The weighted mean is an average in which individual values contribute unequally according to assigned weights, allowing observations with greater frequency, importance, exposure, probability, sample size, or statistical precision to influence the result more strongly than observations with smaller weights. Instead of adding every value and dividing by the number of values as in the ordinary arithmetic mean, the weighted mean multiplies each value by its corresponding weight, adds those weighted products, and divides by the total weight. The general formula is x̄w = Σwᵢxᵢ/Σwᵢ, and when the weights have already been normalized so that they sum to 1, the formula simplifies to x̄w = Σwᵢxᵢ. Weighted means appear throughout statistics, probability, education, survey analysis, finance, measurement, meta-analysis, grouped data, and any calculation in which some observations represent more of the total than others. Their usefulness depends entirely on whether the weights have a defensible interpretation, because inappropriate weights can produce a numerically correct calculation that answers the wrong question.
Within descriptive statistics and the broader Statistics & Probability framework, the weighted mean generalizes the arithmetic mean while preserving its basic balancing interpretation. When every observation receives equal weight, the weighted mean reduces exactly to the ordinary arithmetic mean, but when weights differ, the result moves toward values carrying greater total weight.
What Is a Weighted Mean?
A weighted mean is an average calculated by giving different values different levels of contribution.
Suppose two values are:
60 and 90
but the second value should count three times as much as the first.
Using weights:
1 and 3
the weighted mean is:
x̄w = [1(60) + 3(90)]/(1 + 3)
= (60 + 270)/4
= 330/4
= 82.5
The ordinary mean would be:
(60 + 90)/2
= 75
but that treats the two values as equally important.
The weighted mean correctly reflects the intended 1:3 contribution.
Weighted Mean Formula
The general weighted mean formula is:
x̄w = Σwᵢxᵢ / Σwᵢ
where:
- xᵢ = individual value
- wᵢ = weight assigned to that value
- Σwᵢxᵢ = sum of weighted values
- Σwᵢ = total weight
- x̄w = weighted mean
The denominator is the sum of the weights, not automatically the number of observations.
That distinction is the core difference between weighted and ordinary averaging.
Weighted Mean With Normalized Weights
If the weights satisfy:
Σwᵢ = 1
then:
x̄w = Σwᵢxᵢ
because dividing by 1 changes nothing.
For example, suppose:
x₁ = 50, w₁ = 0.20
x₂ = 70, w₂ = 0.30
x₃ = 90, w₃ = 0.50.
Since:
0.20 + 0.30 + 0.50 = 1,
calculate:
x̄w = 0.20(50) + 0.30(70) + 0.50(90)
= 10 + 21 + 45
= 76.
Therefore:
Weighted mean = 76.
Why Weights Matter
Weights determine how much influence each value has on the final average.
Suppose values are:
10 and 100.
With equal weights:
Mean = 55.
With weights:
9 and 1,
the weighted mean becomes:
[9(10) + 1(100)]/10
= 190/10
= 19.
The value 10 dominates because it represents:
90%
of the total weight.
The weighted mean therefore answers a different question from the unweighted mean.
How to Calculate a Weighted Mean Step by Step
Begin by listing each value and its corresponding weight.
Multiply:
wᵢ × xᵢ
for every pair.
Add the products:
Σwᵢxᵢ.
Add the weights:
Σwᵢ.
Then divide:
x̄w = Σwᵢxᵢ/Σwᵢ.
Finally, check whether the result is reasonable given the values carrying the largest weights.
For nonnegative weights, the weighted mean should lie between the smallest and largest values receiving positive weight.
Simple Weighted Mean Example
Suppose three scores are:
70, 80, 95
with weights:
1, 2, 3.
Calculate the weighted products:
1(70) = 70
2(80) = 160
3(95) = 285.
Sum:
70 + 160 + 285
= 515.
Total weight:
1 + 2 + 3
= 6.
Therefore:
x̄w = 515/6
≈ 85.83.
The result lies closer to 95 because 95 has the largest weight.
Weighted Mean vs Arithmetic Mean
The arithmetic mean is:
x̄ = Σxᵢ/n.
The weighted mean is:
x̄w = Σwᵢxᵢ/Σwᵢ.
If:
w₁ = w₂ = … = wₙ = c
for any positive constant c, then:
x̄w = cΣxᵢ/(nc)
= Σxᵢ/n
= x̄.
Therefore, the arithmetic mean is a special case of the weighted mean in which every observation receives equal weight.
Equal Weights Example
Suppose:
x = 4, 8, 12
and all weights equal:
w = 5, 5, 5.
Then:
x̄w = [5(4) + 5(8) + 5(12)]/(5 + 5 + 5)
= (20 + 40 + 60)/15
= 120/15
= 8.
The ordinary mean is:
(4 + 8 + 12)/3
= 8.
Multiplying all equal weights by the same constant does not change the answer.
Scaling All Weights by the Same Constant
Weighted means are unchanged when every weight is multiplied by the same positive constant.
Suppose:
x̄w = Σwᵢxᵢ/Σwᵢ.
Replace each weight with:
cwᵢ.
Then:
x̄w = Σcwᵢxᵢ/Σcwᵢ
= cΣwᵢxᵢ/cΣwᵢ
= Σwᵢxᵢ/Σwᵢ.
Therefore, weights:
1, 2, 3
and:
10, 20, 30
produce the same weighted mean.
Only their relative proportions matter.
Normalizing Weights
Weights that do not sum to 1 can be converted into normalized weights.
Define:
pᵢ = wᵢ/Σwᵢ.
Then:
Σpᵢ = 1.
The weighted mean becomes:
x̄w = Σpᵢxᵢ.
For weights:
2, 3, 5
the total is:
Normalized weights are:
0.20, 0.30, 0.50.
This conversion changes the representation of the weights but not the weighted mean.
