Weighted Tail and Threshold Summaries

Why Threshold Summaries Matter

Many survey questions are about how many people meet a rule, not about the average value of a measurement. A health agency may want the share of adults whose blood pressure exceeds a policy threshold; a program may track the share of households below an income cutoff; a clinical report may describe the share of patients whose measurement falls inside a target range. These are all summaries of a characteristic after applying a clearly defined threshold.

In a weighted survey, respondents can represent different numbers of people in the population. Counting respondents above a cutoff gives an unweighted sample proportion, which may not represent the population share. A weighted threshold summary instead adds the weights of respondents who meet the rule and compares that total with the weights of the eligible respondents. If the population size is known, the resulting estimated proportion can also be expressed as an estimated number of people.

Definition: Let $s$ be the set of sampled records included in a summary, let $w_i$ be the survey weight for record $i$, and let $x_i$ be its measured value. For a condition $A$, define the indicator $I_i(A)$ to equal 1 when record $i$ satisfies $A$, and 0 otherwise. The weighted proportion satisfying $A$ is $$ \hat{p}_A=\frac{\sum_{i\in s}w_i I_i(A)}{\sum_{i\in s}w_i}. $$ The denominator must use the same eligible records and weight definition as the numerator. Here, weights are assumed nonnegative and their sum over the included records must be positive.

Above and Below a Threshold

A threshold summary begins with the condition to be counted. “Above $c$” can mean $x_i>c$, while “at or above $c$” means $x_i\ge c$. The analogous lower-tail conditions are $x_i<c$ and $x_i\le c$. The distinction matters whenever one or more records equal the cutoff. The inequality should follow the policy or clinical definition, rather than being chosen after looking at the data.

For an upper-tail share at or above $c$, each record’s indicator is $I_i(x_i\ge c)$. The weighted numerator is therefore the sum of the weights for records at or above $c$. Dividing by the sum of weights gives the weighted proportion; multiplying that proportion by 100 gives the percentage. The same construction applies below a cutoff, with the indicator changed to match the stated rule.

Worked Example: Estimated Share at or Above a Policy Threshold

Suppose a hypothetical survey records a measure, in units, for eight eligible respondents. The stated policy rule is “at or above 80 units.” The table gives each respondent’s value and survey weight.

Respondent	Value (units)	Weight	Meets value ≥ 80?
A	62	10	No
B	74	15	No
C	80	20	Yes
D	68	10	No
E	91	15	Yes
F	75	10	No
G	58	10	No
H	83	10	Yes

The total weight is $10+15+20+10+15+10+10+10=100$. Respondents C, E, and H meet the inclusive threshold, and their weights sum to $20+15+10=45$. Thus,

\hat{p}_{x\ge80}=\frac{45}{100}=0.45=45\%.

The estimated weighted share at or above 80 units is 45%. Respondent C is included because the rule is inclusive: a value exactly equal to 80 meets it.

From a Weighted Proportion to a Population Count

A proportion is a share; a count is a number of people or other population units. If the population size $N$ for the target population is known and the weighted proportion estimates the share meeting condition $A$, the estimated count is $N\hat{p}_A$. This is an estimate, not necessarily the exact number in the population. It may be non-integer because it is calculated from an estimated share.

Do not assume that the sum of the weights in the sample is the known population size. The ratio defining $\hat{p}_A$ uses the weight sum as its denominator, while the count calculation uses the separately known $N$. If weights have been constructed so that their sum over the relevant population represents $N$, then the weighted numerator itself gives the corresponding total. In general, when $N$ is supplied independently, multiply $N$ by the weighted proportion.

Worked Example: Converting a Weighted Share to an Estimated Count

Continue with the hypothetical threshold summary above. Suppose the target population contains a known $N=2{,}400$ people. The weighted proportion at or above 80 units was $\hat{p}_{x\ge80}=0.45$. The estimated population count is

\hat{T}_{x\ge80}=N\hat{p}_{x\ge80} =2{,}400(0.45)=1{,}080\text{ people}.

The estimate is 1,080 people in the target population meeting the threshold rule. It is not the number of sampled respondents who met the rule, which was three, and it is not the sum of the sample weights meeting the rule, which was 45. Those are quantities on different scales.

Below a Clinical Cutoff

A lower-tail summary is calculated in the same way, but the indicator must represent the intended cutoff rule. For a strict rule such as “below 120,” records with value exactly 120 are not counted. If the rule instead says “120 or below,” those tied records belong in the numerator. A sound report states the cutoff, inequality, and measurement units together.

Worked Example: Share Below a Clinical Cutoff

In a hypothetical survey, seven eligible respondents have a measurement in millimetres of mercury (mm Hg) and the following weights. The clinical rule for this example is strictly below 120 mm Hg.

Respondent	Measurement (mm Hg)	Weight	Below 120?
A	110	10	Yes
B	118	15	Yes
C	120	20	No
D	125	10	No
E	135	10	No
F	115	5	Yes
G	128	10	No

All weights sum to $10+15+20+10+10+5+10=80$. The weights for measurements strictly below 120 mm Hg sum to $10+15+5=30$. Therefore,

\hat{p}_{x<120}=\frac{30}{80}=0.375=37.5\%.

