Line graphs and scatter plots — basic interpretation: Visual walkthrough

Step 1 — A dot is two facts glued together

WHAT. Before any line, we must be crystal clear what a single dot is. Look at the figure. A student studied for a number of hours (a fact on the horizontal road) and got a test score (a fact on the vertical wall). We slide right by the hours, then climb up by the score, and drop a dot where those two meet.

WHY. A line of best fit is a promise about many dots at once. If we do not know what one dot means, the whole page collapses. So we anchor the picture first.

PICTURE.

In symbols, the dot for the $i$ -th student is written $(x_{i}, y_{i})$ . The little $i$ underneath is just a name-tag — "student number $i$ " — so we can talk about student 1, student 2, and so on without confusion.

Step 2 — Many dots make a cloud, and a cloud has a tilt

WHAT. Now we scatter seven students onto the same picture. Notice with your eyes only: the cloud leans uphill — students on the right (more hours) sit higher (more score).

WHY. This lean is the whole reason a line is useful. If the dots leaned uphill we call it positive correlation; downhill is negative correlation; no lean at all is no correlation. The line of best fit is nothing more than the honest straight-line version of that lean.

PICTURE.

Step 3 — Guess a line, then measure how wrong it is

WHAT. Draw any straight line through the cloud — a guess. For each dot, measure the vertical gap from the dot up (or down) to the line. That gap is the residual: how much the line missed that student by.

WHY. "Best fit" is a competition, and every competition needs a score to beat. The residual is our score. A line that sits far from the dots has big gaps; a good line has tiny gaps. We measure the gap vertically (not slanted) because we are predicting $y$ from $x$ — the question is "given the hours, how far off is our predicted score?"

PICTURE.

Step 4 — Why we square the misses (not just add them)

WHAT. We want a single number for the whole line's badness. First instinct: add up all the residuals. The figure shows why that fails — a dot far above ( $e = + 4$ ) and a dot far below ( $e = - 4$ ) cancel to $0$ , pretending the line is perfect when it is awful.

WHY. We need every miss to count as bad, whether above or below. Two classic fixes: take the size (absolute value) $∣ e_{i} ∣$ , or square each residual $e_{i}^{2}$ . We square because (a) squaring kills the sign — $(- 4)^{2} = 16$ is just as bad as $(+ 4)^{2} = 16$ ; and (b) squaring punishes big misses harder than small ones, which forces the line to avoid disasters. This total is the sum of squared residuals:

$S = \sum_{i = 1}^{n} e_{i}^{2} = \sum_{i = 1}^{n} (y_{i} - m x_{i} - c)^{2}$

Reading it: the $\sum$ (a capital Greek "S", for sum) means "add up, for every student $i$ from $1$ to $n$ "; inside, $e_{i}^{2}$ is one student's squared miss; $S$ is the grand total badness of the whole line.

PICTURE.

Step 5 — Sliding to the smallest area: the balance point

WHAT. Imagine holding the line's tilt fixed and just sliding it up and down. Too high — every dot is below, huge area. Too low — every dot is above, huge area. Somewhere between is a sweet spot. The figure shows that the winning line always passes through the mean point $(\overset{x}{ˉ}, \overset{y}{ˉ})$ — the "centre of gravity" of the cloud.

WHY. The bar over a letter means average: $\overset{x}{ˉ}$ is the average of all the $x$ -values, $\overset{y}{ˉ}$ the average of all the $y$ -values. $\overset{x}{ˉ} = \frac{x _{1} + x _{2} + \dots + x _{n}}{n}, \overset{y}{ˉ} = \frac{y _{1} + y _{2} + \dots + y _{n}}{n}$ Here $n$ is how many dots there are, and dividing the total by $n$ shares it out evenly — that is exactly what an average is. Making $S$ smallest forces the positive misses and negative misses to balance around this centre — like a see-saw settling at its pivot. (This connects to 1.3.04-Mean-median-mode-range — the mean is literally the balance point.)

PICTURE.

Step 6 — What the slope counts: rise over run

WHAT. The slope $m$ is the tilt. Pick any two points on the line; go right by some run $Δ x$ and see how far the line rises, $Δ y$ . The figure marks this staircase.

WHY. The symbol $Δ$ (Greek "delta") just means "change in". The slope answers the exact question a scatter plot exists to answer: "if I study one more hour, how many more marks does the trend promise?" $m = \frac{Δ y}{Δ x} = \frac{rise in score}{run in hours}$ $Δ y$ on top is the vertical rise; $Δ x$ below is the horizontal run; the fraction is "score gained per hour". We use a ratio (not a difference) because we want a rate that stays the same no matter how big a step we take along a straight line — that steadiness is what makes a line a line.

