Predictive HR AnalyticsMastering the HR metric

1. 1Name the outcome as a measurable dependent variable before touching any model. The source's examples use customer loyalty, customer reinvestment, team productivity, sales performance and sickness absence. Write the exact metric, its unit and its source system. Predictive HR analytics is defined as modelling existing people data to inform judgements about possible causal factors driving key performance indicators, so the outcome comes first.
2. 2Audit what data you actually hold. The source lists the usable inputs: skills and qualifications, competency measures, training attended, engagement levels, customer satisfaction data, performance appraisal records, and pay/bonus data. It says the data you have, and the data you lack, determines which business questions you can answer. Mark each candidate predictor as available, missing or unusable.
3. 3Choose the level of analysis: individual or group/team. Customer outcomes such as loyalty and reinvestment usually sit at the level of a customer group, branch or team, while predictors such as engagement may be individual scores. Aggregate or match them so each row of the dataset means one thing, whether an employee, a team or a branch. Chapter 3 treats group-level versus individual-level data as an explicit decision.
4. 4Classify every variable as categorical (binary, nominal, ordinal) or continuous, and fix the roles: the business or customer outcome is the dependent variable and the employee-behaviour or attitude measures are the independent variables. This classification determines the test. The source's performance chapter uses multiple linear regression for continuous outcomes such as loyalty and reinvestment.
5. 5Prepare the data before modelling. The source stresses that getting data into the right format takes time and considerable file manipulation. Join the HR file to the customer or business file on a shared key, check data integrity, and handle missing cases explicitly.
6. 6Run the regression with the outcome as dependent variable and your HR predictors as independent variables. Read statistical significance first, then effect size, to see which predictors matter. Treat the output as evidence about possible causal factors, not proof of causation.
7. 7Turn the model into scenarios. Chapter 10 does this with a customer reinvestment example: change a predictor in the fitted model and read off the implied change in the outcome. Present this as the business case for action, with the assumptions stated.
8. 8Stress-test the metric before it goes live. Chapter 12 raises the 'analytic sample', 'the missing group' and 'the missing factor', and warns that a metric can become the behaviour driver (IMOB). Check who is absent from the data, what omitted factor could explain the result, and whether publishing the outcome as a target would distort behaviour. Pair it with other measures in a balanced scorecard.

The source's performance chapter runs three linked examples on customer outcomes in a financial services organization: multiple linear regression to predict customer loyalty (Example 1a), customer reinvestment (Example 1b), and, in Chapter 10, scenario modelling of customer reinvestment. The passages supplied give only these titles, not the coefficients, so no results are quoted here. The design is as follows. The dependent variable is a customer-level measure, loyalty in one model and reinvestment in the next. The independent variables come from the HR data types the source lists, such as engagement, competency measures, training attended and customer satisfaction data, aggregated to the unit that serves the customer. You run the same predictors against each outcome, compare which ones are significant, and then use the fitted reinvestment model for scenarios: if a predictor rises by a stated amount, what change in reinvestment does the model imply? That scenario output becomes the business case. Before presenting it, you check which branches or customers are missing from the sample and whether any unmeasured factor could be driving both the HR scores and the customer result.

• If the outcome is continuous (loyalty, reinvestment, sales, productivity), use multiple linear regression as in the source's performance examples. If it is binary or categorical, move to the categorical toolkit in Chapter 3.
• If predictors are individual-level but the outcome sits at team or branch level, aggregate the predictors to that level first. Do not mix levels in one row.
• If a needed variable is missing from your data, narrow the business question to what the data can answer. The source says available and missing data determine what can be answered.
• If a significant predictor is about to become a published target, anticipate IMOB: add a balancing metric before it is tracked.
• If the sample leaves out a group (a region, a customer segment, leavers), report the finding as applying only to the analysed sample, not the whole organization.

• Treating a regression association as proven causation. The tell: the recommendation says 'X will raise loyalty' with no mention of possible causal factors or omitted variables.
• Mismatched units of analysis. The tell: individual employee scores are matched to branch-level customer data, so rows are duplicated or arbitrary.
• Choosing the outcome because it is easy to pull, not because it reflects the business result. The tell: the metric is an HR activity count with no link to customer or financial results.
• Ignoring the missing group or missing factor. The tell: results look clean but whole teams, regions or leavers have dropped out of the file, or a strong plausible driver was never measured.
• Letting the metric drive the behaviour (IMOB). The tell: after the metric is published, scores improve while the underlying customer experience does not.