Percentage Weights
Percentage weights are normalized weights expressed on a 0% to 100% scale.
Suppose a final grade is based on:
Homework = 20%
Midterm = 30%
Final exam = 50%.
The percentages sum to:
100%.
As decimal weights:
0.20, 0.30, 0.50.
If scores are:
Homework = 80
Midterm = 70
Final = 90,
then:
Weighted grade = 0.20(80) + 0.30(70) + 0.50(90)
= 16 + 21 + 45
= 82.
Therefore:
Final weighted grade = 82.
Percentage Weights That Do Not Sum to 100%
Suppose reported weights are:
20%, 30%, 40%.
Their total is:
90%.
If these are intended as relative weights rather than literal shares, calculate:
x̄w = [20x₁ + 30x₂ + 40x₃]/90.
Do not divide by:
100
unless the remaining 10% represents another explicitly defined component.
The denominator must equal the total weight actually included.
Weighted Mean From Frequencies
A frequency table naturally produces a weighted mean because each distinct value represents multiple observations.
If value xᵢ occurs fᵢ times:
x̄ = Σfᵢxᵢ/Σfᵢ.
Here:
frequency = weight.
This formula gives exactly the same arithmetic mean that would be obtained by writing every repeated observation individually.
The frequency-weighted form is simply more efficient.
Frequency-Weighted Example
Suppose:
| Value | Frequency |
|---|---|
| 10 | 2 |
| 20 | 5 |
| 30 | 3 |
Total frequency:
n = 2 + 5 + 3
= 10.
Weighted total:
10(2) + 20(5) + 30(3)
= 20 + 100 + 90
= 210.
Therefore:
x̄ = 210/10
= 21.
The value 20 has the greatest influence because it occurs most frequently.
Weighted Mean From Relative Frequencies
A relative frequency can also serve as a normalized weight.
If relative frequencies rᵢ satisfy:
Σrᵢ = 1,
then:
x̄ = Σrᵢxᵢ.
Using the previous frequency table:
r = 0.20, 0.50, 0.30.
Therefore:
x̄ = 0.20(10) + 0.50(20) + 0.30(30)
= 2 + 10 + 9
= 21.
The raw-frequency and relative-frequency methods are mathematically equivalent.
Weighted Mean and Expected Value
Expected value is fundamentally a probability-weighted mean.
For a discrete random variable X with possible values xᵢ and probabilities pᵢ:
E(X) = Σpᵢxᵢ
where:
Σpᵢ = 1.
For example, suppose:
| Outcome | Probability |
|---|---|
| 0 | 0.20 |
| 10 | 0.50 |
| 20 | 0.30 |
Then:
E(X) = 0(0.20) + 10(0.50) + 20(0.30)
= 0 + 5 + 6
= 11.
Expected value is therefore a weighted mean with probabilities as weights.
Weighted Mean of Group Means
A common use of weighted means is combining averages from groups of different sizes.
Suppose:
Group A mean = 70, n_A = 20
Group B mean = 90, n_B = 80.
The combined mean is not:
(70 + 90)/2
= 80
because the groups do not have equal sizes.
Instead:
Combined mean = [20(70) + 80(90)]/(20 + 80)
= (1400 + 7200)/100
= 86.
The larger group contributes proportionally more information to the combined total.
Why Averaging Averages Can Be Wrong
Suppose three classes have:
Class A: mean 60, n = 10
Class B: mean 70, n = 20
Class C: mean 90, n = 70.
An unweighted average of class means gives:
(60 + 70 + 90)/3
≈ 73.33.
But the true combined mean is:
[10(60) + 20(70) + 70(90)]/100
= (600 + 1400 + 6300)/100
= 8300/100
= 83.
The simple average substantially understates the overall mean because the highest-mean class contains most observations.
Weighted Mean and Sample Size Weights
When combining group means representing nonoverlapping sets of observations, sample sizes are natural weights:
x̄_combined = Σnᵢx̄ᵢ / Σnᵢ.
This formula reconstructs the total sum because:
Group total = nᵢx̄ᵢ.
Summing group totals and dividing by the total number of observations gives the combined mean.
This works exactly when the group means and group sizes refer to the same quantity and groups do not overlap.
Weighted Mean With Frequency Data
Frequency weights represent literal replication.
If observation:
x = 12
has weight:
5,
frequency weighting treats it as though:
12, 12, 12, 12, 12
were present in the data.
This interpretation is exact for counts.
Other types of weights do not necessarily represent repeated observations, which is why weighting conventions must be distinguished carefully.
Frequency Weights vs Importance Weights
Suppose value A receives weight 5 because it occurred five times.
That is a:
frequency weight.
Suppose value B receives weight 5 because a decision-maker considers it five times as important.
That is an:
importance weight.
Both use the same arithmetic formula, but the interpretations differ.
A weighted mean is meaningful only when the weighting rule itself is meaningful.
Probability Weights
Probability weights satisfy:
wᵢ ≥ 0
and usually:
Σwᵢ = 1.
They describe how probability mass is distributed across possible outcomes.
The weighted mean then represents an expected value.
For probabilities:
0.1, 0.2, 0.7
and outcomes:
5, 10, 20,
the expected value is:
0.1(5) + 0.2(10) + 0.7(20)
= 0.5 + 2 + 14
= 16.5.
The largest-probability outcome exerts the greatest influence.
Reliability Weights
Measurements can sometimes be given larger weights when they are considered more reliable.
Suppose two instruments estimate the same quantity:
Estimate A = 100
Estimate B = 110.
If B is much less precise, assigning both equal weight may be inappropriate.
A statistical method can instead assign larger weight to the more precise estimate.
The correct reliability weight depends on the measurement-error model rather than subjective preference alone.
