Someone on your team slides a chart across the table: "For every extra dollar we spend on ads, we get $4.20 back, look, the line fits beautifully." That sentence is a regression. So is "users who finish onboarding are 3× more likely to renew," and so is the churn-risk score your tool flashes red next to an account. Regression is the quiet engine under most of the numbers a leader is handed, and knowing how it works is mostly about knowing when not to trust it.

The quick version

  • Regression fits a line (or curve) through data to estimate how one thing moves when another changes, and to predict the next point.
  • Linear = a straight-line relationship. Non-linear = a curve (diminishing returns, S-curves). Logistic = predicting a yes/no outcome as a probability (will they churn, convert, default?).
  • A good fit is not proof of cause. A model can be accurate and still mislead you about what to actually do.
  • Always look at the picture, not just the summary number. Identical-looking statistics can hide wildly different realities.

The idea in depth

Regression is older, and more humbling, than the dashboards make it look. The word itself comes from a discovery that surprised the man who made it.

Where the word comes from, and the trap built into it

In 1886 the Victorian polymath Francis Galton measured the heights of 205 sets of parents and their 928 adult children and published "Regression towards Mediocrity in Hereditary Stature" (Journal of the Anthropological Institute, vol. 15, 1886). He expected tall parents to produce equally tall children. Instead, exceptionally tall parents tended to have children closer to the average, and very short parents likewise. Extremes drifted back toward the middle. He called that drift regression, and the name stuck to the whole technique, even though most of the time we now use it to draw a predictive line rather than to study heredity.

That finding is still one of the most useful, and most ignored, ideas in management. It's now called regression to the mean: any measurement that's part skill and part luck will, when it's extreme, tend to be less extreme next time. Daniel Kahneman devotes a chapter of Thinking, Fast and Slow (2011) to it, retelling how Israeli Air Force instructors were sure that praising a good landing made the next one worse and screaming at a bad one made the next one better. The data say otherwise: an unusually good flight is usually followed by a more average one regardless of what you say, because the luck doesn't repeat. The instructors had pinned a cause (their feedback) on a pattern that was pure statistics. So the move is: before you credit the new playbook or the reorg for a rebound after a terrible quarter, ask whether an average month would have looked like improvement anyway.

A model can be accurate about the past and still lie to you about the future.

Three flavours, one idea

The mechanics are the same throughout: pick an outcome you care about (Y), pick the things you think drive it (the X's), and let the maths find the line or curve that sits closest to the dots. What changes is the shape you allow and the kind of outcome you're predicting.

flowchart TD
    Q("What kind of outcome are you predicting?") --> N("A number: revenue, time, headcount")
    Q --> C("A category: yes/no, churn/stay")
    N --> L("Is the relationship a straight line?")
    L -->|Roughly yes| LIN("Linear regression")
    L -->|It curves / plateaus| NON("Non-linear regression")
    C --> LOG("Logistic regression, outputs a probability 0 to 1")
Choosing the right shape starts with one question: is the outcome a number or a category? Leaders Loop

Linear regression draws the best straight line through the cloud of points and reports a slope: "each extra unit of X is associated with this much change in Y." It's the workhorse, cheap, transparent, easy to explain to a board. Its weakness is that it assumes the world is a straight line, and the world rarely is.

Non-linear regression lets the relationship bend. The most leader-relevant bend is diminishing returns: the first salespeople you hire lift revenue a lot, the twentieth much less. A straight line through that data will overpromise at the top end, it will keep predicting gains long after the curve has flattened. Fitting a curve instead tells you where the plateau is, which is exactly the budgeting question you actually have.

