MerQuR+Statistics Glossary

Statistics Glossary → Weighted Total (Survey Total)

Weighted Total (Survey Total)analysis

An analysis that estimates the population total of a variable (such as a total number of people, total area or total spending), together with its design-correct standard error, from sample data collected under a complex sampling design.

Concept

Some questions are about totals rather than means: how many pupils in the province use school buses, what is the total cultivated area of the region's farms, what is households' total annual health spending? Such questions feed directly into planning and budget decisions. When the whole population cannot be counted, the population total is estimated from a sample, which requires weights showing how many population units each sampled unit represents.

The total is estimated by multiplying each observation by its weight and summing: T̂ = Σ wᵢyᵢ (the Horvitz-Thompson estimator). Unlike a mean, a total depends directly on the scale of the weights: the weights must add up to the population size, or the total will be wrong. The standard error comes from the stratum and cluster structure; similarity of units within a cluster increases uncertainty, while stratification and a high sampling fraction (the finite population correction) reduce it.

The relative precision of a total can be judged by the ratio of its standard error to the total (the coefficient of variation). For variables with a few very large values (e.g. the turnover of a handful of giant firms), the total can be overly sensitive to those units. Subgroup totals should be computed with the design kept intact, and differences between subgroup totals tested with design-correct standard errors.

When to use it

  • When a total quantity in the population (number of people, area, output, spending) has to be estimated from weighted survey or sample data.
  • If the interest is the value per person, use the weighted mean; if it is a category's share of the population, use the weighted frequency.
  • If records exist for the entire population (a full census), no estimation is needed; the total is computed directly.

An everyday example

To plan next year's school-bus budget, a provincial education directorate selects a stratified sample of middle schools and records how many pupils at each selected school use the bus. Suppose each sampled school in the rural stratum represents 4 schools and each in the urban stratum 10. Multiplying each school's number of bus users by its weight and adding up gives an estimate of about 14,500 bus users in the province. If the standard error is, say, 900, the directorate can plan for demand of roughly 12,700 to 16,300 pupils.

Parameters in MerQur

Weight (required)

Concept: The sampling weight is how many population units a sampled unit represents and is the inverse of the selection probability. In estimating a total, the scale of the weights feeds straight into the result: if the weights are scaled to the sample size rather than the population size (relative weights), the total is meaningless.

Everyday example: If an urban sampled school has a weight of 10, its 120 bus users enter the total as 1200 pupils. If the weights have been rescaled to average 1, the 'total' only reproduces the number of pupils in the sample, not the provincial total.

What it does: The sampling weight (inverse of the selection probability; the final weight with nonresponse or calibration adjustment if any). The total is computed as a weighted sum, the SE by Taylor linearization.

Effect on the result: The total and SE; the weight scale changes the total directly.

On the validation data: IceCream stratified total 36565.2.

→ Sampling weight (design weight)

Stratum

Concept: Stratification divides the population into non-overlapping groups and samples each separately. Because differences between strata do not add to sampling uncertainty, the standard error of the total shrinks when the strata differ.

Everyday example: If bus use is low in urban schools and heavy in rural ones, making town and countryside separate strata estimates each sector's contribution to the total separately and more precisely.

What it does: Computes the variance from the deviations of PSU totals within strata; design degrees of freedom = PSU - strata.

Effect on the result: SE, confidence interval and degrees of freedom; the point estimate does not change.

On the validation data: SE of the apistrat enroll total: unstratified 117625, stratified by school type 117319.

→ Stratified sampling (design stratum)

Cluster / PSU

Concept: The cluster (PSU) is the higher-level unit within which units are selected as a group. Because units in the same cluster are similar, effective information falls and the standard error of the total usually grows.

Everyday example: If districts were selected first and schools then drawn from within them, the district is the cluster. Schools in the same district have similar bus needs because of similar road conditions, so giving the cluster variable lets the standard error reflect this.

What it does: Sets the primary sampling unit; the variance is computed from PSU totals.

Effect on the result: The SE usually grows because of within-cluster similarity; degrees of freedom = PSU - strata.

On the validation data: SE: apistrat (no clusters) 117625; apiclus1 (clustered by dnum) 941611.

→ Primary sampling unit (PSU) and cluster sampling

FPC (population)

Concept: The finite population correction shrinks the variance by multiplying it by (1 − n/N) when a substantial part of the population is sampled. It does not change the total itself, only its uncertainty.

Everyday example: If 10 of the 40 rural schools were sampled, a quarter of that stratum has already been observed, and the stratum's contribution to the variance is reduced by a quarter.

What it does: Finite population correction; the variance is multiplied by (1 - n_h/N_h).

Effect on the result: The SE and confidence interval shrink; the total itself does not change.

On the validation data: SE: apistrat with FPC 114642 (117319 without FPC); apiclus1 cluster + FPC 932235.

→ Finite population correction (FPC)

Single-PSU stratum (No contribution (certainty, default), Centered at the average (adjust))

Concept: If a stratum contains only one cluster, the between-cluster variation within it cannot be computed. The option decides whether the stratum contributes nothing to the variance or whether a more cautious contribution is computed relative to the average of all clusters.

Everyday example: If one district stratum had only a single school sampled, how much bus use varies between that district's schools cannot be known. The cautious option adds this unknown to the standard error of the total.

What it does: Sets the variance contribution of a single-PSU stratum (no contribution or centred at the average).