The weighted estimate is that 37.5% of the eligible population is below 120 mm Hg. The respondent recorded at exactly 120 mm Hg is excluded because the stated condition is $x_i<120$. If the known population size were 4,000 people, the corresponding estimated count would be $4{,}000(0.375)=1{,}500$ people.

Shares Within a Specified Range

A range summary counts observations satisfying both a lower and an upper bound. A range can include both endpoints, exclude both, or use a mixed convention; its definition should make this explicit. For example, “between 18.5 and 24.9, inclusive” means $18.5\le x_i\le24.9$. The numerator adds the weights for all records meeting both inequalities, and the denominator remains the sum of weights for the same eligible records.

Result: For an inclusive range from $a$ to $b$, where $a\le b$, the weighted proportion is

\hat{p}_{a\le x\le b} =\frac{\sum_{i\in s}w_i I_i(a\le x_i\le b)} {\sum_{i\in s}w_i}.

The estimated count in a known population of size $N$ is $N\hat{p}_{a\le x\le b}$. For a range with different endpoint rules, change the indicator to match those rules.

Worked Example: Share Within an Inclusive Target Range

Suppose a hypothetical survey records body mass index (BMI, in kg/m²) for seven eligible respondents. The summary sought is the weighted share with BMI from 18.5 through 24.9, inclusive.

Respondent	BMI (kg/m²)	Weight	Within 18.5–24.9?
A	17.8	5	No
B	18.5	10	Yes
C	21.2	15	Yes
D	24.9	10	Yes
E	25.0	8	No
F	23.4	7	Yes
G	30.0	5	No

The denominator is the total weight, $5+10+15+10+8+7+5=60$. The in-range records are B, C, D, and F. Their weights sum to $10+15+10+7=42$, so

\hat{p}_{18.5\le x\le24.9}=\frac{42}{60}=0.70=70\%.

The estimated share in the inclusive target range is 70%. Both endpoint values in the table, 18.5 and 24.9, are included. If the known population size were 12,000 people, the estimated count would be $12{,}000(0.70)=8{,}400$ people.

Why the Indicator Formula Works

A threshold condition converts each measured value into a binary outcome: it either meets the rule or does not. A weighted mean of these zero-one outcomes is exactly the weighted proportion meeting the rule. To see this, let $z_i=I_i(A)$. In the weighted sum, each record satisfying $A$ contributes $w_i(1)=w_i$; each record not satisfying $A$ contributes $w_i(0)=0$. Therefore the numerator is precisely the total weight of records meeting the condition:

\frac{\sum_{i\in s}w_i z_i}{\sum_{i\in s}w_i} = \frac{\text{weight of eligible records satisfying }A} {\text{weight of all eligible records}}.

This also explains why a threshold summary is not generally the same as the unweighted fraction of sampled records meeting the rule. The unweighted calculation gives every sampled record equal influence, whereas the weighted calculation gives each record influence proportional to its survey weight. In the first worked example, three of eight respondents met the rule, so the unweighted sample fraction is $3/8=37.5\%$, while the weighted estimate is 45%. The two describe different calculations; the weighted estimate is the one that accounts for the supplied survey weights.

Practical Checks and Common Mistakes

Before reporting a threshold summary, verify four details: the target group whose records belong in the denominator, the measurement units, the exact inequality at each endpoint, and whether a count is requested in addition to a proportion. If the summary is restricted to an eligible subgroup, the numerator and denominator must both refer to that subgroup. A threshold share is descriptive; it does not by itself quantify sampling uncertainty or establish why individuals fall on one side of the cutoff.

Ignoring ties. A cutoff value can occur in the data. Decide whether equality qualifies, and encode that decision with $<$, $\le$, $>$, or $\ge$.
Using an unweighted fraction. Counting qualifying respondents and dividing by $n$ ignores variation in weights. Use the weighted numerator and denominator for the weighted estimate.
Calling a share a count. A value such as 0.45 is a proportion, not 45 people. Multiply by the known population size to estimate a count.
Using the wrong count multiplier. The population count is $N\hat{p}_A$ when $N$ is known. Do not substitute the number of sampled records or assume the sample weight sum equals $N$.
Leaving the range ambiguous. “Between $a$ and $b$” does not always communicate whether endpoints are included. Write the inequalities explicitly.

Key takeaway: Define the threshold or range as a condition, turn that condition into a zero-one indicator, and calculate its weighted mean. This yields the weighted proportion; multiplying by a known target-population size yields the corresponding estimated count.

Check Your Understanding

For each question, state the relevant condition and show how the weighted summary would be formed.

A rule counts values strictly greater than 50. What happens to a record exactly equal to 50, and which inequality defines the indicator?
In a weighted sample, 28 of 70 total weight units meet a threshold. Calculate the weighted proportion as a percentage.
If the estimated weighted proportion is 0.32 and the known target population contains 7,500 people, calculate the estimated count and state its units.
For a range from 10 through 20 inclusive, write the condition a record must satisfy. Are values equal to 10 or 20 included?
Explain why the number of sampled respondents meeting a rule need not equal the weighted estimated proportion multiplied by the sample size.