Answer

Centering vs standardizing predictors in regression: when to do which?

Inspired by a question on Cross Validated · September 24, 2026

regressionnormalization

What the two transformations actually change

Centering shifts a variable so that a chosen value becomes zero, usually its mean: x - mean(x). Standardizing centers the variable and then rescales it to units of one standard deviation: (x - mean(x)) / sd(x), often called a z-score.

In an ordinary least squares model, both are linear re-expressions of the predictors, so the model sees exactly the same information. The fitted values, residuals, R-squared, overall F test and the test of the highest-order term (an interaction, or the squared term in a quadratic) come out identical. What changes is the meaning of the individual coefficients, and with it their size, and for lower-order terms also their standard errors and p-values. So the choice is about interpretation and computation, not about getting a better model.

When to center

• Zero is not a meaningful value. The intercept is the predicted outcome when every predictor equals zero. If a predictor is age, blood pressure or years of schooling, zero lies far outside the data and the intercept describes nobody. After centering at the means, the intercept is the predicted outcome for a case that is average on every predictor.
• The model has an interaction. In y ~ a * b, the coefficient on a is the slope of a when b equals zero, not an average effect. With raw variables that is often an impossible situation. Centering b turns the coefficient on a into the slope of a at the typical value of b, which is usually what people want to report.
• The model has polynomial terms. A variable and its square are strongly correlated when the variable is all positive. Centering before squaring reduces that correlation, makes the linear coefficient the slope at the mean, and can improve numerical accuracy.
• Multilevel models. Centering at the grand mean or at each cluster's mean changes which question the coefficient answers (overall versus within-cluster effects), so choose deliberately.

You do not have to center at the mean. Any value that makes zero meaningful works, such as age 18, a clinical cutoff or the scale midpoint. In a model with no interactions or polynomial terms, centering changes only the intercept; every slope and its test stay the same.

When to standardize

• You want effects on a common, unit-free scale. A standardized coefficient gives the change in the outcome for a one-SD increase in that predictor, holding the others fixed. That helps when predictors come in unrelated units, such as dollars and minutes. Standardizing the outcome too gives the fully standardized weights that SPSS labels Beta.
• The method penalizes coefficient size or uses distances. Lasso, ridge and elastic net shrink every coefficient by the same rule, so a predictor recorded in grams is treated differently from the same predictor in kilograms. K-nearest neighbours, k-means, support vector machines and principal component analysis share the problem. Standardizing puts every predictor on an equal footing.
• The fitting algorithm struggles. Gradient-based optimizers, neural networks and some mixed or nonlinear models converge faster and more reliably when inputs are on similar scales.

Standardizing does not reveal which predictor is objectively more important. A one-SD change depends on how much the variable happens to vary in your sample, so the same effect looks larger in a diverse sample and smaller in a narrow one. Report raw units whenever they mean something to your readers, and think twice before standardizing a 0/1 dummy variable, whose SD depends on how the groups are split.

See it in R and Python

This simulation fits the same interaction model three times: with raw predictors (age in years and weekly work hours), with centered predictors, and with standardized predictors. The R version uses base R only; the Python tab runs the same analysis with NumPy.

set.seed(24)
n <- 200
age   <- rnorm(n, mean = 45, sd = 10)   # years
hours <- rnorm(n, mean = 35, sd = 8)    # weekly work hours
y <- 20 + 0.10 * age + 0.30 * hours - 0.02 * (age - 45) * (hours - 35) + rnorm(n, sd = 3)

# Three versions of the same predictors
age_c <- age - mean(age);   hours_c <- hours - mean(hours)   # centered
age_z <- age_c / sd(age);   hours_z <- hours_c / sd(hours)   # standardized

raw <- lm(y ~ age * hours)
ctr <- lm(y ~ age_c * hours_c)
std <- lm(y ~ age_z * hours_z)

tab <- function(m) round(summary(m)$coefficients[, 1:3], 3)   # estimate, SE, t
tab(raw); tab(ctr); tab(std)

# Same fit every time: identical R-squared and fitted values
round(sapply(list(raw = raw, centered = ctr, standardized = std),
             function(m) summary(m)$r.squared), 4)
all.equal(fitted(raw), fitted(ctr))

# Collinearity between a main effect and its interaction term
round(c(raw = cor(age, age * hours), centered = cor(age_c, age_c * hours_c)), 2)

