Sample vs Sampling Distribution

Section one

Key concepts

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

Key concepts

Deviations are the currency of every variance.

Every sum of squares in regression is built by measuring a distance, squaring it, and adding it up. Only the reference point changes:

y_i − y = deviation — distance from the mean
y_i − ŷ_i = e_i = residual — distance from the regression line
ŷ_i − y = explained — distance the model moved you off the mean

var(y) = ∑(y_i − y)²n − 1

There are two different variances, and they answer different questions.

	Sample distribution	Sampling distribution
What varies	the data	the estimate
Question	How spread out are the observations?	How much would my estimate move in a new sample?
Centered on	the center of the data	the “true” parameter
Measured in	standard deviations	standard errors
You can see it	yes — it is your data	no — it is hypothetical

The variance of Y partitions into an explained and an unexplained part.

Total SS: ∑(y_i − y)² (deviations)
Regression SS: ∑(ŷ_i − y)² (explained)
Error SS: ∑(y_i − ŷ_i)² (residuals)
TSS = RSS + ESS

Total SS becomes the variance of Y. Error SS becomes the variance of the slope.

The regression line is a conditional mean.

Standard deviation: on average, how far is the mean from each data point.
SD of the residual: on average, how far is the conditional mean from each data point.

Same sentence, one word changed. The regression line is what the mean of Y becomes once you let it depend on X.

The standard error of the slope is built out of the unexplained variance.

SSE = ∑e_i² σ²_ε = SSEn − 2 SE_b₁ = √σ²_ε∑(x_i − x)²

Or in words:

Standard Error of the Slope ≈ residualsample size · variance X

Standard error: on average, how far off is our best guess of the slope from the true slope?

The Central Limit Theorem is what makes the sampling distribution usable.

Across repeated samples the estimated slope is approximately normally distributed and centered on the true slope. That shape is what licenses:

β₁ = b₁ ± t · SE_b₁

Inference happens in the sampling distribution, not in the sample.

unexplained variance → Error SS → standard error → confidence interval → p-value

Section two

Key relationships

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

The whole deck in one sentence

Sample variance describes your data; sampling variance describes your estimate — and the second is manufactured out of the first.

That is the subtle part. These are not two unrelated topics that happen to share the word variance. The unexplained portion of the variance in your data is the raw material the standard error is made from. Noisy data produce an unstable estimate. The deck alternates between the two on purpose: slides labeled “this is sample variance” and “this is sampling variance” are asking you to notice which one you are looking at, because the pictures look similar and the meaning is completely different.

Two distributions, two centers, two rulers

The side-by-side slide near the end is the thesis of the lecture. Read it carefully:

The sample distribution puts heart rates on the axis. Its center is the center of the data. The red segments are deviations, and their typical size is a standard deviation.
The sampling distribution puts slopes on the axis. Its center is the true slope. The intervals are estimates from repeated samples, and their typical spread is a standard error.

Same picture-making logic, different unit of observation. In the first, a dot is a person. In the second, a dot is an entire study.

You only ever get one sample

This is why the simulation slides exist. In real work you draw one sample, compute one slope, and that is all you will ever see. The sampling distribution is counterfactual — it is the spread of the estimates you would have gotten from all the samples you did not draw.

The simulations make that invisible object visible: each frame draws a new sample from the same population, fits a new line, and drops one more slope onto the sampling distribution. Watching the line jump around from sample to sample is sampling variance. The standard error is just the width of that jumping, calculated rather than simulated.

Three deviations, three sums of squares

Every quantity in the deck comes from one of three distances:

Distance	Name	Sum of squares	Becomes
y_i − y	deviation	Total SS	variance of Y
ŷ_i − y	explained	Regression SS	model fit (R²)
y_i − ŷ_i	residual	Error SS	variance of the slope

The third row is the one this deck is really about. Everything that determines whether your result is statistically significant flows out of the residual.

Notation warning. This deck writes TSS = RSS + ESS where RSS = Regression SS and ESS = Error SS. Many textbooks use exactly the opposite (RSS = Residual, ESS = Explained). One slide reconciles it explicitly — “Error SS, same as the Residual SS.” When you see ESS in these slides, read error, not explained.

Why the regression line is “just” a mean

Slide 5 is doing more work than it looks like. It slices caffeine into bins and plots the average heart rate in each — 74, 78, 82, 86, 90, 95, 99, 103, 107, 111 — against a flat “mean of Y” line at 90.

