People Analytics & Text Mining with R

1. 1Define the outcome before the leader variable. Pick one or two business outcomes that leadership plausibly moves, such as turnover, engagement, or sales. The source ties leadership to turnover: 'People leave their bosses, not their jobs' holds true in research, and employees happy with their supervisor stay longer.
2. 2Build one row per team or manager. Columns: engagement result (survey), manager tenure (the PostNL case in the source pairs manager tenure with team engagement), team absenteeism, team lateness, team tenure, and the business outcome. Merge survey results with business data, as the iNostix case does.
3. 3Pull the early-warning indicators per team. The source names absenteeism as the strongest indicator of turnover intentions, together with tenure. Lateness signals demotivation. Low performance predicts exit, and so does very long exceptional performance with no challenge. A team where these cluster under one manager is a leadership flag.
4. 4Check expectation-setting directly. The source says that when job expectations and reality are badly misaligned, people may leave within a few months. Compare exit timing for new hires by manager. Early exits under one manager point to expectation-setting, not pay.
5. 5Add the senior-leadership exposure lens where sales is involved. The VoloMetrix finding is that more exposure to senior leadership correlates with successful sales results. Treat relationship with management as an internal-network variable and measure it.
6. 6Run correlation first, then multiple regression in R with the leadership variable alongside controls. Use summary() and read each coefficient's Pr(>|t|): below 0.05 is statistically reliable. Compare Adjusted R-squared, not R-squared, when you add variables.
7. 7Control for omitted variables. The causal-modelling passage warns that models must account for all influential variables to prevent omitted variable bias. Include team tenure, compensation satisfaction, and co-worker factors before you credit or blame the manager.
8. 8Translate to money and get a decision. Express the finding the way Best Buy does ('a 0.1% engagement increase = $100,000 income'). Then ask your sponsor whether they agree and who owns implementation. Success is measured by impact, not report count.

The source does not publish a manager-level dataset, so this example shows how to read the regression output the source does give. The one-predictor model (Performance ~ Test_scores) gives Multiple R-squared 0.6403 and Adjusted R-squared 0.6146. Residual Standard Error is 7.365 on a mean Performance of 21.6579 (14 df), so predictions are off by about 34%. Adding a second predictor (Personality) lifts R-squared to 0.7121 and Adjusted R-squared to 0.6678. RSE becomes 6.838 on a mean of 17.9953 (13 df), a 38% error. Apply the same read to a leadership model by swapping in a manager variable such as manager tenure or senior-leadership exposure. Keep it only if its p-value is below 0.05 and Adjusted R-squared rises. If Adjusted R-squared falls or p exceeds 0.05, the manager variable is not earning its place. Then size the effect in dollars, the way Best Buy does with engagement.

• If a team shows high absenteeism and short tenure under one manager, then treat it as a leadership flag first and check pay and job content second. The source ranks absenteeism with tenure as the strongest turnover indicator.
• If new hires leave within a few months under a specific manager, then audit the expectation gap between what was promised and the actual job before changing anything else.
• If adding the manager variable raises R-squared but not Adjusted R-squared, then drop it. R-squared always rises with more variables.
• If the manager coefficient has p ≥ 0.05, then do not claim a leadership effect. Report it as not reliable.
• If the audience is executives, then lead with the dollar or outcome impact. Give three main points (Rule of Three) and name an owner for the action.

• Blaming the boss for what is pay satisfaction. The tell: the exit reasons cite comparison with a colleague's higher pay, and the manager variable loses significance once pay satisfaction is controlled.
• Omitted variable bias. The tell: the manager coefficient is large, but you left out team tenure, job content, or co-worker factors, so the effect shifts when you add them.
• Reading R-squared instead of Adjusted R-squared. The tell: you keep adding variables and the fit always looks better.
• Treating correlation as proof that a manager caused the outcome. The tell: you cite a correlation, such as senior-leadership exposure and sales results, with no causal model or control variables.
• Reporting activity instead of impact. The tell: the deck lists surveys run and dashboards built but no outcome change, owner, or sponsor agreement.

• ▸Team / manager ID: one row per team so variance between managers is visible.
• ▸Manager tenure (months): the variable the PostNL case pairs with team engagement.
• ▸Team engagement score: from the latest engagement survey.
• ▸Absenteeism rate: the strongest turnover indicator alongside tenure.
• ▸Lateness rate: a demotivation signal.
• ▸Early-exit count (first few months): flags an expectation gap set at hiring.
• ▸Senior-leadership exposure: a relationship-with-management measure, relevant for sales teams.
• ▸Business outcome (turnover, sales, or income): the dependent variable in the regression.
• ▸Regression read: record p-value (<0.05?) and Adjusted R-squared before and after adding the manager variable.
• ▸Action owner and sponsor agreement: who implements the recommendation, and did the sponsor approve?

1. 1Define the construct as three measurable behaviors, using the VoloMetrix set: time with customers, size of internal network, and time with managers and senior leadership. Add the Genpact finding that top engineers hold more meetings and relationships. Write each metric's unit before you pull any data (hours per week, count of distinct contacts, hours with senior leaders).
2. 2Choose the outcome you want to predict (sales results, or a performance rating) and a population with a comparable role. VoloMetrix found the pattern held 'regardless of region, territory, or sales role', but you should still compare like with like, since consultative and transactional sales differ on breadth versus depth.
3. 3Identify the data sources that can show relationships and communication. The passages point to organizational network analysis (NodeXL in Excel) and email and meeting data (the 'What Emails Reveal About Your Performance At Work' case). Check access before you commit. The passages warn that mismatched employee IDs across systems (SAP vs Payroll), privacy laws, vendor-held data, and knowledge-hiding colleagues all raise effort.
4. 4Build one row per employee: network size, hours with managers and senior leaders, hours with customers, and the outcome. Then run a correlation, since the passages treat correlation as the entry-level test (they list 'Correlation with R' for predicting performance). Treat the result as an association and not as proof of cause.
5. 5Rank the employees by network metrics and look at top performers against the rest. Ask whether top performers show larger networks and more leader time. Note the reasoning offered for why: a larger network exposes people to more ideas, gives easier access to expertise, and builds a better understanding of what is happening.
6. 6Convert the findings into actions you can monitor. The source calls these metrics 'highly actionable': use them in territory design and in coaching, for example by scheduling more manager and senior-leader contact for low-network salespeople.
7. 7Handle privacy before you share results. Introduce the analytics as an opportunity, not a looming threat, and explain how working conditions and employee satisfaction can improve. In unionized settings, be ready to justify how you will use the data.
8. 8Tell the story to the sponsor with visuals. Then ask whether they agree with the recommendations and who owns implementation.

VoloMetrix, reported in HBR's 'What Makes Great Salespeople', examined salesperson behavior and found three metrics highly correlated with top performers: time with customers, larger internal networks, and time with managers and senior leadership. The pattern held across region, territory, and sales role. VoloMetrix then built a broader framework from the three findings (two were combined, and one was added). The framework included customer engagement, covering accounts touched, time per account, frequency of interaction, and depth and breadth of relationships. It also included internal networks, where salespeople with bigger relationship sets are exposed to more ideas, have easier access to expertise, and better understand what is happening. The source calls these metrics strong predictors of sales results and says they are actionable to monitor, to consider in territory design, and to use to improve sales performance. For customer engagement, the source adds that depth beats breadth for top performers, though the right balance depends on what is sold: depth for consultative sales, breadth for transactional. Genpact separately found that its best-performing engineers had more meetings and relationships than others.

• If the role is sales, track time with customers, internal network size, and time with managers and senior leaders together, because those three are the documented predictors.
• If the sales is consultative, favor depth of customer relationships; if it is transactional, favor breadth.
• If employee IDs differ across systems or data access is blocked, then count that as high effort in the Action Priority Matrix and do not treat the project as a Quick Win.
• If this is your first project, pick a high-impact, low-effort question (Quick Win) and deliver what the business head asked for before adding extra insights.
• If staff would read network monitoring as a threat, then frame it as a way to improve working conditions and satisfaction before you collect anything.

• Treating correlation as cause: you tell low-network salespeople to simply 'network more' without testing it. The tell is that no one has checked whether top performers got larger networks as a result of success.
• Counting only volume: you reward time with many customers, or many contacts, regardless of depth. The tell is shallow relationships spread across many accounts when the sales model is consultative.
• Ignoring data linkage and access: IDs do not match across systems or colleagues withhold information. The tell is that the project stalls at data assembly, or the results rest on partial records.
• Skipping privacy and trust: you roll out communication tracking with no explanation. The tell is resistance from unions, staff, or HR colleagues.
• Poor storytelling: you present correlations without a narrative or visuals. The tell is that the sponsor does not understand the findings, has no named implementation owner, and no action follows.

