Predictive HR Analytics

1. 1Frame POS and justice as testable hypotheses, not sentiments: write them in the ARHAT If/Then form, e.g. 'If we increase the proportion of the workforce participating in profit-related pay, then trust in management will rise.'
2. 2Separate the constructs into measurable sub-variables before collecting data: perceived contribution value (POS), perceived fairness of pay distribution (distributive justice), fairness of the process (procedural justice), and supervisor support. Assign each a short rating scale so employees rate them 0-10.
3. 3Start with data you already have (compensation records, engagement/eNPS surveys, trust-in-management items) and evaluate quality before collecting new items; use regression or averages to estimate missing values rather than dropping cases blindly.
4. 4Build a service-climate proxy the way Ployhart et al. did: have employees rate their own unit's climate on a short scale so you can correlate 'good service is valued, facilitated, and rewarded' against the outcome you care about.
5. 5Test each pay lever against well-being outcomes separately (job satisfaction, organizational commitment, trust in management), because incentives that look identical (performance-related vs profit-related vs share ownership) move well-being in opposite directions.
6. 6Add participation breadth as a moderator variable: measure what percentage of the workforce actually takes part in profit-related pay, and split high vs low participation groups before drawing conclusions.
7. 7Correlate the justice/POS measures with a business KPI (customer NPS, attrition) so sponsors see value; recall ISS found employee engagement correlated with customer experience at 0.55.
8. 8Present findings in the three-act structure — setup (why fairness matters to the KPI), confrontation (which lever backfires), resolution (the specific action) — with a call to action for sponsors.

A study in the Human Resource Management Journal tested how three incentive types affect employee well-being measured by job satisfaction, organizational commitment, and trust in management. Performance-related pay was positively associated with all three outcomes — a clean win. Profit-related pay, however, only produced higher satisfaction, commitment, and trust when participation was high AND profits were perceived to be equitably distributed; where participation was low to medium, the same scheme produced lower levels of all three. Share ownership showed a negative relationship with job satisfaction and no relationship with commitment or trust, even at high uptake. The operational read: distributive justice ('equitably distributed') and participation breadth are moderators, not decorations. If you had run a single blanket regression pooling all incentive types, you would have concluded incentives are neutral — the truth only surfaces when you split by pay type and by participation level.

• If profit-related pay reaches only a low-to-medium share of the workforce, then expect it to lower satisfaction, commitment, and trust — widen participation before expanding it.
• If employees perceive profits are not equitably distributed, then treat distributive justice as the binding constraint and fix perceived fairness before adding more incentive money.
• If the lever is share ownership, then do not expect well-being gains — model it as neutral-to-negative and justify it on other grounds.
• If an incentive is 'working' on paper but you see complaints about unfair pay distribution or overwork, then the intervention is backfiring — measure procedural and distributive justice, not just payout size.
• If you lack a fairness metric in existing data, then add a short self-rating climate/justice scale before running the regression, as Ployhart et al. did for service climate.

• Treating all incentive pay as one variable. The tell: your regression shows a flat/near-zero coefficient because performance-related gains and share-ownership losses cancel out.
• Ignoring participation breadth as a moderator. The tell: you roll out profit-related pay company-wide expecting the 'high participation' benefit but only a small group actually enrolls, and trust drops.
• Assuming more incentive money automatically raises well-being. The tell: rising complaints about unfair distribution, overwork, and stress alongside higher payouts.
• Measuring outcomes as a single 'engagement' blob instead of job satisfaction, commitment, and trust in management separately. The tell: you miss that share ownership hits satisfaction but not commitment.
• Chasing fairness with no link to a business KPI, so sponsors withdraw support. The tell: the project isn't tied to a problem the business unit's KPIs care about, and it stalls (Step 1 warns not to investigate what isn't a real problem).

• ▸Incentive type: name it (performance-related / profit-related / share ownership).
• ▸Participation breadth (%): what share of the workforce actually takes part.
• ▸Distributive justice rating (0-10): employee-rated fairness of how rewards are distributed.
• ▸Procedural justice rating (0-10): employee-rated fairness of the decision process.
• ▸Perceived supervisor support / POS rating (0-10): does the org value and care about them.
• ▸Outcome 1 — Job satisfaction score: current measured level.
• ▸Outcome 2 — Organizational commitment score: current measured level.
• ▸Outcome 3 — Trust in management score: current measured level.
• ▸Linked business KPI + hypothesis: state 'If {X} then {Y}' tying fairness to a KPI (e.g. attrition, cNPS 0.55-type link).

1. 1Define the strain construct operationally before measuring it: capture worry, lost sleep, and self-reported stress as the psychological response, then trace them back to the two upstream drivers — high job demands and low control.
2. 2Map role-level demand sources named in the AIHR turnover model: role ambiguity, role overload, and role conflict (when a person holds two contradictory positions). Score each employee or team on these three.
3. 3Cross-reference strain against interacting turnover predictors. Because predictors interact — some effects enhanced, some reduced — segment by demographics (age, tenure, marital/parental status) so you don't misread who is at risk.
4. 4Test whether your incentive design is adding strain rather than performance. Check whether performance-related pay is producing complaints of unfair distribution or overwork, and whether profit-related pay reaches a high proportion of the workforce (low uptake correlates with lower well-being).
5. 5Deploy control-restoring interventions and measure the delta. Flexible work arrangements are the lever the passages support: flextime, remote work, compressed weeks (four 10-hour shifts), job sharing.
6. 6Re-survey on a short cadence to catch movement. Best Buy moved engagement surveys from annual to quarterly once they saw the size of the impact; apply the same frequency logic to strain measures.
7. 7Quantify the downstream cost so strain reduction competes for budget: link it to turnover, quality errors, and shareholder return using the engagement figures as your comparison baseline.

