1. 1Start with a thorough literature review, then pose your research question(s) explicitly; match the question form to the method (case study fits 'how' and 'why' questions about contemporary events you cannot manipulate).
2. 2Classify your intended case study as explanatory/causal, descriptive, or exploratory, because the type dictates what theory or framing you need before collecting data.
3. 3Develop theoretical propositions before data collection — the second half of Yin's definition states case study 'benefits from the prior development of theoretical propositions to guide data collection and analysis.' Write these propositions down so they specify which variables of interest matter (there will be more variables than data points).
4. 4Decide your epistemological orientation openly: a realist orientation assumes a single observer-independent reality, while a relativist/interpretivist orientation assumes multiple observer-dependent realities; state which you hold because it changes how you frame 'theory' in the design.
5. 5Use the propositions to define the unit of analysis (the 'case') and to build the link between data and propositions — this is a formal design component, not an afterthought.
6. 6Frame the study so it aims for analytic generalization (generalizing to theoretical propositions) rather than statistical generalization (extrapolating to a population); build in replication logic if using multiple cases.
7. 7Specify criteria for interpreting findings in advance, and identify plausible rival explanations to test throughout design, collection, analysis, and interpretation.

Consider the exploratory 'what' question the passages name: 'What can be learned from a study of a startup business?' Because this is exploratory, its justifiable purpose is to develop pertinent hypotheses and propositions for further inquiry rather than to test settled ones — so the theory here is thin and generative by design. If instead you asked a 'how' or 'why' question about the same startup and you cannot manipulate its behavior and the events are contemporary, a case study becomes the preferred method. You would then, per Yin's twofold definition, develop prior theoretical propositions to guide which of the many variables of interest to observe, and rely on multiple converging sources of evidence. The framing choice — exploratory versus explanatory — determines whether theory precedes data (explanatory) or emerges from it (exploratory). The Lipset, Trow, and Coleman (1956) single case study is cited as aiming for a 'generalizing' rather than a 'particularizing' analysis, illustrating that even one case is framed toward theoretical propositions.

• If your question is 'how' or 'why' about a contemporary event you cannot control, then use a case study framed with prior theoretical propositions.
• If your question is an exploratory 'what' (e.g., 'what can be learned from...'), then treat theory as generative — build hypotheses and propositions as output, not input.
• If you assume a realist orientation, then frame theory as capturing a single observer-independent reality; if relativist/interpretivist, then frame theory to accommodate multiple observer-dependent meanings.
• If you want to generalize, then aim for analytic (theoretical) generalization via replication logic, never statistical generalization to a population.
• If you or your department are predisposed to a method, then rewrite the study question to honestly match that method rather than force a mismatch.

• Confusing case study with a data-collection tactic (ethnography, participant-observation) and skipping proposition development — the tell is a study with rich fieldwork but no stated theory guiding what data matter.
• Treating the case as a 'sample' and trying to generalize to a population — the tell is language about representativeness and probabilities instead of theoretical propositions.
• Collecting data before framing propositions, so more variables of interest than data points overwhelm the analysis — the tell is a massive, unreadable document with no interpretive spine.
• Mismatching question type and method (e.g., forcing a 'why' question into a survey) — the tell is a gross misfit where the method cannot address the question's form.
• Leaving epistemological orientation implicit, so 'theory' is used inconsistently — the tell is realist assumptions in design colliding with relativist claims in interpretation.

• ▸Research question (write it as who/what/how/why): state the exact question
• ▸Question type: mark exploratory / descriptive / explanatory-causal
• ▸Case study type justification: why this type fits the question
• ▸Epistemological orientation: realist or relativist/interpretivist, and why
• ▸Prior theoretical propositions: list each proposition guiding data collection
• ▸Unit of analysis (the case): define precisely and its boundaries
• ▸Variables of interest vs. data points: name key variables and evidence sources
• ▸Generalization aim: analytic (to which propositions) not statistical
• ▸Rival explanations to test: list substantive rivals to examine throughout

1. 1Adopt the researcher code of ethics as a non-negotiable baseline before fieldwork: integrity and humility, protection of subjects, privacy, no distortion of findings, distance from respondents, acknowledgment of collaborators, and transparency of financial support. Treat these as accreditation-level obligations, not options.
2. 2Anonymize your instruments at design stage: build interview schedules and questionnaires so respondent names and identities can be safeguarded, separating identifiers from response data so information disclosed cannot later be traced back to harm a person's career or social position.
3. 3Screen your funding: make any financial support public knowledge in the report, and reject unethical support outright. Decide in advance that no sponsor will be permitted to shape or distort findings.
4. 4Set consent and privacy boundaries before contact: do not force anyone to answer, do not probe deep into inner lives, and for any study involving children, secure guardian or parental awareness and agreement first. Never threaten a respondent to extract information.
5. 5Manage the interaction distance deliberately: stay engaged enough to gather rich data but not so enmeshed that you 'go native' and internalize respondents' worldview, which would make validity and reliability impossible to test.
6. 6Report honestly: present data and findings fully and without distortion, do not cook data, and do not lure respondents with money or promises of immediate benefit, since inducement distorts what they tell you.
7. 7Expose your work to other scholars for intellectual scrutiny to minimize your own subjectivity, since topic choice and interpretation are shaped by temperament, cultural milieu, and political orientation.
8. 8Acknowledge every contributor and agency in the final report, from initial stages to completion, since successful research is not one person's job and credit should be shared.