• ▸Role/population: the job group studied (e.g., sales, engineers) so comparisons are like-for-like.
• ▸Outcome metric: the performance or sales measure you are predicting.
• ▸Network size: count of distinct internal relationships per employee, with data source.
• ▸Time with managers and senior leadership: hours per period, with data source.
• ▸Time with customers (and depth vs breadth): hours, accounts touched, and relationship depth, tagged consultative or transactional.
• ▸Data access check: ID matching, privacy law, vendor data, and who may be withholding data.
• ▸Finding: correlation between each metric and the outcome, plus top-vs-rest comparison.
• ▸Action: the concrete change (coaching, territory design, manager time) and the monitoring metric.
• ▸Communication plan: how you frame this to staff, the sponsor's agreement, and the named implementation owner.

1. 1Pull the demographic and commute fields from your HRIS: age, gender, marital status, tenure, home address and commute time. Treat this as the descriptive layer ("What is John's marital status, salary market-ratio, performance rating, traveling time to office?"). The HRIS reports what happened but cannot find patterns, so it only supplies the inputs.
2. 2Turn the descriptive question into a predictive one: "What variables affect John's tenure?" and "Which Sales employee is likely to leave next quarter?" Define the outcome (leaves within a set window) before you pick predictors.
3. 3Convert home addresses into commute time. Use bands like the ones Kenexa uses, from "less than 10 minutes" to "more than 45 minutes". Then look for your own tipping point. FlightNetwork found 1 hour 15 minutes in Toronto, and the source says to use home postal codes to find the equivalent in your city.
4. 4Encode tenure as a curve rather than a straight line. The source reports that employees are more likely to leave in their fourth or fifth year than in their first, so compare cohorts by year of tenure. Encode age as negatively related to turnover (younger employees job-hop more).
5. 5Add context to the demographics: moves and time at current address (frequent movers are more likely to leave), plus triggering events such as having kids, marriage, divorce, caring for a sick family member, a close colleague leaving, or a reorganization.
6. 6Run correlation first to check whether each attribute relates to turnover, then multiple regression or a decision tree to see which variables carry the most weight. Correlation shows that a relationship exists; regression shows which variable has the most impact.
7. 7Combine the attributes into a short at-risk screen alongside pay and performance signals: no title change or raise in over a year, high recent ratings, a commute past your tipping point, a recent life change. Give the flagged list to managers with a stated reason for each name.
8. 8Restrict access to the scores to trained senior management and tell employees how the data is used. HP's Flight Risk score raised privacy problems because employees may not know their data was collected or agree with its use.

FlightNetwork's Chief HR Officer, Omer Aziz, reduced turnover prediction to five criteria: no title change for more than 1 year, no salary increase for more than 1 year, high ratings in the last 2 performance reviews, a commute over 1 hour 15 minutes (his Toronto tipping point), and a life change in the past year. Across 270 employees, 12 met the first four criteria. He contacted their managers, explained the method and warned them of the flight risk. A few days later, one flagged employee asked his manager for a raise. The manager, already warned, negotiated an increase, and the employee stayed. The commute and life-change criteria work as conditions on the pay and promotion signals. Aziz's own illustration is an employee with a 1-hour commute and a newborn, whose commute now cuts into family time.

• If your applicants or staff commute 30 to 45 minutes or more, treat commute as a primary turnover predictor, since one manufacturer saw quitting probability above 92 percent in that band. Then calibrate your own tipping point from postal-code data rather than borrowing Toronto's 1 hour 15 minutes.
• If an employee is highly rated but has had no raise or title change in over a year and also has a long commute or a recent life event, flag them and brief the manager before they ask for a raise.
• If a long commute is the driver in high-volume entry roles (call-center, fast-food), move the fix upstream into hiring. Screen for transport access and distance, as Gate Gourmet did, taking unwanted turnover from about 50 percent to 27 percent.
• If you are tempted to score individuals on age, gender or marital status alone, then use them only as context for the model and restrict who sees the output. Demographics shift the baseline, whereas behavior and pay conditions give you something to act on.
• If tenure is in the model, then check years 4 and 5 specifically instead of reading it as a straight-line effect.

• Treating a borrowed threshold as universal. The tell: your commute cutoff matches a published figure (1 hour 15 minutes, 35 minutes) but you never tested it against your own city's postal-code data.
• Reading demographic correlations as causes or as individual verdicts. The tell: managers say "she's young, so she'll leave" instead of acting on a specific at-risk reason.
• Looking at one attribute at a time. The tell: promotions are assumed to reduce flight risk, but HP found that a promotion without a substantial pay raise still raised the likelihood of quitting.
• Collecting and scoring data without disclosure or access control. The tell: employees do not know a Flight Risk score exists, and it circulates beyond trained senior management.
• Stopping at a descriptive HRIS dashboard. The tell: you can report tenure and age by team but cannot say which employees are likely to leave next quarter or why.

• ▸Employee/role group: name the population you are screening (e.g., call-center new hires).
• ▸Outcome and window: define what counts as leaving and the time horizon (e.g., unwanted exit within 6 months).
• ▸Commute band: record home-to-work time using bands from under 10 minutes to over 45 minutes, plus your tested tipping point.
• ▸Tenure year: record years in role and flag the fourth-to-fifth-year window.
• ▸Age, gender, marital status: record the values as context variables only, noting what they are being compared against.
• ▸Moves/time at address: record time at current address and number of moves.
• ▸Triggering events (past 12 months): list kids, marriage, divorce, family caregiving, colleague departure, reorganization.
• ▸Pay and performance conditions: record months since last raise, months since last title change, and the last 2 ratings.
• ▸Action and owner: name the manager briefed, the retention step planned, and who is authorized to see the score.

1. 1Start with the business question, not the model. Ask how turnover stops a business unit from hitting its KPIs, and don't investigate attrition if it isn't a real problem. Get the sponsor's agreement on scope, timeframe and budget before you pull data.
2. 2Run a literature review before choosing variables. The passages offer a starting list: demographics (age, gender, marital status, tenure), commute time, performance rating, salary market-ratio, personality trait, behavioral signs (increasingly tardy, participating less in meetings, vacation days not taken) and triggering events (having kids, divorce or marriage, caring for a sick family member, a close colleague leaving, reorganization).
3. 3Add attitudinal drivers to the list: pay satisfaction (perceived pay versus peers, not absolute pay), co-worker satisfaction, work satisfaction and the climate culture rating. Pay satisfaction matters because people compare themselves with a friend or colleague in a similar job who earns much more.
4. 4Write each hypothesis as a testable relationship, such as 'longer commute is associated with a higher chance of leaving.' Local thresholds need local data. FlightNetwork's 1 hour 15 minute commute 'tipping point' is specific to Toronto, so derive yours from home-address postal codes.
5. 5Gather the data and define the outcome as a left/stayed flag. Use correlation in R to screen which predictors move with flight risk. Then fit a logistic regression to estimate each employee's probability of leaving. HP's Flight Risk score used two years of data to produce a probability for each of 300,000 employees.
6. 6Check for interactions before reading single coefficients. At HP, higher pay, promotions and better ratings were negatively related to flight risk, yet an employee promoted without a substantial raise was still more likely to quit.
7. 7Score the population and name the key risk factors per employee, so managers get a prompt for action rather than a bare number. Restrict access to the scores to trained senior management, because employees may not know their data is being used.
8. 8Tell the story with people first, then data. Say 'one-third of our new hires in Sales leave' rather than 'attrition is high.' Cast turnover as the villain and your retention action as the hero, and state the cost: lost productivity, onboarding time, and departing employees taking knowledge, network and customers with them.

FlightNetwork's Chief HR Officer, Omer Aziz, built a five-criteria 'at-risk' screen instead of a model. An employee is flagged if they have had no title change in more than 1 year, no salary increase in more than 1 year, high ratings in the last 2 performance reviews, a commute over 1 hour 15 minutes (the Toronto 'tipping point'), and a life change in the past year. Picture a well-rated employee who has not been promoted or given a raise in 14 months, commutes 1 hour 20 minutes each way, and recently had a child. That employee meets all five criteria and is the profile to approach first. The logic is that the screen targets valued people (high ratings) who have stalled (no title or pay change) and face strain (commute, life event). Aziz suggests setting the commute threshold from postal codes for your own city. You can turn the same five variables into a logistic regression on a left/stayed outcome. That gives you data-derived weights and probabilities, which replace his equal-weight rule of thumb.

• If attrition isn't blocking a business unit's KPIs, then don't start the project; pick the problem the sponsor cares about.
• If a driver is a perception such as pay, measure satisfaction with it (a survey item) rather than the absolute figure, since satisfaction is what predicts turnover.
• If a threshold comes from another company or city (such as a 1h15m commute), then re-derive it from your own data before using it.
• If an employee shows promotion without a meaningful raise, then raise their risk score, since HP found the combination still predicts quitting.
• If you produce individual flight-risk scores, then limit access to trained senior managers and treat confidentiality as a launch condition.