Inverse-Variance Weighting
A common precision-weighting method uses:
wᵢ = 1/vᵢ
where:
vᵢ = variance of estimate i.
The weighted estimate becomes:
θ̂w = Σwᵢθ̂ᵢ/Σwᵢ.
An estimate with smaller variance receives a larger weight because it is more precise.
If the independent estimates are unbiased and their variances are known, inverse-variance weighting has particularly useful efficiency properties.
Inverse-Variance Example
Suppose two independent estimates are:
θ̂₁ = 10, variance = 4
θ̂₂ = 14, variance = 16.
Weights:
w₁ = 1/4 = 0.25
w₂ = 1/16 = 0.0625.
Weighted mean:
θ̂w = [0.25(10) + 0.0625(14)]/(0.25 + 0.0625)
= (2.5 + 0.875)/0.3125
= 3.375/0.3125
= 10.8.
The result lies much closer to 10 because the first estimate has one quarter of the variance of the second and therefore receives four times the weight.
Weighted Mean and Standard Error
A standard error can be calculated for a weighted estimator when the weighting and variance structure are specified. If independent observations Xᵢ have fixed weights wᵢ and variances σᵢ², then:
Var(x̄w) = Σwᵢ²σᵢ²/(Σwᵢ)².
Therefore:
SE(x̄w) = √[Σwᵢ²σᵢ²]/|Σwᵢ|.
For ordinary nonnegative weights, the denominator is simply the positive total weight.
The familiar:
s/√n
formula should not be applied automatically to every weighted mean.
Standard Error Under Equal Independent Variance
Suppose independent observations all have the same variance:
Var(Xᵢ) = σ².
Then:
Var(x̄w) = σ² Σwᵢ²/(Σwᵢ)².
Therefore:
SE(x̄w) = σ√Σwᵢ²/Σwᵢ
for nonnegative weights.
If all weights equal 1:
Σwᵢ² = n
and:
Σwᵢ = n.
Then:
SE = σ√n/n
= σ/√n,
which recovers the ordinary sample-mean standard error.
Effective Sample Size for Unequal Weights
Under certain independent equal-variance weighting settings, an effective sample size is defined as:
n_eff = (Σwᵢ)²/Σwᵢ².
Then:
SE(x̄w) = σ/√n_eff.
If all weights are equal:
n_eff = n.
If weights are highly unequal:
n_eff < n.
This shows why a weighted data set with many records can contain substantially less effective information when a small number of observations carry most of the total weight.
The formula should be interpreted within the weighting assumptions that justify it rather than as a universal definition for every survey or model.
Unequal Weight Example
Suppose four observations have weights:
1, 1, 1, 7.
Then:
Σw = 10
and:
Σw² = 1 + 1 + 1 + 49
= 52.
Therefore:
n_eff = 10²/52
= 100/52
≈ 1.923.
Although there are four observations, the highly concentrated weighting makes the equal-variance effective information roughly comparable to fewer than two equally weighted observations.
Extreme weighting can therefore reduce precision substantially.
Weighted Mean and Sample Variance
A sample variance measures dispersion around the sample mean, but weighted data require care because there is no single correction formula appropriate to every type of weight. Frequency weights, probability weights, analytic weights, reliability weights, and survey weights represent different statistical structures.
A simple weighted second central moment can be written:
m₂,w = Σwᵢ(xᵢ − x̄w)²/Σwᵢ.
However, this quantity is not automatically the unbiased sample variance estimator under every weighting design.
Finite-sample corrections depend on what the weights mean.
Weighted Population Variance
If normalized weights represent true population probabilities or proportions:
Σwᵢ = 1,
a weighted population variance is:
σ²_w = Σwᵢ(xᵢ − μ_w)²
where:
μ_w = Σwᵢxᵢ.
If weights are not normalized:
σ²_w = Σwᵢ(xᵢ − μ_w)²/Σwᵢ.
This is the natural weighted second central moment.
Its interpretation depends on the weights defining the population distribution being summarized.
Weighted Mean and Skewness
Skewness can also be defined using weights when weights genuinely represent the distributional mass of observations. A weighted mean determines the center, after which weighted second and third moments describe spread and asymmetry.
For example:
μ_w = Σwᵢxᵢ/Σwᵢ
m₂,w = Σwᵢ(xᵢ − μ_w)²/Σwᵢ
m₃,w = Σwᵢ(xᵢ − μ_w)³/Σwᵢ.
A weighted moment skewness can then be calculated from:
m₃,w/m₂,w^(3/2).
Thus, changing weights can alter not only the center but also the apparent shape of the weighted distribution.
Weighted Mean and Range
The range depends only on the minimum and maximum observed values and does not depend on the size of ordinary nonzero weights.
Suppose values are:
10, 20, 100
with weights:
100, 1, 1.
The weighted mean lies close to 10.
However:
Range = 100 − 10
= 90
regardless of those weights.
The range describes endpoint span, while the weighted mean describes weighted central location.
Weighted Mean and Outliers
An extreme numerical value can strongly influence a weighted mean, especially if it also receives a large weight.
Suppose:
x₁ = 10, w₁ = 1
x₂ = 20, w₂ = 1
x₃ = 1000, w₃ = 10.
Then:
x̄w = [10 + 20 + 10,000]/12
= 10,030/12
≈ 835.83.
The extreme observation dominates because it is both numerically distant and heavily weighted.
Weighting does not make a mean robust.
Small Weight on an Extreme Value
Using the same values but weights:
1, 1, 0.01
gives:
x̄w = [10 + 20 + 0.01(1000)]/2.01
= 40/2.01
≈ 19.90.
The extreme value now has very little influence because its weight is tiny.
Therefore, an observation’s effect on a weighted mean depends jointly on:
- its numerical value;
- its assigned weight.
