An engineer has no revenue signal. A salesperson has no code review history. A support lead has no sprint velocity. None of that is a performance finding. It is a description of what those jobs produce.
Most workforce scoring systems handle this the same way. They run every person through one scorecard, and where an input is absent they fill it with a low value, usually a zero, sometimes a population average. The number that comes out reads like a measurement of the person. A meaningful part of it is a measurement of how closely that person's job resembles the job the scorecard was designed around.
We decided early that Levos would not do this. The reasoning is not proprietary and it is not new. The statistics literature settled it more than twenty years ago. It simply has not made the trip into people analytics.
What counts as missing data in employee scoring?
Four different situations produce an identical blank cell, and they carry four different meanings.
The role does not generate the artifact. An account executive has no pull requests. A recruiter has no deal velocity. A site reliability engineer has no quota attainment. The signal is not late or lost. It was never going to exist.
The source tool is not connected. The signal exists inside a system nobody has authorized yet, or inside a system the team stopped using in March. The work happened. The evidence is sitting somewhere the platform cannot see.
The history is short or interrupted. New hires, part-time employees, people returning from leave, contractors, and anyone who moved teams last month all carry thinner records than a tenured employee in a stable role. Thin is not low.
The work happens somewhere nothing observes. Whiteboard sessions, hallway decisions, customer conversations held outside a recorded system, mentoring. Some of the most valuable work in any organization leaves the faintest trace, and it is not evenly distributed across roles or levels.
Four causes, one blank cell
A model that cannot separate these four is not measuring performance. It is measuring instrumentation coverage and calling the result a person. That is the whole argument, and everything below is the evidence for it.
Why a zero measures the job, not the person
A zero is a claim. A blank is the absence of one. Collapsing the second into the first is the point where a scoring system stops describing the workforce and starts describing itself.
Consider what the same fill rule does to two employees in the same company.
| The absent input | What a zero-filled scorecard records | What is actually true |
|---|---|---|
| Engineer, no pipeline record | Weak revenue contribution | The role does not carry pipeline |
| Salesperson, no code review data | Weak output quality | Quality shows up in proposals and calls |
| New hire, six weeks of history | Below-average across most families | Six weeks of evidence, no basis for the comparison |
| Team on an unconnected tool | Low delivery signal | The delivery record exists, in a system nobody authorized |
| Employee on parental leave | Sharp decline | Protected leave |
Every row produces a defensible-looking number attached to an indefensible comparison. Because the number carries the same formatting, decimal places, and dashboard treatment as a fully evidenced score, nothing downstream flags it. The manager preparing a review sees a figure. The finance lead building a headcount case sees a figure. Neither sees the gap the figure was built to cover.
What the statistics literature settled in 2002
This is not a novel objection, and it is worth citing the source rather than asserting it.
Schafer and Graham, in "Missing Data: Our View of the State of the Art," published in Psychological Methods in 2002, reviewed the standard approaches and were direct about substitution: mean imputation "may accurately predict missing data but distort estimated variances and correlations." That fragment is the entire problem in eleven words. The imputed value can look plausible at the level of the individual cell while corrupting every relationship the dataset was built to reveal.
They also quantified the damage in a way that should trouble anyone shipping a workforce score. With roughly 25% of data missing, they reported that a nominal 95% confidence interval delivered approximately 86% true coverage. The interval says the error rate is 5%. The actual error rate is closer to 14%, nearly three times what the label claims.
The damage lands on the confidence, not only on the score
State the limit honestly: that is a general result about imputation in statistical estimation, not a study of employee performance data. We are not going to convert it into a workforce benchmark, and anyone who quotes it back at you as one is overreaching. What transfers is the shape of the failure, and the shape is what matters here. Reliability degrades faster than completeness does. A dataset that is three quarters complete does not produce a score that is three quarters as trustworthy. It produces a score whose stated confidence is inflated, which is worse than a score that is simply low, because a low score invites a question and an inflated confidence closes one.
Missing data is rarely random, and that is the expensive part
Everything above holds even if the gaps fall evenly. They do not.
The classification still used for this comes from Rubin, writing in Biometrika in 1976, which separated data missing completely at random, data missing at random given what else is observed, and data missing not at random, where the reason a value is absent is related to the value itself. Only the first case is safe to ignore. Workforce data is frequently in the third.
The three types, and which one you have
The examples below are ours, offered as illustration rather than as findings from that paper.
| Type | Workforce illustration | What a fill rule does |
|---|---|---|
| Missing completely at random | A connector outage drops one week for everyone | Adds noise, distributes it evenly |
| Missing at random | Sales roles lack code signals, and role is recorded | Correctable, if the model uses role |
| Missing not at random | A struggling team quietly stops logging work | Reads the silence as a low score and buries the reason for the silence |
That third row is where the real cost sits. The people least likely to supply an input are often the people the input would have described least favorably, which means the missingness itself carries the signal being scored, and filling it with a zero or a mean both hides the signal and misstates the person.