• Treating pay, promotion and rating effects as independent and additive. The tell: promoted-but-underpaid employees are scored low-risk and then resign.
• Importing another firm's cut-offs without testing. The tell: your flag list is dominated by people who never leave, or misses people who do.
• Building the model before agreeing on the problem with sponsors. The tell: sponsors won't give data access or act on recommendations.
• Releasing scores widely to line managers without training. The tell: employees learn they are being scored, and the data is misread or breaches confidentiality.
• Presenting rates instead of people. The tell: the audience nods at 'attrition is high' and no one commits to action.

• ▸Business question: how does turnover block a named KPI in a named unit?
• ▸Sponsor and agreed scope/timeframe/budget: who approves, who supplies data?
• ▸Outcome definition: what counts as 'left' and over what window?
• ▸Candidate predictors: tick demographics, commute, rating, salary market-ratio, personality, behavioral signs, triggering events.
• ▸Attitudinal measures: pay satisfaction, co-worker satisfaction, work satisfaction, climate culture rating (source and survey tool).
• ▸Local thresholds: your city's commute tipping point, years without a title or pay change.
• ▸Interaction checks: for example, promotion without a substantial raise.
• ▸Model and result: correlation findings, logistic regression variables and each employee's probability.
• ▸Access and privacy plan: who may see individual scores, who is trained, how employees are informed.
• ▸Story line: the villain, the hero action, and the people-first headline statistic.

1. 1State the question as a testable link: 'Does employee engagement (or a personality trait, or organizational climate) predict customer satisfaction?' The source frames three predictors: engagement, personality traits such as conscientiousness and extraversion, and organizational climate and autonomy. Pick one per project. This is Step 1 (Ask) of the ARHAT approach.
2. 2Review the literature before you touch data (Step 2). The source lists three cases for this construct: engagement correlating with customer experience, conscientiousness and extraversion affecting customer service, and climate and autonomy affecting customer service. Use them to choose your predictor and to state the direction of the effect you expect.
3. 3Write the hypothesis as a directional statement (Step 3), for example 'Units with higher engagement scores have higher customer satisfaction scores.' Decide the unit of analysis (store, team, or branch) before you collect anything.
4. 4Gather matched data (Step 4a). Take the predictor from engagement or pulse surveys, personality assessments, or climate and autonomy items. Take the outcome from customer-side measures such as cNPS. Merge the two into one table, one row per unit. iNostix's approach is to merge engagement survey outcomes with business data.
5. 5Run correlation first. The book teaches R's correlation function (Excel works too) to see whether the predictor and customer satisfaction move together, and which personality traits mark top-performing customer service staff.
6. 6Run multiple regression to isolate the predictor's effect (Step 4b). Add other drivers such as personality, climate, or engagement, so you can see which one carries the effect. The book shows the prediction form with its diversity-index example: plug in a diversity index of 2.7 and $90 of advertising to predict sales. Do the same with an engagement score and your other drivers to predict customer satisfaction.
7. 7Tell the story (Step 5). Name the audience, build a narrative from the finding to the action it justifies, and chart it simply. Express the result as a business consequence. Best Buy's example is that a 0.1% increase in engagement predicts a $100,000 increase in income.
8. 8Add qualitative diagnosis when the number moves. The book's tools are pulse surveys, focus groups, Glassdoor reviews, sentiment analysis, and word clouds. Use them to learn why engagement or customer sentiment is low.

The passages give no customer-satisfaction coefficients, so this example applies the method to the one real quantified engagement result they contain. Best Buy can predict that a 0.1% increase in employee engagement results in a $100,000 increase in income. That figure is store income, not satisfaction, but it shows the output form you are aiming for: a small, specific engagement change translated into a business number. To apply it to customer satisfaction, follow the ARHAT sequence. Ask: does engagement predict customer experience? Review: the engagement-and-customer-experience study (Case 2 in the book's engagement analytics section) already says it does. Hypothesize: higher-engagement stores earn higher cNPS. Analyze: put each store's engagement score and cNPS in one table, correlate them, then run multiple regression with climate or personality scores as additional predictors. Tell the story: 'each X-point engagement gain predicts a Y-point cNPS gain.' You fill in X and Y from your own regression; the source supplies neither.

• If you want to know whether two measures move together, then run correlation. If you want to isolate one driver while holding others constant, or to produce a prediction, then run multiple regression.
• If the engagement-to-satisfaction link is weak or absent, then test the other two source-named drivers, personality (conscientiousness, extraversion) and organizational climate and autonomy, before dropping the idea.
• If leaders say 'happy employees mean happy customers,' then require the data. The source warns that making employees happy does not mean they will work hard for you, so test engagement against the customer outcome rather than assuming it.
• If the audience is non-technical, then lead with the business consequence in dollars or score points, and show a simple chart rather than regression output.
• If the statistics show a gap but not its cause, then use pulse surveys, focus groups, or sentiment analysis of comments to diagnose it.

• Measuring satisfaction on one side only. You correlate engagement with itself, or with an employee proxy, and never merge in customer data. The tell: no customer-side column such as cNPS in your table.
• Reporting a correlation as proof that engagement causes satisfaction. The tell: the narrative says 'will raise' with no regression controls or theory behind it. The source's own wording is 'predict' and 'correlates'.
• Starting with data instead of a question. You run correlations across everything and report whatever is significant. The tell: no written hypothesis from Step 3.
• Handing over statistical output instead of a story. The tell: the audience asks 'so what?' because there is no consequence stated in business terms.
• Treating personality as a hiring shortcut without testing it locally. The tell: you cite conscientiousness and extraversion from the literature but never correlate them with your own customer service staff's results.

• ▸Question: the one-sentence question linking a driver to customer satisfaction.
• ▸Driver tested: engagement, personality trait (conscientiousness or extraversion), or climate/autonomy.
• ▸Customer outcome metric: cNPS or other satisfaction score, with its source and period.
• ▸Unit of analysis: store, team, or branch, and the number of units.
• ▸Hypothesis: direction and expected sign (higher driver gives higher satisfaction).
• ▸Data sources merged: survey, assessment, or climate data, plus the customer data, and the join key.
• ▸Correlation result: r between driver and outcome.
• ▸Regression result: coefficients with the other drivers included, and the predicted outcome for one scenario value.
• ▸Story line: audience, one-sentence business consequence, and chart type.

1. 1Write the performance question as one sentence with a named outcome, for example 'Do conscientiousness and extraversion predict customer service performance in our store staff?' This is ARHAT Step 1 (Ask Questions). Name the role group, because the passages show the same driver behaves differently by job type.
2. 2Review the literature before touching data (ARHAT Step 2). Collect the findings that bear on your role group: personality traits (conscientiousness, extraversion), service climate and perceived autonomy (Ployhart et al. 2011), engagement (ISS), and training (Ployhart et al. 2011, restaurants). Record each finding with its job context so you can check whether it transfers.
3. 3State each hypothesis in a directional, testable form (Step 3). Example: 'Higher conscientiousness scores correlate with higher performance ratings among customer service staff.' Write one hypothesis per predictor. Keep the performance rating or productivity measure as the single dependent variable.
4. 4Define the outcome column up front. Use the individual job performance or productivity rating that already exists, and record who rated it and over what period. Merge it with the predictor columns (engagement survey result, training attended, personality score, inclusion measure) into one data frame with one row per employee.
5. 5Import the merged file into R (read.csv with header=TRUE) and plot first. Run hist() on the performance column and check the shape before any statistics. A graph shows patterns that raw numbers hide.
6. 6Run a correlation of each predictor against performance as the first screen. Then run multiple regression with the predictors together, so each effect is read with the others held constant. The source's own sales example enters the diversity index and advertising spend in the same model.
7. 7Translate results into business units for the audience (Step 5, Tell the Story). Follow the Best Buy pattern: a 0.1% increase in engagement is tied to a $100,000 increase in income. State the predictor change and the outcome change in one sentence, then add the visual.
8. 8Before using any model to select or rank people, check the legal context. The book lists Griggs v. Duke Power Co. (1971) and Johnson v. Transportation Agency (1987) under starting an HR analytics project. Get legal review for any hiring use.

You run people analytics for a retail chain and want to know whether training lifts customer service performance. Step 1: 'Does service training raise individual service performance ratings?' Step 2: you find Ployhart et al. (2011), who studied restaurants and found that service training did not lead to higher service levels. Employees in less complex service jobs, such as waiting tables, seemed to benefit less than those in more complicated jobs such as software sales. The source's explanation is that attending training does not mean the employee learns or applies the skills. Step 3: you split the hypothesis by job complexity. 'Training correlates with performance ratings for complex-selling roles but not for simple service roles.' You also add two rival predictors from the same literature: conscientiousness and extraversion, and service climate and perceived autonomy. Step 4: you merge rating, training attendance, personality scores and climate scores per employee, then correlate and run multiple regression by role group. If training only shows an effect in the complex group, you report it that way. Step 5: you give the audience one visual per role group and a sentence in business units, rather than a single company-wide training verdict.