• ▸Business or customer outcome: name the exact metric (e.g. customer loyalty, reinvestment, sales, productivity, sickness absence) and its unit.
• ▸Outcome source and owner: state the system and the person who vouches for the data.
• ▸Level of analysis: individual, team or branch, and the key used to join HR and business data.
• ▸Candidate predictors available: list the HR variables you hold (engagement, competencies, training, appraisal, pay) and their variable type.
• ▸Predictors missing: list what you wish you had and how this limits the question.
• ▸Test selected and why: state the test (e.g. multiple linear regression) and its link to the variable types.
• ▸Result summary: record significant predictors and direction. Do not claim causation.
• ▸Scenario and business case: describe the predictor change tested, the implied outcome change, and the assumptions.
• ▸Sample and bias check: note who is missing from the sample, which unmeasured factor could explain the result, and which balancing metric guards against IMOB.

1. 1Name the practice and classify it: recruitment/selection, training/development, performance management or onboarding. Write one sentence stating the outcome it is meant to move, such as performance, turnover, stress or customer results. If you cannot name the outcome, you cannot evaluate the practice.
2. 2Test the case for the practice against data and research, not against what other firms do. The source cites Rousseau and Barends (2011) on moving from 'copycat practices' to advising leadership through an assessment of available data and research. Record what evidence exists before you launch or keep the practice.
3. 3Link the data sources before you analyse. HR database fields (role, tenure, performance rating, gender, leaver information), attitude survey results and operational or customer data sit in separate systems. Join them at employee or team level so the practice can be tied to an outcome, not just reported.
4. 4For selection, check reliability and validity first. Run the pool checks: are gender and undergraduate (UG) proportions consistent across the applicant pool, and do gender and UG influence shortlisting and offers? Then validate each technique by predicting performance (multiple linear regression) and turnover from the selection data.
5. 5For training, onboarding or any other intervention, take a baseline before it starts and measure again after. The book's Chapter 9 pattern is before/after on the whole group, then the same comparison split by subgroup (gender, department).
6. 6Define your analytic sample explicitly. Ask who is missing (for example leavers or non-respondents) and which factor is missing (a variable that could explain the result). Set aside time to be sceptical and interrogate the results before you report them.
7. 7Before scaling, model the business case. Use scenario modelling to estimate the potential impact of the training programme or an induction day and the return on the investment, then compare it with cost.
8. 8Guard the metric itself. Choose a balanced set of metrics so no single number becomes the behaviour driver (Institutionalized Metric-Oriented Behaviour). Check data quality and privacy, and remember that every metric is linked to human beings.

The source's Chapter 9 shows the pattern. Chapter 9 titles an intervention study 'stress before and after intervention', repeated 'by gender'. A value-change initiative is studied overall, then 'by department'. A supermarket checkout training course is analysed, then revisited ('Redux'). The passages give only these titles, not the results, so the figures below are placeholders to fill with your own data. Take a stress-reduction programme. Step 1: pull each employee's pre-programme survey stress score and their post-programme score. Step 2: test whether the overall change is statistically significant. Step 3: repeat the test by gender, because a group average can hide a subgroup that did not move. Step 4: check who is missing from the post-survey sample before drawing conclusions. Step 5: if the effect holds, use scenario modelling to estimate the benefit of extending it, as Chapter 10 does for a training programme and an induction day.

• If a practice has no data or research behind it and is only being adopted because others have it, treat it as a copycat practice and gather evidence before committing budget.
• If a selection method has not been shown to predict performance or turnover, do not treat it as validated. Run the regression, and keep it as a hard gate only once it earns that status.
• If an intervention shows an overall effect, then split by gender and by department before declaring success. The group result may not hold in every subgroup.
• If a single metric is being used to judge a practice, then add balancing metrics, because the metric will start to drive behaviour.
• If results look surprisingly strong or weak, then check the analytic sample, the missing group and the missing factor before acting.