Suppose you are a male researcher studying why families break down, interviewing female respondents. The ethical risk here is failure of distance: if you sympathize too heavily, a respondent may begin to see you as a 'substitute' for her husband, which contaminates the interaction and the data. You manage this by striking a social balance — engaged enough to gather full accounts, aloof enough to keep validity testable. You anonymize her responses on the schedule so nothing she discloses about the marriage can be traced back and used to jeopardize her position. You do not probe deeper into her private life than the questions require, and you do not offer money or promised benefits. In the report you present her account as given, without distortion, and you acknowledge those who helped arrange access.

• If the study involves children, then do not proceed until guardians or parents are aware and agree.
• If you find yourself internalizing respondents' views, fears, and aspirations, then you have gone native — pull back, because the data's reliability is now questionable.
• If a sponsor or offered funding is unethical or would distort findings, then reject it and make all accepted support public knowledge.
• If you are tempted to offer money or immediate benefits to respondents, then stop — inducement distorts findings; gather information in a free, natural environment of interaction.
• If you carry prejudicial stereotypes about the community you study (for example, an Eastern or Central Province researcher working in Nyanza), then expose your interpretation to other scholars to check for the 'idols of the tribe.'

• Cooking data — presenting distorted or fabricated findings. Tell: results conveniently match the researcher's political persuasion or the sponsor's interest rather than what respondents said.
• Going native — becoming fully meshed with respondents. Tell: the researcher describes the group's worldview as their own and can no longer test the validity of what was disclosed.
• Letting a sponsor shape conclusions. Tell: financial support is hidden from the report and findings soften around the funder's stake.
• Breaching privacy — probing inner lives, forcing answers, or threatening respondents. Tell: respondents seem coerced, and identifiers on schedules are left unprotected so information can reach a respondent's enemies.
• Misrepresenting your own competence — no integrity or humility. Tell: the researcher claims certainty beyond their limits and is unwilling to learn from respondents, unlike Newton, who felt like a boy picking pebbles by the shore.

• ▸Consent status: For children/vulnerable groups, guardian awareness and agreement obtained? (yes/no)
• ▸Anonymization: How are respondent identities separated and safeguarded on instruments?
• ▸Privacy boundary: Which questions are off-limits to avoid probing inner lives or forcing answers?
• ▸Distance plan: How will you stay engaged yet avoid going native with this population?
• ▸No-inducement declaration: Confirm no money or promised benefit offered to respondents.
• ▸Funding disclosure: Name all financial support to be made public; flag any unethical source to reject.
• ▸Honest-reporting pledge: Confirm findings will be presented fully, undistorted, sponsor-proof.
• ▸Scrutiny plan: Which scholars will review your interpretation to reduce subjectivity/stereotype bias?
• ▸Acknowledgments log: List every individual and agency who contributed, from start to final report.

1. 1Consolidate people data into a single employee view: extract, transform, load, and convert data from payroll, HRMS, attendance legacy systems, and the LMS into one repository so leadership stops reconciling across silos in Excel.
2. 2Link people data to business data (revenue, cost, profit per FTE) so decisions align with organizational outcomes rather than staying operational metrics.
3. 3Retire outdated operational metrics (FTE count, cost of rewards, time to fill, raw attrition) and select the critical few lead and lag indicators that connect to shareholder value.
4. 4Build predictive models on structured and unstructured data — e.g., attrition early-warning systems — that name likely leavers, their persona traits, and the key predictors that, if addressed, reduce churn.
5. 5Convert skeptics with evidence: demonstrate the predictive power with data (turnover trends, underlying reasons, likely leavers) rather than asserting it.
6. 6Run periodic follow-up: track which predictions actually came true, and attach an explicit ROI to every people practice and investment.
7. 7Present through visual intelligence, not Excel dumps: crisp insight representation crashes decision cycle time and wins fast CEO buy-in given limited attention spans.

Take talent retention, which organizations treat as a key outcome of People Analytics investment. You feed both structured and unstructured data through data-mining tools to build a predictive attrition model that flags, with reasonable accuracy, which individuals are likely to exit, the persona traits of likely leavers, and the key predictors driving the trend. Because replacing top talent in senior and critical roles is time-consuming and expensive, you install early-warning triggers linked to engagement and attrition signals. You then act on the addressable predictors to cut churn in the short and medium term. Critically, you follow up periodically to confirm which of these exit predictions came true, and you quantify the ROI of the interventions — this validation is what converts the gut-feel skeptic and proves the model earns its investment.

