Research Design

Testing the Validity of the Counterfactual

A reading guide for lecture p-02 — the selection problem, the two tests every study should report, and why the Bonferroni correction changes your decision rule.

Section one

Key concepts

The selection problem, the balance tests, and the correction that makes them honest.

Key concepts

Nature gives us correlations, not experiments.

The deck opens with the city-scaling example from Radiolab. Walking speed correlates with patents, crime, and salary. So does population.

  • If you increased the walking speed of a city, would you get more patents?
  • Obviously not — population drives both.

Four ways a correlation can appear without X causing Y:

  1. X and Z both affect Y (no correlation between X and Z)
  2. Z affects both X and Y — high correlation, no causal relationship
  3. Causal chain X → Z → Y — no direct relationship, high correlation
  4. Reverse causalityY = b0 + b1X + e and X = b0 + b1Y + e are both highly significant. Nothing in the regression output tells you which direction is causal.

The counterfactual is valid when groups differ *only* by the treatment.

Stated as the deck states it: the framework is valid and robust when the groups differ by the treatment and are otherwise identical. When that holds, post-treatment differences in outcomes can be read as caused by the treatment. Everything else in this lecture is machinery for checking that condition.

Selection is the biggest problem in impact evaluation.

Those who participate in a program are different from those who do not. The microfinance example makes it concrete. Suppose loans have zero effect on income, and:

  Not entrepreneurial Entrepreneurial
No loan 30 people @ $10 15 people @ $20
Takes a loan 20 people @ $10 35 people @ $20
  • No loan: (30·$10 + 15·$20) / 45 = $13.33
  • Loan: (20·$10 + 35·$20) / 55 = $16.37

The loan appears to work — and we built the example so that it does nothing. People who know they are good at business are more likely to borrow. That is selection, not impact.

The estimator you choose decides whether selection fools you.

Same data, two estimators:

  • Post-test only (T2 − C2) — requires the groups to be balanced before the intervention. In the microfinance example they are not, so it reports a false effect.
  • Difference-in-difference ([T2 − T1] − [C2 − C1]) — the pre-treatment imbalance cancels out, so it correctly reports no impact. The cost is that it needs pre-treatment data, which is often the thing you do not have.

More robust, more data-intensive. That trade recurs for the rest of the course.

Selection also happens on the way *out* — attrition.

Reflexive design, average household income $2.00 before and after a program that does nothing:

  • Random attrition — people drop out at every income level; mean stays $2.00; T2 − T1 = 0. The study still gets the right answer.
  • Non-random attrition (low earners leave) — mean rises to $2.50; the program looks effective.
  • Non-random attrition (high earners leave) — mean falls to $1.50; the program looks harmful.

Attrition is natural. The question is never whether it happened but whether it was random.

Randomization "fails" 5 times in 100, by construction.

If you randomly split 100 people into two groups of 50, how often do the average weights differ? Mathematically, always — they are never exactly equal. So “different” has to mean statistically different, which means a t-test at some α.

At α = 0.05, each measured characteristic will read as significantly different 5 times out of 100 even when both groups came from the same population. The deck calls this unhappy randomization — not a failure of the process, just a bad draw.

The Bonferroni correction fixes the multiple-comparison problem.

The balance table reports many contrasts, and each gets its own 5% chance of a false alarm. With n contrasts the chance of at least one p-value below 0.05 is roughly n × 0.05, not 0.05.

So to stay 95% confident about the groups rather than about each contrast, divide:

  • 10 contrasts → new decision rule is 0.05 / 10 = 0.005
  • 7 contrasts → new α = 0.05 / 7 = 0.0071; a smallest p-value of 0.04 is 0.04 > 0.0071, so do not reject — the groups are equivalent

Section two

Key relationships

The mental map. Selection in, selection out, and which estimator survives each.

The lecture has two halves that mirror each other

Selection into the study group (who joins) and selection out of it (who leaves) are the same problem at two different moments, and they get the same fix: compare the traits of the groups you actually have, and test whether the differences are larger than chance would produce.

The balance table is the most important table in any study

The deck says this outright. Before you read a single result, read the table comparing treatment and control characteristics. For the counterfactual to hold, the groups can only differ by the treatment — so this table is where a study either earns its causal language or does not.

Two things to check: which traits were measured (unmeasured traits cannot be balanced, only assumed), and whether the decision rule was corrected for the number of contrasts.

The four attrition tests, and the two that get reported

The deck’s attrition diagram splits each group into stayers and leavers, producing four possible tests:

Most studies report only 0 and 1. That is defensible — test 1 is agnostic about the type of attrition and only asks whether balance survived it — but it is weaker than test 2.

The deck is careful about a nuance worth keeping: attrition can be non-random and still harmless if it is non-random in the same way in both groups. If all high performers leave both arms, the groups stay balanced. Test 3 checks exactly that.

All four tests are the same regression

This is the practical payoff of the lecture. Subset the data to isolate the two groups you want to compare, build a dummy for one of them, and run:

Y = b0 + b1(dummy) + e

A significant b1 means the groups are not balanced (test 1), attrition is non-random (test 2), or attrition is non-equivalent (test 3). Y can be the pre-treatment outcome or any participant trait. It is the same tool you used in Lab 2, pointed at different subsets.

RCTs, natural experiments, and quasi-experiments arrive at balance differently

All three are trying to satisfy the same condition. Only the mechanism differs.

How to read the slides

Slides 3–19 are the correlation-is-not-causation setup and can be skimmed if the first course is fresh. Slides 21–29 are the two worked examples and are the heart of the lecture. Slides 34–43 are the statistical argument about balance testing, and slide 42 (Bonferroni) is the one you will actually need to apply in the lab.

Section three

What should be clear in my mind?

The checklist to run against any study's balance table.

What the deck asks you to be able to do

Key takeaways