• If your role is a simple, low-complexity service job, then do not assume training lifts performance. Test it against personality and climate predictors first, since the restaurant evidence shows training did not raise service levels.
• If your role is complex and sales-like, then include training as a predictor, but split the analysis by job type rather than pooling roles.
• If a correlation looks strong in one predictor, then rerun it in multiple regression with the other predictors (for example advertising spend alongside diversity) before claiming an independent effect.
• If the result will inform hiring or selection, then pause and obtain legal review against the case law the book lists (Griggs v. Duke Power Co., Johnson v. Transportation Agency) before use.
• If the audience is non-technical, then lead with a graph and a one-line business-unit statement (as in the Best Buy 0.1% to $100,000 example), not the regression table.

• Treating training attendance as skill applied. The tell: attendance is high, but ratings do not move for simple-service roles, as in the restaurant study.
• Pooling all roles into one model. The tell: a weak, muddy overall correlation, while the source shows effects depend on job complexity.
• Starting with data instead of a hypothesis, so every predictor gets run and any hit is reported. The tell: you cannot say which literature finding each variable tests.
• Using ratings without knowing their origin. The tell: you cannot state who rated, over what period, or whether the rating measures productivity or opinion, so the outcome column is untraceable.
• Reporting statistics without the story step. The tell: stakeholders ask 'so what?' because there is no visual and no business-unit sentence.

• ▸Research question: one sentence naming the performance outcome and the role group.
• ▸Role group and job complexity: simple or complex, and why. Used to decide whether to split the analysis.
• ▸Performance measure: the rating or productivity metric, who rated it, and the period covered.
• ▸Literature findings: each finding with source and job context (e.g. Ployhart et al. 2011, restaurants).
• ▸Hypotheses: one directional, testable statement per predictor.
• ▸Predictor columns: engagement, training, communication, inclusion, personality; the column name and data source for each.
• ▸Analysis plan: histogram check, then correlation per predictor, then multiple regression with all predictors.
• ▸Legal check: whether results touch selection or hiring, and who reviewed it against the listed case law.
• ▸Business-unit statement: 'A change of X in the predictor is associated with a change of Y in performance', plus the planned visual.

1. 1Write the business question first, in the ARHAT 'Ask' step. Name the financial outcome (store income, revenue, sales per employee, profit margin, EBIT) and the sponsor who cares about it. The source warns against analytics projects the business heads neither want nor support.
2. 2Pick one people driver per hypothesis from the four the source ties to financial outcomes: engagement, diversity, training, compensation. Write it as a directional claim, e.g. 'Higher store engagement scores are associated with higher store income.' This is the 'Hypothesis Formulation' step.
3. 3Review the literature before touching data (ARHAT step 2). The source's Sales and Profitability cases are Best Buy (engagement → store income), engagement plus customer advocacy → profit, diversity → market share, revenue and EBIT, and sales-trait studies (Rentokil, Deloitte, HBR). Use them to choose which variables to include and what direction to expect.
4. 4Build one table with one row per unit, such as a store, branch or business unit, and one period. Put the financial outcome in one column and the people measures in the others. The source's method is to merge engagement survey outcomes with business data. Include controls you can defend, such as the advertising spend used in the source's diversity example.
5. 5Convert categorical mixes into a single number before regression. The source turns an ethnicity mix (70% Chinese : 20% Malay : 10% Indian) into a diversity index (e.g. 2.7). The passages do not give the index formula, so document whichever formula you choose and apply it identically to every unit.
6. 6Run correlation first to screen each driver against the outcome. Then run multiple regression with the drivers together, because the source's examples predict sales from several inputs at once (diversity index plus advertising spend; training hours plus other variables). Both are a few commands in R.
7. 7Turn the model into a prediction and a prescription. Plug in a scenario, as the source does with 'diversity index 2.7 and $90 on advertising'. Express the result as a unit effect, like Best Buy's 'a 0.1% increase in engagement results in a $100,000 increase in income'. That framing moves from predictive to prescriptive analytics.
8. 8Tell the story (ARHAT step 5). Name the audience, build the narrative around the money figure, and use a visual that shows the driver-to-outcome relationship. Leave the regression table in an appendix.

Best Buy is the source's clearest case. The company merged engagement survey results with store financials. From that it can predict that a 0.1% increase in employee engagement results in a $100,000 increase in income. Here is how to reproduce the logic on your own data. Question: does engagement move store income? Data: one row per store, with engagement score and store income. Screen with correlation, then run regression with controls. The slope is the headline: income change per unit of engagement. Before extrapolating, test whether the relationship holds across the range of scores you observed. The source gives the 0.1% to $100,000 figure only for small changes, so do not scale it to a 5% jump without checking. The finished slide reads 'Each 0.1% engagement gain is worth about $100,000 in store income', and the decision it supports is how much to invest in engagement.

• If the outcome is a financial figure and your driver is a survey score, merge the two at the same unit level (store with store). Never compare company-wide engagement with store-level income.
• If you have more than one plausible driver (diversity, advertising, training), use multiple regression rather than separate correlations. A single correlation can credit one driver with another's effect.
• If a driver is a category mix such as an ethnicity breakdown, convert it to a single index first. Document the formula, because the passages do not specify one.
• If stakeholders cannot name the decision the number will inform, return to the 'Ask' step. Do not run the analysis.
• If you want to say 'raise X by this much, gain Y', state the unit and the range tested, as Best Buy's 0.1% → $100,000 does. Do not extend the claim beyond that range.

• Starting with data instead of a business question. Tell: the deliverable is a dashboard of correlations with no named decision or sponsor.
• Treating correlation as proof that engagement, diversity or training caused the profit. Tell: the claim is stated as 'engagement drives profit' with no control variables and no literature support (the source lists causal modelling as a separate technique).
• Mismatched units of analysis. Tell: company-level survey averages are regressed on store-level or individual-level revenue.
• Undocumented index construction. Tell: a diversity index like 2.7 is reported but nobody can say how it was computed, so results cannot be replicated or compared across periods.
• Extrapolating a small-increment effect to large changes. Tell: a 0.1% to $100,000 style figure is multiplied up to a big engagement jump and presented as a forecast.

• ▸Business question: one sentence naming the financial outcome (revenue, sales per employee, profit margin, EBIT, store income) and the decision it informs.
• ▸Sponsor: the business head who wants the answer and will act on it.
• ▸Driver(s): engagement, diversity index, training, compensation, or others, each with its unit of measure.
• ▸Index formula (if any): how a category mix was converted to a number; paste the exact formula.
• ▸Unit of analysis and period: store, branch or unit, and the time window, identical for outcome and drivers.
• ▸Control variables: e.g. advertising spend; note why each is included.
• ▸Hypothesis: directional statement, e.g. 'Higher X is associated with higher Y.'
• ▸Result: correlation and regression slope per driver, expressed as 'a unit change in X relates to Z change in outcome'.
• ▸Valid range and caveat: the range of driver values observed, and the limit on extrapolation.

1. 1Define the metric before you model it: absenteeism is the frequency of unscheduled absence and sick days per employee per year. Fix the unit (days or hours per employee per year) so every driver is converted into the same terms. Deloitte's finding is expressed as roughly one day, or 6.5 hours, per employee per year.
2. 2Pull the three driver families the source ties to absenteeism: diversity and inclusion (share of employees who feel included), engagement (engagement score by unit), and learning and development (share who say they have had opportunities to learn and grow). Collect one value per unit, such as a store or team, for each driver.
3. 3Pair each unit's driver values with its actual absence record. The Deloitte finding is a correlation between inclusion and actual absenteeism, not between inclusion and intended attendance. Use recorded absence, not survey intent.
4. 4Run a correlation of each driver against absence, then compare quartiles. Marks & Spencer compared the top engagement quartile with the bottom quartile and found absenteeism 25% lower. Split your units into engagement quartiles and compare mean absence between top and bottom.
5. 5Translate the effect into attendance terms. Use the Deloitte rule of thumb as a benchmark only: a 10-percentage-point rise in employees who feel included corresponds to about one more day of attendance per employee per year in one organization. Multiply days gained by headcount to size the effect, then price it with your own cost per absent day.
6. 6Benchmark engagement against absence at the individual level. The source cites 2.7 sick days per year for engaged employees versus 6.2 for disengaged ones. Segment your own population by engagement and compare your gap with that figure.
7. 7Frame the result as a story: absenteeism is the villain, the inclusion, engagement or learning programme is the hero, and the data is the supporting character. State it in people terms, such as 'two out of three are engaged', rather than as a bare score.
8. 8Treat these findings as hypotheses to test in your own data. The sources are single-organization or cited results, and the book's framework is ask, review literature, hypothesise, analyze, tell the story.