Take the resignation model in the passages: a logistic regression predicts whether staff will resign from compa-ratio, age, and travel time, using Logit(L) = 12.117 – 5.452*Compa-ratio – 0.163*Age + 0.814*Travel time. Travel time carries a positive coefficient of 0.814 — the farther the commute, the higher the resignation probability — which maps directly onto strain from a demanding context the employee cannot control. For a staff member with compa-ratio 0.7, age 30, and travel time coded 1 (far), you compute the logit, take e^L, then convert to P(X). Because commute is a strain driver the passages tie to retention (Dr. John Sullivan, TLNT), the practical response is a control-restoring intervention: offer remote or flexible arrangements, since PGi found 82 percent of telecommuters showed lower stress and 69 percent showed lower absenteeism. You then re-run the model after the intervention to see whether predicted resignation probability drops. This turns an abstract strain reading into a testable, cost-linked action.

• If demands are high AND control is low, treat strain as elevated even when self-reported stress is moderate — the combination is the risk, not either factor alone.
• If the strain source is role conflict or role overload, redesign the role first; adding pay does not resolve contradictory positions and can add overwork stress.
• If profit-related pay participation is low-to-medium, expect lower well-being; only push it as a lever when uptake is high and profits are seen as equitably distributed.
• If an at-risk employee has a long commute or inflexible schedule, offer flexible work before other retention spend — the evidence links it to lower stress, higher morale, and lower absenteeism.
• If the employee is young and low-tenure, weight strain more heavily as a turnover trigger, since younger and fourth-to-fifth-year staff job-hop more.

• Measuring only self-reported stress and ignoring worry and lost sleep — the tell is a survey that shows 'acceptable' stress while sleep complaints and turnover rise.
• Assuming pay fixes strain. The tell is performance-related pay that 'backfires' into complaints of unfair distribution and overwork instead of higher satisfaction.
• Reading predictors in isolation. The tell is a model that flags a demographic group but misses that marriage is a significant predictor only when people are younger — interactions were ignored.
• Rolling out share ownership expecting a morale lift. The tell is flat well-being scores; the passages found share ownership has a negative relationship with job satisfaction and no relationship with commitment or trust even at high uptake.
• Surveying strain annually. The tell is that you discover a spike quarters late — Best Buy moved to quarterly precisely because slow cadence hides fast-moving impact.

• ▸Strain symptoms observed: note worry, lost sleep, reported stress levels per employee/team
• ▸Demand rating (1-5): rate role overload and workload intensity
• ▸Control rating (1-5): rate autonomy over schedule, method, and location
• ▸Role driver present: mark role ambiguity / role overload / role conflict
• ▸Interacting predictors: record age, tenure year, marital/parental status, commute/travel time
• ▸Incentive check: is pay adding strain? (unfair-distribution or overwork complaints Y/N; profit-pay uptake high/med/low)
• ▸Control-restoring intervention selected: flextime / remote / compressed week / job share
• ▸Downstream cost linked: expected effect on turnover, quality errors, or TSR
• ▸Re-measure date: set quarterly review to compare pre/post strain scores

1. 1Frame job satisfaction as measurable so you can correlate it: decide which components you are capturing — emotional (survey sentiment), cognitive (rating scales), and behavioural (turnover, absenteeism, engagement scores) — because satisfaction connects to outcomes you can quantify.
2. 2Gather sentiment inputs alongside numeric ratings: use company-wide and pulse surveys, focus groups, and Glassdoor reviews to capture the emotional and cognitive appraisal, then convert attitude ratings to numbers you can analyze in Excel.
3. 3Pair each satisfaction/engagement variable with a business outcome variable in two aligned columns (e.g., engagement score per store vs store income, or department satisfaction vs turnover rate).
4. 4Run correlation using Excel's correl() function — place both data sets inside the function separated by a comma — to test whether satisfaction moves with the outcome (positive: engagement up, sales up; negative: satisfaction up, turnover down).
5. 5Read the coefficient: it ranges from -1 to +1; treat 0.5 or above in either direction as significant, and the closer to 1 the stronger the relationship.
6. 6When a significant relationship exists, run Multiple Regression to quantify impact — predict the dependent outcome (revenue, turnover) from the satisfaction/engagement independent variable so you get a dollar or percentage figure, not just a direction.
7. 7Contextualize the numbers with qualitative analysis: pull focus-group quotes or survey verbatims to explain WHY the relationship exists, turning a coefficient into a story leaders act on.
8. 8Guard against reading correlation as causation — check for a third variable and gather more evidence before recommending an intervention.

Best Buy analyzed the link between employee engagement (a behavioural correlate of satisfaction) and store financial performance. They found the relationship was strong and directional enough to state it in business terms: a 0.1% increase in employee engagement results in a $100,000 increase in store income. This is a correlation-then-regression pattern — first establishing that engagement and income move together, then quantifying the dollar impact per unit change. The power of framing it this way is that it converts a soft attitude metric into a budget argument a project sponsor understands. To replicate it, you would place per-store engagement scores in one column and per-store income in the next, confirm the correlation, then regress income on engagement to read off the coefficient. The qualitative layer — verbatims from engaged stores — explains why engaged staff drive income, giving the number a narrative.

• If the correlation coefficient is below 0.5 in either direction, then do not present it as a significant relationship — the association is too weak to act on.
• If satisfaction and turnover move in opposite directions (satisfaction up, turnover down), then treat it as a negative correlation and use it to justify retention spend.
• If correlation is significant but you cannot rule out a third variable, then stop at 'implied relationship' and do more analysis before claiming cause and effect.
• If you need a directional 'does it move together' answer, use correlation; if you need to predict the level or dollar/percent impact of the outcome, use regression.
• If leaders won't act on a coefficient alone, then attach a qualitative quote or story so the number connects emotionally.

