Regression Coefficients: Formula, Properties and Interpretation

Every time you fit a straight line to a scatter of points, you’re really answering one specific question: for every one-unit change in one variable, how much does the other change? That number is the regression coefficient. It looks like a single letter, b, but it carries the entire story of how two variables move together, and getting comfortable with it unlocks everything else in bivariate analysis, from predicting sales using ad spend to estimating how rainfall affects crop output.

Because you can predict either variable from the other, every pair of variables actually gives you two regression coefficients, not one. This post walks through how these coefficients are computed, what their properties mean in plain terms, and how the two coefficients relate to each other and to correlation.

What exactly is a regression coefficient? 🔗

When you fit a line of Y on X, written as Y = a + bX, the constant b tells you how much Y changes for a one-unit change in X. This is called bYX, the regression coefficient of Y on X. Flip the roles, and you get a second line, X = a’ + b’Y, where bXY tells you how much X changes for a one-unit change in Y.

These two coefficients are not interchangeable, and they are almost never equal. bYX assumes X is doing the explaining and Y is being explained; bXY assumes the opposite. A single dataset of paired observations on X and Y, such as advertising spend and sales, or hours studied and exam scores, will always produce two distinct regression lines unless the correlation between the variables is perfect.

How to compute regression coefficients 🔗

There are two equivalent ways to arrive at these values, and knowing both makes cross-checking your work much easier.

Using deviations from the mean 🔗

If you take deviations of each observation from its own mean, x = X − X̄ and y = Y − Ȳ, the regression coefficient of Y on X is:

bYX = Σxy / Σx²

and the regression coefficient of X on Y is:

bXY = Σxy / Σy²

Notice that the numerator, Σxy, is identical in both formulas. This term captures the joint variation between X and Y, while the denominator anchors each coefficient to the spread of the variable being treated as the predictor. This is essentially the same least-squares logic that underlies simple linear regression, where the line is chosen to minimise the sum of squared errors between observed and predicted values.

Using correlation and standard deviation 🔗

A quicker route, especially when you already know the correlation coefficient r and the standard deviations σX and σY, is:

bYX = r × (σY / σX)

bXY = r × (σX / σY)

These formulas are useful because they show, at a glance, why the two coefficients are rarely equal: they depend on the ratio of the two standard deviations, not just on how strongly the variables are related.

A quick worked example 🔗

Say a district’s agricultural department is comparing monthly rainfall (X) with crop yield (Y) across several seasons. If Σxy works out to 2704, Σx² to 5398, and Σy² to 2224, then bYX = 2704/5398 ≈ 0.50 and bXY = 2704/2224 ≈ 1.22. The first tells you that for every additional unit of rainfall, yield rises by about 0.50 units on average. This kind of coefficient is exactly what agricultural forecasting models use in practice: studies estimating crop yield from rainfall and temperature rely on the same slope logic, just extended to more variables, and government meteorological research routinely uses regression models built on district-level weather and yield data to forecast production before harvest.

Interpreting the coefficient correctly 🔗

A regression coefficient is not just a number, it’s a rate. bYX = 0.50 means “yield increases by 0.50 units, on average, for every one-unit rise in rainfall, holding everything else constant.” The word “average” matters. No single data point will match this exactly; the coefficient describes the overall trend, not a guarantee for any one observation.

It’s also worth remembering that a coefficient’s size depends on the units you’re measuring in. A coefficient of 500 isn’t automatically “more important” than a coefficient of 0.5 if the two are measured on completely different scales. This is one of the most common misreadings of regression output, and researchers who study how regression is taught point out that unstandardised coefficients should always be interpreted alongside the units and scale of the original variables, not compared directly across different studies.

Important properties of regression coefficients 🔗

Once you’ve calculated bYX and bXY, a set of properties consistently hold true, and they’re worth memorising because they double as quick sanity checks on your calculations.