• If a metric is operational (time to fill, FTE count) and does not link to shareholder value or business outcomes, then drop it in favor of a critical-few indicator that does.
• If a leader relies on gut feel or corporate belief systems, then use analytics not to override them but to validate or refute the intuition with relevant data and sharp insight.
• If you have made a prediction (e.g., who will leave), then schedule a periodic follow-up to check which predictions came true before claiming credibility.
• If people data and business data sit in non-interacting systems, then prioritize a single-version-of-the-truth consolidation before attempting correlation analysis.
• If you cannot articulate the ROI of a people practice, then treat the investment case as incomplete regardless of how good the insight feels.

• Stopping at prediction without follow-up — the tell: no record of which forecasts actually came true, so credibility never accumulates.
• Leaving data in silos — the tell: analysts still spend most of their time reconciling payroll, HRMS, and attendance data in Excel, and by the time they finish the underlying data has already changed.
• Reporting the trivial many operational metrics — the tell: dashboards full of FTE counts and time-to-fill that distract and don't connect to business deliverables.
• Presenting insight as dense spreadsheets — the tell: CEOs disengage because the representation doesn't fit their limited attention span and buy-in stalls.
• Ignoring the politics of transparency — the tell: business leaders quietly resist because hard evidence exposes decisions they preferred to keep on gut feel.

• ▸Decision in question: state the people decision (hire, exit, promote, reward, engage).
• ▸Prior basis: name the current gut/habit assumption you are testing.
• ▸Data sources to unify: list the source systems (payroll, HRMS, attendance, LMS) feeding the single employee view.
• ▸Business data linkage: identify the outcome metric (profit/revenue per FTE, churn cost) you connect to.
• ▸Critical-few indicators: the lead and lag measures that drive this outcome.
• ▸Prediction + trigger: state the predictive output (e.g., likely leavers) and the action trigger.
• ▸Follow-up date: when you will check which predictions came true.
• ▸ROI statement: the expected return and how you will measure it.
• ▸Visual for buy-in: the one crisp insight visual you will show the CEO.

1. 1Decompose every survey item against the four-component model before fielding: comprehension, retrieval, judgment, and reporting (mapping the judgment to the response format). Ask which components a given question loads most heavily.
2. 2Audit comprehension at three levels the book separates: syntax (word order, tangled constructions), semantics (vague, unfamiliar, or ambiguous terms and false presuppositions), and pragmatics (what the respondent infers the question is really for).
3. 3For factual/behavioral items, trace the retrieval path: identify whether the target event is recallable directly, must be reconstructed from autobiographical memory organization, or will be estimated. Expect frequency questions to trigger estimation, not enumeration.
4. 4For attitude items, apply the belief-sampling model: assume the respondent samples whatever considerations are currently accessible rather than reading off a fixed attitude, so control what earlier items make accessible (context effects).
5. 5Design the reporting stage deliberately: choose open items versus rating scales versus unordered categories knowing each induces its own artifacts (rounding on open numerics, anchor effects on scales, satisficing on unordered lists).
6. 6Add an editing check for sensitive topics: anticipate that respondents edit true answers before reporting them, and select a data-collection mode whose privacy properties reduce that editing.
7. 7Budget for speed: respondents answer typical attitude questions in under 5 seconds, so assume shortcuts and partial overlap of stages rather than methodical deliberation, and pretest for the shortcut respondents will actually take.

Take the NCS item: 'Was anything stolen from you while you were away from home, for instance at work, in a theater or restaurant, or while traveling?' Comprehension is loaded here: the 'for instance' list is meant to jog scope, but it can be read as restricting theft to those three settings (a semantic/pragmatic effect). Retrieval requires scanning autobiographical memory across an unbounded reference period, so minor thefts are under-recalled and reconstruction dominates. Judgment forces the respondent to decide whether a borderline event 'counts' as stolen. Reporting is a yes/no, and if the topic feels sensitive or self-incriminating the respondent may edit toward 'no.' The cognitive model tells you the biggest risk is not the response scale but the retrieval and scope-comprehension stages, so a fix targets the examples and reference period, not the answer format.

• If the question asks about opinions, then assume belief sampling and control prior-item context; if it asks about counts, then assume estimation and provide reference periods and anchors.
• If a term is vague, unfamiliar, or carries a presupposition, then define it inline before the respondent reaches the period, because comprehension often closes at the end of the sentence.
• If reported behavior is sensitive, then move toward a more private administration mode, since misreporting is an editing/reporting-stage failure, not a memory failure.
• If an open numeric answer clusters on round values, then treat it as rounding at the reporting stage and read precision claims cautiously.
• If earlier items shift answers to later ones, then diagnose it as a context effect at retrieval/judgment and either reorder, buffer, or randomize.