# Confidence interval for the standardized age effect
round(confint(std)["age_z", ], 2)

import numpy as np

rng = np.random.default_rng(24)
n = 200
age = rng.normal(45, 10, n)     # years
hours = rng.normal(35, 8, n)    # weekly work hours
y = 20 + 0.10 * age + 0.30 * hours - 0.02 * (age - 45) * (hours - 35) + rng.normal(0, 3, n)

# Three versions of the same predictors
age_c, hours_c = age - age.mean(), hours - hours.mean()                  # centered
age_z, hours_z = age_c / age.std(ddof=1), hours_c / hours.std(ddof=1)    # standardized

def fit(a, b):
    # OLS for y ~ a * b; returns estimate, SE, t table plus R-squared and fitted values
    X = np.column_stack([np.ones(n), a, b, a * b])
    beta, *_ = np.linalg.lstsq(X, y, rcond=None)
    fitted = X @ beta
    resid = y - fitted
    sigma2 = resid @ resid / (n - X.shape[1])
    se = np.sqrt(np.diag(sigma2 * np.linalg.inv(X.T @ X)))
    r2 = 1 - resid @ resid / np.sum((y - y.mean()) ** 2)
    return np.column_stack([beta, se, beta / se]), r2, fitted

raw, ctr, std = fit(age, hours), fit(age_c, hours_c), fit(age_z, hours_z)
for name, (table, r2, _) in zip(("raw", "centered", "standardized"), (raw, ctr, std)):
    print(name, "R-squared", round(r2, 4))
    print(np.round(table, 3))   # rows: intercept, a, b, a:b; columns: estimate, SE, t

# Same fit every time: identical fitted values
print(np.allclose(raw[2], ctr[2]))

# Collinearity between a main effect and its interaction term
print(round(np.corrcoef(age, age * hours)[0, 1], 2),
      round(np.corrcoef(age_c, age_c * hours_c)[0, 1], 2))

The figures below come from one seeded run of the R code. Python uses a different random number generator, so its figures differ slightly, but the pattern is the same.

• The fit never changes. R-squared is 0.5912 in all three models, and the fitted values of the raw and centered models are identical.
• The interaction test never changes. Its t value is -8.039 every time. In raw and centered units the coefficient is -0.022 per year-hour; standardized, it is -1.791 per SD-by-SD.
• The main effects change meaning. With raw variables the age coefficient is 0.875 (t = 9.046): the age slope for someone who works zero hours a week, far outside the data. After centering it is 0.122 (t = 5.486): the age slope at average hours.
• Standardizing only rescales. The standardized age effect is 1.182 points per SD of age, 95% CI [0.76, 1.61], with the same t of 5.486 as the centered model. The intercept, 34.702, is the predicted score for someone of average age working average hours.
• Centering removes nonessential collinearity. The correlation between age and the age-by-hours product is 0.70 with raw variables and 0.03 after centering.

Common mistakes

• Believing centering fixes multicollinearity. It removes the artificial correlation between a variable and its own product or square, which can matter for numerical stability. It does nothing about real correlation between two different predictors, and it does not change the test of the interaction.
• Reading raw main effects in an interaction model as average effects. Without centering they are conditional effects at zero on the other variable.
• Standardizing the product term. For standardized interaction coefficients, standardize each predictor first and then multiply them, as in the example, rather than z-scoring the finished product.
• Comparing standardized coefficients across studies. Samples with different spreads produce different standardized values for the same underlying effect.
• Forgetting to reuse the training values. When you predict new data, center and scale with the mean and SD from the data the model was fitted to, not from the new data.

How to report standardized regression coefficients in APA style (7th edition)

Say in the method section which variables were centered or standardized, and around what value, because the coefficients cannot be interpreted without it. Using the standardized model above as an example:

"Age and weekly work hours were standardized before forming their product term. Holding work hours at their mean, a one-SD increase in age was associated with a 1.18-point higher score, b = 1.18, 95% CI [0.76, 1.61], t(196) = 5.49, p < .001. The model explained 59% of the variance, R² = .59."

If you standardize the outcome as well, label the coefficients β and say so. In a regression table, give b, SE, the 95% CI and p for each term, and add a table note stating how each predictor was centered or scaled.

Related tools and guides

• APA 7 formatter for regression
• Which statistical test should I use?
• How to normalize data to a 0-1 range
• Is R-squared useful or misleading?
• Standard score (Wikipedia)

More answered questions

• Is R-squared useful or misleading?
• Logit vs probit: what is the difference, and which should you use?
• AIC vs BIC: which model selection criterion should you use?

Written by AskStats with AI assistance. This is general information, not advice for your specific data or study. When the results matter, check your approach with a qualified statistician.