The flat line is the unconditional mean: your best guess about anyone’s heart rate if you know nothing about them. The rising sequence of bin averages is the conditional mean: your best guess once you know their caffeine intake. The regression line is a smooth version of that second sequence.

Once you see it that way, the parallel interpretations fall out for free:

Standard deviation = how far the data sit from the mean.
SD of the residual = how far the data sit from the conditional mean.

Regression did not introduce a new kind of dispersion. It moved the reference point from a flat line to a tilted one.

What makes a standard error small

The approximation slide is worth memorizing because it tells you what to do as a researcher:

SE_b₁ ≈ residualsample size · variance X

Three levers, and only three:

Less residual (numerator) — a better-specified model with less unexplained noise. This is what control variables and specification buy you.
A larger sample (denominator) — more cases, more precision.
More spread in X (denominator) — because ∑(x_i − x)² = (n − 1) · var(x), a treatment that barely varies cannot produce a precise estimate of its effect, no matter how many cases you have.

That last one surprises people. If everyone in your study got roughly the same dose, you cannot learn much about dosage.

The chain that ends in a p-value

The final slide is the payoff, and it is worth being able to recite:

unexplained variance → Error SS → standard error → confidence interval → p-value

Read left to right, this says that statistical significance is downstream of model fit. Read right to left, it says a p-value is never a standalone fact — it inherits everything from how much of Y your model failed to explain. This is also why the Central Limit Theorem appears in the middle of the deck: it is the step that turns a spread of hypothetical estimates into a probability statement.

How to read the slides

Slides	What they are doing	What to take away
1–2	Framing: deviations as the common currency	Every variance is squared distances from some reference point.
3–6	Sample variance, four ways	The mean, then the conditional mean, then the partition into explained and residual.
7	Partitioning the variance of Y	The formula sheet for the whole deck. Note the ESS/RSS convention.
8–10	The Central Limit Theorem and SE_b₁	The standard error is the residual variance, scaled by n and the spread of X.
11–12, 15	Simulations	The sampling distribution made visible. Each frame is a study you did not run.
13–14	Back to sample variance	Deliberate whiplash. Ask yourself which variance you are looking at.
16	Sample vs sampling, side by side	The thesis slide. Two centers, two rulers.
17	Deviations vs residuals	The conditional-mean parallel.
18	Unexplained variance → p-value	The chain. Be able to recite it.

The alternation between “this is sample variance” and “this is sampling variance” is not disorganization — it is the exercise. If you can label each slide correctly without reading the title, you have the distinction.

Section three

What should be clear in my mind?

The four things this deck exists to settle.

The difference between the distribution of the sample and the sampling distribution. One describes the data you collected and is measured in standard deviations; the other describes how your estimate would behave across repeated samples and is measured in standard errors. You observe the first and can only infer the second.
The difference between a deviation and a residual. A deviation is the distance from a data point to the mean. A residual is the distance from a data point to the regression line — that is, to the conditional mean. Both are “distances from a best guess”; the second just has a better guess.
Where the standard error of the slope comes from. It is the variance of the residual, divided by the spread of X, square-rooted. Unexplained variance in your model becomes uncertainty in your estimate.
Why the confidence interval and p-value depend on unexplained variance. Because the whole chain — Error SS to standard error to interval to p-value — starts at the residual. Improving fit narrows the interval.

Key takeaways

Two variances, one word. Sample variance is about spread in the data; sampling variance is about spread in the estimate. Always ask which one is on the axis.
A dot means different things in the two pictures. In the sample distribution a dot is an observation. In the sampling distribution a dot is a whole study.
The sampling distribution is counterfactual. You draw one sample; the standard error is a calculated stand-in for the ones you did not draw.
The regression line is the mean conditioned on X. This is why the interpretation of the residual SD mirrors the interpretation of the ordinary SD word for word.
Total SS splits into explained and error, and it is the error half that determines the precision of your slope.
Three levers shrink a standard error: less residual, more cases, more variation in X. A treatment with little variation cannot yield a precise estimate.
Significance is downstream of fit. Unexplained variance → Error SS → standard error → confidence interval → p-value.
Watch the abbreviation. In this deck ESS means Error SS, the opposite of the common textbook convention.