• Blaming a bad rating scale when the real defect is comprehension. Tell: rewording the scale changes little while rewording the question stem moves responses.
• Assuming attitudes are stable and pre-formed. Tell: answers swing with question order and immediately preceding topics.
• Treating frequency answers as counted rather than estimated. Tell: heaping on round numbers and sensitivity to the offered response ranges.
• Ignoring the editing stage on sensitive topics and attributing low prevalence to genuine behavior. Tell: prevalence rises when the same question is asked in a more private mode.
• Assuming respondents move through comprehend-retrieve-judge-report as clean sequential stages. Tell: sub-5-second answers and evidence that retrieval begins before comprehension finishes.

• ▸Item text: paste the exact question as respondents will see it.
• ▸Question type: mark factual/behavioral or attitude.
• ▸Comprehension risks: list syntax, semantic (vague/unfamiliar/presupposed), and pragmatic issues.
• ▸Retrieval demand: state reference period and whether recall, reconstruction, or estimation is expected.
• ▸Judgment load: note any borderline 'does it count' or accessibility-driven sampling.
• ▸Reporting format: record open/scale/categorical and the artifact it invites (rounding/anchor/satisficing).
• ▸Sensitivity/editing: flag whether respondents may edit the true answer and the mode chosen to reduce it.
• ▸Context exposure: name preceding items that could shift this answer.
• ▸Fix applied: record the single change targeting the heaviest-loaded component.

1. 1Separate the two questions science answers: what generates new theories (imagination, speculation) versus what eliminates old ones (pitiless comparison of theory to evidence). Anchor your practice on the second — the capacity to remove options permanently, not to invent them.
2. 2Adopt the iron rule: in professional argument, admit only observable empirical evidence. Set aside philosophical, theological, aesthetic, moral, and cultural considerations when adjudicating between theories — not because they are worthless, but because banning them forces convergence.
3. 3Exploit even minute discrepancies between theory and evidence. Treat small mismatches as decisive, not as noise to be waved away, because science's critical power comes from unbending pursuit of these gaps.
4. 4Engineer motivation, not just logic. Method tells you what to do with data, but is useless without plenty of the right data; build in a driver (competitive refutation appetite or paradigm-squeezing zeal) that induces people to endure mundane, largely negative labor.
5. 5Compartmentalize deliberately: cultivate obliviousness to broader intellectual and spiritual matters during the work, so nothing pulls you off the empirical path. The norm's whole function is to suppress the impact of your being in the world on your knowing of the world.
6. 6Accept that objectivity lives at the institutional level, not in each individual interpretation. Expect gut reactions, local bargains, and after-the-fact rationalization in the lab; rely on the shared iron rule and competition to grind out convergence over the long term.
7. 7Sustain intense, prolonged focus on details of little intrinsic interest — the 99 percent that is boring and frustrating — because the 1 percent of dramatic results depends entirely on the mass of tedious data behind it.

Through the 1960s the endocrinologists Roger Guillemin and Andrew Schally raced to determine the molecular structure of TRH, a hypothalamic hormone. This was not a flash of genius but an epic slog: literally tons of sheep and pig brain tissue had to be mashed and processed to yield just 1 milligram of TRH for analysis. Several rivals dropped out, unable to countenance the immense amount of hard, dull, costly, repetitive work. As Schally put it, the key factor was not money but the will. The two finished in a dead heat and shared the 1977 Nobel Prize. The case shows the institution at work: competition supplied the motivation, and single-minded, almost inhuman focus on mundane data — not clever theorizing — produced the knowledge.

• If a claim rests on beauty, philosophical elegance, or spiritual fit rather than observable evidence, then it is inadmissible in the professional argument — set it aside.
• If theory and evidence disagree even slightly, then treat the discrepancy as significant and pursue it rather than explaining it away.
• If a promising research path demands only speculation and debate, then distrust it; if it demands prolonged production of experimental data, then it is likely the productive one.
• If your team lacks a motivational driver (refutation appetite, competitive race, or paradigm-squeezing) then expect data production to stall regardless of how sound the logic is.
• If you catch yourself broadening the work toward humanist or metaphysical meaning during the inquiry, then narrow back — obliviousness protects the empirical path.

• Defining science by one surface feature — 'controlled experiment' or 'advanced math' — which excludes whole fields; the tell is that cosmology, evolutionary biology, or classical genetics fall outside your definition.
• Treating natural philosophy as science because it too was methodical, evidence-attentive, and creative; the tell is that competing hypotheses (water, air, fire, apeiron) never get eliminated and knowledge doesn't accumulate.
• Expecting individual scientists to weigh evidence by shared objective rules; the tell is Latour and Woolgar's finding — 'local tacit negotiations, constantly changing evaluations,' no explicit appeal to norms — mistaken for a failure of science rather than its normal texture.
• Concluding that because the lab looks chaotic and rationalized after the fact, there is no method and science is no more truth-directed than anything else; the tell is Latour-style skepticism that ignores institution-level convergence.
• Believing science would flourish if scientists cared more about the rest of existence; the tell is Snow/Wilson-style calls to join the cultures, which would actually pull practitioners off the iron rule's empirical path.

