Regression Foundations

Interpreting Program Impact

A reading guide for lecture p-05 — why a program is a bet, and why statistical significance answers a much narrower question than people assume.

Section one

Key concepts

The definitions and decision rules the slides depend on, each stated as a claim.

Key concepts

A program is a bet. The coefficient is the expected value; the interval is the risk.

  • b1 = the point estimate = our best guess of the payoff
  • the confidence interval = the range of plausible outcomes
  • the cost of the program = the price of placing the bet

Choosing between programs is choosing between bets, and that always balances expected payoff against uncertainty.

Statistical significance tells you one thing only: whether the sign is certain.

Statistical significance tells us if we can be certain about the direction of our program effects (the confidence interval does NOT contain zero).

  • CI excludes zero → “significant” → all plausible outcomes have the same sign
  • CI contains zero → “not significant” → you cannot rule out the opposite sign

It says nothing about how large the effect is, or whether the program is worth buying.

The p-value is the largest confidence interval you can draw before it contains the null.

  • p = 0.03 → the 97% interval just touches zero
  • A 95% interval might exclude zero while the 99% interval from the same model contains it

Significance is therefore a function of the threshold you picked, not a property of the finding.

Anatomy of a coefficient plot.

  • the vertical line at b1 = 0 = the null hypothesis (no impact)
  • the dot = the slope estimate (best guess)
  • the whiskers = the confidence interval
  • not significant if the interval crosses the line

Estimates should be unbiased and efficient.

  • Unbiased = accurate. Estimates cluster around the true value.
  • Efficient = precise. Estimates cluster tightly.

A model can be accurate but imprecise — right on average, but any single study lands far off.

Power is the ability to detect an effect that is really there.

  • Type I error — claiming impact when there is none (false positive). Usually caused by bias.
  • Type II error — missing impact that exists (false negative). Usually caused by low power, i.e. large standard errors.

Confidence has an increasing marginal cost.

IntervalCapturesGain
±1.0 SD68.2%
±1.5 SD86.6%+18.4
±2.0 SD95.4%+8.8
±2.5 SD98.8%+3.4
±3.0 SD99.8%+1.0

Going from 95% to 99.9% confidence costs far more interval width than the first steps did. The t-value for a 95% two-sided interval is 1.96.

Section two

Key relationships

The mental map. How the pieces connect, and how to read the deck.

The bet is the organizing metaphor, and it is doing real work

Most of this deck is not about regression at all — it is about four bets. That is deliberate, and the mapping is exact:

In the betIn the regression
price of the betcost of the program
expected valuethe coefficient b1
range of outcomesthe confidence interval
chance of losing moneythe interval crossing zero
rounds of playsample size

Once the mapping is in place, questions that sound technical become questions you already know how to answer. Would you rather have a bet with a $1,400 expected value and no chance of loss, or one with a $2,500 expected value and a one-in-four chance of losing $2,000? There is no universally right answer — and that is the point the deck is making about programs.

Note the aside on slide 8, which is more careful than it needs to be: with a single two-outcome bet the expected value is theoretical and unachievable. It only becomes the likely outcome across many plays — which is exactly the relationship between one sample and the sampling distribution. Rounds of play are sample size.

Significance answers a narrower question than people think

This is the argument the deck builds across slides 14 through 19, and it is worth stating bluntly:

“Statistically significant” means the confidence interval does not contain zero. That is all it means.

It does not mean the effect is large. It does not mean the program is worthwhile. It does not mean the result is important. It means every plausible value has the same sign, so you can be confident about the direction of the effect.

The deck proves this by constructing cases where significance and value point in opposite directions. Bet #2 has the higher expected value ($2,500) but is “not significant” because the range spans zero. Bet #4 has a much lower expected value ($1,050) but is “significant” because every outcome is positive. If significance were the decision rule, you would take the worse bet.

Hence the deck’s own warning: statistical significance should NOT be the first or only piece of information to consider.

The p-value as a dial, not a verdict

Slides 36 to 38 give a definition of the p-value that is more useful than the textbook one, and it follows directly from the coefficient plot.