Effect on the result: SE and interval only when a single-PSU stratum exists.

On the validation data: apiclus1 (stratum dnum) SE: no contribution 161733; centred at the average 162002.

→ Primary sampling unit (PSU) and cluster sampling

Subgroup (by)

Concept: A domain (subgroup) total estimates the total for a subpopulation while keeping the full design information. The subgroup totals add up to the overall total.

Everyday example: If the number of bus users is wanted district by district, district is given as the subgroup. Each district gets its own total and confidence interval, and together they add up to the provincial total.

What it does: Computes subgroup (domain) totals.

Effect on the result: A total per domain.

On the validation data: apiclus1, enroll total domain E 2.10972e+06, SE 631349.

→ Domain (subpopulation) estimation

Confidence Level (%90, %95, %99, Other (custom))

Concept: Sets how much assurance the confidence interval gives of covering the true population total. A higher level gives a wider interval.

Everyday example: If the number of bus users in the province lies between about 12,700 and 16,300 with 95% confidence, at 99% the range widens to about 12,200 to 16,800.

What it does: Sets the level of the confidence intervals.

Effect on the result: Intervals only.

On the validation data: apistrat total lower bound: 95% 3.4611e+06, 90% 3.49772e+06, 99% 3.38899e+06 (the same in the three programs).

→ Confidence level and confidence interval

Additional output: Unweighted n and sum of weights (N̂)

Concept: The unweighted n gives the number of actual sampled units the estimate rests on, and the sum of weights the estimated size of the population they represent. A sum of weights close to the known population size is a good sign that the weights are correctly scaled.

Everyday example: If the sample has 60 schools and the weights sum to 410, these 60 schools represent about 410 middle schools in the province. If the province is known to have 412 middle schools, the weights are consistent.

What it does: Gives the valid unweighted number of observations and the sum of weights per domain.

Effect on the result: Additional table only.

On the validation data: Domain E sum of weights 4873.97 (the three programs the same).

Additional output: Differences between domains (contrasts)

Concept: Gives the difference between subgroup totals with a design-correct standard error, so that you can judge whether two subgroups' totals really differ.

Everyday example: If the northern districts have 2000 more bus users than the southern ones, the contrast gives the standard error of that difference and helps decide how to split the fleet between the two regions.

What it does: Gives the difference in totals and its SE for pairs of domains.

Effect on the result: Additional table only.

On the validation data: apiclus1 E - H difference 1.57412e+06, SE 543747.

→ Domain (subpopulation) estimation

Variance method (Taylor (default), Jackknife (JK1 / JKn), BRR (balanced half-samples), Fay BRR, Bootstrap (Rao-Wu), Supplied replicate weights) and related fields (Fay coefficient (rho), Bootstrap replicates, Random seed, Replicate weight columns, Replicate type (JK1, JKn, BRR, Fay, bootstrap, other), Scale, Replicate scales (rscales), Deviations from)

Concept: The standard error can be computed by formula-based Taylor linearisation or by replication methods (jackknife, balanced repeated replication, bootstrap). Replication methods recompute the total in subsamples formed in line with the design and derive the uncertainty from how much those totals vary.

Everyday example: In the jackknife, clusters are left out one at a time, the remaining schools' weights are adjusted, and the provincial total is re-estimated. If the total changes a lot when a cluster is removed, the result is sensitive to that cluster and the standard error comes out large.

What it does: Computes the SE by resampling instead of Taylor (jackknife JK1/JKn, BRR, Fay BRR, Rao-Wu bootstrap, supplied replicate weights; details in 082). Deviations are taken from the replicate mean (MerQur default) or from the full-sample estimate.

Effect on the result: SE, confidence interval and degrees of freedom; the total does not change.

On the validation data: apistrat (stratum stype) JKn, 200 replicates: enroll total 3687177.53, SE 117319.09; api_stu total 3086008.63, SE 101841.20.

→ Taylor linearization and replication variance methods

Example on real data

Data: California API apistrat, R survey package

Variables: targets enroll (enrolment) and api_stu (students tested), weight pw, stratum stype (school type)

  • Question: From the stratified school sample (weight pw, strata stype), what are the population totals of enrolled students (enroll) and students taking the API test (api_stu)?
  • Design: n = 200 schools, 3 strata, no clusters; design df = 197; total T̂ = Σ w·y, Taylor SE.
  • Result (enroll): T̂ = 3,687,177.53 (SE = 117,319.09), 95% CI [3,455,815.02; 3,918,540.04].
  • Result (api_stu): T̂ = 3,086,008.63 (SE = 101,841.20), 95% CI [2,885,169.74; 3,286,847.52].
  • Additional finding: The total of students taking the API is about 84% of the total enrolment; the relative precision of the two totals is similar (SE / T̂ about 3.2% and 3.3%).
  • APA: enroll: T = 3,687,177.53, SE = 117,319.09, 95% CI [3455815.02, 3918540.04]; api_stu: T = 3,086,008.63, SE = 101,841.20, 95% CI [2885169.74, 3286847.52] (N = 200).

Related concepts

Standard errorDomain (subpopulation) estimationPrimary sampling unit (PSU) and cluster samplingTaylor linearization and replication variance methodsSampling weight (design weight)Finite population correction (FPC)Stratified sampling (design stratum)Stratified analysis (stratum variable)

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.