasi-Experiments & Difference-in-Differences

AI does not remove the need for causal thinking. It makes quasi-experiments faster to design, easier to stress-test, and more dangerous if used blindly.

1. Better control-group search: ML can scan many candidate stores, cities, branches, or customer cohorts and suggest controls with similar pre-treatment patterns. The analyst still has to defend why they are comparable.
2. Event-study automation: AI-assisted analytics tools can generate pre-trend plots, placebo tests, and robustness checks quickly. This helps teams see whether the DiD assumption is plausible before presenting the result.
3. Unstructured confounder detection: LLMs can scan news, policy updates, earnings-call transcripts, app-release notes, or campaign calendars to identify concurrent shocks that could bias the estimate.

Interview Relevance

Common Mistake

What to Revise Next

Now move from one-time causal measurement to continuous learning. Revise Multi-Armed Bandits and Continuous Optimisation to understand adaptive experimentation, then Building an Experimentation Culture and a Test Backlog to learn how teams turn causal questions into a repeatable operating system.