Deloitte Australia and the Victorian Equal Opportunity and Human Rights Commission studied inclusion and absenteeism in one organization. They found that the more included an employee feels, the more likely that employee is at work. If 10 percent more employees feel included, attendance rises by about one day per year per employee, or 6.5 hours. Apply that to a hypothetical unit (these figures are illustrative, not from the source): 400 employees and an inclusion gain of 10 points implies about 400 extra attendance days a year. Multiply by your own cost per absent day to size the benefit. Cross-check against Marks & Spencer, where top-quartile engagement stores had 25% lower absenteeism than bottom-quartile stores. If your own quartile gap is far from 25%, investigate before reporting.

• If you have inclusion survey data and recorded absence by unit, then correlate them first; that is the Deloitte-style test and the cheapest to run.
• If engagement scores exist by store or team, then compare top and bottom quartiles on absence rather than relying on the average, as Marks & Spencer did.
• If your measured effect is much larger than the cited benchmarks (about one day per 10-point inclusion gain, 25% quartile gap, 44% drop from the Gallup learning scenario), then check data quality and confounders before presenting it.
• If learning opportunities are low (the source cites only 4 in 10 employees saying they have them), then treat raising that share as a testable lever, since Gallup suggests going to 8 in 10 could cut absenteeism by 44%.
• If stakeholders need a decision, then convert days gained into cost and present the people-first story, not the correlation coefficient.

• Treating a correlation as proof of causation: the source reports correlations, and a claim that inclusion 'will increase' attendance is a hypothesis. The tell is a savings figure with no test or pilot behind it.
• Using survey intent or self-reported attendance instead of actual absence records. The tell is a driver-to-outcome link built entirely from survey items.
• Copying benchmarks as your own result: the one-day-per-10-percent figure comes from data from one organization. The tell is a forecast that cites Deloitte but contains no local data.
• Comparing averages and hiding the spread: the 25% gap came from top versus bottom quartiles, so averaging masks it. The tell is a flat correlation that disappears when you split by quartile.
• Presenting data as the main character: leading with a score rather than people loses the audience. The tell is a slide that opens with 'average engagement is seventy percent' instead of 'two out of three are engaged'.

• ▸Absence unit of measure: state days or hours per employee per year and the period covered.
• ▸Unit of analysis: list the stores, teams or sites being compared.
• ▸Inclusion measure: percentage of employees who feel included, per unit.
• ▸Engagement measure: engagement score per unit, plus its quartile (top or bottom).
• ▸Learning measure: percentage saying they have had opportunities to learn and grow, per unit.
• ▸Actual absence per unit: recorded days from HR records, not survey intent.
• ▸Observed gap: top versus bottom quartile absence difference, compared with the 25% Marks & Spencer benchmark.
• ▸Days gained and cost: estimated attendance days gained times your own cost per absent day.
• ▸Story line: villain (absenteeism), hero (the programme), and one people-first sentence stating the finding.

1. 1Define the outcome before the drivers. Pick one safety measure (incident count per employee) and one health measure (sick days per year). Keep them separate: the passages treat safety incidents and health/sick days as different outcomes with different drivers.
2. 2Build one row per employee or team with the outcome plus the candidate drivers the source names: age group, tenure band (the accident case contrasts staff with 5 years of service and below against longer-serving staff), supervisor communication score, engagement score, and workplace air quality or building type.
3. 3Cut incidents by age group first, then by age group and supervisor communication score together. This is the two-step view in the IBM Kenexa Talent Insights example. The second cut shows whether a risky age group is really a poor-communication group.
4. 4Cut incidents by tenure band, using 5 years of service and below as the first split. Check whether incidents concentrate in the shorter-tenure band before you design any intervention.
5. 5Quantify engagement links with correlation, then regression. Correlate engagement with incident counts and with sick days. Then fit a regression with engagement plus age, tenure and communication score, so you see each driver's effect with the others held constant.
6. 6Use the fitted model to predict. In R, the `predict(model, data.frame(...))` pattern takes two or more hypothetical teams and returns an expected outcome for each. Use it to size the effect of a change before you fund it.
7. 7Add the environment as a separate test. Compare health complaints (headaches, respiratory complaints) and a performance measure between better-air and standard-air locations. The Harvard/Syracuse green-building study is the benchmark: 30% fewer headaches and respiratory complaints, about 27% better cognitive-task performance.
8. 8Price the result through the indirect path. Multiply the safety-to-profit correlation by the upstream driver's correlation with safety, as in the turnover-quality/safety-profit chain. Brief business heads in advance that the result may be surprising, inconclusive, or a confirmation of what they already suspected.

The source decomposes how staff turnover reaches profit. The total turnover-to-profit correlation is -.03, which is modest. Quality/safety outcomes correlate .21 with profitability, and the source gives -.12 as the turnover-to-quality/safety link. The indirect path is .21 × -.12 ≈ -.025. That is roughly 80-plus percent of the -.03 total (0.025 / 0.03 ≈ 83%), so most of turnover's profit damage runs through quality and safety. Practical reading: if you must defend a safety investment to finance, show this path rather than a stand-alone safety correlation. The health side works the same way. Engaged employees average 2.7 sick days a year and disengaged employees 6.2, a gap of 3.5 days per person. Illustratively, across 100 employees that is 350 days, which you can put in front of operations next to the incident cuts by age and supervisor communication score.

• If incidents cluster in one age group, then cut that group again by supervisor communication score before you act. A communication problem needs a supervisor fix, not an age-targeted program.
• If incidents cluster at 5 years of service and below, then direct onboarding and early-tenure safety effort there first.
• If engagement correlates with incidents or sick days, then run a regression with age, tenure and communication in the model before you claim engagement is the driver.
• If the direct correlation looks small (like -.03), then test indirect paths through quality/safety before you dismiss the factor.
• If you plan a building or air-quality change, then measure headaches, respiratory complaints and a cognitive-task or performance measure before and after, since those are the outcomes the green-building study moved.

• Reading an age-group spike as an age effect. The tell: the spike disappears or shifts once you split by supervisor communication score.
• Treating a small correlation as unimportant. The tell: the direct figure is about -.03 and nobody has checked the mediated path (.21 × -.12).
• Mixing health and safety outcomes in one metric. The tell: one 'wellbeing' number moves while incidents and sick days move in opposite directions.
• Claiming causation from single-variable correlations. The tell: no regression with age, tenure and communication alongside engagement, so the engagement effect cannot be separated from them.
• Presenting without managing expectations. The tell: stakeholders react badly to an inconclusive or surprising result because you never discussed possible outcomes in advance.

• ▸Outcome metric: name the safety measure (incidents per employee) and the health measure (sick days per year); keep them separate.
• ▸Time window and population: state the period and which employees or sites are included.
• ▸Age-group cut: list incidents per group; flag the highest group.
• ▸Age × supervisor communication cut: note whether the age pattern holds within communication-score bands.
• ▸Tenure cut: compare incidents for ≤5 years of service against longer tenure.
• ▸Engagement link: record the correlation of engagement with incidents and with sick days; note the regression result with age, tenure and communication included.
• ▸Environment comparison: list headache/respiratory complaints and a performance measure by building or air-quality condition.
• ▸Indirect-path check: write the chain (driver → quality/safety → profit) with both correlations and their product.
• ▸Expected result and surprises: record what stakeholders expect and what outcomes you told them to be prepared for.

1. 1Fix the decision before you build the story. Write one sentence: what do you want the audience to know or do (make a decision, or discuss a topic)? Tie it to a business problem and KPI agreed with the project sponsor in Step 1, so the story answers a question someone already cares about.
2. 2Profile the audience. Name who is in the room, their needs and biggest challenges, then classify each as Neutral, Critic, Uninformed, Subject-Matter Expert, or Top Management. Each type changes content: critics need a common ground, uninformed audiences need a decision on how much to explain, experts need something new, top management needs it concise.
3. 3Assemble the three elements and check each pairing. Data plus narrative explains (context, source, relevance, what you did, why it matters). Data plus visuals enlightens (trends and outliers invisible in raw numbers). Narrative plus visuals engages. All three together drive change. If one is missing, the story is incomplete.
4. 4Build the narrative in three acts. Setup (why should I care?): give background and context, hook the audience, and cast the villain (the problem, e.g. high staff turnover, absenteeism, performance issues) and the hero (the solution, e.g. an incentive or training programme). Confrontation (how will this make my life better?): detail the problem, then the solution and supporting points. Resolution (what action do I need to take now?): give a road map to a better future and a call to action.
5. 5Place your data points as evidence inside the acts, not as the story itself. Different people read the same data differently because data carries no context, so every number needs a sentence telling the audience what it means for them.
6. 6Develop the visuals after the narrative is written. Visualize the emotions: add human elements and natural transitions, and choose images and charts that carry the plot, not decoration.
7. 7Close with the ask. After presenting recommendations, ask the project sponsor whether they agree and who is responsible for implementing them. Record the owner's name before you leave the meeting.
8. 8Prepare for resistance. If the change affects people, plan to involve those affected in deciding how to make the change, and frame analytics as an opportunity, not a looming threat.

