Organizations are not people.
These are public organization-level aggregates. They cannot explain an individual employee’s experience, identify subgroups, or support individual-level predictions. Avoid the ecological fallacy.
Public evidence. Human questions.
An independent HR analytics studyPUBLIC SURVEY ATLAS
An independent HR analytics studyPUBLIC SURVEY ATLAS
THEENGAGEMENTATLAS.
Engagement is a signal.
Understanding it takes a wider view.
The premise
How does the employee experience differ across organizations, change over time, and inform what we ask next?
An exploration of fully public, organization-level survey aggregates. No individual records. No simulated survey results.
01The view
The view
across time.
Averages tell one story. The spread between organizations tells another. Start with both, and check who is represented each year.
Engagement, year by year
INDEX · 0–100Mean engagementMedianContinuously observed agency names · meanShaded band: organization min–max
Annual summaries are across included agencies, not an estimate for all employees. Changing coverage can change the mean. The continuous cohort, when shown, uses reporting names observed in every source year—not a guarantee of unchanged agency boundaries or workforces.
Read the annual data table
02Every point.
Every point.
A different context.
Look beyond the leaderboard. Explore how a measured aspect of work sits alongside engagement, one public agency at a time.
Each dot is one published agency aggregate; all dots have equal size. Search filters the view. Missing values are excluded from the plot, not replaced. Intent to leave, when available, is self-reported survey intention—not observed attrition.
Who is represented in the intention-to-leave supplement?
Read the filtered organization table
03Related.
Related.
Not independent.
The engagement index contains Leaders lead, Supervisory support, and Intrinsic work experience. Their correlations with the whole are partly mechanical—not discoveries of independent drivers.
Feature ↔ engagement
Spearman rank correlation (ρ), with the supplied p-values and multiple-testing-adjusted q-values. EEI components share content by construction; global satisfaction comes from the same survey. Shared responses and omitted context also contribute. None of these relationships identifies a causal effect.
Read all correlation statistics
04The path matters
The path matters
as much as the point.
A single snapshot hides direction. Compare published organization trajectories without treating year-to-year differences as proven change in individual employees.
Lines connect consecutive survey years only. Missing years remain gaps; organization naming or survey changes may affect comparability. Default examples are alphabetical, not selected for performance.
Read the selected trajectories
What changes together within agencies?
FIRST DIFFERENCES · DESCRIPTIVEAdjusted associations in annual engagement change per one standard deviation of feature change, with year effects. EEI components are part of the outcome itself: a close fit is partly compositional, not validation of independent drivers or interventions.
Read adjusted change estimates and uncertainty
Read change correlations and cluster-bootstrap intervals
Within-agency change methodology
05A landscape.
A landscape.
Not a taxonomy.
Multiple survey dimensions, seen together. A two-dimensional PCA projection makes similarities visible; descriptive clusters organize questions, not people.
Multivariate survey landscape
Coordinates are principal-component scores, not index points. Proximity in this projection is an imperfect summary of the full feature space. Cluster numbers have no rank.
Read the PCA coordinates and clusters
How to interpret these descriptive clusters
Feature standardization gives survey dimensions comparable influence on the projection. Cluster profiles below return to the original 0–100 scale. A silhouette score is a geometric diagnostic, not evidence of distinct types of organization; neither PCA nor clustering identifies causes.
06The future.
The future.
With a reality check.
A next-year engagement forecasting tournament, judged on an actual temporal holdout. More sophisticated models must earn their place against a simple baseline.
The model tournament
MAE · INDEX POINTS · LOWER IS BETTERRead the full holdout scorecard
Does each ML model improve on persistence? Read the paired comparison
Prediction meets observation
Selected model predictions against actual held-out engagement. The diagonal is perfect agreement, not a fitted line. These are retrospective forecasts, not future promises.
Read held-out predictions and baseline
Inside the explanatory ML model
Supplied permutation importance: change in model error when a feature is shuffled. This is predictive relevance, not causal leverage; correlated features can share or obscure importance. The explained ML model need not be the tournament winner.
Read permutation importance statistics
Open the forecasting methodology
07Better questions.
Better questions.
Not easy verdicts.
Responsible people analytics makes the limits visible. The purpose of this atlas is to guide evidence-seeking conversations—not to automate judgments about workforces.
A pattern is not a lever.
Correlation, clustering, and predictive importance are descriptive or predictive. None identifies the effect of an HR intervention. No causal uplift, business impact, or “what-if” promises are made here.
Context before action.
Check survey definitions, response coverage, uncertainty, and local context. Pair the findings with employee listening and domain expertise before designing or evaluating a response.
Where to take the conversation
Data quality, in plain sight
Read feature missingness
Measurement and comparability
OPM FEVS engagement (EEI) is a percent-positive index on a 0–100 scale, composed of Leaders lead, Supervisory support, and Intrinsic work experience. Global satisfaction and the other survey percentages also use 0–100. A difference in index points is not a relative percentage change. Intent to leave is self-reported, not an actual exit rate.
Published respondent counts and response rates provide context but do not eliminate nonresponse bias. Changes in agency boundaries, instruments, wording, and response composition may compromise comparisons across years. Agency means should not be interpreted as employee-weighted federal estimates. See OPM documentation for exact item definitions and survey methodology.
What this analysis cannot establish
Aggregate data cannot diagnose individuals, demonstrate subgroup fairness, recover within-agency distributions, or identify causal intervention effects. Forecast evaluation in historical years does not guarantee transportability to later years, other sectors, or changing survey instruments.
Analysis design and reproducibility
Use the linked repository for source ingestion, cleaning decisions, feature definitions, temporal evaluation, and reproducible code. Source publications, not this visualization, remain authoritative.
Public-data boundary
Only fully public agency-level survey aggregates belong in this portfolio. No confidential telemetry, respondent-level records, de-anonymized customer lists, or synthetic substitute survey data are used. Public availability does not remove the obligation to respect source licensing and avoid stigmatizing interpretations.
08Open evidence.
Open evidence.
Open workings.
Good analysis should be inspectable. Follow the publication, understand the terms, and examine the analytical choices behind the visuals.
A portfolio, not an official publication
This is an independent analytical interpretation of public survey releases. It is not endorsed by the source organization and should not be used as a substitute for official statistics or local interpretation.