• ▸Theory under test: state the specific hypothesis being adjudicated
• ▸Admissible evidence: list only the observable data bearing on it
• ▸Inadmissible influences flagged: name any beauty/philosophy/theology/politics creeping in to be excluded
• ▸Discrepancies found: record every theory-evidence mismatch, however small
• ▸Data-production burden: describe the tedious labor required and whether it's being sustained
• ▸Motivation driver: identify what keeps practitioners at the mundane work (refutation, competition, paradigm)
• ▸Compartmentalization check: confirm broader intellectual/spiritual concerns are set aside during the work
• ▸Convergence status: note whether rival theories are being eliminated or merely multiplied

1. 1Classify motivations by their goals, not by the instigating drives or the observable behavior — Maslow proves goals are the only stable centering point, since one act carries many motivations and one need reaches consciousness through many cultural paths.
2. 2Locate the person on the prepotency hierarchy by asking which need is currently unsatisfied and dominant: physiological, then safety, then love/belonging, then esteem, then the cognitive desire to know and understand. The dominant need is the one whose absence the person feels 'keenly, as never before.'
3. 3Test whether apparent behavior is a channel for a different underlying need. Ask whether the person who 'thinks he is hungry' is actually seeking comfort or dependence, and whether the stated goal is a means rather than an ultimate end.
4. 4Read intensity as a signal of deprivation: when a need dominates, capacities, memory, intelligence and philosophy all reorganize around satisfying it. Gauge intensity by how completely other interests have been 'pushed into the background.'
5. 5Separate motivation from the other determinants of behavior — biological, cultural, and situational. Interpret the field in terms of the organism, never substitute field description for motivation.
6. 6Distinguish goal-directed behavior from expressive behavior. Do not read the bass voice, the springy walk, or rote conventional acts as motivated; expressiveness and goal-directedness coexist in average behavior but are not identical.
7. 7Once a lower need is fairly gratified, expect the next need to 'emerge' and the whole striving cycle to repeat with the new need at its center — plan for satisfaction to reveal wants rather than end them, because man is a perpetually wanting animal.

Take Maslow's own case: 'the person who thinks he is hungry may actually be seeking more for comfort, or dependence, than for vitamins or proteins.' You do not classify this by the eating behavior (that is only a channel) nor by a somatic hunger drive (that would be atypical), but by the ultimate goal. You test the surface claim by checking whether food actually satisfies him or whether comfort-substitutes do the work. Contrast this with the genuinely and 'dangerously hungry man,' for whom 'no other interests exist but food' — he dreams, remembers, and emotes only about food, and defines Utopia as a place with plenty of food. Only in that extreme, rare emergency case may you speak of 'pure hunger drive.' In the normal peaceful society such emergencies are rare, because culture itself is an adaptive tool whose function is to make physiological emergencies come less often.

• If the person is missing everything in an extreme fashion, then treat the physiological need as dominant — it is the most prepotent, and all other needs become non-existent or pushed into the background.
• If both physiological and safety needs are fairly well gratified, then expect love, affection, and belongingness needs to emerge as the new center of striving.
• If an act appears to serve one drive, then assume it is multidetermined — sexual behavior, for example, is driven not only by sex but chiefly by love and affection needs.
• If you are tempted to build a list of fundamental needs, then stop — lists get you nowhere, and physiological needs alone 'can come to almost any number' depending on the specificity of description.
• If behavior is rote, habitual, automatized, conventional, or stimulus-bound, then do not assume it expresses character structure or is motivated at all.

• Centering the theory on hunger or any somatic, localizable drive. The tell: your model treats the atypical, isolable drive as the typical case, when hunger, sex, and thirst are the exceptions, not the pattern.
• Carrying the rat-picture over to humans. The tell: you assume motivations are almost entirely physiological because rats have few others and most research used rats — a 'pseudo-simplicity' fallacy.
• Classifying by instigating drive or observable behavior instead of goal. The tell: the same act gets sorted many ways, because typically an act has more than one motivation and there are various cultural paths to one goal.
• Treating the hierarchy as rigid and fixed. The tell: you are surprised when needs are pursued at great cost to safety, or when the order reverses — Maslow explicitly says it is not nearly as rigid as implied.
• Assuming satisfaction ends wanting. The tell: you expect the fed or loved person to be content, forgetting that a hungry man who once sneered at love will later hunger for it 'more than anything else' — gratification only surfaces the next need.

• ▸Observed behavior — describe the act without interpreting its cause
• ▸Stated/conscious goal — what the person says they want
• ▸Ultimate/underlying goal — the basic, often unconscious end the act may channel
• ▸Dominant unsatisfied need — mark one: physiological / safety / love-belonging / esteem / know-understand
• ▸Intensity evidence — note which other interests have been pushed into the background
• ▸Multidetermination check — list other needs this same act may simultaneously satisfy
• ▸Non-motivational determinants — biological, cultural, situational factors also shaping the act
• ▸Expressive vs goal-directed — is any of this style/character rather than striving?
• ▸Next-need forecast — if the dominant need is gratified, which need will emerge as the new center?