• Treating a strong correlation as proof of causation — the tell is a recommendation to 'increase satisfaction to raise sales' with no check for a confounding third variable.
• Reporting a coefficient under 0.5 as meaningful — the tell is calling a 0.3 correlation 'a clear relationship.'
• Measuring only the cognitive rating and ignoring emotional and behavioural components — the tell is a satisfaction score that doesn't move with actual turnover or absenteeism.
• Presenting raw numbers with no narrative — the tell is a data slide packed with bullet points that leaves the audience unmoved and the project unfunded.
• Assuming making employees happy automatically means they work hard — the tell is expecting satisfaction to convert to performance without engagement drivers in between.

• ▸Satisfaction variable: name the metric and its component (emotional / cognitive / behavioural).
• ▸Data source: survey, pulse, focus group, or Glassdoor — one per row.
• ▸Unit of analysis: employee, department, or store — keep it consistent across both columns.
• ▸Outcome variable: the business result (turnover %, income, absenteeism days).
• ▸Correl() result: paste both data sets into correl(), record the coefficient (-1 to +1).
• ▸Significance check: mark YES only if coefficient is 0.5 or above in either direction.
• ▸Direction: positive or negative, stated as a plain sentence.
• ▸Quantified impact: regression coefficient expressed as $ or % change per unit of satisfaction.
• ▸Story note: one qualitative quote or verbatim that explains the why.

1. 1Assemble a performance criterion you can measure objectively — appraisal rating, sales figures, customer feedback score, or a productivity rate like supermarket checkout scan speed — and record it per employee alongside candidate/employee predictor variables (traits, skills, experience).
2. 2Pull descriptive data from your HRIS first (demographics, tenure, costs, current performance) to establish 'what happened' before attempting any prediction; know that HRIS alone cannot surface correlations.
3. 3Formulate a hypothesis linking predictors to performance (e.g. specific personality traits predict good customer service or sales) before touching the data, so the analysis is structured rather than a fishing expedition.
4. 4Run a simple Correlation in Excel between each candidate trait and the performance criterion to see which traits move together with high performance (Section 15.5 / 16.4 approach).
5. 5For multi-variable predictions, build a Multiple Regression in Excel to estimate performance from traits plus experience plus training; for a categorical outcome (good vs poor performer) use Logistic Regression.
6. 6Interpret variable interactions — remember predictors enhance or dampen each other — rather than reading each coefficient in isolation.
7. 7Package findings for the audience: know the audience, develop the narrative, then build visuals, and run the data-visualization checklist before presenting to HR partners or business heads.
8. 8Validate the model against a fresh group before acting on hiring or promotion decisions, then feed results back to refine the criterion and predictors.

Rentokil Initial (Case 4, Section 16.4 / 26.1) wanted to hire salespeople who would actually sell more. Instead of relying on gut feel, they identified the personality traits of their existing high-performing salespeople and used those traits as the predictor for future hires. The book's approach mirrors this: take Customer Service and Sales staff, record their x and y personality traits, and run a simple correlation against a performance criterion (sales figures or customer feedback). The correlation reveals which traits move with strong performance, so you can say 'staff with x and y traits are likely to be good performers if hired.' HBR's 'Three Behaviors That Drive Successful Salespeople' (Case 15.4) confirms that specific, observable behaviors — not general impressions — separate top from average performers. The payoff is a predictive performance model built on traits, skills, and experience that raises your odds of hiring and promoting the right people.

• If you only need to know which single trait tracks with high performance, then use Correlation; if you must predict a performance number from several inputs, then use Multiple Regression.
• If the performance outcome is categorical (good performer / poor performer, will succeed / won't), then use Logistic Regression rather than linear regression.
• If an employee performs exceptionally well over a long period, then flag them as a flight risk — sustained top performers leave because they lack challenge.
• If two predictors interact (e.g. marriage and age), then interpret them jointly, because one effect is enhanced while another is reduced.
• If you cannot state a hypothesis before analyzing, then stop and formulate one — unstructured problems without a framework are why most analytics efforts fail.

• Relying on gut feel about who is a strong performer instead of a measured criterion — the tell is that no one can point to the sales figures, appraisal rating, or scan rate behind the judgment.
• Treating HRIS dashboards as prediction — the tell is you can describe what happened and what's happening now but cannot surface any correlation or forecast.
• Reading regression coefficients in isolation — the tell is a confident single-variable claim that ignores how predictors amplify or dampen each other.
• Confusing high performance with retention security — the tell is assuming your best performers are safe when sustained top performers are actually likely to leave for lack of challenge.
• Skipping audience and narrative work and dumping charts on stakeholders — the tell is a deck full of visuals with no story, presented without running the data-visualization checklist.

• ▸Performance criterion: name the measurable outcome (appraisal rating, sales $, customer feedback, scan rate)
• ▸Population: which role/team the model covers (e.g. Customer Service, Sales)
• ▸Predictor 1: candidate trait/skill/experience to test
• ▸Predictor 2: second predictor (add interaction note if relevant)
• ▸Data source: where each value comes from (HRIS, sales system, survey)
• ▸Method: Correlation / Multiple Regression / Logistic Regression
• ▸Hypothesis: the trait-to-performance link you expect before running it
• ▸Result: correlation strength or coefficient, plus plain-language reading
• ▸Action: hiring/promotion decision this supports and validation date

