Why Weight a Frequency Table?
A frequency table summarizes how many units fall into each category of a categorical variable. In a survey, however, respondents may represent different numbers of population members. A table of raw respondent counts gives each respondent equal influence, so it may not describe the population well when the survey weights differ. A weighted frequency table uses the weights to estimate category totals and the corresponding population shares.
The interpretation depends on two choices: which records belong in the analysis, and what total the percentages use as their denominator. For example, percentages may describe shares of the full target population, shares among records with a valid response, or shares within a stated population domain. A percentage without its denominator is incomplete.
When the weights are expansion or final survey weights, \(\widehat{T}_k\) estimates the population count in category \(k\). A weighted count need not be an integer: it is an estimate based on the survey weights, not a count of respondents. If weights have instead been normalized to a scale that does not represent population counts, their weighted category totals should not be described as population counts without appropriate rescaling. The ratio defining a share uses the weights in both numerator and denominator.
Constructing a One-Way Weighted Table
For a one-way table, each included record contributes its weight to exactly one category, provided the categories are mutually exclusive and every included record has a valid category value. The estimated total represented by those records is \(\widehat{T}=\sum_i w_i\). If the categories form a complete partition of the included records, their weighted counts add to that total. Dividing each count by the same total gives shares that add to 1, or percentages that add to 100%, subject to rounding.
The key step is that, for a complete and mutually exclusive set of categories, each included record belongs to one and only one category. Thus \(\sum_k I(y_i=k)=1\) for every included record. Dividing the identity by \(\widehat{T}\) shows that the category shares sum to 1. If missing values are excluded, this result applies to the valid records and their weighted total—not automatically to the full target population.
Worked Example: Weighted Counts and Shares for Three Categories
Suppose a hypothetical survey records preferred contact method for eight respondents. Their weights represent population counts, and all eight have a valid response.
| Respondent | Category | Weight |
|---|---|---|
| 1 | Phone | 2,400 |
| 2 | 1,600 | |
| 3 | Phone | 1,000 |
| 4 | 2,000 | |
| 5 | 1,200 | |
| 6 | Phone | 1,800 |
| 7 | 900 | |
| 8 | 1,100 |
The estimated total represented population is \(2{,}400+1{,}600+1{,}000+2{,}000+1{,}200+1{,}800+900+1{,}100=12{,}000\). The weighted count for Phone is \(2{,}400+1{,}000+1{,}800=5{,}200\); for Email it is \(1{,}600+1{,}200+1{,}100=3{,}900\); and for Mail it is \(2{,}000+900=2{,}900\).
Using the full represented total as the denominator gives:
The counts sum to 12,000, and the percentages sum to 100.00%. In this hypothetical survey, the weighted table estimates that 43.33% of the represented population prefers phone contact, 32.50% prefers email, and 24.17% prefers mail. These are weighted population shares, not the percentages of the eight respondents in each category.
Choose and Label the Denominator
A denominator is defined by the population the percentage is intended to describe. For a full-population distribution, use the weighted total for the full target population represented by the included records. If the target is a particular subset, restrict both the category count and the denominator to that subset. If some category values are missing, decide whether the table treats missingness as a category or excludes those records; the resulting denominator must be stated.
For example, if the question is “What share of all eligible people is insured?”, records without an insurance response do not belong in the numerator for “insured,” but they remain part of the full-population denominator if they are eligible members of the target population. If the question is instead “Among people with a known insurance status, what share is insured?”, the denominator is the weighted total of valid responses. Those percentages answer different questions, even when they use the same weighted insured count.
Worked Example: Full Eligible Population Versus Valid Responses
In a hypothetical survey about insurance status, the weighted count is 1,700 insured, 900 uninsured, and 400 records with missing status. Assume these three groups together represent 3,000 eligible people. The 400 missing-status records are not a substantive insurance category.
