Suppose a property portal wants to understand monthly rents for two-bedroom flats in one part of Bengaluru. It has a few hundred listings with rent, carpet area and distance to the nearest metro station. A linear regression fitted to this data might produce an equation like this: rent equals ₹5,000, plus ₹18 for every square foot, minus ₹1,200 for every kilometre from the metro.
For a 1,000 square foot flat 2 km from a station, the model predicts ₹5,000 + ₹18,000 − ₹2,400 = ₹20,600. That equation is the whole output of a regression in compact form, and reading it correctly is the core skill.
What the coefficients mean
Linear regression finds the straight-line (or flat-plane, with more inputs) relationship that minimises the sum of squared differences between actual and predicted values. This method is called ordinary least squares.
Each coefficient describes the expected change in the outcome for a one-unit change in that input, holding the other inputs constant. The ₹18 per square foot is the difference in rent between two flats that are the same distance from the metro but differ by one square foot. That “holding constant” clause is the reason regression is more useful than looking at one correlation at a time: it separates the effect of size from the effect of location, which are tangled together in raw data if larger flats tend to be further out.
The intercept, ₹5,000, is the predicted rent when every input is zero. A zero-square-foot flat at a metro station is meaningless, so the intercept here is just an anchor for the line. Do not try to interpret it unless zero is a realistic value for every input.
Coefficients describe association in the data. Whether the ₹1,200 per kilometre is a causal effect of distance depends on whether something else that varies with distance, such as the age of buildings, was left out of the model.
R-squared: share of variation explained
Rents vary across listings. R-squared tells you what fraction of that variation the model accounts for. Formally, it is one minus the ratio of the leftover (residual) variation to the total variation around the average rent. An R-squared of 0.64 means the model explains 64% of the variance in rents; the remaining 36% comes from things the model does not include, such as floor, furnishing, building amenities and plain randomness in what landlords ask.
What counts as a good R-squared depends entirely on the field. Physical processes measured precisely can give values above 0.95. Models of human behaviour, such as predicting which customers will spend more next month, are often useful at 0.2 or 0.3. A low R-squared does not make the coefficients wrong; it means individual predictions will be imprecise.
A high R-squared is not proof of a good model either. Two variables that both trend upward over time, like monthly revenue and headcount, can give an R-squared above 0.9 with no meaningful link. And R-squared never falls when you add an input, even a useless one, because the model can always fit the sample a little more closely. Adjusted R-squared corrects for this by penalising extra inputs. With 50 observations, three inputs and an R-squared of 0.64, adjusted R-squared is about 0.62. If adding an input lowers adjusted R-squared, that input is not pulling its weight.
For judging prediction quality, the error in the outcome’s own units is usually more useful to a business audience than R-squared. “Predictions are typically within about ₹2,500 of the actual rent” is something a manager can act on.
When inputs overlap: multicollinearity
Now the analyst adds a third input: number of bathrooms. In this market, bigger flats almost always have more bathrooms, so carpet area and bathroom count are strongly correlated. This is multicollinearity: two or more inputs carry largely the same information.
The consequence is subtle. The model’s overall predictions may barely change. But the individual coefficients become unstable, because the model cannot tell how much of the rent difference to credit to area and how much to bathrooms. They always move together, so there are few listings where one varies and the other does not. Symptoms include:
- coefficients that swing dramatically when you add or drop a few rows or another input;
- signs that make no sense, such as rent falling with each extra square foot once bathrooms are included;
- large standard errors, so neither input looks statistically significant on its own even though together they clearly matter.
The standard diagnostic is the variance inflation factor. For each input, regress it on all the other inputs and take that R-squared; the VIF is one divided by one minus that value. If area is 90% explained by the other inputs, its VIF is 10, and the standard error of its coefficient is about 3.2 times (the square root of 10) larger than it would be without the overlap. Rules of thumb treat VIFs above 5 or 10 as worth investigating, though those cut-offs are conventions rather than laws.
Does it matter?
That depends on why you built the model. If the goal is pure prediction, such as estimating a fair rent for a new listing, multicollinearity is often harmless, provided new data has the same pattern of overlap as the training data.
If the goal is explanation, such as telling landlords how much an extra bathroom adds to rent, multicollinearity is a serious problem because the very number you want to report is unreliable. Options include dropping one of the overlapping inputs, combining them into a single measure, collecting data where they vary independently, or using regularised methods like ridge regression, which trade a little bias for much more stable coefficients.
A related trap is the “dummy variable trap”: encoding a category such as city as one 0/1 column per city and also including an intercept. The city columns always add up to one, which makes them perfectly collinear with the intercept. Dropping one city as the reference level solves it, and most libraries do this automatically.
Written by the RP Strativa Team. The rent model and all figures are illustrative.
Leave a Reply