The weighted arithmetic mean is similar to an ordinary arithmetic mean (the most common type of average), except that instead of each of the data points contributing equally to the final average, some data points contribute more than others. The notion of weighted mean plays a role in descriptive statistics and also occurs in a more general form in several other areas of mathematics.
If all the weights are equal, then the weighted mean is the same as the arithmetic mean. While weighted means generally behave in a similar fashion to arithmetic means, they do have a few counterintuitive properties, as captured for instance in Simpson's paradox.
Given two school classes, one with 20 students, and one with 30 students, the grades in each class on a test were:
- Morning class = 62, 67, 71, 74, 76, 77, 78, 79, 79, 80, 80, 81, 81, 82, 83, 84, 86, 89, 93, 98
- Afternoon class = 81, 82, 83, 84, 85, 86, 87, 87, 88, 88, 89, 89, 89, 90, 90, 90, 90, 91, 91, 91, 92, 92, 93, 93, 94, 95, 96, 97, 98, 99
The mean for the morning class is 80 and the mean of the afternoon class is 90. The unweighted mean of the 80 and 90 is 85, so the unweighted mean of the two means is 85. However, this does not account for the difference in number of students in each class (20 versus 30); hence the value of 85 does not reflect the average student grade (independent of class). The average student grade can be obtained by averaging all the grades, without regard to classes (add all the grades up and divide by the total number of students):
Or, this can be accomplished by weighting the class means by the number of students in each class. The larger class is given more "weight":
Thus, the weighted mean makes it possible to find the mean average student grade without knowing each student's score. Only the class means and the number of students in each class are needed.
Convex combination example
Since only the relative weights are relevant, any weighted mean can be expressed using coefficients that sum to one. Such a linear combination is called a convex combination.
Using the previous example, we would get the following weights:
Then, apply the weights like this:
which expands to:
Therefore, data elements with a high weight contribute more to the weighted mean than do elements with a low weight. The weights cannot be negative. Some may be zero, but not all of them (since division by zero is not allowed).
The formulas are simplified when the weights are normalized such that they sum up to , i.e.:
For such normalized weights the weighted mean is then:
Note that one can always normalize the weights by making the following transformation on the original weights:
Using the normalized weight yields the same results as when using the original weights:
The ordinary mean is a special case of the weighted mean where all data have equal weights, .
The standard error of the weighted mean (unit input variances), can be shown via uncertainty propagation to be:
The weighted sample mean, , is itself a random variable. Its expected value and standard deviation are related to the expected values and standard deviations of the observations, as follows. For simplicity, we assume normalized weights (weights summing to one).
If the observations have expected values
then the weighted sample mean has expectation
In particular, if the means are equal, , then the expectation of the weighted sample mean will be that value,
For uncorrelated observations with variances , the variance of the weighted sample mean is
whose square root can be called the standard error of the weighted mean (general case).
Consequently, if all the observations have equal variance, , the weighted sample mean will have variance
where . The variance attains its maximum value, , when all weights except one are zero. Its minimum value is found when all weights are equal (i.e., unweighted mean), in which case we have , i.e., it degenerates into the standard error of the mean, squared.
Note that because one can always transform non-normalized weights to normalized weights all formula in this section can be adapted to non-normalized weights by replacing all .
For the weighted mean of a list of data for which each element potentially comes from a different probability distribution with known variance , one possible choice for the weights is given by the reciprocal of variance:
The weighted mean in this case is:
and the standard error of the weighted mean (with variance weights) is:
Note this reduces to when all . It is a special case of the general formula in previous section,
The equations above can be combined to obtain:
The significance of this choice is that this weighted mean is the maximum likelihood estimator of the mean of the probability distributions under the assumption that they are independent and normally distributed with the same mean.
Correcting for over- or under-dispersion
Weighted means are typically used to find the weighted mean of historical data, rather than theoretically generated data. In this case, there will be some error in the variance of each data point. Typically experimental errors may be underestimated due to the experimenter not taking into account all sources of error in calculating the variance of each data point. In this event, the variance in the weighted mean must be corrected to account for the fact that is too large. The correction that must be made is
where is the reduced chi-squared:
The square root can be called the standard error of the weighted mean (variance weights, scale corrected).
When all data variances are equal, , they cancel out in the weighted mean variance, , which again reduces to the standard error of the mean (squared), , formulated in terms of the sample standard deviation (squared),
where . Further simplification leads to
Weighted sample variance
Typically when a mean is calculated it is important to know the variance and standard deviation about that mean. When a weighted mean is used, the variance of the weighted sample is different from the variance of the unweighted sample.
The biased weighted sample variance is defined similarly to the normal biased sample variance :
where , which is 1 for normalized weights. If the weights are frequency weights (and thus are random variables), it can be shown that is the maximum likelihood estimator of for iid Gaussian observations.
For small samples, it is customary to use an unbiased estimator for the population variance. In normal unweighted samples, the N in the denominator (corresponding to the sample size) is changed to N − 1 (see Bessel's correction). In the weighted setting, there are actually two different unbiased estimators, one for the case of frequency weights and another for the case of reliability weights.
If the weights are frequency weights, then the unbiased estimator is:
This effectively applies Bessel's correction for frequency weights.
For example, if values are drawn from the same distribution, then we can treat this set as an unweighted sample, or we can treat it as the weighted sample with corresponding weights , and we get the same result either way.
If the frequency weights are normalized to 1, then the correct expression after Bessel's correction becomes
where the total number of samples is (not ). In any case, the information on total number of samples is necessary in order to obtain an unbiased correction, even if has a different meaning other than frequency weight.