1. 1Measure before you manage. Redeem your Q12 access code at q12.gallup.com/breaktherules and administer the survey to your team; the 12 items measure what matters most to employees and tie directly to productivity, profitability, retention and turnover.
2. 2Respect the order of the 12 items. There is a deliberate progression (Chapter 1's 'Mountain Climbing' establishes why the items are ranked), so address foundational needs before higher ones rather than jumping to development questions with employees whose basic expectations are unmet.
3. 3Identify each person's talents, not just their skills. Have people complete the Clifton StrengthsFinder 2.0 to surface their top five themes across the three categories — striving (the why), thinking (the how), and relating (the who) — so you deploy them where their four-lane highways are, not their wastelands.
4. 4Build roles around strengths. People who focus on their strengths every day are six times as likely to be engaged. Assign work that matches the recurring behaviors talent produces, and stop trying to recarve character that neuroscience says is fixed past the mid-teens.
5. 5Manage around weaknesses instead of fixing them. Turn a harmful weakness into an irrelevant nontalent by redesigning the role, partnering the person with a complementary talent, or shifting the task — don't invest in remediation that yields only a 'thin path.'
6. 6Spend the most time with your best people. Great managers 'play favorites' — study your top performers to learn the whys, hows, and whos behind their excellence, then select and coach for those same talents.
7. 7Find the right fit rather than promoting up a ladder. Avoid promoting people to their level of incompetence; create heroes in every role so a person can grow in place instead of being pushed into work that mismatches their talents.
8. 8Re-measure to close the loop. Use Q12 results to track engagement over time, aiming toward the elite ratio of nine engaged employees for every one actively disengaged.

The passages contrast best, average, and worst performers in the same role to show engagement is about talent-fit, not effort. Take nurses: contrary to the belief that professionals should stay detached, the best nurses form strong emotional relationships with patients and use those emotions to take control and smooth the patient's world. The worst are overwhelmed by the same emotions, and the average protect themselves by keeping their distance — they are 'emotionally disengaged.' A manager who fixates on teaching detachment would be coaching everyone toward the average, disengaged state. Instead, the great manager studies why the best nurse succeeds (a relating talent for empathy channeled into control), then selects and positions for that talent. The same logic appears with salespeople: the best salesperson still feels call reluctance daily, but a talent for confrontation and love of sparring pulls them through, while the average just woodenly follows a six-step script.

• If someone excels at a task, study the whys/hows/whos behind it and select for those same talents — don't assume the behavior can simply be trained into others.
• If a weakness is harmful, redesign the role to make it an irrelevant nontalent; if it can only be improved to a 'thin path,' don't invest heavily in fixing it.
• If a person focuses on their strengths every day, expect roughly six times the engagement likelihood — route work toward strengths accordingly.
• If a promotion would move someone away from the talents that made them excellent, create a hero role in place instead of promoting them to incompetence.
• If basic Q12 expectations are unmet, address foundational items before attempting higher-order development conversations, because the 12 items have a required order.

• Trying to fix weaknesses instead of managing around them — the tell is a development plan dominated by remediation of what someone struggles with rather than deployment of what they do best.
• Applying the Golden Rule uniformly — treating every employee the same the way average waiters give 'every table the same droning spiel'; the tell is standardized management with no tailoring to individual talent.
• Coaching toward detachment or 'professional distance' — the tell is your average performers looking disengaged (like the average nurse who keeps distance) while your best quietly break the rule.
• Promoting your best performers out of the role they excel in — the tell is a strong individual contributor faltering after a title bump, i.e., promoted to their level of incompetence.
• Skipping measurement or measuring once and stopping — the tell is managing engagement by anecdote; without repeated Q12 data you can't link actions to productivity, profitability, and turnover.

• ▸Employee & role: name the person and the exact role performed.
• ▸Top five talent themes: list from Clifton StrengthsFinder 2.0.
• ▸Striving/thinking/relating notes: capture the why, how, and who this person naturally shows.
• ▸Best-performer study: what do your top people in this role do differently (whys/hows/whos)?
• ▸Strength-deployment plan: which daily tasks let this person use strengths (target: strengths used every day).
• ▸Weakness to manage around: one harmful weakness + the redesign/partner/reassign move to make it a nontalent.
• ▸Right-fit check: is growth-in-place a hero role, or is this a promotion toward incompetence?
• ▸Q12 baseline & re-measure date: current engagement reading and next survey date.

