Simple i.i.d. case

When treating the weights as constants, and having a sample of n observations from uncorrelated random variables, all with the same variance and expectation (as is the case for i.i.d random variables), then the variance of the weighted mean can be estimated as the multiplication of the unweighted variance by Kish's design effect (see proof):

${\displaystyle \operatorname {Var} ({\bar {y}}_{w})={\hat {\sigma }}_{y}^{2}{\frac {\overline {w^{2}}}{{\bar {w}}^{2}}}}$

With ${\displaystyle {\hat {\sigma }}_{y}^{2}={\frac {\sum _{i=1}^{n}(y_{i}-{\bar {y}})^{2}}{n-1}}}$, ${\displaystyle {\bar {w}}={\frac {\sum _{i=1}^{n}w_{i}}{n}}}$, and ${\displaystyle {\overline {w^{2}}}={\frac {\sum _{i=1}^{n}w_{i}^{2}}{n}}}$

However, this estimation is rather limited due to the strong assumption about the y observations. This has led to the development of alternative, more general, estimators.

Survey sampling perspective

From a model based perspective, we are interested in estimating the variance of the weighted mean when the different ${\displaystyle y_{i}}$ are not i.i.d random variables. An alternative perspective for this problem is that of some arbitrary sampling design of the data in which units are selected with unequal probabilities (with replacement).^[1]: 306

In Survey methodology, the population mean, of some quantity of interest y, is calculated by taking an estimation of the total of y over all elements in the population (Y or sometimes T) and dividing it by the population size – either known (${\displaystyle N}$) or estimated (${\displaystyle {\hat {N}}}$). In this context, each value of y is considered constant, and the variability comes from the selection procedure. This in contrast to "model based" approaches in which the randomness is often described in the y values. The survey sampling procedure yields a series of Bernoulli indicator values (${\displaystyle I_{i}}$) that get 1 if some observation i is in the sample and 0 if it was not selected. This can occur with fixed sample size, or varied sample size sampling (e.g.: Poisson sampling). The probability of some element to be chosen, given a sample, is denoted as ${\displaystyle P(I_{i}=1\mid {\text{Some sample of size }}n)=\pi _{i}}$, and the one-draw probability of selection is ${\displaystyle P(I_{i}=1|{\text{one sample draw}})=p_{i}\approx {\frac {\pi _{i}}{n}}}$ (If N is very large and each ${\displaystyle p_{i}}$ is very small). For the following derivation we'll assume that the probability of selecting each element is fully represented by these probabilities.^{[2]: 42, 43, 51} I.e.: selecting some element will not influence the probability of drawing another element (this doesn't apply for things such as cluster sampling design).

Since each element (${\displaystyle y_{i}}$) is fixed, and the randomness comes from it being included in the sample or not (${\displaystyle I_{i}}$), we often talk about the multiplication of the two, which is a random variable. To avoid confusion in the following section, let's call this term: ${\displaystyle y'_{i}=y_{i}I_{i}}$. With the following expectancy: ${\displaystyle E[y'_{i}]=y_{i}E[I_{i}]=y_{i}\pi _{i}}$; and variance: ${\displaystyle V[y'_{i}]=y_{i}^{2}V[I_{i}]=y_{i}^{2}\pi _{i}(1-\pi _{i})}$.

When each element of the sample is inflated by the inverse of its selection probability, it is termed the ${\displaystyle \pi }$-expanded y values, i.e.: ${\displaystyle {\check {y}}_{i}={\frac {y_{i}}{\pi _{i}}}}$. A related quantity is ${\displaystyle p}$-expanded y values: ${\displaystyle {\frac {y_{i}}{p_{i}}}=n{\check {y}}_{i}}$.^{[2]: 42, 43, 51, 52} As above, we can add a tick mark if multiplying by the indicator function. I.e.: ${\displaystyle {\check {y}}'_{i}=I_{i}{\check {y}}_{i}={\frac {I_{i}y_{i}}{\pi _{i}}}}$

In this design based perspective, the weights, used in the numerator of the weighted mean, are obtained from taking the inverse of the selection probability (i.e.: the inflation factor). I.e.: ${\displaystyle w_{i}={\frac {1}{\pi _{i}}}\approx {\frac {1}{n\times p_{i}}}}$.