Note that the estimator can be unbiased only if the weights are not standardized nor normalized, these processes changing the data's mean and variance and thus leading to a loss of the base rate (the population count, which is a requirement for Bessel's correction).
If the weights are instead non-random (reliability weights), we can determine a correction factor to yield an unbiased estimator. Assuming each random variable is sampled from the same distribution with mean and actual variance , taking expectations we have,
where . Therefore, the bias in our estimator is , analogous to the bias in the unweighted estimator (also notice that is the effective sample size). This means that to unbias our estimator we need to pre-divide by , ensuring that the expected value of the estimated variance equals the actual variance of the sampling distribution.
The final unbiased estimate of sample variance is:
The degrees of freedom of the weighted, unbiased sample variance vary accordingly from N − 1 down to 0.
The standard deviation is simply the square root of the variance above.
As a side note, other approaches have been described to compute the weighted sample variance.
Weighted sample covariance
In a weighted sample, each row vector (each set of single observations on each of the K random variables) is assigned a weight .
Then the weighted mean vector is given by
And the weighted covariance matrix is given by:
Similarly to weighted sample variance, there are two different unbiased estimators depending on the type of the weights.
If the weights are frequency weights, the unbiased weighted estimate of the covariance matrix , with Bessel's correction, is given by:
Note that this estimator can be unbiased only if the weights are not standardized nor normalized, these processes changing the data's mean and variance and thus leading to a loss of the base rate (the population count, which is a requirement for Bessel's correction).
In the case of reliability weights, the weights are normalized:
(If they are not, divide the weights by their sum to normalize prior to calculating :
Then the weighted mean vector can be simplified to
and the unbiased weighted estimate of the covariance matrix is:
The reasoning here is the same as in the previous section.
Since we are assuming the weights are normalized, then and this reduces to:
If all weights are the same, i.e. , then the weighted mean and covariance reduce to the unweighted sample mean and covariance above.
The above generalizes easily to the case of taking the mean of vector-valued estimates. For example, estimates of position on a plane may have less certainty in one direction than another. As in the scalar case, the weighted mean of multiple estimates can provide a maximum likelihood estimate. We simply replace the variance by the covariance matrix and the arithmetic inverse by the matrix inverse (both denoted in the same way, via superscripts); the weight matrix then reads:
The weighted mean in this case is:
For example, consider the weighted mean of the point [1 0] with high variance in the second component and [0 1] with high variance in the first component. Then
then the weighted mean is:
which makes sense: the [1 0] estimate is "compliant" in the second component and the [0 1] estimate is compliant in the first component, so the weighted mean is nearly [1 1].
Accounting for correlations
In the general case, suppose that , is the covariance matrix relating the quantities , is the common mean to be estimated, and is a design matrix equal to a vector of ones (of length ). The Gauss–Markov theorem states that the estimate of the mean having minimum variance is given by:
Decreasing strength of interactions
Consider the time series of an independent variable and a dependent variable , with observations sampled at discrete times . In many common situations, the value of at time depends not only on but also on its past values. Commonly, the strength of this dependence decreases as the separation of observations in time increases. To model this situation, one may replace the independent variable by its sliding mean for a window size .
Exponentially decreasing weights
In the scenario described in the previous section, most frequently the decrease in interaction strength obeys a negative exponential law. If the observations are sampled at equidistant times, then exponential decrease is equivalent to decrease by a constant fraction at each time step. Setting we can define normalized weights by
where is the sum of the unnormalized weights. In this case is simply
approaching for large values of .
The damping constant must correspond to the actual decrease of interaction strength. If this cannot be determined from theoretical considerations, then the following properties of exponentially decreasing weights are useful in making a suitable choice: at step , the weight approximately equals , the tail area the value , the head area . The tail area at step is . Where primarily the closest observations matter and the effect of the remaining observations can be ignored safely, then choose such that the tail area is sufficiently small.
Weighted averages of functions
- Gatz, Donald F.; Smith, Luther (June 1995). "The standard error of a weighted mean concentration—I. Bootstrapping vs other methods". Atmospheric Environment. 29 (11): 1185–1193. doi:10.1016/1352-2310(94)00210-C.
- "GNU Scientific Library – Reference Manual: Weighted Samples". Gnu.org. Retrieved 22 December 2017.
- "Weighted Standard Error and its Impact on Significance Testing (WinCross vs. Quantum & SPSS), Dr. Albert Madansky" (PDF). Analyticalgroup.com. Retrieved 22 December 2017.
- Price, George R. (April 1972). "Extension of covariance selection mathematics" (PDF). Annals of Human Genetics. 35 (4): 485–490. doi:10.1111/j.1469-1809.1957.tb01874.x.
- Mark Galassi, Jim Davies, James Theiler, Brian Gough, Gerard Jungman, Michael Booth, and Fabrice Rossi. GNU Scientific Library - Reference manual, Version 1.15, 2011. Sec. 21.7 Weighted Samples
- James, Frederick (2006). Statistical Methods in Experimental Physics (2nd ed.). Singapore: World Scientific. p. 324. ISBN 981-270-527-9.
- G. H. Hardy, J. E. Littlewood, and G. Pólya. Inequalities (2nd ed.), Cambridge University Press, ISBN 978-0-521-35880-4, 1988.
- Jane Grossman, Michael Grossman, Robert Katz. The First Systems of Weighted Differential and Integral Calculus, ISBN 0-9771170-1-4, 1980.
- Bevington, Philip R (1969). Data Reduction and Error Analysis for the Physical Sciences. New York, N.Y.: McGraw-Hill. OCLC 300283069.
- Strutz, T. (2010). Data Fitting and Uncertainty (A practical introduction to weighted least squares and beyond). Vieweg+Teubner. ISBN 978-3-8348-1022-9.