• Copycat adoption: a practice is launched because peers use it. The tell is that no outcome or evidence is written down.
• Reporting without inference: dashboards count hires, training completions or ratings, but no statistical model links the practice to an outcome. The tell is that nobody can say what the practice changed.
• Missing group: only people still present (for example survey respondents who stayed) are analysed, which flatters the intervention. The tell is that the sample size is smaller than the cohort.
• Averaging away bias or non-effect: a whole-population result is accepted without a gender or department split. The tell is that the same number is quoted for everyone.
• Metric gaming: the tracked number becomes the target and behaviour bends toward it. The tell is that the metric improves while related outcomes stay flat.

• ▸Practice and type: name it and mark selection, training, performance management or onboarding.
• ▸Target outcome: one sentence on the outcome it should move (performance, turnover, stress, customer result).
• ▸Evidence for adopting it: the data or research that supports it, not 'peers do it'.
• ▸Data sources linked: list the HR database, survey and operational sources and the join key.
• ▸Baseline measure and date: the pre-intervention value and when it was taken.
• ▸Post measure and test used: the follow-up value and the statistical test applied.
• ▸Subgroup splits checked: gender, department or others, with the result for each.
• ▸Analytic sample, missing group, missing factor: who is included, who is absent, and what variable might be omitted.
• ▸Balancing metrics and ethics/privacy notes: the metrics paired with the main one, and data protection checks.
• ▸Business case and decision: the modelled impact and return, then scale, adjust or stop.

1. 1Build the data as one row per employee and one column per characteristic. Each cell holds one value, for example 42, female, performance rating 5. If the data set is team-level, label it as team-level before you start, because individual and team analyses answer different questions.
2. 2Name each characteristic as its own variable (age, gender, tenure, ethnicity, education, personality score). Give each a unique name, and set its attributes in the variable view (SPSS) or the equivalent. This avoids duplicate and rejected names later.
3. 3Classify each characteristic by measurement level: binary (gender coded as two groups), nominal (ethnicity, education type), ordinal (job grade), or interval/ratio (age, tenure in months, a personality scale score). The level decides which tests are allowed, so do it before any analysis.
4. 4Write the question and assign roles before touching software. Decide which variable is the dependent variable (DV: turnover, performance, shortlisting, engagement) and which characteristics are independent variables (IVs). Then state the question in one line, e.g. 'Does the gender mix differ across job grades?'
5. 5Choose the test from the DV type and the IV type. Categorical DV with one categorical IV: chi-square on a crosstab. Categorical DV with several predictors: logistic regression. Time to an event: survival analysis. Continuous DV with a two-group IV: independent samples t-test. Same group at two points: paired samples t-test. Continuous DV with several predictors: multiple linear regression.
6. 6For continuous DVs, check the distribution before reading results. Parametric tests assume roughly normal data. If the data is skewed or too peaked or flat, switch to a test that is robust to violations of normality.
7. 7Run the descriptive cut first (frequency tables, means by group), then the inferential test, then the predictive model. Dashboards typically stop at descriptive frequencies such as percent male vs female by role. The value of this construct lies in testing whether a characteristic predicts or is linked to an outcome.
8. 8Where diversity categories may combine, test them together by interacting them in the model, for example gender and ethnicity, rather than reading each one alone. Before circulating findings on age, gender or ethnicity and performance, apply an ethical review of what the pattern can and cannot justify.

The source's case studies show how one characteristic maps to a test. Take 'gender and job grade'. Each employee is a row, with gender coded binary and job grade coded ordinal, so both variables are categorical. With a categorical DV and a single categorical IV, the source's test table points to chi-square on a crosstab of frequencies. The output shows whether the proportion of men and women varies across grades. Next, move to turnover, where the source uses Kaplan-Meier survival analysis to compare expected length of service for men versus women. The DV is the time between start date and end date, so survival analysis is the right tool rather than a t-test on tenure. For a continuous outcome such as a supermarket checkout scan rate, the source links employee demographic data to performance. It also includes a worked analysis on whether high-performing age groups exist. The passages give the structure of these analyses but not their numeric results, so any figures must come from your own data.

• If the DV is categorical and you have one categorical characteristic (e.g. gender by job grade), then run chi-square on a crosstab.
• If the DV is binary (stay/leave, shortlisted/not) and you have several characteristics plus other predictors, then use logistic regression to estimate likelihood.
• If the question is how long someone stays, or the chance of surviving to 3, 6 or 12 months, then use survival analysis, not a comparison of average tenure.
• If the DV is continuous and the characteristic splits two groups, then use an independent samples t-test. If it is the same group measured at two times, use a paired samples t-test.
• If a continuous DV is clearly skewed, too peaked or too flat, then use a test robust to normality violations before drawing conclusions about the wider population.