Influence of One Observation
The normalized weight:
pᵢ = wᵢ/Σwᵢ
shows the direct contribution structure.
Then:
x̄w = Σpᵢxᵢ.
If one observation has:
pᵢ = 0.60,
it controls 60% of the linear weighting in the mean.
This does not mean the final result equals 60% of that value alone, because the remaining observations contribute the other 40%, but it makes the source of influence transparent.
Nonnegative Weights and the Minimum–Maximum Property
If:
wᵢ ≥ 0
for every observation and at least one weight is positive, then:
Minimum ≤ x̄w ≤ Maximum.
The weighted mean is a convex combination after weights are normalized.
For example, if values range from:
20 to 80,
a properly normalized nonnegative weighted mean cannot equal:
A result outside the observed range indicates either negative weights or a calculation problem.
Negative Weights
Some advanced statistical, financial, optimization, or signal-processing methods can generate negative weights.
If negative weights are permitted, the weighted mean can fall outside the minimum and maximum observations.
For example:
x₁ = 10, w₁ = 2
x₂ = 20, w₂ = −1.
Then:
Σw = 1
and:
x̄w = [2(10) − 1(20)]/1
= 0.
The result:
0
lies outside the range:
10 to 20.
Negative weights therefore fundamentally change the ordinary averaging interpretation.
Zero Weights
An observation with:
wᵢ = 0
has no effect on the weighted mean.
Suppose:
x = 10, 20, 1000
with weights:
1, 1, 0.
Then:
x̄w = [1(10) + 1(20) + 0(1000)]/(1 + 1 + 0)
= 30/2
= 15.
The third observation is mathematically excluded from the average.
Zero weights should therefore be used only when that complete exclusion is intentional.
Total Weight Cannot Be Zero
The weighted mean formula requires:
Σwᵢ ≠ 0.
If:
Σwᵢ = 0,
the denominator vanishes and the ordinary weighted mean is undefined.
This can occur with combinations of positive and negative weights.
With standard nonnegative weights and at least one positive weight, the problem cannot occur.
Checking the denominator is essential in generalized weighting systems.
Weighted Mean With Missing Values
Missing observations require a defined treatment.
Suppose values and intended weights are:
80 with weight 0.20
missing with weight 0.30
90 with weight 0.50.
If the missing component is excluded and remaining weights are renormalized:
Total retained weight = 0.70.
Then:
Weighted mean = [0.20(80) + 0.50(90)]/0.70
= 61/0.70
≈ 87.14.
Simply treating the missing value as zero would produce a very different and usually invalid answer.
Renormalizing After Exclusion
When one weighted component is legitimately omitted, surviving weights can be normalized:
wᵢ* = wᵢ/Σ retained weights.
In the previous example:
New weight for 80 = 0.20/0.70
≈ 0.2857.
New weight for 90:
0.50/0.70
≈ 0.7143.
Then:
x̄w = 0.2857(80) + 0.7143(90)
≈ 87.14.
Renormalization preserves the relative weighting among retained components.
Weighted Mean for Grades
Suppose a course uses:
| Component | Score | Weight |
|---|---|---|
| Homework | 90 | 20% |
| Midterm | 75 | 30% |
| Final | 85 | 50% |
Calculate:
90(0.20) = 18
75(0.30) = 22.5
85(0.50) = 42.5.
Therefore:
Weighted grade = 18 + 22.5 + 42.5
= 83.
The final examination affects the result more strongly because it carries half of the total weight.
Why the Ordinary Grade Average Is Wrong
The simple average would be:
(90 + 75 + 85)/3
= 250/3
≈ 83.33.
That is close to the weighted result here only by coincidence.
The correct result remains:
83
because the components have unequal weights.
With more extreme scores or weights, the difference can become much larger.
When percentages are supplied, those percentages determine contribution.
Weighted Mean for Investment Returns
Suppose a portfolio allocates:
50% to Asset A
30% to Asset B
20% to Asset C.
If the respective returns for one period are:
8%, 4%, −2%,
the portfolio return, ignoring additional cash-flow complications, is:
0.50(8%) + 0.30(4%) + 0.20(−2%)
= 4% + 1.2% − 0.4%
= 4.8%.
The allocation proportions act as weights.
The result is not the simple average of the three returns because the capital allocations differ.
Weighted Average Price
Suppose a buyer purchases:
10 units at $20
and:
30 units at $25.
The weighted average purchase price is:
[10(20) + 30(25)]/(10 + 30)
= (200 + 750)/40
= 950/40
= 23.75.
A simple average of prices:
(20 + 25)/2 = 22.50
would ignore that three times as many units were bought at $25.
Quantity must serve as the weight.
Weighted Mean for Rates
Rates can require particular caution because the appropriate weight depends on what is being averaged.
Suppose two machines process work at different rates.
A simple weighted arithmetic mean may be appropriate if rates are being averaged over equal exposure units, but travel speeds over equal distances often require a harmonic-type calculation instead.
The key question is:
What quantity is additive?
Weighted arithmetic means are valid when the chosen weights correctly reconstruct the total numerator and denominator associated with the rate.
Combined Average From Totals
A powerful way to verify a weighted mean is to reconstruct totals.
Suppose groups have:
mean x̄ᵢ
and:
size nᵢ.
Then group total is:
Tᵢ = nᵢx̄ᵢ.
Total across all groups:
T = Σnᵢx̄ᵢ.
Total observations:
N = Σnᵢ.
Therefore:
Overall mean = T/N
= Σnᵢx̄ᵢ/Σnᵢ.
This confirms why group size is the correct weight when combining ordinary subgroup means.
Weighted Mean From a Two-Way Table
Suppose average spending differs by customer group:
| Group | Customers | Mean Spending |
|---|---|---|
| A | 50 | 40 |
| B | 150 | 60 |
| C | 300 | 80 |
Weighted total spending is:
50(40) + 150(60) + 300(80)
= 2,000 + 9,000 + 24,000
= 35,000.