Take the passages' own attrition material. Audience: a business head who is a Top Management type, so keep it concise. Setup: staff turnover is the villain, and you tie it to a KPI the sponsor already owns. Confrontation: Gallup finds 6 in 10 employees know what is expected of them at work; raising that to 8 in 10 could achieve a 14% reduction in staff turnover and a 7% increase in productivity. The hero is a programme that clarifies expectations. You can add that Genpact's people analytics team found leavers became significantly less engaged in their communications up to six months in advance, so early signs let you act before resignation. Resolution: ask the sponsor to approve a clear-expectations programme and name the person who will implement it. Note the passages also cite Stanford professor Chip Heath's finding that 63% remember stories but only 5% remember a single statistic, which is why the numbers sit inside the narrative.

• If the audience is Top Management, then keep it concise and lead with the recommendation and the action required.
• If the audience is Critics, then identify why they disagree and open with a common ground before presenting the proposal.
• If the audience is Uninformed, then decide how much they know and size the background accordingly; if they are Subject-Matter Experts, then lead with what is new to them.
• If a chart or number does not support the setup, confrontation, or resolution, then cut it, because data alone is less memorable than story and open to varied interpretations.
• If the sponsor does not agree or no implementer is named at the end, then treat the recommendation as unfinished and resolve ownership before closing.

• Presenting data without a story: the audience interprets the numbers differently or asks 'so what?'. The tell is that the meeting ends in discussion with no decision.
• Treating storytelling as an afterthought: a good project fails because recommendations are poorly expressed. The tell is that analysis took months but the story was assembled the night before.
• One-size-fits-all deck: the same content goes to executives, critics, and experts. The tell is that executives lose patience and experts hear nothing new.
• Dropping the call to action: the resolution act is missing or vague. The tell is nobody is named as responsible for implementation.
• Framing analytics as a threat: employees or unions see data use as surveillance. The tell is pushback about privacy or data use rather than the findings.

• ▸Business question and KPI: the problem agreed with the sponsor and the KPI it affects.
• ▸Audience and type: who is in the room, and whether Neutral, Critic, Uninformed, Subject-Matter Expert, or Top Management.
• ▸Audience needs and challenges: what they care about and what worries them.
• ▸Desired decision or action: what you want them to know or do, in one sentence.
• ▸Setup (why): background, context, the villain (problem) and the hero (solution).
• ▸Confrontation (how): key data points, each paired with its meaning, and how the solution improves their situation.
• ▸Resolution (what): the road map and the specific call to action.
• ▸Visuals: the chart or image that carries each act, and the emotion or human element it conveys.
• ▸Owner and sponsor agreement: who implements, and whether the sponsor agreed.

1. 1Start with a business question, not a survey. Using the book's ARHAT order (Ask, Review literature, Hypothesize, Analyze, Tell the story), write one outcome question such as 'Does engagement relate to customer experience, store income, or profitability?' The book's engagement cases cover exactly these outcomes (ISS profitability, ISS customer experience, Best Buy store income).
2. 2Review the literature before touching data. The book frames engagement as repeatedly linked to customer experience, profit, and store income, and it also runs separate cases on drivers: flexible work and morale (Upwork) and manager tenure and team engagement (PostNL). Decide whether your question is engagement as a driver of outcomes or engagement as an outcome of management conditions.
3. 3Write the hypothesis as a testable relationship, for example 'higher unit engagement is associated with higher unit customer-experience scores'. This sets up the null hypothesis (no relationship) that the regression output will reject or fail to reject.
4. 4Secure a sponsor and data access first. The book states that analytics projects without a strong sponsor are likely to fail, and that they fail even with a good hypothesis if you cannot get the data. Agree scope, timeframe and budget with the sponsor, and identify the data owners you depend on for business data.
5. 5Gather and merge. Take engagement survey outcomes, including eNPS where you use it, and merge them with business data at the same unit level (store, team, region). This is the iNostix case: 'merging engagement survey outcomes with business data'. Without a shared key, no correlation is possible.
6. 6Run the correlation first, then multiple regression. Use R's correlation to see if engagement moves with the outcome. Then use multiple regression to add other predictors (the book's sales example adds a second predictor, advertising spend) so engagement is not credited with another variable's effect.
7. 7Read the regression output in a fixed order. Check the Estimate and Std. Error, then the t value (Estimate divided by Std. Error; far from 0 means a relationship), then Pr(>|t|) (below 0.05 is statistically significant). Then check the Residual Standard Error as a percentage of the mean outcome, and Multiple/Adjusted R-squared for fit.
8. 8Tell the story for the audience. Follow the book's Know the Audience, Develop the Narrative, Develop the Visuals sequence. Lead with the business outcome and the direction of the relationship, and show the sponsor what decision it supports.

The passages give no engagement-specific regression numbers, so this uses the book's own worked regression output (Test_scores predicting Performance) as the template for reading an engagement-to-outcome model. Swap 'engagement score' in for Test_scores and a unit outcome such as store income or customer experience in for Performance. In the book's output the predictor's t-statistic is 2.625 against a standard error of 0.1740. That is far enough from zero to reject the null hypothesis, so a relationship exists. The Residual Standard Error is 6.838 on 13 degrees of freedom against a mean outcome of 17.9953. Any prediction is therefore off by about 38% (6.838/17.9953). Multiple R-squared is 0.7121, so the model explains 71% of the variance in the outcome. The read-out is that the relationship is significant and the fit is decent, but individual predictions stay loose. The book's earlier output shows the same pattern with a t of 4.992, a mean of 21.6579, and a 34% error. In the story for the sponsor, report both the significance and the 34–38% prediction error.

• If the p-value is below 0.05 and the t value is far from zero, treat the engagement-outcome relationship as reliable. If not, report no demonstrated relationship rather than a weak positive one.
• If the Residual Standard Error divided by the mean outcome is large (the book's examples show 34% and 38%), use the model to show direction and relative size, not to forecast a single unit's result.
• If the correlation holds but another plausible driver exists (the book's sales model also includes advertising spend), move to multiple regression before claiming engagement's effect.
• If you cannot get the business data or a senior sponsor, pause the project. The book says projects fail without data access even when the hypothesis is sound.
• If the question is about what shapes engagement rather than what engagement shapes, model drivers such as flexible work arrangements or manager tenure, following the Upwork and PostNL cases, instead of an outcome model.

• Running survey results alone and never merging them with business data. The tell: the report describes scores and trends but cannot state an effect on sales, profit or customer experience.
• Treating a significant p-value as a precise forecast. The tell: a headline like 'engagement predicts store income' with no mention of the 34–38% typical error or the R-squared.
• Reading correlation as proof of cause. The tell: the narrative says 'raising engagement will raise income' when only association was tested and no other variables were controlled.
• Starting without a sponsor or data owner. The tell: the hypothesis is clear but the extracts are stuck, delayed, or incomplete across units.
• Presenting regression statistics to a business audience unedited. The tell: slides show t values and Signif. codes instead of a plain statement of the finding and the decision it supports.

• ▸Business question: one sentence naming the outcome (e.g. customer experience, store income, profitability).
• ▸Hypothesis: state the expected relationship between engagement and the outcome, and the null.
• ▸Sponsor and data owners: name the senior sponsor and who controls survey and business data.
• ▸Agreed scope, timeframe, budget: record what the sponsor signed off before work started.
• ▸Engagement measure: survey index or eNPS, plus the unit level it is reported at.
• ▸Outcome and merge key: the business metric and the shared identifier (store, team, region) used to merge.
• ▸Other predictors to control for: variables to add in multiple regression (e.g. advertising spend in the book's example).
• ▸Results: predictor t value, p-value, Residual Standard Error as % of mean outcome, Adjusted R-squared.
• ▸Audience, narrative and visual: who decides, the one-line finding, and the chart that shows it.