• Stopping at dashboard-style descriptives. The tell: your outputs are percentages by gender, age band or function, with no test, no model and no predicted outcome.
• Choosing the test from habit rather than from data type. The tell: you cannot say whether your IV and DV are binary, nominal, ordinal or interval before running the analysis.
• Mixing units of analysis. The tell: individual-level characteristics (one row per employee) are blended with team-level data (one row per team), and the conclusions slide between 'people' and 'teams'.
• Ignoring distribution assumptions on continuous outcomes. The tell: t-tests and regressions are reported without any check of normality, skew or peakedness.
• Reporting demographic-performance patterns without an ethical caveat. The tell: a finding such as an age group scoring higher goes into a deck with no discussion of fairness, bias or what action is justified.

• ▸Characteristic: name the attribute (e.g. gender, tenure, ethnicity, education, age, personality trait).
• ▸Variable name: give a unique, software-safe column name.
• ▸Measurement level: binary, nominal, ordinal, interval or ratio.
• ▸Role in model: DV or IV, with the outcome it relates to (turnover, performance, engagement, selection).
• ▸Unit of analysis: one row per employee, or team-level; state which.
• ▸Research question: one sentence, e.g. 'Does X differ across Y?'
• ▸Selected test: chi-square, logistic regression, survival analysis, independent/paired t-test, or multiple regression, and why it fits the DV/IV types.
• ▸Assumption check: for continuous DVs, note the normality result and any robust alternative used.
• ▸Ethical and interpretation note: what the result can and cannot justify, and who reviews it before action.

1. 1Define the construct as a set of measurable survey items, not a vague mood. The passages list the attitude-survey content that captures it: job-strain, engagement, job satisfaction, person-organization fit and perceptions of justice. Add inclusion questions, which Gartner frames as a work environment where all employees feel respected, accepted, supported and valued. Write down which items stand for leadership, integrity/support, and justice/fairness before you collect anything.
2. 2Decide who owns the instrument. If an external provider runs it (the passages name Willis Towers Watson, Glint, Culture Amp, Medallia, Qualtrics), ask what they measure and how they build the index. If you run it in-house, you carry that design work yourself. Either way, get the item-level export (often Excel from the survey program) rather than only the headline index.
3. 3Check that items belong together before you aggregate them. Run factor analysis to see whether items load onto the intended constructs, then reliability analysis on each scale. The passages cover both, including a four-item engagement scale. Only after that build the index, in the same way organizations build an 'engagement index' or 'inclusion index'.
4. 4Link the perception data to other sources. Join survey results to employee profile data (age, gender, role, tenure, country, team leader, etc.), and to customer satisfaction and operational data where you can. A perception score alone shows how people feel; the linked data shows where and with what consequence.
5. 5Cut the data by group first. Compare groups with the independent samples t-test (two groups) or one-way ANOVA (more than two). Examples from the passages: differences between functions, teams or countries. Look for pockets rather than organization-wide averages.
6. 6Track change over time on the same people. For two time points use the paired samples t-test. For three or more use the one-way repeated measures ANOVA (SPSS: Analyze > General Linear Model > Repeated Measures). If you need to compare how change differs between groups, use the within-and-between ANOVA (general linear model).
7. 7Look for drivers with multiple regression, for example to predict team-level engagement from perception measures. Treat the output as 'potential predictors', not proven causes. The passages point to Chapter 12 for the causality caution.
8. 8Read the free-text survey comments alongside the numbers, then tie findings to an action and the business context. Target the intervention to the broken part. The passages' image is not putting a plaster cast on the whole foot when only the middle toe is broken.

The passages describe an HR director whose function has gone through a fundamentally changing work environment over an extended period and who wants to track stress levels every six months over a two-year period. Because the same employees are measured at three or more points, the right test is the one-way repeated measures ANOVA, not a series of separate t-tests. In SPSS you choose Analyze, then General Linear Model, then Repeated Measures. The dependent variable is continuous (a stress or job-strain score) and the within-subject factor is the time point. A significant result tells you the mean changed across waves; it does not tell you why. To find out whether, say, one team or country drove the change, you move to the within-and-between ANOVA, which adds a group factor. You then read the survey comments and link the finding to other data. The passages give a parallel linking example: if customer data shows ethnic minority clients in branches outside major cities are the most dissatisfied, that points toward unconscious bias or diversity awareness training in those branches rather than a company-wide programme.