For a table showing shares of the full eligible population, the denominator is 3,000. The insured share is \(1{,}700/3{,}000\times100=56.67\%\), and the uninsured share is \(900/3{,}000\times100=30.00\%\). The remaining \(400/3{,}000\times100=13.33\%\) have missing status. If missingness is displayed as a separate table row, the three percentages add to 100%.
For a table describing insurance status among valid responses, the denominator is \(1{,}700+900=2{,}600\), not 3,000. The valid-response percentages are \(1{,}700/2{,}600\times100=65.38\%\) insured and \(900/2{,}600\times100=34.62\%\) uninsured. These add to 100% because missing-status records are excluded from both numerator and denominator.
The 56.67% figure estimates the share of all eligible people known to be insured based on the weighted count, while 65.38% describes the distribution among those with known status. Excluding missing records changes the population being described; it does not simply tidy the same percentage.
A table may also state a particular margin as its denominator. For instance, a distribution restricted to rural residents should use the weighted rural total, not the total for all residents. The table heading or percentage label should make that base explicit, such as “Percent of rural residents.” This tutorial concerns one categorical variable at a time; when a table has multiple categorical dimensions, each percentage’s row or column base requires additional care.
Worked Example: Percentages Within a Stated Population Domain
Suppose a hypothetical survey records preferred travel mode and whether each sampled person is in the rural or urban domain. The weights below represent population counts. The requested table describes travel mode among rural residents, so its denominator is the rural margin.
| Domain | Travel mode | Weight |
|---|---|---|
| Rural | Bus | 800 |
| Rural | Car | 600 |
| Rural | Bus | 400 |
| Urban | Car | 1,000 |
| Rural | Walk | 200 |
| Urban | Walk | 700 |
The rural weighted total is \(800+600+400+200=2{,}000\). Within that domain, the weighted counts are 1,200 for Bus, 600 for Car, and 200 for Walk. Therefore the rural percentages are \(1{,}200/2{,}000\times100=60.00\%\) for Bus, \(600/2{,}000\times100=30.00\%\) for Car, and \(200/2{,}000\times100=10.00\%\) for Walk.
These percentages describe the distribution among represented rural residents. The urban records do not enter either the rural category counts or the rural denominator. Using the all-domain total of 3,700 instead would answer a different question: it would express rural category counts as shares of the combined rural and urban population, not as the distribution within the rural margin.
Checks and Interpretation
Before reporting a weighted frequency table, verify that the category definitions are exhaustive and mutually exclusive for the records included, and check that all records use the intended weights. Then verify that the weighted category counts sum to the denominator represented by the table. When percentages are meant to describe a full distribution over those same records, confirm that they sum to approximately 100%; small discrepancies can arise from rounding.
For each percentage, name the population or margin in words. “Email is 32.50%” is incomplete; “Email accounts for an estimated 32.50% of the represented full population” identifies its interpretation. Likewise, a percentage based on valid responses should say so. Weighted descriptive percentages summarize the represented population under the weighting and record-inclusion choices; they do not by themselves state the uncertainty of the estimate.
Common Mistake
A frequent error is to divide every weighted category count by the sum of the weights in the dataset, even when the table is meant to describe only valid responses or a particular domain. The result may be arithmetically correct but answer the wrong question. The correction is to define the target records first, calculate category counts within those records, and divide by the weighted total of those same records. Another error is to use raw respondent counts for the numerator while using weighted totals for the denominator; numerator and denominator must use consistent weights and record rules.
Check Your Understanding
For each question, identify the appropriate weighted count or denominator and explain the interpretation.
- A category has weighted count 750, and the full target population has weighted total 5,000. What is its full-population percentage?
- A table excludes records with missing category values. Which weighted total should be used to make the valid-category percentages sum to 100%?
- Why can a weighted category count be non-integer even though people are counted in whole numbers?
- A report gives the share of a category among residents in one domain. Which records belong in the numerator and denominator?
- What does it indicate if category weighted counts sum to less than the stated full-population denominator and missing values were not shown as a category?