Variance of the weighted sum (pwr-estimator for totals)

If the population size N is known we can estimate the population mean using ${\displaystyle {\hat {\bar {Y}}}_{{\text{known }}N}={\frac {{\hat {Y}}_{pwr}}{N}}\approx {\frac {\sum _{i=1}^{n}w_{i}y'_{i}}{N}}}$.

If the sampling design is one that results in a fixed sample size n (such as in pps sampling), then the variance of this estimator is:

${\displaystyle \operatorname {Var} \left({\hat {\bar {Y}}}_{{\text{known }}N}\right)={\frac {1}{N^{2}}}{\frac {n}{n-1}}\sum _{i=1}^{n}\left(w_{i}y_{i}-{\overline {wy}}\right)^{2}}$

Proof

The general formula can be developed like this:

${\displaystyle {\hat {\bar {Y}}}_{{\text{known }}N}={\frac {{\hat {Y}}_{pwr}}{N}}={\frac {{\frac {1}{n}}\sum _{i=1}^{n}{\frac {y'_{i}}{p_{i}}}}{N}}\approx {\frac {\sum _{i=1}^{n}{\frac {y'_{i}}{\pi _{i}}}}{N}}={\frac {\sum _{i=1}^{n}w_{i}y'_{i}}{N}}.}$

The population total is denoted as ${\displaystyle Y=\sum _{i=1}^{N}y_{i}}$ and it may be estimated by the (unbiased) Horvitz–Thompson estimator, also called the ${\displaystyle \pi }$-estimator. This estimator can be itself estimated using the pwr-estimator (i.e.: ${\displaystyle p}$-expanded with replacement estimator, or "probability with replacement" estimator). With the above notation, it is: ${\displaystyle {\hat {Y}}_{pwr}={\frac {1}{n}}\sum _{i=1}^{n}{\frac {y'_{i}}{p_{i}}}=\sum _{i=1}^{n}{\frac {y'_{i}}{np_{i}}}\approx \sum _{i=1}^{n}{\frac {y'_{i}}{\pi _{i}}}=\sum _{i=1}^{n}w_{i}y'_{i}}$.^[2]: 51

The estimated variance of the pwr-estimator is given by:^[2]: 52

$${\displaystyle \operatorname {Var} ({\hat {Y}}_{pwr})={\frac {n}{n-1}}\sum _{i=1}^{n}\left(w_{i}y_{i}-{\overline {wy}}\right)^{2}}$$

where ${\displaystyle {\overline {wy}}=\sum _{i=1}^{n}{\frac {w_{i}y_{i}}{n}}}$.

The above formula was taken from Sarndal et al. (1992) (also presented in Cochran 1977), but was written differently.^{[2]: 52 [1]: 307 (11.35)} The left side is how the variance was written and the right side is how we've developed the weighted version:

$${\displaystyle {\begin{aligned}\operatorname {Var} ({\hat {Y}}_{\text{pwr}})&={\frac {1}{n}}{\frac {1}{n-1}}\sum _{i=1}^{n}\left({\frac {y_{i}}{p_{i}}}-{\hat {Y}}_{pwr}\right)^{2}\\&={\frac {1}{n}}{\frac {1}{n-1}}\sum _{i=1}^{n}\left({\frac {n}{n}}{\frac {y_{i}}{p_{i}}}-{\frac {n}{n}}\sum _{i=1}^{n}w_{i}y_{i}\right)^{2}={\frac {1}{n}}{\frac {1}{n-1}}\sum _{i=1}^{n}\left(n{\frac {y_{i}}{\pi _{i}}}-n{\frac {\sum _{i=1}^{n}w_{i}y_{i}}{n}}\right)^{2}\\&={\frac {n^{2}}{n}}{\frac {1}{n-1}}\sum _{i=1}^{n}\left(w_{i}y_{i}-{\overline {wy}}\right)^{2}\\&={\frac {n}{n-1}}\sum _{i=1}^{n}\left(w_{i}y_{i}-{\overline {wy}}\right)^{2}\end{aligned}}}$$

And we got to the formula from above.