• If the same employees or teams are measured at two time points, then use the paired samples t-test; if at three or more, use the one-way repeated measures ANOVA.
• If you want to know whether change differs between groups (e.g. teams, functions), then use the within-and-between ANOVA (general linear model) rather than pooling everyone.
• If you are about to report an index (engagement, inclusion, justice), then run factor and reliability analysis first; if items do not hang together, do not aggregate them.
• If an overall perception score looks acceptable but linked data (customer, operational, demographic) is available, then cut by group and location before acting, because the problem may sit in one segment.
• If a regression identifies 'drivers' of perceptions or engagement, then call them potential predictors and check the causality caveat before committing budget to an intervention.

• Treating a single survey snapshot as the answer. The tell: no repeated measures and no comparison across time points, even though the environment is changing (e.g. a merger or restructure).
• Using an external provider's index as a black box. The tell: you cannot say which items feed it or whether they were tested with factor or reliability analysis.
• Staying inside HR data. The tell: perception scores are never linked to customer, sales or operational data, so no one can show a link to revenue or efficiency.
• Reporting only organization-wide averages. The tell: a flat headline score hides a dissatisfied group, such as one location or demographic, and the response is a blanket programme.
• Reading predictors as proven causes. The tell: findings are written as 'X causes lower justice perceptions' on correlational regression output alone.

• ▸Perception items in scope: list the survey items for leadership quality, organizational integrity/support, justice/fairness, and inclusion (e.g. job-strain, job satisfaction, person-organization fit, perceptions of justice).
• ▸Scale check: record factor analysis result and reliability result for each scale before building any index.
• ▸Data sources linked: list employee profile, customer, sales and operational data joined to the survey file.
• ▸Grouping variables: name the splits to compare (team, function, country, location, demographic group).
• ▸Time design: state the number of waves and whether the same participants appear in each (decides paired t-test vs repeated measures ANOVA vs within-and-between ANOVA).
• ▸Test chosen and why: independent t-test, one-way ANOVA, repeated measures, or regression, with one line of justification.
• ▸Finding and affected segment: state the result and exactly which group or location it concerns.
• ▸Qualitative check: summarize what the survey comments say about the same segment.
• ▸Targeted action and caveat: name the narrow intervention and note that predictors are potential, not proven, causes.

1. 1Pick one written definition before you touch data. The source's anchor is Schaufeli and Bakker (2003): a state of mind characterized by vigour, dedication and absorption, measured with the UWES. The consultancy definitions (Culture Amp, Mercer, Hay, Willis Towers Watson) mix in enthusiasm, discretionary effort, commitment and an enabling environment. If your survey items follow a different definition from the one you report against, every later analysis inherits the mismatch.
2. 2Separate engagement from the neighbouring constructs and record each as its own variable. Job satisfaction has emotional, cognitive and behavioural components (Locke, 1976). Job strain is a psychological response to high demands combined with perceived low control (House and Rizzo, 1972). Person-organization fit and perceived organizational support (POS) are further distinct constructs. Do not collapse them into one 'wellbeing' score before you test whether they behave as one.
3. 3Interrogate the measure before you analyse outcomes. Run factor analysis on the item-level responses to test whether the engagement scale measures one construct or several. If the items split into two sets, either the scale was built from two sub-constructs or some badly designed items are measuring something other than engagement.
4. 4If you find two factors, name them from the item content. For example, mental concentration items versus organizational commitment items. Then decide whether to report two sub-scores or fix or drop the stray items. Do not average them into one number.
5. 5Run reliability analysis on each retained scale or sub-scale (the source's example is a four-item engagement scale) before you build composites. Only then compute the composite score for each person, on the same 1-to-5 metric as the items.
6. 6For group-level data (teams, units), run Principal Components Analysis on the group-level engagement data rather than assuming the individual-level structure carries over. Then compare groups with the independent samples t-test (for example, between two populations) to see where engagement differs.
7. 7Model engagement as a mediator between conditions and outcomes. Use multiple regression to predict team-level engagement from work-condition variables. The source's HR data set holds POS, distributive justice, procedural justice, job satisfaction and value fit (all 1–5 composites), plus gender and tenure. Use multiple regression to predict an outcome such as sickness absence (the source's Example 6) from the strain and satisfaction measures.
8. 8Read the free-text survey comments alongside the numbers (survey comments analysis). Then tie the findings to actions and business context. A statistic with no owner and no decision attached does not change anything.