1. 1Frame turnover as a business problem before analyzing: ask how it prevents the business unit from achieving its KPIs, and confirm with the sponsor that attrition is real and worth resources — don't investigate attrition if it isn't hurting the business.
2. 2Measure at two levels. Individual level: for each person, record leaver vs. stayer and length of service before leaving. Team level: compute a separation rate percentage (leavers / headcount) so you can compare units and quantify replacement cost.
3. 3Build a simple attrition model first rather than waiting for perfect data. Following Nielsen, start with ~20 points of employee data (gender, tenure, age, manager rating) and refine later (e.g., add CSR-program time and participation).
4. 4Choose the technique to the data: use a Decision Tree to segment resigners, Correlation to test which variables move with flight risk, and Logistic Regression to predict the categorical outcome (resigns vs. stays) for each employee.
5. 5Score flight risk per employee, then rank so managers can act on the highest-risk names — the goal is prevention, not a report.
6. 6Put programs in place to fix the identified drivers (e.g., a pay-raise conversation, incentive programmes, culture fixes), then measure the impact and share the findings with the sponsor.
7. 7Track leading indicators alongside the model: Climate Culture Rating (percent giving top ratings to 'good place to work'), Depletion Ratio (annual percent of top talent lost), and Accession Ratio (new+replacement hires / total employment) to catch rising cost and falling engagement early.
8. 8Present findings as a story: cast high turnover as the villain and the retention program as the hero, lead with people not numbers ('one-third of our new hires in Sales left within a year' rather than 'attrition is high').

At Nielsen Holdings, a senior business leader asked People Analytics VP Piyush Mathur a direct question: do you know why people are leaving my team? Mathur and technologist Doug Shagam refused to chase perfect data and instead built a basic attrition model on ~20 points of employee data — gender, tenure, age, and manager rating — to identify what was driving rising company-wide attrition. They deliberately anchored the work to a critical business question ('if you're not answering a really critical business question, it's not worth doing'). Within months they identified the primary drivers of voluntary attrition, put programs in place, and measured the impact. The result: voluntary attrition cut by nearly fifty percent, saving Nielsen millions of dollars. Over time they refined the model with richer variables such as time and participation in CSR programs.

• If turnover doesn't demonstrably block a business unit's KPIs, then don't start the project — get sponsor agreement on scope, timeframe, and budget first.
• If your outcome is categorical (resigns vs. stays), then use Logistic Regression; if you want to segment leavers into groups, use a Decision Tree; if you're screening which drivers matter, use Correlation.
• If a flight-risk signal appears (e.g., an employee's likely to seek a raise), then intervene early — pre-warn the manager so retention action beats the resignation, as Namely's case shows.
• If Climate Culture Rating is falling or Depletion Ratio is rising, then expect higher turnover and lower profitability, and prioritize the culture/top-talent drivers.
• If you're tempted to wait for perfect data, then ship a simple 20-variable model now and refine it iteratively — perfect data delays impact.

• Investigating attrition that isn't a real business problem. Tell: no sponsor is invested, no KPI is named, and the project stalls for lack of support.
• Assuming only voluntary turnover matters. Tell: research shows involuntary (dismissing poor performers) has about the same negative association with profits (-.03 overall) as voluntary — treating dismissals as 'free' misreads the data.
• Overstating turnover's direct profit impact. Tell: the direct correlation with profits is only about -.03 (explaining ~0.1% of profit variation); if you claim turnover alone tanks profits you'll lose credibility. Much of the effect runs indirectly through quality/safety.
• Perfectionism on data collection. Tell: months pass with no model; Nielsen's win came from acting on 20 rough variables, not waiting.
• Leading a presentation with numbers instead of people. Tell: you say 'attrition rate is high' or 'engagement is seventy percent' and the audience disengages, instead of 'one-third of our new hires left.'

• ▸Business question — state how turnover blocks a specific unit KPI, confirmed by sponsor
• ▸Measurement level — individual (leaver/stayer, length of service) and team separation rate %
• ▸Model variables — list ~20 data points (gender, tenure, age, manager rating, CSR participation...)
• ▸Technique chosen — Decision Tree / Correlation / Logistic Regression, with reason
• ▸Flight-risk score — per-employee probability of resigning, ranked highest first
• ▸Identified drivers — top voluntary-attrition causes surfaced by the model
• ▸Intervention & owner — retention program (raise, incentive, culture fix) and who acts
• ▸Leading indicators — Climate Culture Rating %, Depletion Ratio, Accession Ratio
• ▸Impact & story — attrition reduction achieved, dollars saved, one people-first headline

1. 1Define your customer-facing outcome variables concretely: customer satisfaction score, expressed loyalty, intention to reinvest, and amount of reinvestment — pick the ones your business actually captures.
2. 2Adopt Net Promoter Score as the loyalty metric: survey with the single question 'How likely is it that you would recommend a company, product or service to a friend or colleague?' scored 0-10.
3. 3Classify responses: Promoters score 9-10, Passives score 7-8, Detractors score 0-6. Calculate NPS = %Promoters − %Detractors. Track both customer NPS (cNPS) and employee NPS (eNPS).
4. 4Pair each customer outcome with candidate drivers you can influence — motivation/engagement of service staff, amount of training and quality of service staff, and the ability to act on customer expectations.
5. 5Run correlation in Excel between eNPS and cNPS to test the employee-engagement-to-customer-loyalty link; a strong result (ISS found 0.55) confirms happy service employees lead to happy customers.
6. 6Use Multiple Regression in Excel to quantify the payoff — e.g. estimate how a 1% increase in employee engagement moves customer satisfaction and revenue, converting the correlation into a dollar/percentage forecast.
7. 7Move up the analytics levels: describe current NPS (descriptive), find why detractors score low by asking them their reason (diagnostic), forecast reinvestment (predictive), then prescribe which customer segment to target to improve profitability (prescriptive).
8. 8Institutionalize monitoring of leading indicators: service-employee eNPS, employee churn, absenteeism, and training hours — the metrics ISS recommended tracking.

