A published survey percentage can differ from the percentage obtained by dividing responding students into groups. That is not necessarily an error. The published estimate may use weights, giving different respondents different contributions to an estimate. A reader needs to distinguish the observed number of responses from the weighted result.

The UK Data Service notes that many social surveys use complex sampling and contain survey weights. That establishes why checking the methodology matters. The following calculation is CollegeVerse's deliberately simple fictional illustration. It is not a reconstruction of any official graduate survey, a recommended weighting scheme, or evidence that a particular university adjusts its results this way.

Reproduce both denominators

Imagine a survey with 100 responding students. Eighty are in Group A and twenty in Group B. Within A, sixty report the outcome of interest; within B, ten do. The unweighted result is therefore seventy out of one hundred, or 70%.

Now suppose the supplied demonstration weights are one for every A respondent and four for every B respondent. Each positive response and each denominator contribution receive the same respondent's weight.

Group Actual respondents Positive responses Weight Weighted denominator Weighted positives
A 80 60 1 80 60
B 20 10 4 80 40
Total 100 70 Not one common weight 160 100

The weighted proportion is 100 divided by 160, or 62.5%. There are still only 100 actual respondents and seventy observed positive responses. The weighted positive total of 100 is not one hundred different people reporting that outcome. Reporting it as a headcount would change the meaning of the data.

The estimate falls here because B has a lower within-group outcome rate and receives more relative weight. No individual response has changed. A different weight structure could produce a different direction. The illustration demonstrates the mechanism, not an argument that weighted results should always be lower or higher.

Check what the weight is intended to represent

Read the producer's description of the target population, sampling design, response adjustments and weight variable. Some datasets provide multiple weights for different analysis populations or time periods. Selecting a variable simply because its name contains “weight” is not a defensible substitute for the instructions.

In the toy example, multiplying all weights by ten leaves the proportion unchanged: 1,000 divided by 1,600 still equals 62.5%. It dramatically changes the sum of weights. This is why a weighted base cannot automatically be interpreted as either the sample size or a population count without understanding its scale and purpose.

If a table lists a percentage beside a base, ask whether the percentage is weighted while the base is unweighted. That presentation can be legitimate and useful when labelled clearly. Do not force the two to reconcile through simple integer multiplication if the publication explicitly uses different bases.

Keep precision and bias questions open

A weighted denominator of 160 does not give this fictional dataset the information of 160 independent respondents. Complex designs and unequal weights matter for uncertainty. Use the producer's appropriate uncertainty estimates and methods; do not calculate a confidence interval by treating the weighted total as an ordinary random-sample size.

Weighting also does not prove that every source of nonresponse bias has disappeared. The available adjustments may not capture all differences between respondents and nonrespondents. Pair this audit with the graduate survey nonresponse guide when interpreting outcome claims.

For a shortlist note, retain three separate fields: actual response count, reported weighted percentage, and the methodology explaining the weights. If those details are unavailable, state the limitation. The goal is to reproduce the producer's intended measure and understand its bounds, rather than replacing it with an easier calculation that answers a different question.