An alternative term, for when the sampling has a random sample size (as in Poisson sampling), is presented in Sarndal et al. (1992) as:^[2]: 182

$${\displaystyle \operatorname {Var} ({\hat {\bar {Y}}}_{{\text{pwr (known }}N{\text{)}}})={\frac {1}{N^{2}}}\sum _{i=1}^{n}\sum _{j=1}^{n}\left({\check {\Delta }}_{ij}{\check {y}}_{i}{\check {y}}_{j}\right)}$$

With ${\displaystyle {\check {y}}_{i}={\frac {y_{i}}{\pi _{i}}}}$. Also, ${\displaystyle C(I_{i},I_{j})=\pi _{ij}-\pi _{i}\pi _{j}=\Delta _{ij}}$ where ${\displaystyle \pi _{ij}}$ is the probability of selecting both i and j.^[2]: 36 And ${\displaystyle {\check {\Delta }}_{ij}=1-{\frac {\pi _{i}\pi _{j}}{\pi _{ij}}}}$, and for i=j: ${\displaystyle {\check {\Delta }}_{ii}=1-{\frac {\pi _{i}\pi _{i}}{\pi _{i}}}=1-\pi _{i}}$.^[2]: 43

If the selection probability are uncorrelated (i.e.: ${\displaystyle \forall i\neq j:C(I_{i},I_{j})=0}$), and when assuming the probability of each element is very small, then:

${\displaystyle \operatorname {Var} ({\hat {\bar {Y}}}_{{\text{pwr (known }}N{\text{)}}})={\frac {1}{N^{2}}}\sum _{i=1}^{n}\left(w_{i}y_{i}\right)^{2}}$

Proof

We assume that ${\displaystyle (1-\pi _{i})\approx 1}$ and that

$${\displaystyle {\begin{aligned}\operatorname {Var} ({\hat {Y}}_{{\text{pwr (known }}N{\text{)}}})&={\frac {1}{N^{2}}}\sum _{i=1}^{n}\sum _{j=1}^{n}\left({\check {\Delta }}_{ij}{\check {y}}_{i}{\check {y}}_{j}\right)\\&={\frac {1}{N^{2}}}\sum _{i=1}^{n}\left({\check {\Delta }}_{ii}{\check {y}}_{i}{\check {y}}_{i}\right)\\&={\frac {1}{N^{2}}}\sum _{i=1}^{n}\left((1-\pi _{i}){\frac {y_{i}}{\pi _{i}}}{\frac {y_{i}}{\pi _{i}}}\right)\\&={\frac {1}{N^{2}}}\sum _{i=1}^{n}\left(w_{i}y_{i}\right)^{2}\end{aligned}}}$$

Variance of the weighted mean (π-estimator for ratio-mean)

The previous section dealt with estimating the population mean as a ratio of an estimated population total (${\displaystyle {\hat {Y}}}$) with a known population size (${\displaystyle N}$), and the variance was estimated in that context. Another common case is that the population size itself (${\displaystyle N}$) is unknown and is estimated using the sample (i.e.: ${\displaystyle {\hat {N}}}$). The estimation of ${\displaystyle N}$ can be described as the sum of weights. So when ${\displaystyle w_{i}={\frac {1}{\pi _{i}}}}$ we get ${\displaystyle {\hat {N}}=\sum _{i=1}^{n}w_{i}I_{i}=\sum _{i=1}^{n}{\frac {I_{i}}{\pi _{i}}}=\sum _{i=1}^{n}{\check {1}}'_{i}}$. With the above notation, the parameter we care about is the ratio of the sums of ${\displaystyle y_{i}}$s, and 1s. I.e.: ${\displaystyle R={\bar {Y}}={\frac {\sum _{i=1}^{N}{\frac {y_{i}}{\pi _{i}}}}{\sum _{i=1}^{N}{\frac {1}{\pi _{i}}}}}={\frac {\sum _{i=1}^{N}{\check {y}}_{i}}{\sum _{i=1}^{N}{\check {1}}_{i}}}={\frac {\sum _{i=1}^{N}w_{i}y_{i}}{\sum _{i=1}^{N}w_{i}}}}$. We can estimate it using our sample with: ${\displaystyle {\hat {R}}={\hat {\bar {Y}}}={\frac {\sum _{i=1}^{N}I_{i}{\frac {y_{i}}{\pi _{i}}}}{\sum _{i=1}^{N}I_{i}{\frac {1}{\pi _{i}}}}}={\frac {\sum _{i=1}^{N}{\check {y}}'_{i}}{\sum _{i=1}^{N}{\check {1}}'_{i}}}={\frac {\sum _{i=1}^{N}w_{i}y'_{i}}{\sum _{i=1}^{N}w_{i}1'_{i}}}={\frac {\sum _{i=1}^{n}w_{i}y'_{i}}{\sum _{i=1}^{n}w_{i}1'_{i}}}={\bar {y}}_{w}}$. As we moved from using N to using n, we actually know that all the indicator variables get 1, so we could simply write: ${\displaystyle {\bar {y}}_{w}={\frac {\sum _{i=1}^{n}w_{i}y_{i}}{\sum _{i=1}^{n}w_{i}}}}$. This will be the estimand for specific values of y and w, but the statistical properties comes when including the indicator variable ${\displaystyle {\bar {y}}_{w}={\frac {\sum _{i=1}^{n}w_{i}y'_{i}}{\sum _{i=1}^{n}w_{i}1'_{i}}}}$.^{[2]: 162, 163, 176}

