From Survey Estimates to a Population Profile
A population profile is a set of complementary estimates that describes a defined target population. It may report how many people meet a condition, the population average of a measurement, the share in a category, a rate per unit of exposure, and features of a distribution such as its median or spread. The aim is not to produce a long list of weighted numbers. It is to show what the survey estimates say about the population, how uncertain those estimates are, and what each number represents.
The weights and the survey design have different roles. A survey weight \(w_i\) indicates how much of the target population record \(i\) represents for a particular analysis. Weighted formulas produce point estimates. Design-consistent standard errors additionally require information about how the sample was selected, such as strata, clusters, and replicate weights. The weights alone do not determine uncertainty.
A total has the units of the quantity being added: a total of people has units of people, and a total of dollars has units of dollars. A mean has the units of the measured variable. For a binary indicator, a weighted total estimates the number of population members in the category, while the weighted mean estimates the proportion in that category. If a rate uses a different denominator, its units must express that denominator, such as visits per person-year.
A Worked Set of Weighted Descriptive Estimates
Consider the following deliberately small, hypothetical survey dataset. Each record represents one eligible adult, the weight is in represented adults, annual visits are for one person-year of observation, and income is recorded in thousands of dollars. The table is for practicing point-estimate arithmetic; it does not include the design information needed to calculate standard errors.
| Record | Weight \(w_i\) | Income (thousands of dollars) | Insured (1 = yes) | Annual visits |
|---|---|---|---|---|
| 1 | 80 | 20 | 1 | 0 |
| 2 | 120 | 25 | 1 | 1 |
| 3 | 60 | 30 | 0 | 0 |
| 4 | 140 | 35 | 1 | 2 |
| 5 | 90 | 40 | 1 | 1 |
| 6 | 110 | 45 | 0 | 3 |
| 7 | 70 | 50 | 1 | 0 |
| 8 | 130 | 60 | 0 | 2 |
| 9 | 100 | 80 | 1 | 4 |
| 10 | 100 | 100 | 0 | 1 |
Worked Example: Totals, a Mean, a Proportion, and a Rate
First, the total represented adult population is estimated by summing the weights:
For income, multiply each recorded value by its weight and add. Because income is recorded in thousands of dollars, the total is in thousands of dollars:
The weighted mean income is therefore \(49{,}150/1000=49.15\) thousand dollars, or \(\$49{,}150\). This is an estimated population average, not a total and not the unweighted average of the ten observed incomes.
For insurance, the weighted total of the binary indicator estimates the number of insured adults. The records with indicator 1 have weights 80, 120, 140, 90, 70, and 100:
The point estimate is 600 insured adults out of an estimated 1,000, or 60.0% of the target population. For annual visits, the weighted total is \(120+280+90+330+260+400+100=1{,}580\) visits. Since each eligible adult contributes one person-year, the estimated rate is \(1{,}580/1{,}000=1.58\) visits per person-year. If exposure differed across people, the denominator would need to be the weighted total of the relevant exposure, rather than simply the number of adults.
Weighted Quantiles and Distribution Displays
A population profile should not rely on averages alone. A weighted distribution function is formed by accumulating weights across the ordered values. For income \(x_i\), the weighted empirical distribution function is \( \hat{F}_w(t)=\sum_i w_i I(x_i\leq t)/\sum_i w_i \), where \(I(x_i\leq t)\) equals 1 when the condition is true and 0 otherwise. Under the inverse-cumulative convention, the weighted quantile at probability \(q\) is the smallest value \(t\) for which \(\hat{F}_w(t)\geq q\). Quantile conventions should be stated when they could affect the result.
Worked Example: Weighted Median and Interquartile Range
The incomes in the table are already ordered. Their cumulative weights are 80, 200, 260, 400, 490, 600, 670, 800, 900, and 1,000. The weighted cumulative proportion first reaches 0.50 at an income of 45 thousand dollars: the cumulative weight through 40 is 490, while through 45 it is 600. Thus, using the inverse-cumulative convention, the weighted median is \(\$45{,}000\).
For the first quartile, the cumulative weight first reaches 250 at an income of 30 thousand dollars (cumulative weight 260). For the third quartile, it first reaches 750 at an income of 60 thousand dollars (cumulative weight 800). The interquartile range is \(60-30=30\) thousand dollars. It describes the distance between the estimated 25th and 75th percentiles; it is not a confidence interval.
A weighted histogram or distribution plot can display where the represented population lies across values. For the income data, consider bins \([0,30)\), \([30,50)\), \([50,80)\), and \([80,110)\), in thousands of dollars. The weighted counts are respectively 200, 400, 200, and 200 adults. They sum to 1,000, the estimated population size. Because the bin widths are not all equal, bar heights intended to represent density should be weighted count divided by bin width; then each bar's area, rather than its height alone, represents weighted population count. The estimates and plotting choices should make the population represented by the graphic clear.
