Why a Weighted Mean Is a Ratio
A sample mean gives each observed unit the same influence. In a complex survey, however, sample units may represent different numbers of population units. A weighted mean accounts for that difference by combining each outcome with its survey weight. Its denominator matters just as much as its numerator: it represents the estimated number of population units contributing to the mean.
The construction follows a simple principle. First estimate the total of the outcome across the target population. Then estimate how many target-population units contributed eligible outcome values. The mean is the first estimated total divided by the second. This ratio form applies to continuous measurements and, with a zero-one outcome, to population proportions.
The weights may be inverse inclusion probabilities or final survey weights that incorporate additional adjustments. In either case, the calculation uses the weights supplied for the survey analysis. The interpretation depends on what population those weights are intended to represent and which sampled units are included in the calculation.
From Estimated Totals to an Estimated Mean
Suppose the target is a finite population of \(N\) units, with outcome \(Y_i\) for unit \(i\). Its population mean is the total outcome divided by the population count:
If the population count and outcome total are not both known, a survey can estimate them from its sample. For a probability sample, let \(\pi_i\) be the probability that population unit \(i\) is included, with \(\pi_i>0\). The inverse-probability weight \(w_i=1/\pi_i\) expands each observed unit to represent its share of the population. The resulting estimated total of the outcome and estimated population count are
Dividing these estimates gives the ratio estimator of the population mean. With other final weights, the same ratio structure applies, though the meaning of the weights depends on how they were constructed. The essential requirement is that the numerator and denominator estimate the outcome total and population count for the same target population.
A Short Justification of the Ratio Form
For a probability sample, the inverse-probability expansion gives an unbiased estimator of each finite-population total. To see the point, let \(I_i\) equal 1 if unit \(i\) is sampled and 0 otherwise. Then \(P(I_i=1)=\pi_i\), and the estimated outcome total can be written as \(\hat{T}_y=\sum_{i\in U}I_iY_i/\pi_i\). Taking its expectation over repeated samples under the sampling design gives
Applying the same argument with every unit’s outcome replaced by 1 shows \(E(\hat{N})=N\). Thus each total estimator is design-unbiased. But their ratio is not generally design-unbiased: in general, the expectation of a ratio is not the ratio of expectations, so \(E(\hat{T}_y/\hat{N})\) need not equal \(T_y/N\). The ratio form is used because it directly estimates the mean and, under suitable sampling conditions, is often consistent as the sample grows. Unbiasedness of both totals alone does not establish unbiasedness of their ratio.
Worked Example: A Continuous Outcome
Suppose a hypothetical survey records weekly hours of unpaid caregiving for three sampled adults. Their final survey weights are shown below. Assume all three adults belong to the target population and have an observed value for the same outcome.
| Adult | Weekly caregiving hours, \(y_i\) | Weight, \(w_i\) | Weighted hours, \(w_i y_i\) |
|---|---|---|---|
| A | 10 | 2 | 20 |
| B | 20 | 1 | 20 |
| C | 30 | 3 | 90 |
The estimated outcome total is \(20+20+90=130\) weekly caregiving hours, and the estimated population count is \(2+1+3=6\) adults. Therefore,
The estimate describes the weighted mean weekly caregiving time for the target population represented by these weights. The unweighted sample mean is \((10+20+30)/3=20\) hours, but it treats each sampled adult as equally representative. Here the adult reporting 30 hours has a larger weight, so the weighted mean is higher.
Binary Outcomes Give Weighted Proportions
A binary outcome is coded 1 when a specified condition is present and 0 when it is absent. For example, define \(y_i=1\) if a person has a particular health-insurance coverage and \(y_i=0\) otherwise. The weighted outcome total then adds the weights of the sampled people with the condition, while the denominator adds the weights of all sampled people in the target population.
The weighted mean of a zero-one outcome is therefore the weighted proportion with outcome value 1. Multiplying that mean by 100 expresses it as a percentage. This interpretation applies only when the coding and the denominator population match the question being asked.