This evidence comes from machine learning, not from HR
Min, Asif, and Vaidya, in "Exploring the inequitable impact of data missingness on fairness in machine learning," published in IEEE Intelligent Systems in 2025, built 150 experimental dataset variants from three real-world datasets. Where missingness correlated with sensitive attributes and with outcomes, it worsened disparities "even for little missingness," and produced a feedback loop in which a decision made on incomplete data reinforces the non-disclosure that made it incomplete.
That is fairness research on machine learning systems, not an employment study, and we are labeling the transfer rather than hiding it. The mechanism is what carries across: correlated missingness plus a fill rule equals compounding distortion, and the compounding runs in the direction of whoever was already least visible.
There is no regulation telling you how to handle this
Worth saying plainly, because several vendors imply otherwise. There is no regulatory guidance on missing data in worker evaluation. The EEOC's 2023 technical assistance on algorithmic tools in employment selection, which has since been removed from eeoc.gov, was scoped to adverse impact under Title VII rather than to how a model treats an absent input. NIST's AI Risk Management Framework does not address it either.
The statistics literature resolved this in 2002. Regulators have not arrived at it. Anyone waiting for a compliance requirement to force the issue will wait through several review cycles, and each one will produce decisions about real people made on scores that quietly counted a blank as a failure.
What Levos does with a signal that is not there
We do not read a missing input as a zero. That is the principle, and it is why our methodology page states the position rather than burying it: "Every score has a confidence rating. Every signal cites its source. We tell you what we know, what we are inferring, and what we cannot tell you yet."
Three consequences follow from that.
Evidence is read for what it is. Levos derives six families of behavioral signals from connected tools. A revenue signal is meaningful for a rep and says nothing about a platform engineer, and reading an unpopulated family as a weak result is the error this pillar exists to name.
Uncertainty is reported, not absorbed. Every signal carries a confidence rating based on data freshness, source quality, and pattern strength, and every family score carries a confidence rating and a source citation. A thin record and a poor record are separate statements, and collapsing them is what most single-number dashboards do. That is the subject of the next pillar in this series.
Aggregates hold a floor. Team and department views require at least 5 people in the group, as set out in how we operate. Below that, a group average is a personal disclosure wearing an aggregate label, and the correct output is nothing at all.
Levos is a Human Capital Operating System: an intelligence layer above the existing stack. The gap between self-reported and observed data shows up in AI adoption as performative AI use, and HRIS records and AI value covers what HRIS data can and cannot show.
What the principle costs
It costs coverage. Some cells stay empty, and some comparisons a customer would like to run cannot be run yet. A platform willing to return "not enough evidence" demos worse than one that renders every slice on request. It is also the only version of this instrument a serious operator should agree to run against their own people.
Levos has no customer results to cite on this yet, and we would rather say so than manufacture one. What we can point at is the method, which is documented, and the position: the platform's job is to be accurate about what it can see and honest about what it cannot. A number invented to complete a row fails at both.
Change how work gets measured. It starts with refusing to score what you did not observe.
Frequently asked questions
What is missing data in employee scoring?
Any input a scoring model expects for a person but does not have. It arises when the role does not generate the artifact, when the source tool is not connected, when the person's history is short or interrupted, or when the work happens where nothing observes it. Four causes, one blank cell, and none of them is evidence of low performance.
Why is treating a missing signal as a zero a problem?
Because a zero is a claim and a blank is not. Schafer and Graham found in 2002 that mean substitution may predict a missing value accurately while distorting estimated variances and correlations. The resulting score carries the same apparent precision as a fully evidenced one, which is exactly why nothing downstream catches it.
How much does missing data affect the reliability of a score?
Schafer and Graham reported that at roughly 25% missing, a nominal 95% confidence interval delivered approximately 86% true coverage: an error rate nearly three times the stated one. That is a general statistical result rather than a workforce study, so read it as direction and rough magnitude, not as a benchmark for your organization.
Is missing workforce data random?
Usually not. Rubin's 1976 classification separates missing completely at random, missing at random, and missing not at random, and workforce data often falls into the third case, where the reason a value is absent is related to the value itself. Min, Asif, and Vaidya found in 2025, across 150 dataset variants built from three real-world datasets, that correlated missingness worsens disparities even at low rates. That is machine learning fairness research, not an employment study.
What should a people analytics platform do with a missing signal?
Report it as missing. Do not read an absent input as a zero, attach a confidence rating to every signal and every score, hold a minimum group size on aggregates, and state openly what the method cannot tell you yet.