ISS studied the link between service-employee engagement and customer experience and found Employee Engagement correlates strongly with Customer Experience at 0.55. They measured both sides with NPS: eNPS captured employee loyalty and cNPS captured customer loyalty, each built from the 0-10 'how likely to recommend' question, with Promoters at 9-10, Passives at 7-8, and Detractors at 0-6. Plotting cNPS against eNPS showed that happier service employees produced happier, more loyal customers. ISS traced the drivers to motivation and engagement of service staff, the amount and quality of training, and staff's ability to act on customer expectations. Their prescriptive recommendation was to monitor eNPS, employee churn, absenteeism, and training hours as the levers that move customer loyalty. In the same book's framing, a 1% increase in employee engagement is estimated to yield a 2% increase in customer satisfaction alongside a $100k revenue increase.

• If your loyalty measure is a single recommend-question, then compute NPS as %Promoters minus %Detractors rather than a raw average — the score only means something in that difference.
• If eNPS and cNPS correlate strongly (around 0.55 or higher), then invest in service-staff engagement, training, and responsiveness because those are the drivers behind customer experience.
• If you have promoters, then ask them what makes them willing to recommend; if you have detractors, then ask them the reason for their low score — both are needed for actionable insight.
• If you can only describe current satisfaction (descriptive), then you cannot yet prescribe which customer segment to target — run correlation and regression first to reach predictive/prescriptive levels.
• If a driver is not something you can act on (motivation, training, responsiveness), then do not treat it as a lever for improving reinvestment.

• Averaging 0-10 NPS scores instead of subtracting %detractors from %promoters — the tell is a 'satisfaction average' near 7 that hides a large detractor block.
• Treating the 0.55 correlation as proof that engagement causes loyalty and skipping the diagnostic step — the tell is no follow-up asking detractors why they scored low.
• Measuring cNPS but never tracking the upstream eNPS, churn, absenteeism, and training hours — the tell is you can report customer scores but cannot explain or move them.
• Stopping at descriptive dashboards of satisfaction and calling it predictive — the tell is you report 'what happened' but never forecast reinvestment amount or run a regression.
• Inventing dollar impacts beyond what your own data supports — the tell is citing '$100k revenue' or '2% satisfaction' figures as your own results rather than validating them with your Excel regression.

• ▸Recommend question (0-10): confirm exact wording used in your survey
• ▸%Promoters (score 9-10): enter percentage of respondents
• ▸%Detractors (score 0-6): enter percentage of respondents
• ▸cNPS = %Promoters − %Detractors: compute customer loyalty score
• ▸eNPS (same method, employees): compute employee loyalty score
• ▸eNPS–cNPS correlation (Excel): enter coefficient and note if strong (≥0.55)
• ▸Driver levers to act on: rate staff motivation, training amount/quality, responsiveness to expectations
• ▸Reinvestment forecast (regression): enter predicted intention and amount of reinvestment
• ▸Detractor reasons captured: log top reasons cited for low scores

1. 1Define the villain before designing the intervention: name the specific problem (high staff turnover, absenteeism, low team engagement) with a data point, e.g. 'one-third of Sales new hires leave within 6 months.' Do not launch a programme without a quantified problem it targets.
2. 2Locate where the problem concentrates using individual and historical patterns, not averages — segment by tenure, first-year status, ethnicity, gender, manager rating, or manager tenure to find the attributes with the highest probability of the bad outcome.
3. 3Match the intervention type to the driver you found: if lateral moves retain people, build a rotation programme; if long manager tenure depresses engagement, revise the training programme for long-tenure managers.
4. 4Score effort vs impact on the Action Priority Matrix before committing — assess your team's analytics capability, data access/quality, and the level of change (people affected). Start with Quick Wins to build credibility.
5. 5Engage the people affected in deciding how the change happens, and bring in unions and data owners early; the best way to reduce resistance is participation, not announcement.
6. 6Track the target metric before and after the intervention so you can demonstrate measurable business results (attrition cut, engagement lifted).
7. 7Tell the story to executives to get them on board: lead with people, then data — give leaders simple, accessible views of the numbers so the message lands.

Nielsen's attrition model revealed that an employee's first year matters most — leaving risk is high before the first performance review — while ethnicity and gender had no impact on tenure, and promotions and lateral moves retained people. Because insights are meaningless without action, Mathur's team built specific interventions rather than a generic retention push. After identifying employees with the highest flight-risk attributes, leaders set up direct chats with them, and 40 percent of those flight-risk employees were transferred into new roles. They created a 'Golden Year' programme to track and support the first year, since attrition peaked there. They also launched a 'Ready to Rotate' group to empower employees to find internal roles, because the data showed lateral moves raised the probability of staying by 48 percent. The programmes were driven by the discovered driver, targeted at the segment where the problem concentrated, and tied to measurable retention outcomes.

• If the attrition/problem risk concentrates in a specific segment (e.g. first-year employees), then build a programme targeting that segment rather than an organization-wide one.
• If your analysis shows a retention lever like lateral moves increases stay-probability, then design the intervention around that lever (a rotation programme), not around assumed causes.
• If the change affects many people and requires high effort, then treat it as a Major Project — invest in stakeholder relationships, communication and training — but if it affects too few, expect low impact.
• If people will resist the change, then get those affected to participate in deciding how the change is made.
• If the company is unionized or in a country with restrictive privacy laws, then delay or reframe the programme and engage unions early with benefit stories before launching.

• Designing the intervention before quantifying the villain — the tell: you can name the programme (training, incentives) but not the specific problem metric it moves.
• Relying on averages, medians and turnover rate instead of individual and historical patterns — the tell: you can't say which employees or which period carry the risk, so the programme hits the wrong people.
• Producing insights without action — the tell: an attrition model exists but no chats, tracking programme, or rotation group followed from it.
• Announcing the change to affected staff instead of involving them — the tell: rising resistance and knowledge-hiding by colleagues who withhold or delay data.
• Skipping the executive story and presenting raw stats — the tell: 'average engagement is 70%' framing instead of 'two out of three people here are engaged,' and leaders don't act.