The source illustrates a scale that looks like one thing but is two. Suppose engagement is defined as a combination of 'organizational engagement' and 'psychological job engagement'. The designer writes six items on mental focus and concentration, and three items on organizational commitment. All nine are reported as one 'engagement' score. Step one is factor analysis on the nine item responses. If the answers on the six concentration items and the three commitment items pattern differently, the analysis returns two factors rather than one. That tells you the scale is measuring two sub-constructs, or two slightly different forms of engagement. You then name the factors from their item content, check reliability on each set separately, and report two scores. A single 'engagement' average would hide the fact that people can be absorbed in their job while feeling little commitment to the organization, or the reverse. Only after this step do you compare groups with a t-test or regress team engagement on its predictors. The source also shows that a two-factor result can mean badly designed items measuring something other than engagement, so you inspect the item wording before deciding which explanation applies.

• If factor analysis returns one factor, then build a single composite after checking reliability. If it returns two, then name each factor from its item content and report the sub-scores separately.
• If an item loads away from the rest of its intended scale, then read its wording. Rewrite or drop it if it is badly designed; keep it as its own sub-construct only if the definition supports that.
• If the question is about teams rather than individuals, then use group-level data (PCA, team-level regression). Do not generalise individual-level results to teams.
• If you compare two groups on one engagement score, then use the independent samples t-test. If you need drivers of engagement or an outcome such as sickness absence, then use multiple regression.
• If a stakeholder asks for 'engagement' without a definition, then fix the definition first (vigour, dedication, absorption, or commitment and discretionary effort). Items that measure a different definition will not answer their question.

• Treating a multi-item scale as one construct without testing it. The tell: a composite score that moves unpredictably, or sub-groups of items that clearly answer differently.
• Reporting engagement when the items actually measure satisfaction or commitment. The tell: the survey wording does not match any stated definition such as vigour, dedication and absorption.
• Skipping reliability analysis and building composites straight from item averages. The tell: no reliability result is documented for the scale.
• Merging engagement, job satisfaction, job strain and fit into one wellbeing number. The tell: you cannot say which component changed when the score moved, and strain effects disappear into the average.
• Relying on numbers alone and ignoring careless responding and open comments. The tell: implausible response patterns go unchecked (Meade and Craig, 2012 is the source's reference on this), and recommendations have no link to the comment themes.

• ▸Working definition: write the single definition you report against (for example, vigour, dedication and absorption).
• ▸Scale and items: list every item and its response range (the source's composites use 1 to 5).
• ▸Intended sub-constructs: state how many constructs the designer meant to measure and which items belong to each.
• ▸Factor analysis result: record the number of factors found and which items group together.
• ▸Item problems: list items that load away from their intended factor and your decision (rewrite, drop, keep separate).
• ▸Reliability by scale: record the reliability result for each retained scale or sub-scale.
• ▸Adjacent constructs tracked: note whether job satisfaction, job strain, POS, fit and justice are held as separate variables.
• ▸Level of analysis: individual or team, and the test chosen (t-test, PCA, multiple regression) with the reason.
• ▸Outcome and action: name the outcome modelled (for example, sickness absence), the finding from the numbers and comments, and the decision owner.