Total customers:
Therefore:
Overall mean = 35,000/500
= 70.
A simple average of:
40, 60, 80
would produce 60 and would be wrong because the groups have unequal sizes.
Weighted Mean and Simpson’s Paradox
Weighted averages can help explain why aggregate results sometimes differ dramatically from subgroup results.
If subgroup sizes differ between comparisons, the overall mean can be driven by the weighting pattern rather than by within-group performance alone.
This is one mechanism behind Simpson’s paradox, where an association observed within several groups can reverse after groups are combined.
The weighted mean itself is not misleading when calculated correctly.
The potential problem is interpreting an aggregate without recognizing how subgroup weights influence it.
Changing Weights Changes the Estimand
Suppose values remain:
20, 40, 80
but weights change from:
1, 1, 1
to:
1, 1, 8.
The ordinary mean is:
46.67.
The new weighted mean is:
[20 + 40 + 640]/10
= 700/10
= 70.
The data values did not change.
The target being summarized changed because the weighting distribution changed.
Weights therefore define the population or importance structure underlying the average.
Weighted Mean and Survey Data
Survey weights are often designed to account for unequal sampling probabilities, nonresponse adjustments, or population calibration.
A weighted population mean estimate can have the form:
x̄w = Σwᵢxᵢ/Σwᵢ.
However, uncertainty estimation for survey-weighted means usually requires the actual sampling design.
Simply computing an unweighted:
s/√n
standard error after applying survey weights can understate or overstate uncertainty.
Point estimation and variance estimation must use compatible methods.
Probability Weights in Surveys
If observation i has inclusion probability:
πᵢ,
a basic inverse-probability weight can be:
wᵢ = 1/πᵢ.
An observation that was less likely to enter the sample receives a larger weight because it represents more members of the target population.
This logic can correct unequal sampling probabilities under suitable designs.
The exact survey estimator may involve additional calibration, stratification, clustering, or finite-population adjustments.
Weighted Mean and Measurement Precision
Suppose several laboratories estimate the same quantity.
An estimate with small standard error is more precise than one with large standard error.
If independent unbiased estimates are combined using inverse-variance weighting:
wᵢ = 1/SEᵢ².
Then the most precise measurements exert the largest influence.
This is not because their numerical values are preferred, but because their uncertainty is smaller.
Weighting should therefore reflect precision before examining whether the values are favorable.
Example Using Standard Errors as Precision Weights
Suppose:
Estimate A = 50, SE = 2
Estimate B = 54, SE = 4.
Variances:
4 and 16.
Inverse-variance weights:
1/4 = 0.25
1/16 = 0.0625.
Weighted estimate:
[0.25(50) + 0.0625(54)]/(0.3125)
= (12.5 + 3.375)/0.3125
= 15.875/0.3125
= 50.8.
The more precise first estimate receives four times as much weight.
Weighted Mean and Transformations
For a linear transformation:
yᵢ = axᵢ + b,
with the same weights:
ȳw = a x̄w + b.
This follows because:
Σwᵢ(axᵢ + b)/Σwᵢ
= aΣwᵢxᵢ/Σwᵢ + bΣwᵢ/Σwᵢ
= ax̄w + b.
Thus, ordinary changes of measurement units transform the weighted mean exactly as expected.
Unit Conversion Example
Suppose a weighted mean temperature is:
20°C.
Convert to Fahrenheit:
F = 1.8C + 32.
Therefore:
Weighted mean Fahrenheit = 1.8(20) + 32
= 68°F.
There is no need to recalculate all weights and individual temperatures if the transformation is linear.
The transformed weighted mean follows directly.
Nonlinear Transformations
For a nonlinear function g:
Weighted mean of g(x)
is generally not equal to:
g(weighted mean of x).
In symbols:
Σpᵢg(xᵢ) ≠ g(Σpᵢxᵢ)
in general.
For example, the weighted average of squared values is not generally the square of the weighted average.
This distinction is central to variance, expected utility, nonlinear risk measures, and many probability calculations.
Example of Nonlinearity
Take:
x₁ = 1
x₂ = 3
with equal normalized weights:
0.5, 0.5.
Weighted mean:
x̄w = 2.
Square the weighted mean:
2² = 4.
Weighted mean of the squares:
0.5(1²) + 0.5(3²)
= 0.5 + 4.5
= 5.
Therefore:
E(X²) ≠ [E(X)]².
Their difference is related directly to variance.
Weighted Mean as a Balance Point
The weighted mean can be interpreted as the balance point of values carrying different masses.
Imagine values positioned along a line and each weight representing physical mass.
The weighted mean is the location at which the system balances:
Σwᵢ(xᵢ − x̄w) = 0.
Equivalently:
Σwᵢxᵢ = x̄wΣwᵢ.
This balancing identity is the weighted counterpart of the ordinary fact that deviations from the arithmetic mean sum to zero.
Weighted Deviations Sum to Zero
Using:
x̄w = Σwᵢxᵢ/Σwᵢ,
multiply both sides by:
Σwᵢ.
Then:
x̄wΣwᵢ = Σwᵢxᵢ.
Rearrange:
Σwᵢxᵢ − x̄wΣwᵢ = 0.
Factor:
Σwᵢ(xᵢ − x̄w) = 0.
Thus, weighted positive deviations and weighted negative deviations balance around the weighted mean.
Weighted Mean Minimizes Weighted Squared Error
For positive weights, the weighted mean minimizes:
Σwᵢ(xᵢ − c)²
over possible constants c.
The minimizing value is:
c = x̄w.
This is the weighted analogue of the ordinary mean’s least-squares property.
It explains why weighted means appear naturally in weighted least squares and estimation problems where some observations are deliberately given greater influence.