Worked Example: Reading a Weighted Distribution Plot
For the four income bins above, the widths are 30, 20, 30, and 30 thousand dollars. The corresponding weighted-count densities are \(200/30=6.67\), \(400/20=20.0\), \(200/30=6.67\), and \(200/30=6.67\) represented adults per thousand dollars. The area of the second bar is \(20.0(20)=400\) adults, matching its weighted count; the other areas are likewise 200, 200, and 200 adults. A plot using weighted counts as heights could still show bin counts, but with unequal widths its bar areas would not represent those counts.
Design-Consistent Confidence Intervals
A confidence interval adds uncertainty to a point estimate. In complex surveys, the standard error \(SE\) must account for the actual sample design, using the available strata, clusters, and other design information or appropriate replicate weights. A conventional large-sample interval for an estimate \(\hat{\theta}\) is \(\hat{\theta}\pm1.96SE(\hat{\theta})\) when a normal approximation is suitable. The design-based standard error, not a calculation that treats the records as a simple random sample, is used in this expression. For proportions near 0 or 1, sparse categories, small numbers of sampled clusters, or quantiles with unstable estimates, a suitable alternative interval procedure may be needed.
The short justification for the interval formula is that, under the relevant sampling-design approximation, the standardized error \((\hat{\theta}-\theta)/SE(\hat{\theta})\) is approximately standard normal. About 95% of a standard normal distribution lies between \(-1.96\) and \(1.96\). Inverting that range gives an interval centered at \(\hat{\theta}\) with margin \(1.96SE(\hat{\theta})\). This is a repeated-sampling coverage statement under the assumed design and approximation, not a probability statement about a fixed parameter after the interval has been calculated.
Worked Example: Intervals for a Population Mean and Proportion
Suppose a hypothetical, larger survey's design file and analysis produce a weighted mean annual income of \(\$49{,}150\) with design-based \(SE=\$2{,}400\). The interval calculation is:
Subject to the design and normal-approximation conditions, the interval is a 95% confidence interval for the target population's mean annual income. Its endpoints are dollars per adult, not total dollars. Now suppose the estimated insured proportion is 0.600 with design-based \(SE=0.055\):
This corresponds to approximately 49.2% to 70.8% of the target population. These standard errors are hypothetical design-based outputs for illustration, not quantities recoverable from the ten-row table above. Their calculation requires the sample-design information; weights by themselves are insufficient.
Worked Example: Interpreting an Interval for a Rate
Suppose a different hypothetical survey estimates 1.58 annual visits per person-year, with a design-based standard error of 0.18 visits per person-year. Using the stated normal approximation:
The interval describes uncertainty about the population rate of visits per person-year, given the survey design and the approximation. It does not say that each person has between 1.2272 and 1.9328 visits, nor does it give the range of individual visit counts.
Putting the Profile Together
A useful profile makes its denominator and target explicit, presents estimates in a consistent form, and pairs important estimates with design-consistent uncertainty. A concise report might show an estimated population size, a mean and selected quantiles for a continuous measure, proportions for key categories, and rates with their exposure units. A plot can make distributional patterns visible, while confidence intervals indicate which estimates are relatively precise. The reporting approach taught in Reporting Weighted Tables and Figures can help organize those outputs; the central task here is to ensure that the quantities form a coherent description of the intended population.
For each estimate, check that the analysis includes only records eligible for its numerator and denominator, uses weights appropriate to the target population, and applies the correct design information for uncertainty estimation. If weights were rescaled, verify that the reported total remains interpretable; rescaling may leave a weighted mean or proportion unchanged while changing the numerical total. A point estimate, its standard error, the interval, and the units should refer to the same estimand and analysis population.
Specify the population, time period, variable, eligibility rules, and any domain being described.
Use weighted totals, means, proportions, rates, quantiles, and distribution summaries as appropriate. Include units and denominators.
Use standard errors and confidence intervals consistent with the sample design, and check whether the approximation is appropriate.
Explain what the estimates say about the target population, how precise they are, and what the survey cannot establish.
Check Your Understanding
Answer each question using the definitions and calculations in this tutorial.
- In the hypothetical table, what does the weighted mean income estimate, and what are its units?
- Why does the weighted total of a binary insurance indicator estimate a population count, while dividing it by the weight total gives a proportion?
- Using the inverse-cumulative convention, why is the weighted median income 45 thousand dollars rather than 40 thousand dollars?
- Why should unequal-width histogram bars use weighted-count density when bar area is intended to represent population count?
- What additional information is needed to calculate a design-consistent standard error that cannot be obtained from the weights alone?
- Interpret a 95% confidence interval for a population rate, and explain why it is not a range of individual outcomes.