1. 1Name the behavior as an observable action, not an attitude. Guest's warning in the passages is that engagement gets muddled as 'an attitude, a behaviour or an outcome'. Write each target as a recorded act: leaving, days absent, a customer loyalty score tied to a person, a checkout scan rate. Keep attitudes (survey engagement) as separate predictor variables.
2. 2Pick the behavior family. The source works with three: performance (customer loyalty, customer reinvestment, individual performance, scan rates), withdrawal (sickness absence, turnover), and change in performance over time. Choose one as the outcome of the analysis.
3. 3Fix the unit of analysis before you pull data. Turnover can be measured at individual level (did this person leave?) or team level (turnover rate for the team). The unit decides the test: individual leaver/stayer data is categorical, team rates are continuous.
4. 4Link the behavior record to other sources, which is the 'lay of the land' step. Join HR records to L&D records, client profile data, survey results and demographics. A behavior metric alone, such as customer satisfaction, gives no prescriptive action. Linked data does.
5. 5Describe before you predict. For turnover, start with frequency tables by region, then chi-square on individual leavers by region. For team-level turnover by country, use one-way ANOVA. This is the day-to-day descriptive layer.
6. 6Match the model to the behavior's form. Use multiple linear regression for continuous outcomes (performance, customer loyalty, sickness absence). Use a predictive model for individual leave/stay. Use Kaplan-Meier survival analysis for expected length of service, for example men vs women. Use stepwise regression to model performance or change in performance over time.
7. 7Test whether the model works the same way for diversity groups. Treat gender, ethnicity and age as predictors or interaction terms, and check for high-performing groups (for example age groups) only with the ethical caveat the source attaches to performance data.
8. 8Close with the 'So what?' test. State what a named manager changes in the departmental plan, daily tasks or team-leader management. For turnover, add the cost model and business case for action.

The source's own illustration starts from customer satisfaction survey results about customer-facing staff. Alone, they only justify a guidance paper to all staff if results are poor. Link them to L&D records and you might find that staff who attended a relationship management course in the previous quarter obtain the best customer service results. That makes the behavior (customer service outcomes) predictable from an action you control (course attendance). Link the same data to client profile information and you might find that ethnic minority clients in branches outside major cities are the most dissatisfied. That points to unconscious bias or diversity awareness training, a different intervention from the course. The same behavior metric yields different prescriptions depending on what you join to it. The source's point is to diagnose before prescribing, so you don't 'put a plaster cast on the entire foot if all that is broken is the middle toe'. For the 'So what?' test, you would present which training or branch-level action to change, not just a significant coefficient.

• If the outcome is leave/stay for individuals and you are comparing categories such as region, then use frequency tables and chi-square. If the unit is team turnover by country, use one-way ANOVA.
• If the question is how long people stay, or whether one group (men vs women) differs in expected length of service, then use Kaplan-Meier survival analysis rather than a simple leaver count.
• If the outcome is continuous (customer loyalty, individual performance, sickness absence), then use multiple linear regression. Use stepwise to select among many candidate predictors, and model change over time as its own outcome.
• If a model shows a significant driver but nobody can say what changes on Monday, then it is not finished. Add the specific action, owner and cost or business case before presenting, because the room will ask 'So what?'.
• If you want to test diversity effects on a behavior, then enter the diversity category as a predictor and check interactions to see whether the model works differently by group. Apply the ethical caveat to performance data.

• Treating engagement as a behavior without defining it. The tell: the same word covers attitude scores, discretionary effort and outcomes in one model, and nobody agrees what it measures.
• Analysing at the wrong level. The tell: you apply individual-level tests to team turnover rates, or you report a team average and make claims about individuals.
• Reporting a behavior metric in isolation. The tell: the recommendation is a generic guidance paper to all staff, because nothing was linked to L&D, client or demographic data.
• Stopping at statistical significance. The tell: the presentation ends in silence or 'So what?' because no one has a change to the plan, tasks or team-leader management.
• Using performance data for demographic comparisons without the ethical check. The tell: age or other group patterns are reported with no discussion of bias or fairness.

• ▸Behavior (observable action): State it as a recorded act, such as leaving, days of sickness absence or scan rate, not an attitude.
• ▸Behavior family: Performance, withdrawal (sickness absence or turnover), or change over time.
• ▸Unit of analysis: Individual or team. This decides the test.
• ▸Data sources to link: List the HR, L&D, client profile, survey and demographic sources and the join key.
• ▸Descriptive step: Frequency tables, chi-square (categorical) or one-way ANOVA (team-level continuous), and the grouping variable.
• ▸Predictive model: Multiple linear regression, stepwise, individual turnover prediction or Kaplan-Meier, with predictors listed.
• ▸Diversity check: Which groups to test as predictors or interactions, and the ethical caveat that applies.
• ▸'So what?' action: Who changes what in their plan, tasks or team management.
• ▸Cost / business case: Cost of the behavior (for example turnover) versus the cost of the proposed action.