1. 1Measure job satisfaction with psychometrically sound instruments (Smith, Kendall, & Hulin, 1969 JDI-style facet measures), not ad hoc single items, so you capture satisfaction with distinct workplace conditions—pay, supervision, shift work, performance reviews, utilization of talent.
2. 2Adopt Hulin's 'standard research design': use a prospective (predictive) design, measure satisfaction at Time 1, then track actual voluntary quits at a later point—not involuntary terminations, which contaminated Weitz & Nuchols (1955).
3. 3Assess quit intentions/withdrawal cognitions alongside satisfaction, since Mobley positioned intentions as the most proximal and strongest turnover antecedent; where quit data are unavailable, use intentions as the surrogate.
4. 4Map the workplace antecedents that drive satisfaction (integration, pay, leader consideration, utilization) rather than treating satisfaction as a root cause—Price's models show satisfaction mediates between environmental drivers and turnover.
5. 5Move from static to dynamic measurement: collect satisfaction repeatedly and model its trajectory (change over time), because Liu et al. (2012) raised explained turnover variance from 5% to 43% by using change scores instead of one-time snapshots.
6. 6Extend the frame to nonwork/wellbeing fit—assess work-family conflict and family embeddedness, since Mobley et al. (1979) flagged nonwork values as moderators of the satisfaction-turnover link.
7. 7Analyze with methods suited to binary outcomes: use logistic regression, survival/hazard functions, or SEM rather than OLS regression, which mis-specifies dichotomous turnover outcomes.

Hulin (1968) evaluated a company program initiated in 1964 designed to improve working conditions. Using a quasi-experimental design, he measured job satisfaction before and after the intervention and tracked subsequent turnover. He concluded that the program 'brought about an increase in job satisfaction' and that 'this increase led to a reduction in turnover in 1966.' This is the operational template: change a concrete workplace condition, measure the satisfaction shift with a sound instrument, and observe voluntary quits at a later date. It demonstrates that satisfaction is not merely correlated with staying—an engineered increase in satisfaction produced a downstream drop in leaving. It also shows why the prospective, individual-level design matters: the time lag lets you attribute the turnover reduction to the earlier satisfaction gain.

• If your criterion includes involuntary terminations, then your dissatisfaction-turnover relationship is contaminated—restrict the outcome to voluntary quits (the Weitz & Nuchols flaw Hulin corrected).
• If quit data are unavailable, then use turnover intentions as a proxy, given their documented predictive superiority as the most proximal antecedent.
• If satisfaction is high but employees still leave, then check subjective expected utility of alternatives—people leave good jobs when they expect higher utility elsewhere (Mobley et al., 1979).
• If you only have one-time satisfaction scores, then expect low explained variance; switch to trajectory/change measures to capture the declining-commitment-to-rising-intentions dynamic (Bentein et al., 2005).
• If nonwork obligations dominate (childcare, family relocation), then weight work-family conflict and family embeddedness rather than assuming job attitudes alone drive the decision.

• Treating satisfaction and its environmental drivers as isomorphic—Weitz & Nuchols warned they are not; the tell is a model that predicts turnover from attitudes but gives managers no leverable action.
• Inflating the satisfaction-turnover relationship by creating equal-sized high/low-tenure subsamples (artifactual 50% quit rate, as in Minor, 1958); the tell is a quit base rate that doesn't match the real workforce.
• Failing to cross-validate, which caused early reviews to overstate validity (Schwab & Oliver, 1974); the tell is impressive coefficients that shrink on a holdout sample.
• Using OLS regression on the binary quit outcome; the tell is non-normal residuals—switch to logistic/survival/SEM.
• Relying on static snapshots and missing declining trajectories; the tell is unexplained turnover variance that dynamic change measures would have captured.

• ▸Satisfaction instrument: name the psychometrically validated facet measure used (e.g., JDI-style).
• ▸Time 1 satisfaction scores: record baseline facet scores (pay, supervision, shift work, utilization).
• ▸Change trajectory: record repeated-measure scores and the direction/slope of change.
• ▸Quit intentions: capture withdrawal cognitions as proximal predictor/proxy.
• ▸Environmental drivers: list workplace antecedents (integration, pay, leader consideration).
• ▸Nonwork/wellbeing factors: note work-family conflict and family embeddedness pressures.
• ▸Outcome definition: confirm criterion is voluntary quits only, with time lag stated.
• ▸Analysis method: specify logistic/survival/SEM (not OLS) and whether cross-validated.

1. 1Define what successful individual performance actually means before you measure it — the finance house executives who had 'no clear criteria of what successful performance in their business, at an individual level, actually meant' show the failure of skipping this step.
2. 2Distinguish inputs from outputs explicitly: assessment methods (psychometrics, interviews, assessment centres) measure inputs to performance, while behaviour and results are outputs 'necessarily not directly measured' by those methods. Map both sides.
3. 3Build a competency model through formal job analysis so task, contextual, and citizenship behaviours each have observable criteria — the assessment centre example interprets responses 'according to a model of competency developed by a formal job analysis'.
4. 4Use multiple methods and multiple assessors on situations of relevance to the role, rather than relying on the intuitive 'warm smile, educational background, manner of speech' clues everyone uses unconsciously.
5. 5Anchor each observed behaviour to a named competency and mark it as positive or negative evidence — as the assessor does when scripting 'sets standards for others'.
6. 6Close the loop between measurement and management: use appraisals, feedback, and day-to-day controls (Chapter 13/14) so assessment predictions convert into 'sustained success' per the Armstrong and Baron definition.
7. 7Treat objectivity as the governing constraint — systematic processes over intuition — to keep the link between individual criteria and organizational outcomes defensible.