This is called a Ratio estimator and it is approximately unbiased for R.^[2]: 182

In this case, the variability of the ratio depends on the variability of the random variables both in the numerator and the denominator - as well as their correlation. Since there is no closed analytical form to compute this variance, various methods are used for approximate estimation. Primarily Taylor series first-order linearization, asymptotics, and bootstrap/jackknife.^[2]: 172 The Taylor linearization method could lead to under-estimation of the variance for small sample sizes in general, but that depends on the complexity of the statistic. For the weighted mean, the approximate variance is supposed to be relatively accurate even for medium sample sizes.^[2]: 176 For when the sampling has a random sample size (as in Poisson sampling), it is as follows:^[2]: 182

${\displaystyle {\widehat {V({\bar {y}}_{w})}}={\frac {1}{(\sum _{i=1}^{n}w_{i})^{2}}}\sum _{i=1}^{n}w_{i}^{2}(y_{i}-{\bar {y}}_{w})^{2}}$.

If ${\displaystyle \pi _{i}\approx p_{i}n}$, then either using ${\displaystyle w_{i}={\frac {1}{\pi _{i}}}}$ or ${\displaystyle w_{i}={\frac {1}{p_{i}}}}$ would give the same estimator, since multiplying ${\displaystyle w_{i}}$ by some factor would lead to the same estimator. It also means that if we scale the sum of weights to be equal to a known-from-before population size N, the variance calculation would look the same. When all weights are equal to one another, this formula is reduced to the standard unbiased variance estimator.

Proof

The Taylor linearization states that for a general ratio estimator of two sums (${\displaystyle {\hat {R}}={\frac {\hat {Y}}{\hat {Z}}}}$), they can be expanded around the true value R, and give:^[2]: 178

$${\displaystyle {\hat {R}}={\frac {\hat {Y}}{\hat {Z}}}={\frac {\sum _{i=1}^{n}w_{i}y'_{i}}{\sum _{i=1}^{n}w_{i}z'_{i}}}\approx R+{\frac {1}{Z}}\sum _{i=1}^{n}\left({\frac {y'_{i}}{\pi _{i}}}-R{\frac {z'_{i}}{\pi _{i}}}\right)}$$

And the variance can be approximated by:^{[2]: 178, 179}

$${\displaystyle {\widehat {V({\hat {R}})}}={\frac {1}{{\hat {Z}}^{2}}}\sum _{i=1}^{n}\sum _{j=1}^{n}\left({\check {\Delta }}_{ij}{\frac {y_{i}-{\hat {R}}z_{i}}{\pi _{i}}}{\frac {y_{j}-{\hat {R}}z_{j}}{\pi _{j}}}\right)={\frac {1}{{\hat {Z}}^{2}}}\left[{\widehat {V({\hat {Y}})}}+{\hat {R}}{\widehat {V({\hat {Z}})}}-2{\hat {R}}{\hat {C}}({\hat {Y}},{\hat {Z}})\right]}$$

.