• ▸Villain (problem): state the issue with one data point (e.g. '1/3 of Sales hires leave in 6 months').
• ▸Affected segment: which attributes/tenure/manager group carry the highest risk.
• ▸Driver found: the specific factor analysis links to the outcome (e.g. lateral moves +48% stay).
• ▸Intervention (hero): the programme designed around that driver.
• ▸Impact/Effort quadrant: Quick Win / Major / Fill In / Thankless — based on capability, data access, level of change.
• ▸Stakeholders to engage early: data owners, domain experts, unions, affected employees.
• ▸Target metric + baseline: the number you will move and its current value.
• ▸Executive story line: people-first framing of the result to secure buy-in.

1. 1Survey your current high performers first to identify which traits actually predict success in the specific role — Rentokil grouped these into interests, conscientiousness, interpersonal skills, and cognitive ability rather than assuming what mattered.
2. 2Validate assumed hiring criteria against data before trusting them; Deloitte's financial services case found top-school pedigree, best grades, and references did NOT predict high sales performance, so audit whether your filters are noise.
3. 3Select an assessment instrument (personality/aptitude test) that empirically correlates with performance in your context, then use it to score candidates against the validated traits — prioritize conscientiousness, the most sought-after Big Five attribute for job performance.
4. 4Compute Quality-of-hire and Success Ratio per source: Success Ratio = number of hires rated at least 'meets-expectations' / total number hired; run it by source to see which schools/companies produce your best hires and cut low-yield sources.
5. 5Track first-year attrition as an onboarding signal — early leavers indicate cultural misfit, job-description mismatch, or overselling by the recruiter; tie each early exit back to a selection or onboarding cause.
6. 6Measure candidate experience with a survey scored by NPS (Promoters 9-10, Passives 7-8, Detractors 0-6) and ask each band a targeted question, since ~60% of candidates report poor experience and 72% of those share it publicly.
7. 7Design onboarding to grow the new hire's internal network — have them meet and mingle with diverse colleagues, assign consistent support staff, and build exposure to senior leadership, which correlates with successful sales results.

Rentokil Initial, the world's leading commercial pest control and hygiene provider, hired external consultant Max Blumberg to find what predicts sales success. His team surveyed the existing sales force to identify traits correlating with performance and grouped them into interests, conscientiousness, interpersonal skills, and cognitive ability. They then searched for a selection test that assessed these traits and found a personality assessment that correlated strongly with sales performance. They used it to select new candidates against those validated traits. As a result, sales rose more than 40% globally with over 300% return on investment. Max estimated the UK business alone could gain £1.5 million a year. The lesson: validate traits from your own top performers, then buy the test that measures them — don't guess.

• If your assumed criteria (top schools, high GPA, references) haven't been validated against actual performance data, then don't trust them — Deloitte found none predicted high performers.
• If first-year attrition is high for a role, then investigate cultural misfit, job-description mismatch, or recruiter overselling before blaming the candidate pool.
• If a source's Success Ratio is low, then review your selection process for that source or drop it; if high, then concentrate sourcing there.
• If constantly dismissing poor hires, then treat it as a selection failure — involuntary turnover drags profit about as much as voluntary (organizational turnover correlates -.03 with profits).
• If choosing between candidates on personality, then weight conscientiousness first (dependable, persevering, orderly) and agreeableness second, per Sackett and Walmsley.

• Hiring on prestige signals (top universities, GPA, references) — the tell is wide performance variation among hires who all had strong resumes, exactly Deloitte's financial services case.
• Adopting a personality/aptitude test without validating it against your own performers — the tell is a test bought off-the-shelf with no local correlation study behind it.
• Treating dismissals of bad hires as 'clean-up' rather than cost — the tell is repeated selection do-overs while performance stays flat; the research shows involuntary turnover hurts profit about as much as voluntary.
• Ignoring candidate experience — the tell is Glassdoor complaints and word-of-mouth damage; 60% of candidates report poor experience and 72% share it.
• Onboarding new hires in isolation with rotating support staff — the tell is slow ramp and low internal network; new hires need diverse colleague contact and consistent support to be effective.

• ▸Role: the specific job you're hiring for
• ▸Predictive traits (from top-performer survey): list traits grouped as interests/conscientiousness/interpersonal/cognitive
• ▸Assessment instrument used: name the validated test and its correlation with performance
• ▸Candidate score per trait: rate the applicant against each validated trait
• ▸Source of candidate: school/company/referral for Success Ratio tracking
• ▸Success Ratio for this source: hires meeting expectations / total hired from source
• ▸First-year attrition flag: note any misfit/mismatch/overselling risk observed
• ▸Onboarding plan: buddy assigned, support-staff continuity, senior-leadership exposure
• ▸Candidate experience NPS band: Promoter/Passive/Detractor plus the targeted question asked

1. 1Pull the structural fields you can actually source from HRIS: function/department, team size, gender mix, geographic location, plus role-level job demands and job control/autonomy. HRIS reports what happened and currently happens, so start descriptive.
2. 2For each characteristic, define a measurable variable — e.g. team size as headcount per manager, gender mix converted into a diversity index, location as region code, autonomy as a survey scale.
3. 3Merge these characteristics with an outcome you care about (engagement, sales, resignation) into one dataset, following the iNostix pattern of merging engagement survey outcomes with business data.
4. 4Run correlation first to test whether each characteristic relates to the outcome, then multiple regression to see which characteristics have the most impact when held together.
5. 5Segment by characteristic — compare high performers against average within a function, as Deloitte did — rather than treating the workforce as one block.
6. 6Contextualize the statistics with qualitative data (focus-group quotes, manager interviews) to explain WHY a characteristic drives the outcome, moving from descriptive to diagnostic.
7. 7Convert the surviving predictors into a prescriptive action (change team size, adjust manager tenure, target a location) and forecast the outcome before rolling it out.
8. 8Tell the story to stakeholders using data, visuals and narrative, tailoring depth to the audience (concise for Top Management).

