AI Glossary / Statistics and probability / Beginner
Deviation scores
Definition
Deviation scores express observations as signed distances from their mean. For an observation x and arithmetic mean x-bar, its deviation score is x - x-bar. A positive value lies above the mean, a negative value below it, and zero exactly at it. The scores retain the original measurement units. They are a basic ingredient of variance, standard deviation and covariance, rather than a complete measure of unusualness by themselves.
What it means
Subtracting the same mean from every observation relocates the numerical origin without changing distances between observations. This operation is also called mean centering. It answers a relative question: how far is this observation above or below the average of the specified reference set? The reference matters. A response time can be above one team's mean and below another team's mean. Neither statement alone says whether the response time meets a service target or represents a statistically unusual event.
How it works
Choose the observations and calculate their arithmetic mean. Subtract that mean from each observation, retaining signs and units. For the same equally weighted set used to calculate the mean, deviations sum to zero apart from numerical rounding. Squaring and summing these deviations produces the sum of squares; dividing by n - 1 gives the usual sample variance when n exceeds one. Dividing each deviation by a nonzero standard deviation instead produces standardized scores. For weighted means, the corresponding weighted deviations sum to zero; their unweighted sum need not. Record which reference mean was used, especially when comparing periods or groups.
Three practical examples
- Reporting: show each store's revenue above or below the average store revenue for a specified month, while retaining currency units.
- Model preparation: center a continuous predictor so that zero represents its training-set mean, making some regression intercepts easier to interpret.
- Quality analysis: calculate measurement deviations as an intermediate step in estimating variation, alongside checks for outliers and process changes.
Think of it this way
Think of moving a ruler's zero mark to the average position. Every object stays the same distance from every other object, but its reading now describes a position relative to that average. Moving the zero mark does not change the objects, make their spread smaller or establish an acceptable operating boundary.
Worked example
Suppose four hypothetical processing times are 6, 8, 10 and 16 minutes. Their mean is 40/4 = 10 minutes. Their deviation scores are -4, -2, 0 and 6 minutes. The signed sum is zero, although the observations plainly vary. Squared deviations sum to 16 + 4 + 0 + 36 = 56 squared minutes. The usual sample variance is 56/3, approximately 18.67 squared minutes, and the sample standard deviation is approximately 4.32 minutes. Notice that the gap between the first and last observations remains ten minutes after centering: 6 - (-4) = 10. If a separate operational target were eight minutes, distances from that target would instead be -2, 0, 2 and 8; these are target differences, not deviations from this sample's mean.
Illustrative case study
Fictional case: Alder Service Analytics prepares a dashboard of task completion times. The analyst adds a column showing deviations from the current team's mean but keeps the original times and agreed service target visible. A manager initially mistakes a negative deviation for a service failure; the team clarifies that shorter times are below average and may be desirable. Before training a prediction model, the analyst calculates centering values using training data and applies those stored values to later records. This invented case shows why reference labels and measurement direction matter; it does not claim a measured productivity improvement.
Common misconceptions and mistakes
A zero average deviation does not mean zero variability: positive and negative values cancel. A deviation score is not a z score unless it has also been divided by a standard deviation. Centering does not make a skewed distribution normal or remove an outlier's influence on the mean. Deviations from a fitted regression prediction are residuals; simple mean deviations coincide with residuals only for an appropriate mean-only model. Changes between a person's two measurements are difference scores, which answer a different question. Avoid silently recomputing a reference mean when comparisons require a fixed baseline.
Business use
Businesses use deviation scores to explain variation and prepare quantitative analyses. Their simplicity makes them useful in training and operational dashboards, provided the mean, population and units are explicit. For machine-learning pipelines, storing the training mean helps reproduce preprocessing and avoids learning transformations from evaluation data. A centered feature can simplify interpretation, but centering alone does not guarantee improved predictions. Review whether differences in size, workload or case mix make a common mean misleading before comparing teams or customers.
People also ask
Must deviation scores be negative?
No. They can be negative, zero or positive depending on an observation's position relative to the reference mean.
Does centering change standard deviation?
Subtracting one constant from every observation leaves its standard deviation unchanged.
Will future observations' deviations average to zero?
Not necessarily. Deviations from a stored training mean sum to zero for the training set, not automatically for later data.
Check your understanding
Why do the four deviations sum to zero while their squared sum equals 56?
Answer: Signed differences cancel around the arithmetic mean. Squaring removes their signs and preserves evidence of spread. The two totals describe different properties, so zero signed total cannot justify saying all processing times are identical.
Related terms
- Mean
- Variance
- Standard deviation
- Z-score
- Covariance
- Residual
- Difference scores
- mean centering
- weighted mean
- reference population
- data preprocessing
Learn next
calculate deviations for a small dataset and verify their signed sum. Compare centering with standardization and distances from a fixed target. Then explore how products of paired deviations form covariance and why a stored training mean should be distinguished from a mean recalculated on new observations.
More in Statistics and probability
- Deterministic imputation - Deterministic imputation, in the statistical-editing sense used by the OECD glossary, fills a missing or designated field when the other accepted values and...
- Deviance information criterion - The deviance information criterion, or DIC, is a Bayesian model-comparison measure that combines posterior average fit with an estimate of effective model...
- Deviance (statistics) - Deviance is a likelihood-based measure of how much a fitted statistical model falls short of a saturated model that can fit the observed responses exactly.
- Dichotomous thinking - Dichotomous thinking reduces a question to two opposing categories while overlooking meaningful degrees, uncertainty or alternative interpretations.
- Dichotomous variables - Dichotomous variables are variables with two possible categories or values.
- Differential entropy - Differential entropy is an information-theoretic quantity defined for a continuous probability distribution with a density.
Sources
David Howell, University of Vermont, statistics glossary: https://www.uvm.edu/~statdhtx/fundamentals9/Glossary/Glossary.html NIST/SEMATECH, Measures of Scale: https://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm NIST/SEMATECH, One-way ANOVA overview: https://www.itl.nist.gov/div898/handbook/prc/section4/prc431.htm
Previous: Deviance (statistics) Next: Diagnosis