PICTURE.

The exact best-fit slope (you meet its full formula in 4.1.01-Correlation-coefficient) is $m = \frac{\sum ( x _{i} - x ˉ ) ( y _{i} - y ˉ )}{\sum ( x _{i} - x ˉ ) ^{2}}$ Do not panic at it — read it as a shape. On top: for each dot, "how far right of centre" $(x_{i} - \overset{x}{ˉ})$ times "how far above centre" $(y_{i} - \overset{y}{ˉ})$ . When a dot is right-and-high or left-and-low, that product is positive → uphill vote. Right-and-low or left-and-high → negative → downhill vote. The bottom just measures how spread out the $x$ 's are, turning the vote total into a fair per-unit rate.

Step 7 — The line assembled, and reading it

WHAT. We now own both pieces: the tilt $m$ from Step 6 and the height $c = \overset{y}{ˉ} - m \overset{x}{ˉ}$ from Step 5. Slot them into the line and we can predict.

WHY. Prediction is the payoff. Give the line an $x$ it has never seen (say $4.5$ hours) and it returns a sensible $y$ . Landing between known dots is interpolation (trustworthy); shooting past the data is extrapolation (risky — see Edge Case below).

PICTURE.

Step 8 — The degenerate & edge cases (never get ambushed)

WHAT. Four ways the neat story bends. Each gets its own panel in the figure.

WHY. A real dataset will eventually hand you every one of these. Know them before they surprise you.

PICTURE.

(a) Flat cloud — no correlation. Dots wander with no lean. The best slope is $m \approx 0$ , so $y \approx \overset{y}{ˉ}$ : the line is horizontal and says "hours tell you nothing about score here." Not a failure — a genuine finding.
(b) A vertical cloud. All dots share (almost) the same $x$ . Then $\sum (x_{i} - \overset{x}{ˉ})^{2} \to 0$ , the denominator of $m$ collapses toward zero, and $m$ blows up — the line tries to stand vertical, which $y = m x + c$ cannot represent. Lesson: you need spread in $x$ for a slope to exist.
(c) An outlier. One rogue dot far from the crowd (parent note's Error case). Because we squared the residuals (Step 4), a far dot's huge square yanks the whole line toward itself. Squaring is powerful but sensitive — always eyeball for outliers first.
(d) Extrapolation trap. The plant that "grows $5$ cm/week" would be $260$ cm in a year only if the line kept going — but the dots end at week 8. Beyond the last dot, the line is a guess about the unknown, not a measurement. Trust inside the cloud; doubt outside it.

The one-picture summary

One frame carries the whole journey: dots → their vertical misses → squared misses whose total area we shrink → the winning line, pinned through the mean point $(\overset{x}{ˉ}, \overset{y}{ˉ})$ , tilted by $m$ (marks per hour), meeting the axis at $c$ .

Recall Feynman retelling — say it back in plain words

Imagine a cloud of dots, each dot a person: how long they studied, sideways, and what they scored, up. The cloud leans uphill. I want one straight line that tells the cloud's story. To judge a line, I measure how far each dot sits above or below it — the miss. To stop ups and downs cancelling, I square every miss and add them into one number, which is really the total area of a bunch of squares. The best line is the one that makes that total area as small as possible. That champion line always threads the average dot — the balance point of the cloud. Its steepness, rise-over-run, tells me how many extra marks one extra hour buys. Where it crosses the vertical axis tells me the score for zero study. I can read a new value off the line between my dots and trust it; past the edges I'm just guessing. And I stay alert: a flat cloud means "no relationship", a stack of dots with no sideways spread has no slope, and one wild dot can drag my whole line because squaring makes it shout.

Reveal-check yourself:

The residual

e_{i}

measures

the vertical gap from dot

i

to the line,

y_{i} - (m x_{i} + c)

We square residuals instead of adding them because

raw residuals cancel (above and below), while squares are always positive and punish big misses more.

The best-fit line always passes through

the mean point

(\overset{x}{ˉ}, \overset{y}{ˉ})

, the balance point of the cloud.

The slope

m

means, in words

the average change in

y

for each one-unit increase in

x

(e.g. marks gained per extra hour).

A vertical cloud breaks the method because

the

x

-values have (near) zero spread, so

m

's denominator collapses and the slope blows up.