The term ${\displaystyle {\hat {C}}({\hat {Y}},{\hat {Z}})}$ is the estimated covariance between the estimated sum of Y and estimated sum of Z. Since this is the covariance of two sums of random variables, it would include many combinations of covariances that will depend on the indicator variables. If the selection probability are uncorrelated (i.e.: ${\displaystyle \forall i\neq j:\Delta _{ij}=C(I_{i},I_{j})=0}$), this term would still include a summation of n covariances for each element i between ${\displaystyle y'_{i}=I_{i}y_{i}}$ and ${\displaystyle z'_{i}=I_{i}z_{i}}$. This helps illustrate that this formula incorporates the effect of correlation between y and z on the variance of the ratio estimators.

When defining ${\displaystyle z_{i}=1}$ the above becomes:^[2]: 182

$${\displaystyle {\widehat {V({\hat {R}})}}={\widehat {V({\bar {y}}_{w})}}={\frac {1}{{\hat {N}}^{2}}}\sum _{i=1}^{n}\sum _{j=1}^{n}\left({\check {\Delta }}_{ij}{\frac {y_{i}-{\bar {y}}_{w}}{\pi _{i}}}{\frac {y_{j}-{\bar {y}}_{w}}{\pi _{j}}}\right).}$$

If the selection probability are uncorrelated (i.e.: ${\displaystyle \forall i\neq j:\Delta _{ij}=C(I_{i},I_{j})=0}$), and when assuming the probability of each element is very small (i.e.: ${\displaystyle (1-\pi _{i})\approx 1}$), then the above reduced to the following:

$${\displaystyle {\widehat {V({\bar {y}}_{w})}}={\frac {1}{{\hat {N}}^{2}}}\sum _{i=1}^{n}\left((1-\pi _{i}){\frac {y_{i}-{\bar {y}}_{w}}{\pi _{i}}}\right)^{2}={\frac {1}{(\sum _{i=1}^{n}w_{i})^{2}}}\sum _{i=1}^{n}w_{i}^{2}(y_{i}-{\bar {y}}_{w})^{2}.}$$

A similar re-creation of the proof (up to some mistakes at the end) was provided by Thomas Lumley in crossvalidated.^[3]

We have (at least) two versions of variance for the weighted mean: one with known and one with unknown population size estimation. There is no uniformly better approach, but the literature presents several arguments to prefer using the population estimation version (even when the population size is known).^[2]: 188 For example: if all y values are constant, the estimator with unknown population size will give the correct result, while the one with known population size will have some variability. Also, when the sample size itself is random (e.g.: in Poisson sampling), the version with unknown population mean is considered more stable. Lastly, if the proportion of sampling is negatively correlated with the values (i.e.: smaller chance to sample an observation that is large), then the un-known population size version slightly compensates for that.

For the trivial case in which all the weights are equal to 1, the above formula is just like the regular formula for the variance of the mean (but notice that it uses the maximum likelihood estimator for the variance instead of the unbiased variance. I.e.: dividing it by n instead of (n-1)).

Bootstrapping validation

It has been shown, by Gatz et al. (1995), that in comparison to bootstrapping methods, the following (variance estimation of ratio-mean using Taylor series linearization) is a reasonable estimation for the square of the standard error of the mean (when used in the context of measuring chemical constituents):^[4]: 1186

${\displaystyle {\widehat {\sigma _{{\bar {x}}_{w}}^{2}}}={\frac {n}{(n-1)(n{\bar {w}})^{2}}}\left[\sum (w_{i}x_{i}-{\bar {w}}{\bar {x}}_{w})^{2}-2{\bar {x}}_{w}\sum (w_{i}-{\bar {w}})(w_{i}x_{i}-{\bar {w}}{\bar {x}}_{w})+{\bar {x}}_{w}^{2}\sum (w_{i}-{\bar {w}})^{2}\right]}$

where ${\displaystyle {\bar {w}}={\frac {\sum w_{i}}{n}}}$. Further simplification leads to

${\displaystyle {\widehat {\sigma _{\bar {x}}^{2}}}={\frac {n}{(n-1)(n{\bar {w}})^{2}}}\sum w_{i}^{2}(x_{i}-{\bar {x}}_{w})^{2}}$

Gatz et al. mention that the above formulation was published by Endlich et al. (1988) when treating the weighted mean as a combination of a weighted total estimator divided by an estimator of the population size,^[5] based on the formulation published by Cochran (1977), as an approximation to the ratio mean. However, Endlich et al. didn't seem to publish this derivation in their paper (even though they mention they used it), and Cochran's book includes a slightly different formulation.^[1]: 155 Still, it's almost identical to the formulations described in previous sections.