Both regression lines pass through (X̄, Ȳ) 🔗

No matter which variable you treat as dependent, both regression lines intersect at the point formed by the mean of X and the mean of Y. This makes sense: the “average” line has to pass through the “average” point.

Both coefficients carry the same sign 🔗

If bYX is positive, bXY will also be positive, and the same holds for negative values. You will never see one coefficient positive and the other negative for the same pair of variables.

The correlation coefficient is their geometric mean 🔗

This is arguably the most elegant property in the whole topic: r = ±√(bYX × bXY). The sign of r matches the sign of both coefficients. This single relationship is why, if you’re given both regression coefficients, you can always recover the correlation between the variables, and vice versa, as shown in standard derivations of this identity using the deviation-based formulas above.

If one coefficient exceeds 1, the other must be below 1 🔗

Since r² can never exceed 1, and r² = bYX × bXY, it follows that if one coefficient is greater than unity, the other has to be a fraction. Both coefficients being greater than 1 at the same time is mathematically impossible.

Regression coefficients are independent of origin but not of scale 🔗

If you shift every value of X and Y by adding or subtracting a constant, the coefficients stay unchanged, since shifting the origin doesn’t affect covariance or variance calculations once deviations are taken from the (new) mean. But if you multiply or divide the values by a constant, the coefficients change proportionally, because you’ve altered the actual scale of measurement, not just the reference point.

How the two coefficients relate to each other 🔗

Beyond the individual properties, bYX and bXY interact with each other and with r in ways that reveal the strength and geometry of the relationship between X and Y.

Their arithmetic mean is never less than r 🔗

(bYX + bXY)/2 ≥ r. This follows directly from the fact that the arithmetic mean of any two positive numbers is always at least as large as their geometric mean, and since r is the geometric mean of the two coefficients, this inequality holds automatically.

What happens at the extremes of correlation 🔗

When r = 0, meaning X and Y are uncorrelated, the two regression lines become perpendicular to each other. This makes intuitive sense: with no linear relationship, knowing X gives you no useful information for predicting Y, so the “best” line of Y on X becomes horizontal, and the “best” line of X on Y becomes vertical.

At the other extreme, when r = ±1, the relationship between X and Y is perfectly linear, and the two regression lines coincide into a single line. There’s no ambiguity left about which variable predicts which, because every point lies exactly on the line.

The angle between the lines reflects the strength of dependence 🔗

Between these two extremes, the angle between the two regression lines narrows as correlation strengthens and widens as it weakens. A small angle signals a strong linear relationship; an angle closer to 90 degrees signals a weak one. This gives you a visual, geometric way to judge how tightly two variables are tied together, without even calculating r directly.

Why this matters beyond the exam 🔗

Regression coefficients aren’t just an algebra exercise. Every time an economist estimates how much consumption rises with income, or a public health researcher estimates how much a disease’s spread responds to vaccination rates, they’re computing and interpreting a bYX or bXY. Understanding the properties covered here means you’ll never mistake a large coefficient for a strong relationship, or forget that correlation and the two slopes are mathematically bound together through a single, elegant identity.

What do you think? If you calculated bYX for two variables and got a value greater than 1, what would that immediately tell you about bXY without doing any further calculation? And if a dataset produces regression lines that are almost perpendicular, what does that suggest about how useful one variable would be for predicting the other?

References

1. https://home.iitk.ac.in/~shalab/regression/Chapter2-Regression-SimpleLinearRegressionAnalysis.pdf
2. https://web.iitd.ac.in/~dharmar/paper/AgriRes2020.pdf
3. https://mausamjournal.imd.gov.in/index.php/MAUSAM/article/download/1041/883/3786
4. https://files.eric.ed.gov/fulltext/EJ863505.pdf
5. http://www.gkgcollege.edu.in/uploads/snstatistics/Module%202%20(1).pdf
6. https://www.kbpimsr.ac.in/Docs/study_data/BBA/STATISTICAL%20TECHNIQUES%20(AECC-C3)%20UNIT%20IV%20B.pdf