Weighted Least Squares Connection
In regression, weighted least squares minimizes an objective of the form:
Σwᵢ(yᵢ − ŷᵢ)².
Observations with larger wᵢ receive greater influence over the fitted model.
When the model contains only an intercept and no predictors, the fitted intercept becomes a weighted mean of the outcomes.
Thus, the weighted mean is not merely a classroom averaging technique; it is the simplest case of a broad weighted optimization framework.
Weighted Mean Is Sensitive to Weight Choice
Suppose values are:
10, 50, 100.
Using weights:
1, 1, 1
gives:
53.33.
Weights:
1, 1, 8
give:
Weights:
8, 1, 1
give:
All calculations are arithmetically valid.
Only the substantive definition of the weights determines which, if any, answers the intended question.
Weighting Cannot Be Chosen After Seeing the Desired Answer
A weighting scheme should be defined by the data-generating process, measurement design, population structure, or decision rule rather than adjusted simply because one set of weights produces a preferred result.
Changing weights can materially change the weighted mean, sometimes more than changing the observations themselves.
Transparent reporting should therefore explain where the weights came from.
A weighted result without a defensible weighting rule can be difficult to interpret.
Weighted Mean With One Positive Weight
Suppose:
x = 10, 20, 30
and weights are:
0, 0, 5.
Then:
x̄w = 5(30)/5
= 30.
The weighted mean equals the only value carrying positive weight.
This is mathematically valid.
It also demonstrates that a weighting scheme can effectively exclude observations even when they remain listed in the data.
Weighted Mean With Repeated Weights
Suppose:
x = 10, 20, 30, 40
and weights:
2, 2, 2, 2.
Because all weights are equal:
x̄w = ordinary mean
= 25.
The actual common weight value does not matter.
Weights:
0.5, 0.5, 0.5, 0.5
would produce exactly the same result.
Only differences among the weights alter the relative influence structure.
Weighted Mean With Decimal Weights
Weights can be decimals.
Suppose:
x = 40, 60, 80
and:
w = 0.5, 1.5, 2.
Weighted total:
0.5(40) + 1.5(60) + 2(80)
= 20 + 90 + 160
= 270.
Weight sum:
Therefore:
x̄w = 270/4
= 67.5.
Weights do not need to be whole numbers unless their interpretation specifically requires counts.
Weighted Mean With Negative Values
Negative data values cause no special difficulty.
Suppose:
x = −10, 0, 20
with weights:
1, 2, 1.
Then:
x̄w = [1(−10) + 2(0) + 1(20)]/4
= 10/4
= 2.5.
The result is determined by the weighted balance of positive and negative values.
The nonnegative weights still guarantee that the answer lies between:
−10 and 20.
Weighted Mean With Percentage Values
Suppose conversion rates from three regions are:
10%, 20%, 30%
with customer counts:
100, 300, 600.
The correctly weighted overall rate is:
[100(10%) + 300(20%) + 600(30%)]/1000
= (10 + 60 + 180)/1000?
When using percentages directly, it is clearer to use decimal rates:
[100(0.10) + 300(0.20) + 600(0.30)]/1000
= (10 + 60 + 180)/1000
= 250/1000
= 0.25.
Therefore:
Overall rate = 25%.
The group denominator must correspond to the exposure behind each rate.
Why Averaging Percentages Can Be Wrong
The simple average:
(10% + 20% + 30%)/3
= 20%
would be correct only if the regions represented equal denominators.
Here, most customers belong to the 30% region, so the actual overall rate is:
25%.
Whenever percentages arise from groups of unequal size, combine the underlying numerators and denominators or use appropriate denominator weights.
Weighted Mean vs Weighted Median
A weighted median identifies a value at which cumulative weight crosses approximately half of total weight.
The weighted mean instead balances all weighted numerical magnitudes.
For values:
10, 20, 100
with weights:
1, 1, 8,
the weighted median is:
100
because most weight lies there.
The weighted mean is:
[10 + 20 + 800]/10
= 83.
The statistics answer different questions.
Weighted Mean vs Mode
The mode identifies the value or category with greatest frequency or probability mass, while the weighted mean combines numerical values according to weights.
Suppose values:
1, 2, 10
have frequencies:
5, 3, 2.
The mode is:
The frequency-weighted mean is:
[5(1) + 3(2) + 2(10)]/10
= 31/10
= 3.1.
Frequency concentration and numerical balance are different properties.
Weighted Mean and Distribution Shape
A weighted mean does not reveal whether the weighted distribution is:
- symmetric;
- skewed;
- multimodal;
- tightly concentrated;
- widely dispersed.
Two weighted distributions can have identical weighted means and entirely different shapes.
Therefore, the weighted mean should often be supplemented with weighted measures of dispersion or graphical summaries when shape matters.
A central estimate is not a complete distributional description.
Weighted Mean and Rounding
Intermediate weighted products should generally retain full available precision until the final result is calculated.
Suppose:
weight = 0.3333
and:
value = 27.84.
Prematurely rounding the weight to:
0.33
changes its contribution.
Across many categories, these small errors can accumulate.
Normalize and calculate using full precision, then round the final weighted mean to an appropriate number of decimal places.
Common Weighted Mean Mistakes
A common mistake is multiplying each value by its weight and then dividing by the number of values rather than by the sum of the weights. Unless weights sum to the number of observations, that produces the wrong average.
Another mistake is averaging subgroup means equally when subgroup sizes differ. Group size must be used as the weight when reconstructing the overall mean from nonoverlapping groups.
Analysts also sometimes assume percentage weights automatically sum to 100%, even when components are missing or excluded. The denominator must represent the total weight actually included.
Another error is using frequency weights, survey weights, reliability weights, and inverse-variance weights as though they had identical statistical meanings simply because the arithmetic formula looks the same.