Take one estimate and start widening the interval. At 95% confidence it might clear zero. At 99% confidence — same data, same model — it might not. Somewhere in between there is a confidence level at which the interval just touches zero. That level is the p-value:

The p-value tells you how large you can draw your confidence interval before it contains the null.

A p of 0.03 means the 97% interval is the widest one that still excludes zero.

Two consequences worth holding onto. First, the same estimate is “significant” or not depending on the threshold you chose in advance — nothing about the finding changed between slides 36 and 37. Second, the p-value is continuous information, and collapsing it to a yes/no at 0.05 throws most of it away.

Effect size and precision are two independent axes

The dartboard slide makes this concrete. An evaluation can be:

Bias and precision are different failures with different causes, and the deck maps them onto the two error types:

The low-power slides make the second case vivid: the true slope is 3, the model is unbiased, the estimates cluster around 3 — and you still fail to reject the null, over and over, because the intervals are too wide. Being right is not enough if you cannot demonstrate it.

Why more confidence gets expensive

Slides 61 through 65 answer a question students rarely ask: why settle for 95%?

Because confidence is bought with interval width, and the exchange rate gets steadily worse. Going from 1.5 to 2.0 standard deviations buys 8.8 points of confidence. The next half-step buys 3.4. The one after that buys 1.0. Near 100%, the interval has to grow without bound.

That is the real reason for conventional thresholds: past a point, extra confidence costs more precision than it is worth. It also links back to the standard error — the only way to get a narrower interval at the same confidence level is to shrink SEb1, which returns you to the three levers from p-03.

Coefficient plots versus regression tables

Slide 55 sets a dense table of coefficients and standard errors against a coefficient plot of the same results. The point is presentational but not trivial: the plot shows effect size and uncertainty simultaneously and in the units of the outcome, while the table forces the reader to do arithmetic in their head to figure out whether an interval covers zero.

Given the deck’s argument — that significance alone is a poor summary — this is consistent. A coefficient plot makes it hard to report only the asterisks.

Looking ahead

Slide 34 flags what comes next: adding control variables changes both the coefficient and the standard error, which means it can change your answer about whether a program works, and in which direction. Everything in this deck is about interpreting a plot; p-06 onward is about whether you have specified the right model in the first place.

How to read the slides

SlidesWhat they are doingWhat to take away
2“Which model is the right one?”The question the whole unit is building toward.
4–13The four betsExpected value vs. range. Build the mapping to regression here.
8, 10The aside on repeated playRounds of play = sample size. This is the CLT again.
14–19Significance vs. valueThe core argument. Significance only fixes the sign.
20, 27Anatomy of a coefficient plotNull line, point estimate, interval.
21–22, 28–33Program comparisonsPractice reading plots. Ask both questions each time.
23–26Effect sizesCaffeine and potato chips: significant and not.
33DartboardsAccurate vs. precise.
34Looking aheadControls move coefficients and standard errors.
35–40What is a p-value?The widest interval that still excludes the null.
43–48Mechanics: t-valuesChoose confidence → get t. 1.96 for 95%.
49R outputA t-test as the two-group version of the same idea.
50–53How often are we wrong?Alpha is the miss rate. 90% intervals are narrower than 95%.
54Unbiased and efficientThe two things an estimate should be.
55Plots vs. tablesPresentation follows from the argument.
56The checklistSix items.
57–66Power and the cost of confidenceType I/II errors; diminishing returns on confidence.

The repeated “which program is better? / what about now?” slides are an exercise, not repetition. Each variation changes one thing — the width of an interval, the position of a point estimate, the confidence level — and asks you to notice that significance and magnitude move independently.

Section three

What should be clear in my mind?

The deck’s own six-item checklist.

  1. Interpreting program impact means understanding both the effect size (the slope) and the precision with which it is estimated (the interval). Neither alone is an answer.
  2. The level of confidence you select determines the t-value, which determines the size of the interval. The width is a choice, not a discovery.
  3. For a program to be statistically significant, the interval around the slope must not contain the null (slope = 0).
  4. You can choose an arbitrary level of confidence such that the interval will not contain the null. Which is precisely why significance is a weak criterion.
  5. The p-value tells you the largest confidence interval you can draw that does not contain the null.
  6. Program investments are bets that balance effect size against confidence.

Key takeaways