Replication-based estimators

Because there is no closed analytical form for the variance of the weighted mean, it was proposed in the literature to rely on replication methods such as the Jackknife and Bootstrapping.^[1]: 321

Other notes

For uncorrelated observations with variances ${\displaystyle \sigma _{i}^{2}}$, the variance of the weighted sample mean is

${\displaystyle \sigma _{\bar {x}}^{2}=\sum _{i=1}^{n}{w_{i}'^{2}\sigma _{i}^{2}}}$

whose square root ${\displaystyle \sigma _{\bar {x}}}$ can be called the standard error of the weighted mean (general case).

Consequently, if all the observations have equal variance, ${\displaystyle \sigma _{i}^{2}=\sigma _{0}^{2}}$, the weighted sample mean will have variance

${\displaystyle \sigma _{\bar {x}}^{2}=\sigma _{0}^{2}\sum _{i=1}^{n}{w_{i}'^{2}},}$

where ${\textstyle 1/n\leq \sum _{i=1}^{n}{w_{i}'^{2}}\leq 1}$. The variance attains its maximum value, ${\displaystyle \sigma _{0}^{2}}$, 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 ${\textstyle \sigma _{\bar {x}}=\sigma _{0}/{\sqrt {n}}}$, i.e., it degenerates into the standard error of the mean, squared.

Because one can always transform non-normalized weights to normalized weights, all formulas in this section can be adapted to non-normalized weights by replacing all ${\displaystyle w_{i}'={\frac {w_{i}}{\sum _{i=1}^{n}{w_{i}}}}}$.

Frequency weights

If the weights are frequency weights (where a weight equals the number of occurrences), then the unbiased estimator is:

${\displaystyle s^{2}\ ={\frac {\sum \limits _{i=1}^{N}w_{i}\left(x_{i}-\mu ^{*}\right)^{2}}{\sum _{i=1}^{N}w_{i}-1}}}$

This effectively applies Bessel's correction for frequency weights.

For example, if values ${\displaystyle \{2,2,4,5,5,5\}}$ 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 ${\displaystyle \{2,4,5\}}$ with corresponding weights ${\displaystyle \{2,1,3\}}$, and we get the same result either way.

If the frequency weights ${\displaystyle \{w_{i}\}}$ are normalized to 1, then the correct expression after Bessel's correction becomes

${\displaystyle s^{2}\ ={\frac {\sum _{i=1}^{N}w_{i}}{\sum _{i=1}^{N}w_{i}-1}}\sum _{i=1}^{N}w_{i}\left(x_{i}-\mu ^{*}\right)^{2}}$

where the total number of samples is ${\displaystyle \sum _{i=1}^{N}w_{i}}$ (not ${\displaystyle N}$). In any case, the information on total number of samples is necessary in order to obtain an unbiased correction, even if ${\displaystyle w_{i}}$ has a different meaning other than frequency weight.

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).

Reliability weights

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 ${\displaystyle \mu }$ and actual variance ${\displaystyle \sigma _{\text{actual}}^{2}}$, taking expectations we have,

${\displaystyle {\begin{aligned}\operatorname {E} [{\hat {\sigma }}^{2}]&={\frac {\sum \limits _{i=1}^{N}\operatorname {E} [(x_{i}-\mu )^{2}]}{N}}\\&=\operatorname {E} [(X-\operatorname {E} [X])^{2}]-{\frac {1}{N}}\operatorname {E} [(X-\operatorname {E} [X])^{2}]\\&=\left({\frac {N-1}{N}}\right)\sigma _{\text{actual}}^{2}\\\operatorname {E} [{\hat {\sigma }}_{\mathrm {w} }^{2}]&={\frac {\sum \limits _{i=1}^{N}w_{i}\operatorname {E} [(x_{i}-\mu ^{*})^{2}]}{V_{1}}}\\&=\operatorname {E} [(X-\operatorname {E} [X])^{2}]-{\frac {V_{2}}{V_{1}^{2}}}\operatorname {E} [(X-\operatorname {E} [X])^{2}]\\&=\left(1-{\frac {V_{2}}{V_{1}^{2}}}\right)\sigma _{\text{actual}}^{2}\end{aligned}}}$