Worked Example: A Binary Outcome
In a hypothetical survey, four sampled adults are asked whether they have health-insurance coverage. Let \(y_i=1\) for “yes” and \(y_i=0\) for “no.” Assume all four adults are eligible members of the target population and all have observed answers.
| Adult | Coverage indicator, \(y_i\) | Weight, \(w_i\) | Weighted indicator, \(w_i y_i\) |
|---|---|---|---|
| A | 1 | 4 | 4 |
| B | 1 | 3 | 3 |
| C | 0 | 2 | 0 |
| D | 0 | 1 | 0 |
The weighted total of covered adults is \(4+3=7\), and the estimated population count is \(4+3+2+1=10\). Thus the weighted proportion is
The estimate is 70%: the weighted estimated proportion of the target population with coverage. It is not the unweighted sample proportion, which is \(2/4=0.50\). The difference arises because the two covered adults together carry 7 of the 10 total weight units.
Defining the Population Behind the Denominator
A ratio is meaningful only when its numerator and denominator refer to compatible units. If the question concerns all adults, both sums should use weights for eligible adults with observed outcomes. If the outcome is defined only for a subgroup, the weighted mean should be calculated among that subgroup rather than dividing its outcome total by a count that includes people outside the subgroup.
The estimated population count \(\hat{N}_w\) need not equal the literal number of records in the sample. It is the sum of the weights for the records included in the calculation. In some analyses a known population count is available, but the ratio estimator introduced here uses the weighted estimated count in its denominator. The numerator and denominator should not be mixed across different weighting schemes or populations.
For records with missing outcome values, the calculation also requires care. A record without an observed \(y_i\) cannot contribute an observed value to the outcome total. If such records are excluded, the denominator should represent the population for which the mean is being estimated, according to the survey’s treatment of missingness and weights. Simply omitting missing outcomes while retaining an incompatible count can change the target or distort the result.
Worked Example: Checking the Target Population
A hypothetical survey estimates average daily travel time among employed adults. Three eligible respondents have observed travel times of 20, 40, and 60 minutes and weights of 2, 1, and 3. A fourth sampled person is not employed and has weight 4. The target is explicitly the employed population, so the non-employed person must not enter either total.
For the employed respondents, the estimated travel-time total is \((2)(20)+(1)(40)+(3)(60)=260\) weighted minutes, and their estimated count is \(2+1+3=6\). Therefore,
Including the non-employed person’s weight in the denominator but not a comparable travel-time value would break the ratio: it would divide a total for employed adults by a count that includes a person outside the target population. The correct calculation restricts both sums to the same eligible group.
Common Pitfalls and Practical Checks
Weighted means are easy to calculate, but several errors can make a precise-looking result answer the wrong question. Before interpreting one, check that the outcome, weights, and denominator all refer to the intended population and records.
- Using the sum of weighted outcomes alone. This is an estimated total, not a mean. Divide by the corresponding sum of weights to obtain the weighted mean.
- Using the number of sampled records as the denominator. The weighted mean divides by \(\sum w_i\), not by \(n\), unless all weights happen to be equal.
- Combining incompatible records. Do not include weights for units outside the population represented by the numerator, or use a denominator from a different group.
- Calling the ratio exactly unbiased because both totals are unbiased. The expectation of a ratio is not generally the ratio of expectations. The ratio is an estimator of the population mean, but exact unbiasedness does not follow from unbiased total estimates.
- Interpreting a binary weighted mean as a count. With zero-one coding, the weighted numerator estimates a total for the coded outcome; dividing by the weighted count gives a proportion, not a number of people.
A useful arithmetic check is that, when all weights are positive, the weighted mean must lie between the smallest and largest observed outcomes included in the calculation. For a binary outcome coded 0 and 1, it must lie between 0 and 1. These checks do not verify that the weights or target population are appropriate, but they can reveal calculation errors.
Check Your Understanding
Use the ratio-of-totals definition and keep the target population explicit.
- A survey has outcome values 5 and 15 with weights 3 and 1. Calculate the weighted mean and state its units if the outcome is measured in kilograms.
- Explain why the weighted mean of a zero-one indicator can be interpreted as a proportion.
- For inverse-probability weights, why is the weighted estimated population count \(\sum_{i\in s}1/\pi_i\) relevant to estimating a population mean?
- Both the weighted outcome total and weighted population count are design-unbiased. Does that fact alone make their ratio design-unbiased? Explain.
- A weighted outcome total is calculated for eligible adults, but the denominator includes weights for all sampled people. What is wrong with this calculation?