Research
Reviewing Attrition Signals Without Predicting Individuals
A method for reviewing aggregate leaving patterns and operating signals in Philippines support lanes without turning small samples into individual predictions.
On this page
Key stats
The method defines eligible leavers, voluntary and involuntary categories, tenure bands, signal windows, and aggregate review cells. It produces no individual risk score, claims no sample, and reports no result.
Methodology
This method note fixes one operating population and review period. The unit is one leaver record linked to a declared exit category, tenure band, role family, schedule, and a defined lookback window of operating signals. Cells below a minimum size are suppressed. Signal review is separated from exit interviews and from employment decisions, and an independent reviewer checks the classification and aggregation.
Key takeaways
Attrition analysis often drifts into prediction: which person is likely to leave, which team is next. That drift creates privacy, fairness, and practical problems. A safer method reviews aggregate patterns in the operating system. It asks whether a cluster of avoidable conditions, such as access delays, unresolved escalations, or schedule churn, appears before a group of exits, and whether those conditions can be repaired.
Define the population as people who left one declared operating population during the review period, plus the comparable population that stayed. The unit is one leaver record linked to the declared exit category, tenure band, role family, schedule, and lookback window. A record without a reliable exit category stays in the population as unclassified. Dropping it would make the remaining pattern look cleaner than the data supports.
Separate voluntary, involuntary, end of contract, transfer, and unresolved categories. These are different events and should not be pooled into a single attrition rate. A contract ending on schedule is not the same as a resignation, and an unresolved category is not a signal to be interpreted. Publish the counts by category with the denominator for each.
Define the observation period and the lookback window before collecting records. A signal window of thirty, sixty, or ninety days is a design choice, not a finding. State which window the review uses and why. If the window changes between cycles, show both readings or label the break so a trend is not created by the measurement rule.
Choose operating signals that are recorded in the normal course of work: access grant and removal dates, schedule changes, unresolved escalation age, first pass acceptance, correction clusters, skipped one to ones, and open tickets at exit. A survey question can add context, but it should not be the only signal. The strongest review uses records that exist whether or not someone is thinking about attrition.
Aggregate before interpreting. A single exit with a slow access grant is an anecdote. A group of exits in one role family with repeated access delays is a pattern worth examining. Set a minimum cell size before the review and suppress smaller cells. Suppression protects privacy and reduces the chance that a manager acts on noise.
Do not build an individual risk score. A score invites action against a person based on a statistical pattern that may reflect the team, the schedule, or the data. If a manager wants to act, the action should be a review of the operating condition: a delayed access queue, an unclear escalation path, a schedule that changes too often, or a review process that creates rework.
Compare leavers with the eligible staying population on the same signals and the same window. A high access delay rate among leavers means little if the rate is equally high among people who stayed. The comparison is the first defense against a dramatic but uninformative finding. Report both groups and the difference, with the cell sizes.
Check the timing of the signal relative to the exit. A signal that appears after the resignation is not evidence of a cause. Record the date of each signal and the date the exit was known, then classify the signal as before or after. This sounds basic, but it removes a large share of misleading narratives.
Exit interviews and conversations can add context, but they are not a representative sample. Participation is uneven, memory is imperfect, and people may not disclose the real reason. Use them as qualitative material with the number of participants and the method stated. Do not convert a handful of comments into a percentage of the population.
Stratify by role family, tenure band, schedule, shift, manager group, and location only where the cell size allows. A pattern that appears only after slicing into six dimensions is likely noise. Start with one or two strata that the operating plan can act on, then expand only when the counts support it. Report the strata that were not analyzed because the cells were too small.
Separate the review from employment decisions. The output is a description of operating conditions and a small set of repairs. It is not a list of people to manage out or a justification for a policy change. Any decision about an individual or a formal policy belongs to the authorized owner, with advice where the law or a contract applies.
Protect the dataset with stable keys, restricted access, and a retention date. Exit records can contain sensitive personal information. Keep only the fields needed for the review and store the key mapping separately. Do not circulate a spreadsheet with names to a broad management group. Aggregate tables should be the default output.
Independent review should test the classification, the aggregation, and the suppression rule. Give a second reviewer the codebook and a protected extract without the first pass labels. Ask them to recompute the counts and identify any cell that should have been suppressed. Differences in exit category are especially important because they can change the denominator for every rate.
Test sensitivity to the exit categories, the signal window, the minimum cell size, and the treatment of unresolved records. If a pattern disappears when unresolved records are included, it is not stable. Present the range and state what decision would change. A finding that only survives one convenient definition is not ready for action.
The limitations are substantial. Attrition is multi causal, the population is small in many lanes, and the records reflect management practice as much as employee experience. A signal may be a consequence of an earlier event rather than a cause. The method describes associations in one population and period. It does not predict an individual or establish causation.
The output should be a small number of operating repairs with owners and review dates. Examples include reducing access wait, clarifying an escalation path, stabilizing a schedule, or fixing a recurring rework cause. Measure the repair with the same signal and window used in the review. If the signal does not move, do not assume the repair failed until the implementation and the data quality are checked.
A useful review also records what was not measured. If a signal was unavailable, or a cell was too small to report, say so in the method note. The absence of a finding is not evidence that the condition does not exist. It may mean the records are too thin to support a conclusion, which is itself a reason to improve the data before acting.
Population and unit
People who left one operating population during the review period, plus the comparable staying population. Unit: one leaver record with exit category, tenure band, and lookback window.
Exit categories
voluntary, involuntary, end of contract, transfer, and unresolved, reported separately
Signals
access dates, schedule changes, escalation age, first pass acceptance, correction clusters, one to one attendance, and open items at exit
Privacy rule
minimum cell size set before review, small cells suppressed, stable keys used, and individual risk scores prohibited
Principal limitation
Small populations, uneven exit interviews, and management practice limit inference. The method describes associations, not causes or individual predictions.
Next step
Use this method to define one bounded work lane while accountable owners retain policy, approval, access, and customer decisions.
FAQs
Does this method identify people likely to leave?
No. It reviews aggregate operating signals and explicitly prohibits individual risk scores.
Why suppress small cells?
To protect privacy and avoid acting on patterns that a handful of records can create by chance.
Can exit interview comments be turned into percentages?
Only with the participant count and method stated. They are qualitative material, not a representative sample.
Who acts on the findings?
The owner of the operating condition, with employment and policy decisions kept with the authorized owner and adviser.
Sources
- https://www.ilo.org/topics-and-sectors/decent-work-and-2030-agenda-sustainable-development
- https://www.worldbank.org/en/country/philippines/overview
- https://www.nist.gov/privacy-framework
- https://www.eeoc.gov/employers/recordkeeping-requirements
- https://www.archives.gov/records-mgmt/policy
- https://www.nist.gov/cyberframework
- https://www.sba.gov/business-guide/manage-your-business
- https://www.bir.gov.ph/