where ${\displaystyle V_{1}=\sum _{i=1}^{N}w_{i}}$ and ${\displaystyle V_{2}=\sum _{i=1}^{N}w_{i}^{2}}$. Therefore, the bias in our estimator is ${\displaystyle \left(1-{\frac {V_{2}}{V_{1}^{2}}}\right)}$, analogous to the ${\displaystyle \left({\frac {N-1}{N}}\right)}$ bias in the unweighted estimator (also notice that ${\displaystyle \ V_{1}^{2}/V_{2}=N_{eff}}$ is the effective sample size). This means that to unbias our estimator we need to pre-divide by ${\displaystyle 1-\left(V_{2}/V_{1}^{2}\right)}$, 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:

${\displaystyle {\begin{aligned}s_{\mathrm {w} }^{2}\ &={\frac {{\hat {\sigma }}_{\mathrm {w} }^{2}}{1-(V_{2}/V_{1}^{2})}}\\[4pt]&={\frac {\sum \limits _{i=1}^{N}w_{i}(x_{i}-\mu ^{*})^{2}}{V_{1}-(V_{2}/V_{1})}},\end{aligned}}}$^[6]

where ${\displaystyle \operatorname {E} [s_{\mathrm {w} }^{2}]=\sigma _{\text{actual}}^{2}}$.

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.^[7]

Frequency weights

If the weights are frequency weights, the unbiased weighted estimate of the covariance matrix ${\displaystyle \textstyle \mathbf {C} }$, with Bessel's correction, is given by:^[8]

${\displaystyle \mathbf {C} ={\frac {\sum _{i=1}^{N}w_{i}\left(\mathbf {x} _{i}-\mu ^{*}\right)^{T}\left(\mathbf {x} _{i}-\mu ^{*}\right)}{V_{1}-1}}.}$

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).

Reliability weights

In the case of reliability weights, the weights are normalized:

${\displaystyle V_{1}=\sum _{i=1}^{N}w_{i}=1.}$

(If they are not, divide the weights by their sum to normalize prior to calculating ${\displaystyle V_{1}}$:

${\displaystyle w_{i}'={\frac {w_{i}}{\sum _{i=1}^{N}w_{i}}}}$

Then the weighted mean vector ${\displaystyle \mathbf {\mu ^{*}} }$ can be simplified to

${\displaystyle \mathbf {\mu ^{*}} =\sum _{i=1}^{N}w_{i}\mathbf {x} _{i}.}$

and the unbiased weighted estimate of the covariance matrix ${\displaystyle \mathbf {C} }$ is:^[9]

${\displaystyle {\begin{aligned}\mathbf {C} &={\frac {\sum _{i=1}^{N}w_{i}}{\left(\sum _{i=1}^{N}w_{i}\right)^{2}-\sum _{i=1}^{N}w_{i}^{2}}}\sum _{i=1}^{N}w_{i}\left(\mathbf {x} _{i}-\mu ^{*}\right)^{T}\left(\mathbf {x} _{i}-\mu ^{*}\right)\\&={\frac {\sum _{i=1}^{N}w_{i}\left(\mathbf {x} _{i}-\mu ^{*}\right)^{T}\left(\mathbf {x} _{i}-\mu ^{*}\right)}{V_{1}-(V_{2}/V_{1})}}.\end{aligned}}}$

The reasoning here is the same as in the previous section.

Since we are assuming the weights are normalized, then ${\displaystyle V_{1}=1}$ and this reduces to:

${\displaystyle \mathbf {C} ={\frac {\sum _{i=1}^{N}w_{i}\left(\mathbf {x} _{i}-\mu ^{*}\right)^{T}\left(\mathbf {x} _{i}-\mu ^{*}\right)}{1-V_{2}}}.}$

If all weights are the same, i.e. ${\displaystyle w_{i}/V_{1}=1/N}$, then the weighted mean and covariance reduce to the unweighted sample mean and covariance above.