A heavily weighted extreme value can dominate the result, so weighting should not be mistaken for robustness.
Finally, a weighted mean should not be reported without enough information to understand why the weights were chosen.
How to Calculate Weighted Mean Correctly
First determine what each weight represents.
Then calculate:
wᵢxᵢ
for every observation or group.
Add the weighted products:
Σwᵢxᵢ.
Add the weights:
Σwᵢ.
Calculate:
x̄w = Σwᵢxᵢ/Σwᵢ.
If the weights are already normalized to sum to 1, simply calculate:
x̄w = Σwᵢxᵢ.
Finally, verify that the result behaves sensibly relative to the most heavily weighted values and check the minimum–maximum property when all weights are nonnegative.
Full Worked Example
Suppose four project scores are:
| Project | Score | Weight |
|---|---|---|
| A | 65 | 10% |
| B | 80 | 20% |
| C | 75 | 30% |
| D | 95 | 40% |
Step 1: Convert Percentages to Decimals
10% = 0.10
20% = 0.20
30% = 0.30
40% = 0.40.
Check:
0.10 + 0.20 + 0.30 + 0.40
= 1.
Step 2: Multiply Each Score by Its Weight
65(0.10) = 6.5
80(0.20) = 16
75(0.30) = 22.5
95(0.40) = 38.
Step 3: Add the Weighted Contributions
6.5 + 16 + 22.5 + 38
= 83.
Therefore:
Weighted mean = 83.
The result is closer to 95 than a simple average would be because the highest score carries the largest weight.
Compare With the Ordinary Mean
The ordinary mean of:
65, 80, 75, 95
is:
(65 + 80 + 75 + 95)/4
= 315/4
= 78.75.
Weighted result:
Difference:
83 − 78.75
= 4.25.
The weighting system raises the average because greater importance is assigned to the higher-scoring components.
Full Frequency-Weighted Example
Suppose a survey records ratings:
| Rating | Number of Responses |
|---|---|
| 1 | 5 |
| 2 | 10 |
| 3 | 20 |
| 4 | 40 |
| 5 | 25 |
Total:
n = 5 + 10 + 20 + 40 + 25
= 100.
Weighted score total:
1(5) + 2(10) + 3(20) + 4(40) + 5(25)
= 5 + 20 + 60 + 160 + 125
= 370.
Therefore:
Weighted mean rating = 370/100
= 3.70.
The frequency-weighted calculation is identical to averaging all 100 individual responses.
Full Combined-Group Example
Suppose three branches report:
| Branch | Employees | Mean Productivity |
|---|---|---|
| A | 20 | 50 |
| B | 30 | 60 |
| C | 50 | 80 |
Total weighted productivity:
20(50) + 30(60) + 50(80)
= 1000 + 1800 + 4000
= 6800.
Total employees:
Therefore:
Combined mean = 6800/100
= 68.
A simple average of branch means would be:
(50 + 60 + 80)/3
≈ 63.33,
which is incorrect for the combined employee population.
Full Inverse-Variance Example
Suppose three independent studies estimate the same parameter:
| Study | Estimate | Standard Error |
|---|---|---|
| A | 10 | 1 |
| B | 12 | 2 |
| C | 15 | 3 |
Variances are:
1² = 1
2² = 4
3² = 9.
Inverse-variance weights are:
w_A = 1
w_B = 1/4 = 0.25
w_C = 1/9 ≈ 0.1111.
Weighted numerator:
1(10) + 0.25(12) + (1/9)(15)
= 10 + 3 + 1.6667
≈ 14.6667.
Total weight:
1 + 0.25 + 0.1111
≈ 1.3611.
Therefore:
Weighted estimate ≈ 14.6667/1.3611
≈ 10.78.
The most precise study exerts the greatest influence.
Standard Error of the Inverse-Variance Mean
Under the independent known-variance model:
SE(θ̂w) = 1/√Σwᵢ.
Using:
Σwᵢ ≈ 1.3611,
we get:
SE ≈ 1/√1.3611
≈ 0.857.
Thus the combined estimate is approximately:
10.78
with standard error:
0.86
under the stated fixed-effect precision-weighting assumptions.
Different meta-analytic assumptions, such as between-study heterogeneity, require different weighting and uncertainty calculations.
Interpreting the Weighted Mean
A weighted mean should always be interpreted in terms of what the weights represent.
If the weights are:
frequencies,
the result is the arithmetic mean across all represented observations.
If they are:
probabilities,
the result is an expected value.
If they are:
group sizes,
the result combines subgroup means into an overall mean.
If they are:
inverse variances,
the result emphasizes statistically precise estimates.
The same formula can therefore represent several different statistical concepts.
When Weighted Mean Is Most Useful
The weighted mean is especially useful when:
- observations occur with unequal frequencies;
- groups have unequal sizes;
- course components have different percentage contributions;
- portfolio components have unequal allocations;
- probabilities weight possible outcomes;
- measurements have unequal statistical precision;
- survey observations represent different amounts of a target population.
It is appropriate whenever an equal-contribution arithmetic mean would misrepresent the structure being summarized.
When an Ordinary Mean Is Better
If every observation genuinely contributes equally, additional weighting is unnecessary.
Assigning arbitrary unequal weights can reduce transparency and may introduce subjectivity or bias.
The ordinary arithmetic mean is simply the weighted mean with equal weights, so there is no benefit in adding complexity unless unequal contribution is substantively justified.
The simplest valid weighting scheme is usually preferable.
How to Report a Weighted Mean
A useful report identifies both the result and the weighting basis.
For example:
“The weighted mean score was 83, using component weights of 10%, 20%, 30%, and 40%.”
For combined groups:
“Weighting branch means by employee counts produced an overall mean productivity of 68.”
