Standard Error of the Slope

Section one

Key concepts

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

Key concepts

This is the one formula to remember.

SE_b₁ = sd_e√(n − 1) · var(x)

Or, stripped to its moving parts:

Standard Error of the Slope ≈ residualsample size · variance X

The deck opens with this and then works backwards to show where it comes from.

The derivation is three steps.

Start from the version that falls out of the variance partition:

SE_b₁ = √σ²_ε∑(x_i − x)²
Note: var(x) = ∑(x_i − x)²n − 1 ⇒ (n − 1) · var(x) = ∑(x_i − x)²
Thus: SE_b₁ = √σ²_ε(n − 1) · var(x)

The only thing that happened is that the sum of squares in the denominator was renamed as sample size times variance. Nothing was approximated.

The standard error is a sampling statistic.

Draw a sample, fit a line, record the slope. Repeat.
The spread of those slopes is the sampling distribution of the slope.
The standard error is the typical distance from one estimate to the true slope — the “average error” of the slope estimate.
Larger samples make that distribution narrower.

Standard deviation and standard error answer different questions.

Standard deviation: how far the data are from the mean, on average.
Standard error: how far our best guess is from the truth, on average.

The mean and the slope draw their noise from different variables.

	SE of the mean	SE of the slope
Formula	SE_x = s_x√n	SE_b₁ = s_ε√(n − 1) · var(x)
Source of variance	X	Y (the residual)
To reduce it	increase sample size	(1) increase sample size (2) explain more variance of Y (add controls) (3) increase variance of X

Section two

Key relationships

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

The deck is built backwards on purpose

Slide 2 states the destination before any derivation has happened: “We want to end up here. Need to work backwards. This is the one formula you need to remember.”

That framing matters. The middle of this deck is algebra, and it is easy to lose the thread and assume the algebra is the point. It is not. The point is that three named quantities — the residual, the sample size, and the variance of X — are the only things that determine how precise your estimate is. The derivation exists to prove that the list is complete, not to be memorized.

If you read nothing else, read slide 2 and slide 15.

The derivation is a renaming, not a discovery

The step that does the work is trivial once you see it. The variance of X is ∑(x_i − x)² divided by n − 1, so multiplying both sides by n − 1 says that the sum of squared deviations in X is sample size times variance. Substituting that into the denominator turns an opaque sum into two things you can actually reason about and control.

This is worth naming as a habit: when a formula is hard to interpret, look for a factor that can be split into a count and a per-unit quantity. The interpretability comes from the split, not from new mathematics.

X and Y play opposite roles

This is the sharpest idea in the deck, and it is easy to skim past on slide 15.

Variance in Y is in the numerator. Noise in the outcome makes your estimate worse.
Variance in X is in the denominator. Spread in the predictor makes your estimate better.

For the standard error of a mean, the variance of the variable itself sits in the numerator — more spread is simply bad. For a slope, that intuition reverses for the explanatory variable. A study where everyone got nearly the same dose has almost no information about what dose does, however many people were in it.

So “more variance” is not good or bad in the abstract. It depends entirely on which variable it is in: unexplained variance in the outcome hurts precision; variance in the treatment helps it.

Three levers, and what each one costs

Slide 15 lists the three ways to shrink a standard error. They are not equally easy, and it is worth knowing what each one asks of you:

Increase the sample size. Always available in principle, expensive in practice, and subject to diminishing returns — the square root means quadrupling n only halves the standard error.
Explain more variance of Y (add controls). A modelling move, and the one that connects this lecture to p-06 and p-09. Free in data-collection terms, but it changes what your coefficient means.
Increase the variance of X. A design move, decided before any data exist. Once collected, you cannot manufacture spread in your treatment.

The ordering is instructive: the cheapest lever to pull after the fact is the second one, which is exactly why control variables get so much attention later in the course.

The simulations are the argument that this is a sampling statistic

Slides 11 through 14 do not add a formula. They show what the formula is measuring.

Each frame draws a fresh sample from the same population, fits a line, and records the slope. Your real study is one of those frames — the one labelled “Our Sample.” The scatter of all the others is the sampling distribution, and the standard error is its typical width.

The pair of simulations at n = 10 and n = 50 is the payoff. Same population, same true slope, and the only change is sample size — the cloud of estimated slopes visibly tightens. That is the square root doing its work, seen rather than derived.

This is also the honest picture of what an estimate is. Your line is not the true line; it is one draw. The standard error is how far off a draw typically lands.

Where this sits in the sequence

p-02 found the standard error hiding in the unexplained half of the variance. This deck cleans it up into a memorable form and shows what it measures. p-04 takes it and builds the confidence interval — the last rung on the road map.

Note that the road map slide now labels the top row sampling variance rather than just variance, which is the vocabulary from the sample-vs-sampling distribution slides. By this point in the course the top of that ladder is explicitly about estimates, not data.

How to read the slides

Slides	What they are doing	What to take away
1–2	The destination, stated first	The one formula. Come back here when the algebra gets thick.
3–5	Variance and the partition, recapped	Nothing new — this is p-02 compressed into three slides.
6–8	The standard error in regression	Two equivalent forms; the second is the interpretable one.
9	The derivation	Note → Thus → Therefore. One substitution, done carefully.
10	The intuitive form	Residual over sample size times variance of X.
11–13	Sampling simulations, n=10 then n=50	Watch the sampling distribution tighten as n grows.
14	The sampling distribution named	Your sample is one draw; the standard error is the average error.
15	Translating concepts	The thesis slide. X and Y play opposite roles; three levers.
16–18	SD vs SE, road map, metaphors	Carried over from p-01. The map now has three rungs built.
19	The checklist	Identical to p-02’s — the two decks share a conclusion.

Section three

What should be clear in my mind?

The deck’s own checklist, plus what p-03 adds to it.

The deck closes with the same two items as p-02, which is a signal that these two lectures are one argument split across two sessions:

We split the variance of Y into explained and unexplained portions with a trick, inserting the regression line ŷ.
The standard error of the slope is derived from the unexplained portion of Y, the residual.
The standard error has exactly three moving parts — the residual, the sample size, and the variance of X — and you should be able to say which direction each one pushes.
The standard error is a property of the sampling distribution, not of your data. It answers “how far off is a typical estimate,” which is only meaningful across samples you did not draw.

Key takeaways

One formula: SE_b₁ = sd_e / √(n−1) · var(x). Everything else in the deck is either where it came from or what it means.
The derivation is a renaming. ∑(x_i − x)² becomes (n − 1) · var(x), turning an opaque sum into two interpretable quantities.
Noise in Y hurts; spread in X helps. The two variances sit on opposite sides of the fraction, and the intuition from the standard error of the mean reverses for the predictor.
Three levers: more cases, more explained variance, more variation in X. The third is a design decision you cannot revisit later.
Sample size has diminishing returns because of the square root — quadruple n to halve the standard error.
The standard error is the “average error” of the slope estimate, measured across repeated samples you will never actually draw.
Your regression line is one draw from the sampling distribution, not the truth.