1. 1Define the dimensions first. The construct covers ethnicity, gender and age, so choose one dimension per index and write down its categories (the book's example is Chinese / Malay / Indian). Do not blend dimensions into one number.
2. 2Pull headcount by category from HR records for a single snapshot date. Calculate each category's proportion p (count ÷ total headcount). Check that the proportions sum to 100%.
3. 3Compute Simpson's: D = Σp². Report one stated form, either 1 − D (probability two random employees differ) or 1/D (effective number of groups). Label which form you used on every chart, because the two scales are not interchangeable.
4. 4Repeat per unit (department, store, region, senior-executive team) and per period. The cited research works at company level and at senior-executive-team level, so decide which level your question needs before you calculate.
5. 5Measure inclusion separately from composition. The index only describes workforce composition. Pair it with an access or outcome measure, such as promotions by group, to test whether equal access exists. Johnson v. Transportation Agency shows the logic: the Agency tracked underrepresentation in a job category (no dispatcher was a woman) and set a measurable annual improvement goal.
6. 6Link the index to a business outcome with multiple regression in R (lm). Put the diversity index and a control such as advertising spend on the right-hand side, and sales on the left. Predict by supplying new values for both, as the book does with an index and $90 advertising.
7. 7Benchmark the result against published effect sizes (Herring, McKinsey) as context only. Then decide on one target, for example Nielsen's goal of raising women in leadership from 34% toward 50% representation, and re-measure on a fixed schedule.
8. 8Tell the story with the audience in mind. Lead with the one outcome-linked finding, show the index trend, and name the action. Check legal context before acting on group characteristics (Griggs: tools must not be designed, intended or used to discriminate).

Take the book's illustrative mix: 70% Chinese, 20% Malay, 10% Indian. Squared proportions are 0.49 + 0.04 + 0.01, so Σp² = 0.54. That gives a Simpson's index of 0.54 (concentration), 1 − D = 0.46 (diversity), and 1/D ≈ 1.85 (effective number of groups out of a maximum of 3). The book quotes a diversity index of 2.7 for this kind of mix as an example. That figure does not follow from the inverse formula on these proportions, so do not copy it; state your own formula and recompute. Next, collect the index for each unit alongside its sales and advertising spend. Fit sales ~ diversity index + advertising in R and read the diversity coefficient. Predict sales for a unit with a given index and $90 of advertising. Compare the direction of your result with the Herring finding (each percentage increase in racial diversity associated with a 9% rise in sales revenue, and 3% for gender) and the McKinsey finding (a 10% increase in racial and ethnic diversity on the US senior-executive team associated with a 0.8% rise in EBIT). Your coefficient stands on your own data; the published figures are only context.

• If you report an index, then state the formula (Σp², 1 − Σp² or 1/Σp²) and the number of categories beside it, because the value is meaningless without them.
• If the question is business impact, then run a regression with controls (e.g. advertising spend) rather than quoting a bare correlation or an external benchmark.
• If the index is high but promotions, pay or attrition differ by group, then treat inclusion as unmet and investigate access, because the index captures composition only.
• If a group is underrepresented in a specific role family, then set a measurable annual improvement target for that role family, following the Johnson v. Transportation Agency plan, rather than a company-wide aggregate target.
• If the audience is executives, then lead with one outcome link (sales, EBIT or absenteeism) and one target, not the index mechanics.

• Reporting the index without naming the formula: two analysts quote 0.46 and 1.85 for the same workforce and each thinks the other is wrong.
• Copying an illustrative number: the book's 2.7 for a 70/20/10 mix does not match the inverse-Simpson result (≈1.85). The tell is an index above what the stated mix can produce.
• Treating composition as inclusion: the index rises after hiring but promotion and retention gaps persist. The tell is a good headline index with no access or outcome metric beside it.
• Reading the cited effects as guaranteed returns: the sales and EBIT figures (15x sales, +9%/+3% per percentage point, +0.8% EBIT per 10%) are study findings from other organisations. The tell is a forecast that cites them instead of your own regression.
• Averaging across the whole company and hiding unit-level gaps: a balanced total can sit on top of homogeneous teams. The tell is no breakdown by unit or by senior-executive team.

• ▸Dimension and categories: name one dimension (ethnicity, gender or age) and list its categories.
• ▸Snapshot date and unit: record the date and the scope (company, department or senior-executive team).
• ▸Headcount per category: enter the count for each category and the total.
• ▸Proportions (p): count ÷ total for each category; confirm they sum to 100%.
• ▸Σp² and chosen index form: enter Σp², then state whether you report 1 − Σp² or 1/Σp².
• ▸Inclusion check: enter one access metric by group (e.g. promotion rate) and flag any gap.
• ▸Outcome variable and control: name the business outcome (e.g. sales) and the control (e.g. advertising spend).
• ▸Regression result: record the diversity coefficient, its direction, and the predicted outcome for your scenario values.
• ▸Target and review date: state one measurable annual improvement goal for a named role family and the next re-measure date.

1. 1Define the business measure the program is meant to move before building it — pick from the passage's list: talent retention/attrition, performance, promotions, engagement, safety, production, quality, innovation, customer satisfaction, market share, sales, profitability, cycle time, or share price.
2. 2Structure evaluation across all five levels: Kirkpatrick's four (reaction, learning, behavior, results) plus Jack Phillips' fifth level, ROI, so you can convert results to a monetary return.
3. 3Establish causal linkage using causal modelling rather than asserting the training caused the outcome — isolate the training's contribution from confounding factors.
4. 4Gather data on training inputs (spend, hours, participation) and the target business outcome (e.g. sales) for enough units to run a regression.
5. 5Run Multiple Regression in R to predict training's impact on the chosen KPI (the passage's example: predict training's impact on sales); interpret the coefficients and check Adjusted R-squared for model fit.
6. 6Answer the CEO's actual question — Phillips & Phillips (2010) found training information given to CEOs does not meet their needs, so map 'what information they have' against 'what information they want' before reporting.
7. 7Tell the story: know the audience, develop the narrative, and use data visualisations that turn regression output into something the sponsor sees as impact, not as a report count.
8. 8Get sponsor approval, implement, then evaluate along the way — measure yourself on the impact made, not the number of reports produced.

Use the two L&D cases named in the passages. Hilton Worldwide University measures business impact of its training, while JetBlue University runs predictive analytics on training. Following the book's method, you would take JetBlue-style training data and run Multiple Regression with R to predict training's impact on sales. You feed the model your training inputs (for example spend or hours) alongside sales figures across stores or units, then read the coefficients to see how much each unit of training moves sales. You check Adjusted R-squared to confirm the model explains meaningful variance. If the coefficient is positive and significant, you now have the evidence to reposition the training function from a cost centre to a driver of results — the whole point of learning analytics per the passage.

• If you cannot connect the program to a critical business measure, then the training function will be seen as a cost centre — build the measurement in before you defend the budget.
• If the CEO is your audience, then report against what they said they want (Phillips & Phillips 2010), not the training data you happen to have.
• If you only measure reaction and learning, then add behavior, results (Kirkpatrick) and ROI (Phillips' fifth level) before claiming business impact.
• If you want to argue the training caused the outcome, then apply causal modelling — a raw correlation is not proof of contribution.
• If engagement/learning opportunity is low, then act on it: moving from 4-in-10 to 8-in-10 employees saying they can learn and grow is tied to a 44% drop in absenteeism and 16% jump in productivity (Gallup).

• Stopping at Kirkpatrick's Level 1 reaction ('smile sheets'). The tell: your reports show high satisfaction scores but no line to sales, retention, or any business KPI.
• Reporting activity instead of impact — counting courses delivered or reports produced. The tell: your dashboard is full of volume metrics; the passage warns you measure on impact made, not reports produced.
• Claiming causation from correlation. The tell: you show training went up and sales went up in the same period, with no causal modelling to isolate the training's contribution.
• Building the analysis without first fixing the business measure. The tell: you have a regression output but the sponsor asks 'so what does this change?' and you have no answer.
• Giving the CEO the information you have rather than what they want. The tell: leadership nods politely and the training budget still gets treated as a cost centre.

• ▸Target business measure — name one KPI from the passage list (e.g. sales, retention, absenteeism)
• ▸Baseline value — current level of that KPI before/without the program
• ▸Training inputs to test — spend, hours, or participation rate you will use as predictors
• ▸Evaluation levels covered — tick Kirkpatrick 1-4 and Phillips ROI (level 5)
• ▸Causal model note — what confounders you are isolating to prove contribution
• ▸Regression output — coefficient(s) and Adjusted R-squared from Multiple Regression in R
• ▸CEO information need — what the sponsor said they want, in their words
• ▸Narrative + visual — the one-sentence story and the chart that carries it
• ▸Approval & owner — sponsor sign-off and who implements the recommendation

1. 1Start with the business question, as the ARHAT framework's Step 1 (Asks Questions) requires. Ask which pay problem the business head cares about, such as turnover cost, lost revenue, or sales productivity. Get that sponsor's support before you touch data. Without it you risk an analytics project nobody backs.
2. 2Define the outcome you will predict. Use a binary one (left / stayed) for flight risk, or a continuous one such as sales or income for productivity. The source ties turnover to lost revenue from productivity loss and onboarding time, and to leavers taking knowledge, network and customers with them. Write that cost down as the business case.
3. 3Build one row per employee. Include salary market-ratio, performance rating, promotion history, tenure, age, gender, marital status, commute time and personality trait. These are the turnover predictors the source lists. Add triggering events (reorganization, a close colleague leaving) where you have them.
4. 4Run a correlation first to see which pay and performance variables move with the outcome. The source teaches R's correlation function for this. Then run logistic regression for the leave/stay outcome, or multiple regression for a continuous outcome such as sales.
5. 5Test pay in combination with other variables, not alone. HP found that higher pay, promotions and better ratings each lowered flight risk. But an employee promoted without a substantial raise was still more likely to quit. Check promotion-with-no-raise as its own group.
6. 6Check time-based triggers. Google found that new salespeople not promoted within four years were much more likely to leave. Look for a tenure or promotion-timing threshold in your own data.
7. 7Turn the model into a per-employee risk score. Give managers the key risk factors behind each score so they can act before the person resigns.
8. 8Restrict access to the scores. The source flags privacy problems: employees may not know their data was collected or agree with its use. Limit access to trained senior management who understand the confidentiality implications.