Logistic regression answers a different kind of question: not "how much?" but "how likely?" When the outcome is binary, the customer renews or doesn't, the candidate accepts or doesn't, the loan defaults or doesn't, a straight line is the wrong tool, because probability can't go below 0 or above 100%. Logistic regression squashes the prediction into that 0–1 range using an S-shaped (logistic) curve and returns a probability. It is the engine behind most "risk scores," lead-scoring models, and churn-prediction tools. The modern form was set out by the British statistician David Cox in "The Regression Analysis of Binary Sequences" (Journal of the Royal Statistical Society, Series B, 1958), building on the "logit" framework that Joseph Berkson had developed for medical bioassay in the 1940s.

flowchart LR
    A(["Inputs: tenure, usage, support tickets…"]) --> B("Logistic model weighs each input")
    B --> C(["S-curve squashes the result to 0–1"])
    C --> D("Output: 'this account has a 78% chance of renewing'")
Logistic regression turns several inputs into a single probability between 0 and 1. Leaders Loop

The honest limitation: a good fit is not the truth

Here is the part the dashboard won't tell you. In 1973 the statistician Francis Anscombe published four small datasets, now called Anscombe's quartet (The American Statistician, vol. 27), that share almost identical summary statistics: the same averages, the same correlation, and the same regression line, y = 3 + 0.5x. Plot them, though, and they look nothing alike. One is a clean line; one is a smooth curve a straight line gets badly wrong; one is a perfect line dragged off course by a single outlier. Anscombe's point was blunt: a tidy regression equation can describe four completely different worlds. The summary number is not the data. The practical rule: never accept a regression result without seeing the scatter plot behind it. If someone shows you a slope but not the picture, that's a yellow flag, not a finding.

The deeper limitation is the one Galton's word smuggles in: regression tells you that two things move together, not that one causes the other. A model can fit beautifully and still be useless for decisions, because the thing it's quietly riding on is a confounder you haven't measured. That's not a footnote; it's the difference between a forecast and a fantasy, and it's why we keep correlation and causation in separate boxes.

A worked example

Say you run customer success at a B2B software company, and your analyst builds a logistic regression to predict renewal. (All figures below are illustrative, to show how to read the output, not real data.) The model lands on three inputs: weekly active usage, number of support tickets, and whether the customer has a named admin. It reports that each extra weekly login raises renewal odds, lots of support tickets lower them, and having a named admin is the single strongest positive signal. Overall it predicts renewal correctly about 80% of the time on accounts it has never seen.

Two readings lead to very different Mondays. The naïve one: "Tickets predict churn, so discourage support contact." That's the flight-instructor error in a suit, acting on a correlation as if it were a lever, and suppressing exactly the signal that lets you save an account. The better one: usage and a named admin are the things you can actually influence, so the move is to drive activation and get every account to name an admin in onboarding, then re-measure next quarter and see if renewal actually shifts. The model points; the experiment proves. And because a new onboarding flow is cheap to undo, it's a low-stakes test, the kind of reversible decision you make fast and learn from.

One last discipline: when the model flags an at-risk account, look at it. If a regression "explains" a number you can't square with the customer in front of you, trust the customer and interrogate the model.

Frequently asked questions

What's the difference between regression and correlation?

Correlation is a single number (between −1 and +1) describing how tightly two things move together. Regression goes further: it gives you an equation, a line or curve, so you can predict Y from X and quantify the relationship ("+$4 of revenue per $1 of spend"). Correlation tells you there's a relationship; regression tries to describe its shape. Neither, on its own, tells you what causes what.

When should I use logistic instead of linear regression?

Use logistic whenever the outcome is a yes/no (or category), not a quantity. "How many dollars?" is linear. "Will they churn, yes or no?" is logistic, because the answer you want is a probability between 0% and 100%, and a straight line would happily predict a −20% or 140% chance of churning, which is nonsense.

How do I know if the relationship is non-linear?

Plot it. If the dots clearly bend, plateau, or form an S, a straight line will systematically miss, over-predicting in some ranges and under-predicting in others. Diminishing returns (each extra unit of spend buys less) and saturation effects are the usual culprits in business data, and both are curves, not lines.

What does "R-squared" mean when someone quotes it?

Roughly, the share of the variation in the outcome the model accounts for, from 0 to 1 (often shown as a percentage). Higher generally means a tighter fit, but a high R-squared does not mean the model is causally correct or will hold next quarter, and a low one isn't always a failure. Anscombe's quartet is the cautionary tale: matching numbers, four different realities. Ask to see the plot.

Do I need to understand the maths to use this?

No, but you need to understand the questions. You don't have to fit the model; you have to know to ask "what's the outcome, what shape did we assume, can I see the scatter plot, and what could be confounding this?" Those four questions catch most of the ways a regression misleads a decision-maker.

Related in the Toolkit

Where to go next