For precision weighting:
“The inverse-variance weighted estimate was approximately 10.78.”
Stating the weighting basis prevents readers from mistaking the result for a simple arithmetic mean.
Frequently Asked Questions About Weighted Mean
What is a weighted mean?
A weighted mean is an average in which values contribute according to assigned weights rather than contributing equally.
What is the weighted mean formula?
x̄w = Σwᵢxᵢ/Σwᵢ
What if the weights sum to 1?
Then:
x̄w = Σwᵢxᵢ.
Is weighted mean the same as arithmetic mean?
Only when all observations receive equal weights.
What does a larger weight do?
It gives the corresponding value more influence on the weighted mean.
Do weights have to sum to 1?
No.
What should I divide by if weights do not sum to 1?
Divide by:
Σwᵢ.
Do percentage weights have to sum to 100%?
If they represent complete percentage shares, yes. If they are merely relative weights, divide by their actual total.
Can weights be normalized?
Yes:
pᵢ = wᵢ/Σwᵢ.
Does normalization change the weighted mean?
No.
Can weights be decimals?
Yes.
Can weights be frequencies?
Yes.
Can relative frequencies be weights?
Yes.
Can probabilities be weights?
Yes.
Is expected value a weighted mean?
For a discrete probability distribution, yes.
How do you combine means from groups with different sizes?
Use group sizes as weights:
x̄ = Σnᵢx̄ᵢ/Σnᵢ.
Can I simply average subgroup means?
Only when the groups contribute equal appropriate weight.
Why is averaging percentages sometimes wrong?
Because percentages based on unequal denominators should be weighted by those denominators.
Can a weighted mean be used for grades?
Yes.
What is a grade example?
If scores are 80, 70, and 90 with weights 20%, 30%, and 50%:
Weighted mean = 82.
Can a weighted mean be used for portfolio returns?
Yes, when component returns are weighted by the relevant portfolio allocations under the intended calculation.
Can a weighted mean be used for prices?
Yes, quantities purchased can be used as weights to calculate an average unit price.
What is a frequency-weighted mean?
A weighted mean in which each weight equals the number of times the corresponding value occurs.
Is a frequency-weighted mean exact?
Yes, when frequencies represent literal repeated observations.
What is inverse-variance weighting?
A method using:
wᵢ = 1/varianceᵢ
so more precise estimates receive larger weights.
Can standard errors determine inverse-variance weights?
Yes:
wᵢ = 1/SEᵢ².
Is inverse-variance weighting the same as frequency weighting?
No. The arithmetic form is similar, but the weights have different meanings.
Can a weighted mean have a standard error?
Yes.
Is its standard error always s/√n?
No.
What is a general fixed-weight variance formula?
For independent observations:
Var(x̄w) = Σwᵢ²σᵢ²/(Σwᵢ)².
What is effective sample size for unequal weights?
Under certain equal-variance independent weighting settings:
n_eff = (Σwᵢ)²/Σwᵢ².
Can unequal weights reduce effective sample size?
Yes.
Can one heavily weighted observation dominate the mean?
Yes.
Is weighted mean robust to outliers?
No.
Can an extreme value with a large weight strongly affect it?
Yes.
Can a small weight reduce an extreme value’s influence?
Yes.
Can weights equal zero?
Yes, but a zero-weight observation contributes nothing.
Can weights be negative?
In advanced applications, yes.
Can negative weights produce a mean outside the observed range?
Yes.
With ordinary nonnegative weights, where must the weighted mean lie?
Between the minimum and maximum positively weighted values.
Can the total weight equal zero?
Not if an ordinary weighted mean is to be calculated.
What happens if Σw = 0?
The formula is undefined.
Does multiplying every weight by 10 change the result?
No.
Why not?
The common factor cancels from numerator and denominator.
Does adding a constant to every value change the weighted mean?
Yes, by the same constant.
What is the transformation rule?
For:
Y = aX + b,
ȳw = ax̄w + b.
Does a nonlinear transformation commute with weighted averaging?
Generally no.
Is weighted mean the same as weighted median?
No.
What is the weighted median?
A rank-based statistic determined by cumulative weight around the 50% point.
Is weighted mean the same as mode?
No.
Can two weighted distributions have the same weighted mean but different variances?
Yes.
Can they have different skewness?
Yes.
Does weighted mean measure spread?
No.
Does it measure skewness?
No.
Does it determine the range?
No.
Can weights change the apparent distribution?
Yes, because they alter how much mass each observation represents.
Can survey weights be used in a weighted mean?
Yes.
Is ordinary variance estimation enough for survey-weighted data?
Not necessarily. The sampling design may require specialized variance estimation.
What happens when a weighted observation is missing?
The treatment depends on the problem; if it is legitimately excluded, remaining weights may need renormalization.
Should a missing value be replaced with zero?
Not unless zero is the actual observed value.
Can weighted mean minimize a loss function?
Yes. With positive weights, it minimizes:
Σwᵢ(xᵢ − c)².
What is the balancing property?
Σwᵢ(xᵢ − x̄w) = 0.
Why is the weighting basis important?
Because different weight definitions produce averages with different statistical interpretations.
What is the biggest mistake when calculating weighted mean?
Using unequal weights without understanding what they represent, or dividing the weighted total by the number of observations instead of the sum of the weights.
What is the main advantage of the weighted mean?
It provides a mathematically coherent average when observations, groups, probabilities, or estimates should not contribute equally.
What is its main limitation?
The result can be dominated by large or heavily weighted values, and its meaning is only as defensible as the weighting scheme used to produce it.
What is the most important rule when using weighted mean?
Match each weight to a clear, defensible interpretation, calculate Σwᵢxᵢ/Σwᵢ, and remember that changing the weights changes the population, importance structure, or precision structure represented by the average.