In 2011 two Hewlett-Packard scientists analyzed two years of data and built a Flight Risk score. It predicted the probability of leaving for each of HP's 300,000 employees. Higher pay, promotions and better performance ratings were each negatively related to flight risk. The key finding was an interaction: an employee who got a promotion but not a substantial pay raise was still more likely to quit. A pay-only model would have missed this. The system also told managers the main risk factors behind each score, which pushed them to build retention strategies. HP's privacy problem was that employees might not know their data was being used. So access should go only to trained senior managers. The same logic applies to your own data. Pair the pay variable (salary market-ratio) with promotion and performance. Look for groups whose pay lags their title or rating, and treat them as the first retention targets.

• If an employee was promoted but their pay did not rise substantially, then flag them as elevated flight risk. HP found that promotion alone does not retain.
• If pay data (market-ratio) is the only predictor in your model, then add performance rating, promotion, tenure and triggering events before acting. Compensation effects are intertwined with these.
• If new hires, especially salespeople, approach four years without a promotion, then review them proactively. Google found this group much more likely to leave.
• If a risk score would be shared widely, then restrict it to trained senior managers. Employees may not know their data is collected.
• If you cannot name the business problem and a sponsor, then do not start the pay analysis. ARHAT Step 1 comes first to avoid projects the business does not support.

• Reading pay as a standalone retention lever. The tell: you give raises or promotions in isolation and the same people still leave, as with HP's promoted-without-a-raise group.
• Skipping the business question. The tell: a finished pay model that no business head asked for and nobody funds.
• Sharing flight-risk scores broadly. The tell: employees discover they are scored without having agreed to it, and managers misread the confidential data.
• Using a model that lists no key risk factors. The tell: managers see a score but cannot say what to change, so no retention action follows.
• Claiming a pay-to-profit link without business data. The tell: the analysis stops at correlation with attrition and never connects to productivity, sales or income. The source's case list (Walmart, Deloitte, HBR) points to that linkage, but its detail is not in the supplied text.

• ▸Business question and sponsor: state the pay problem (e.g. turnover cost, productivity) and who backs it.
• ▸Outcome variable: leave/stay (logistic regression) or sales/income (multiple regression).
• ▸Pay variable(s): salary market-ratio per employee, and the source of the market data.
• ▸Control variables: performance rating, promotion history, tenure, age, gender, marital status, commute time, personality trait.
• ▸Triggering events tracked: reorganization, colleague departures, life events, if available.
• ▸Interaction checks: promotion with no substantial raise; no promotion by year four (new salespeople).
• ▸Top risk factors found: the factors the model shows for each risk tier, to give managers.
• ▸Cost of turnover: lost productivity, onboarding time, and knowledge and customers lost.
• ▸Access and privacy plan: who may see scores, what training they get, and how employees are informed.

1. 1State the hypothesis in the book's testable format before touching data: 'If {employees score higher on conscientiousness / extraversion} → {customer service performance will be higher}.' Share it with your sponsor and get agreement first. Name the role group (customer service, sales) so the claim is narrow.
2. 2Define each trait operationally from the source wording. Conscientiousness = hardworking, responsible, organized, dependable (does what is expected). Extraversion = sociable, active, talkative, high-energy with customers. Use these definitions to check that your assessment's scales actually map to them.
3. 3Start with data you already have. Check whether assessment scores exist for current staff and whether you hold an outcome for the same people (customer service rating, sales result, tenure). Collect new data only if the existing data is too fragmented, outdated, or missing. For gaps, remove the records or estimate the value with regression or an average, as the book advises.
4. 4Build one data frame, one row per employee: trait scores plus one outcome column. Save it as a CSV, load it with read.csv(..., header=TRUE), and plot a histogram of each variable with hist() to check shape and outliers before any statistics.
5. 5Run R's correlation function between each trait and each outcome. Compare top performers with the rest on those same traits. This is how the book proposes identifying the traits of top-performing customer service and sales staff.
6. 6Add context variables before you interpret a trait effect. The source passages show that service climate and perceived autonomy (Ployhart et al., 2011), engagement (ISS, correlation 0.55), and training also affect customer service. If you can, move from correlation to multiple regression so the trait effect is read alongside these factors.
7. 7Treat other companies' findings as hypotheses, not answers. The book states that what works elsewhere may not work in your company, so test the conscientiousness and extraversion claim on your own staff, role by role.
8. 8Tell the story in three parts: know the audience, build the narrative, build the visuals. Lead with the finding for the role, the trait, and the outcome. Before using scores in hiring decisions, review the court rulings on analytics cases the book lists (Griggs v. Duke Power Co., 1971; Johnson v. Transportation Agency, 1987).

The book's customer-service case (Case 2, drawing on Erik van Vulpen's AIHR article) holds that more conscientious and extraverted employees tend to give a better service experience. Conscientious people do what is expected of them, and extraverts bring high energy to customer interactions. You turn that literature claim into a local test. Hypothesis: 'If a service rep scores higher on conscientiousness and extraversion → their customer service result will be higher.' You pull assessment scores and a service outcome for the same reps from existing systems and load them into one data frame. You run hist() on each column, then R's correlation function for each trait against the outcome. The passages give no trait-level coefficients, so your own numbers are the evidence. For scale, the only coefficient the book cites in this area is ISS's 0.55 between employee engagement and customer experience. You then check context. Ployhart et al. (2011) found that stores with a higher service climate and more perceived autonomy had better customer service performance, so you note each rep's store environment before crediting the trait. If the traits correlate in your data, you report them as a screening input for service roles. If they do not, you report that the literature finding did not replicate in your business.

• If the role is customer service, test conscientiousness and extraversion first. These are the two traits the source ties to service experience. If the role is sales, run the same correlation approach against sales results, as in the Rentokil case the book lists.
• If you hold assessment scores and outcomes for the same people already, analyze them before collecting anything new. If they are missing or fragmented, fix or estimate the gaps (remove the records, or use regression or an average) before testing.
• If a trait correlation is weak or absent in your data, report it as such and do not defer to the published literature. The book says other companies' results must be tested in yours.
• If two stores or teams differ in service climate or autonomy, don't attribute the performance gap to personality alone. Add those variables to a multiple regression.
• If scores will feed hiring or selection decisions, review the court rulings on analytics cases (Griggs v. Duke Power Co.; Johnson v. Transportation Agency) before using them.

• Importing a published finding as if it were proven for your workforce. The tell: no local correlation is ever run, and the slide cites only the literature.
• Crediting personality for what climate or autonomy produces. The tell: high performers cluster in particular stores and nobody compared service climate or perceived autonomy.
• Assuming training will fix a trait gap. Ployhart et al. (2011) found restaurant service training did not raise service levels in simple service jobs, so the tell is a plan that responds to low scores with generic training.
• Trait definitions that drift from the assessment. The tell: the scale being called 'conscientiousness' does not measure being organized, responsible, and dependable, so the correlations are uninterpretable.
• Running statistics on dirty or unmatched data. The tell: unexplained blanks, duplicate IDs, or scores and outcomes from different periods that were joined without checking.

• ▸Role group: name the single job family tested (e.g., customer service or sales).
• ▸Hypothesis: write 'If {trait score is higher} → {outcome} will be higher.'
• ▸Sponsor agreement: record who agreed to the hypothesis and the date.
• ▸Trait definitions: paste the operational definition for each trait and the assessment scale mapped to it.
• ▸Outcome measure: name the single performance, service, or retention metric and its source system.
• ▸Existing data check: list what you already hold, its gaps, and how missing values are handled (remove, regression, or average).
• ▸Correlation results: record each trait-outcome coefficient exactly as R returns it, and whether it supports or fails the hypothesis.
• ▸Context variables: record service climate, perceived autonomy, engagement, and training status for the group tested.
• ▸Legal and use review: record whether the court rulings on analytics cases were reviewed before scores are used in decisions.