Check who answered which question before interpreting a score. In a small team, individual answers can move the result substantially, while nonresponse can limit the picture.
Four facts to report with a score
- The number invited and eligible.
- Valid response count for the specific question.
- The distribution, not just an average or eNPS.
- Dates, wording and relevant organisational changes.
A 70% participation rate does not guarantee representativeness. Nonrespondents may have different experiences. Inviting the whole workforce does not remove nonresponse bias.
A small-group eNPS example
For 20 answers, 10 promoters, 6 passives and 4 detractors give +30. Moving one answer from detractor to promoter produces +40. One response changed although the chart shows a ten-point increase.
This is arithmetic, not proof of random variation or statistical significance. The eNPS formula includes passives in the denominator.
Has the situation improved?
Check that wording, scales and groups are comparable. Workforce changes and participation shifts can affect a trend. Anonymous waves usually compare respondent groups rather than paired answers from the same individuals.
Uncertainty intervals require a justified model. Do not automatically apply a random-sample margin-of-error calculator to a voluntary workforce survey. AAPOR requires clear assumptions for precision estimates in non-probability studies.
Segmentation: group responses for a research purpose
Background fields may include department, tenure band or location. Choose details relevant to a decision, not every field available in HR records. Tenure may help investigate onboarding; work location may help investigate access to tools. Do not add sex, age or health fields without a justified purpose and review of legal requirements and identification risk.
Define group membership before collection: reference date for tenure, department transfers and people working across teams. Bands should not overlap and should cover the intended answers. A missing background answer is missing data, not a characteristic to infer.
A response without background fields can still contribute to the overall score if it otherwise meets the survey rules. It cannot enter a breakdown requiring the missing field. Explain why displayed groups may not cover all respondents. Do not remove someone from overall eNPS simply because they omitted their department.
Calculate group and organisation scores correctly
Use the same eNPS formula within a group, with that group’s valid response counts. Do not divide promoters by all employees on the roster. Missing eNPS answers are still not zeros.
Illustrative example: two non-overlapping groups, one question and one wave:
| Group | Valid answers | Promoters | Passives | Detractors | eNPS |
|---|---|---|---|---|---|
| A | 10 | 8 | 0 | 2 | +60 |
| B | 90 | 27 | 18 | 45 | −20 |
| Combined | 100 | 35 | 18 | 47 | −12 |
The unweighted average (+60 − 20) / 2 = +20 is not the eNPS of these 100 respondents. The correct result is 100 × (35 − 47) / 100 = −12. Weighting group scores by valid counts gives the same answer for exhaustive, non-overlapping groups with consistent definitions. Use original counts to avoid rounding errors.
Valid counts also matter when combining means for the same item. Do not extend this directly to different indices, scales or overlapping groups. The “average department” and “average employee response” are different measures; name the measure you calculate.
When must a breakdown be withheld?
Consider disclosure risk and valid answers to the particular item, not just department headcount. Seven respondents may provide only two answers to one question. Set reporting rules before analysis; GDPR does not prescribe a universal five-person minimum.
Review small answer categories, comments and complementary cells. Hiding one team is ineffective if its result can be reconstructed from totals and other teams. Combine meaningful categories, restrict breakdowns or withhold the relevant report section. Even exact counts may need withholding if they identify people. See privacy examples and safeguards.
Comparing departments and many breakdowns
Changes in department participation can improve the overall score without improvement within any department. Check respondent composition and changes within comparable groups. A higher average does not automatically establish an HR intervention’s effect.
Inspecting many breakdowns increases the chance of selecting a random extreme. Plan the main comparisons in advance and label additional analyses exploratory. A difference invites investigation; it does not alone justify ranking managers. Do not add weights to “fix participation” without a justified model and population information.
When does a benchmark help?
When its population, period, sample size, instrument and aggregation are documented. An average of company scores differs from a score pooling all respondents. Across countries, also consider language and context. Without documentation, do not label an illustrative line an “industry average”.
Write a careful conclusion
Instead of “the action improved engagement”, write: “Ratings of expectation clarity increased in this wave; we will discuss the context with the team.” Separate observation, possible explanation and decision.
Review small-group confidentiality, then prepare an action plan. Create an account to organise the survey process.
Form completion and valid rating counts are different. An explicit N/A may satisfy a mandatory item, but does not add a numeric rating or become a zero. Participation or completed status does not establish complete evidence for every metric.
Sources and scope
These sources support the principles discussed here. Examples and schedules are starting points to adapt, not standards or evidence that our templates have been validated.