PostNL analyzed a team characteristic — manager tenure — against team engagement (Case 5). This treats a contextual feature of the team (how long the manager has led it) as an independent variable and engagement as the dependent variable. The correct move is correlation: does an increase in manager tenure correspond to a change in team engagement? If tenure and engagement move together positively, that is a positive correlation the analytics team can then diagnose (why does a longer-tenured manager lift engagement?) and act on — for instance, reducing manager churn in low-engagement teams. Because ISS studies link engagement to profitability and customer experience, a tenure-to-engagement finding chains through to a business outcome. The lesson: a single structural variable, measured cleanly and merged with an outcome, produces an actionable prediction.

• If you only have HRIS demographics and counts, you are at descriptive level — do not claim prediction until you have run correlation or regression.
• If two characteristics both correlate with the outcome, run multiple regression to see which one actually has the most impact when the other is controlled.
• If a characteristic correlates but you cannot explain why, add qualitative data before recommending action — correlation alone is not a cause.
• If the audience is Top Management, cut to the resolution and the recommended action; keep statistics minimal because they don't have patience.
• If a long-held hiring belief (top schools, best grades) shows zero correlation with performance, treat it as a myth to debunk, as Deloitte did, not a rule to keep.

• Stopping at HRIS dashboards and calling it analytics — the tell: your output only answers 'what happened,' never 'why' or 'what's likely.'
• Confusing correlation with regression — the tell: you report that a characteristic 'relates' to sales but can't say which characteristic matters most when several are present.
• Keeping traditional selection filters (top universities, high GPA) that data shows don't predict performance — the tell: your high performers look nothing like your hiring criteria, as at the Deloitte financial services firm.
• Reporting numbers with no narrative — the tell: stakeholders forget the finding, mirroring the Stanford result that 5% remember a statistic but 63% remember a story.
• Ignoring color/contrast and audience type in the final presentation so a sound analysis fails at the 'tell the story' stage.

• ▸Characteristic: name the structural/contextual feature (e.g. team size, manager tenure, location, autonomy).
• ▸Measurement: how you quantify it (headcount, diversity index, tenure in months, survey scale).
• ▸Outcome variable: the state/result you're testing against (engagement, sales, resignation).
• ▸Data source: where each variable comes from (HRIS, engagement survey, business data).
• ▸Correlation direction: positive / negative / zero, from your correlation test.
• ▸Regression impact rank: does this characteristic have the most impact when others are held constant?
• ▸Diagnostic note: the qualitative reason this characteristic moves the outcome.
• ▸Prescribed action: the specific change to make and its forecasted outcome.
• ▸Audience/story angle: hero, villain and one-line message for the stakeholder telling.

1. 1Pull demographic fields from HRIS as descriptive analytics first: age, gender, marital status, tenure, ethnicity, education. HRIS reports what happened and what is currently happening about employee demographics — start there before modeling.
2. 2Decide the role each demographic variable plays: predictor (e.g. age and tenure predicting turnover) or outcome (e.g. gender/ethnicity representation as the fairness measure). Label each variable explicitly so you don't conflate the two in one model.
3. 3For diversity-as-outcome work, convert your company's diversity into an index so it becomes a single measurable, model-ready number rather than raw counts (per the 'How to Convert Diversity Into An Index' method).
4. 4Code categorical demographics for regression: gender as a 2-level categorical variable, ethnicity (underrepresented status) as a categorical with 2-3 levels, then run Multiple Regression to test its impact on an outcome like sales, revenue or EBIT.
5. 5When demographics feed a turnover model, combine them with non-demographic predictors — commute time, performance rating, salary market-ratio, personality, behavioral signs, triggering events — because demographics alone rarely explain flight risk.
6. 6Guard access before deploying: restrict flight-risk / demographic scores to senior management trained in interpretation, and confirm employees are aware their data is collected and how it is used.
7. 7For representation goals, benchmark current composition against the proportion of women and minorities in the area's labor force, then set measurable annual improvement targets (the Johnson v. Transportation Agency logic).

Section 21.3 runs a Multiple Regression predicting ethnicity diversity's impact on sales. You take a company's ethnic composition and first convert it into a diversity index so it is a single continuous predictor. That index becomes the X variable and sales the Y variable in an Excel multiple regression. The output lets you state an operational finding — for example, that increasing ethnic diversity is associated with higher sales, and Case 16.3 extends this to market share, revenue and EBIT. Here demographics function as the predictor and business performance as the outcome. The reverse framing appears in turnover modeling, where demographics (age, gender, marital status, tenure) become predictors of the flight-risk outcome — e.g. 'an employee aged 30, who stays more than xx km from the company, rated average for performance, has a 90% probability of resigning in her 3rd year.'

• If the demographic is the thing you want to change or protect (representation, fairness), then treat it as the OUTCOME and benchmark it against the local labor-force proportion; if it explains another result, treat it as a PREDICTOR.
• If a demographic variable is categorical (gender, ethnicity), then dummy-code it for regression (2 levels = 1 categorical variable, 3 levels = one variable with 3 categories) rather than forcing it into a continuous slot.
• If you are analyzing diversity's business impact, then convert composition into a single index before modeling so results are interpretable as one coefficient.
• If demographic data feeds a scoring model (flight risk), then restrict access to trained senior management and confirm consent — otherwise you inherit the HP-style privacy problem.
• If considering sex or minority status in a promotion decision, then use it as one factor among qualified candidates in a flexible case-by-case approach, not as a rigid quota (Johnson v. Transportation Agency standard).

• Using only demographics to predict turnover and ignoring commute time, performance, pay ratio, behavioral signs and triggering events — the tell is a model that flags whole age or gender groups but misses individuals whose life situation just changed.
• Collecting and scoring employee demographic data without their awareness or agreement, then distributing scores widely — the tell is managers acting on a flight-risk index that employees never consented to (the HP privacy problem).
• Reporting demographic composition from HRIS and stopping at 'what happened' while calling it insight — the tell is dashboards of headcount by gender/ethnicity with no diagnostic or predictive follow-through, since HRIS cannot find patterns or correlations.
• Treating diversity as raw counts in a regression instead of converting it to an index — the tell is uninterpretable multi-variable output you can't summarize as a single impact statement.
• Applying a rigid demographic quota to hiring/promotion instead of a moderate, case-by-case factor — the tell is passing over a well-qualified candidate purely on group membership, which invites the reverse-discrimination lawsuit seen in Johnson.

• ▸Variable name: list one demographic field (age / gender / ethnicity / tenure / education)
• ▸Role: mark PREDICTOR or OUTCOME for this analysis
• ▸Type: continuous or categorical (state number of levels if categorical)
• ▸Coding plan: how it enters the model (raw value, dummy code, or converted index)
• ▸Paired outcome/predictor: the Y (or X) it is being related to (turnover, sales, EBIT, representation)
• ▸Benchmark: local labor-force proportion or target used if this is a representation outcome
• ▸Consent & access: is employee aware, and who is authorized to see the resulting score
• ▸Fairness check: is this variable used as a flexible factor or a rigid quota

1. 1Define engagement as a latent construct, not a single number: assemble multiple attitude-survey items covering effort intensity, focus, absorption, pride and belonging so no one item carries the whole measurement.
2. 2Add an eNPS item — 'Using a scale of 0 to 10, how likely will you recommend this company as a place to work?' — and classify respondents into Promoters (9-10), Passives and Detractors following Reichheld's grouping.
3. 3Merge engagement survey outcomes with hard business data (store income, quality parts-per-million, customer experience scores, TSR) at the same unit level so you can correlate attitude against outcome — this is the iNostix approach in the source.
4. 4Run a Correlation in Excel between the engagement score and the target business metric to size the relationship; the ISS study reported an engagement-to-customer-experience correlation of 0.55, so treat 0.5-ish as a strong, actionable signal.
5. 5Run a Multiple Regression to isolate engagement's contribution while holding other drivers (tenure, manager tenure, demographics) constant, since job satisfaction shifts with age and with time-in-job.
6. 6Convert the coefficient into money the way Best Buy does: quantify the income change per 0.1% engagement move so leadership sees a dollar figure, not a survey rating.
7. 7Segment before you act: split into high-engagement vs low-engagement groups (as DDI did) to expose gaps large enough to justify intervention.
8. 8Tell the story last — know the audience, build the narrative, then build the visuals, and run them through the storytelling checklist before presenting to HR partners.

Best Buy quantified engagement as a store-level metric and merged it with each store's annual income. Their model predicts that a 0.1% increase in employee engagement results in an increase of over $100,000 in that store's annual income. To reproduce this you would place store-level engagement scores in one column and annual store income in the next, run Excel's Correlation to confirm the relationship is strong, then run a Multiple Regression with income as Y and engagement as X. The regression coefficient tells you the income change per unit of engagement, which you scale down to a 0.1% move. Because the output is a dollar figure per store, leadership can compare the cost of an engagement program against the projected income lift across all stores. This converts a soft survey number into a capital-allocation decision.

• If a single item is being reported as 'engagement', then reject it and rebuild the measure from multiple items across effort, absorption, pride and belonging — the construct is latent.
• If the engagement-to-outcome correlation is around 0.5 or higher (e.g. ISS's 0.55 with customer experience), then treat the link as strong enough to model and act on.
• If you need to prove business value to executives, then convert the coefficient into money per unit of engagement (Best Buy's $100,000 per 0.1%) rather than reporting the correlation alone.
• If other drivers like age, job tenure or manager tenure could confound the result, then use Multiple Regression to hold them constant instead of a bare correlation.
• If comparing groups, then contrast high vs low engagement segments — DDI found quality errors of 5,658 ppm for low-engagement vs 52 ppm for high-engagement — because the extremes reveal the payoff.

• Reporting engagement as one average score and losing the multi-item structure; the tell is a dashboard with a single percentage and no item-level breakdown.
• Stopping at correlation and claiming causation; the tell is a slide that says engagement 'drives' income with only an r-value and no regression controlling for confounders.
• Ignoring demographic confounders — satisfaction rises with age but falls the longer someone stays in a job; the tell is a model that never includes tenure or age yet claims a clean engagement effect.
• Leaving the finding in analyst language; the tell is presenting a coefficient with no dollar translation, so executives can't decide whether to fund the intervention.
• Skipping the storytelling checklist and audience framing; the tell is a data-dense deck built for the analyst, not for the HR partner receiving it.

• ▸Survey items used: list the multi-item attitude questions covering effort, absorption, pride, belonging
• ▸eNPS item + grouping: record the 0-10 recommend question and Promoter/Passive/Detractor split
• ▸Unit of analysis: state the level (store, team, individual) at which engagement and business data are merged
• ▸Business outcome variable (Y): name the metric — income, customer experience, quality ppm, TSR
• ▸Correlation coefficient: enter the Excel correlation value and flag if ~0.5 or higher
• ▸Regression coefficient + controls: record the coefficient and the confounders held constant (age, tenure, manager tenure)
• ▸Dollar translation: convert coefficient into money per engagement unit (model on Best Buy $100k/0.1%)
• ▸High vs low segment gap: enter outcome for high-engagement vs low-engagement groups
• ▸Audience + narrative: name the audience and the one-sentence story before building visuals