A professional actor in an assessment centre exercise makes a scripted remark: 'I know they have to say that to cover themselves, but it's really just window dressing. If we paid attention to their rule book every time we would never get anything finished.' The participant being assessed challenges the actor and reiterates the standard expected regarding breaks for people working at screens. The assessor writes down the participant's exact response verbatim during the exercise rather than judging on impression. Later, reviewing the script against the competency model, the assessor marks that response as positive evidence of the competency 'sets standards for others.' This is individual performance captured operationally: a role-relevant situation, an observable behaviour, and interpretation tied to a formally derived competency rather than a general feeling about the candidate.

• If you cannot state what successful individual performance means in the role, then stop and run a job analysis before assessing — undefined criteria correlate with organizational decline (the finance house losing 75 per cent market value).
• If a method measures an input (e.g. personality, verbal reasoning), then do not treat its score as a direct measure of output behaviour — predict, then verify against actual performance.
• If you are relying on day-to-day intuitive clues (smile, punctuality, speech), then flag it as unsystematic and require corroborating systematic evidence.
• If behavioural evidence does not map to a named competency, then it is not scorable — discard it or refine the model.
• If assessment findings never feed appraisal, feedback, or reward, then you have selection without performance management — connect the two.

• Assessing before defining success: teams that 'had no clear criteria' of individual performance measure activity, not proficiency; the tell is scores with no agreed standard behind them.
• Confusing inputs with outputs: treating a test result as performance itself; the tell is nobody can say what behaviour the score predicts.
• Separating methods instead of joining them — the author warns his own earlier books risked 'contributing to separation rather than integration'; the tell is psychometrics, interviews and centres run in isolation with no shared competency model.
• Substituting intuition for discipline: leaning on the 'warm smile' and 'educational background'; the tell is rater confidence high, documentation absent.
• Measuring but never managing: assessment stops at selection with no appraisal or feedback loop; the tell is predictions made and never checked against realized performance.

• ▸Role/level: name the job the criteria apply to.
• ▸Success criteria: state in one line what successful performance at the individual level means here.
• ▸Competencies (task/contextual/citizenship): list the named behaviours drawn from formal job analysis.
• ▸Input measures used: note which psychometrics, interviews, or exercises feed each competency.
• ▸Output behaviours observed: record the actual behaviour or verbatim response, not an impression.
• ▸Evidence mapping: mark each observation positive/negative against a named competency.
• ▸Assessors: list the multiple assessors and which exercises each covered.
• ▸Management link: state how findings feed appraisal, feedback, or reward.
• ▸Review date: when predictions get checked against realized performance.

1. 1Map the triad before diagnosing any behavior: name the Person (dispositions, virtues, vulnerabilities brought in), the Situation (the immediate behavioral setting and its force field), and the System (the political, economic, historic, cultural matrix that creates and legitimizes the situation). Do not stop at the person.
2. 2Audit the situational force field for the specific compliance drivers named in the SPE: assigned roles, deindividuating uniforms, obedience to authority, group chants/norms, self-justification, and the passivity of bystanders. List which are active.
3. 3Trace command complicity upward. When abuse appears at the bottom, ask who built the 'bad barrel' — locate the 'absentee landlord' leadership and the chain of command that set rules, secrecy, and permissions.
4. 4Replace the medical model with a public-health approach: instead of removing or 'curing' the bad individual, modify or teach people to avoid the situational forces producing the behavior.
5. 5Identify the dehumanization step explicitly, because it is the central process: watch for the 'Other' being constructed as enemy, savage, or less-than-human, which then licenses torment.
6. 6Inoculate against unwanted influence by teaching the mental and social tactics of resistance: label the tactic being used (authority, group dynamics, persuasive appeal, compliance strategy), affirm you can be vulnerable, and act as an informed, critical agent.
7. 7Build the exit: reframe the person not as a passive target but as a potential 'hero of the moment' who can act decisively when the call sounds, and reward that resistance rather than treating it as deviance.

Zimbardo served as expert witness for Sergeant Ivan 'Chip' Frederick, an MP who worked the Tier 1A night shift from 4 P.M. to 4 A.M. every night for forty nights without a break. Rather than accept the 'bad apple' account, he investigated the whole barrel: interviews with Frederick, his family, his corrections and military background, and every civilian and military investigating-committee report, which he had to memorize because notes were barred from the trial. Those independent reports laid blame on senior officers and 'absentee landlord' leadership — a system that had built a 'bad barrel.' Yet the prosecutor and judge refused to consider that situational forces could shape individual behavior, holding the standard dispositional view that the fault was Frederick's freely chosen rational decision. Frederick received a harsh sentence, and Zimbardo, doubly frustrated, became informal counselor to Chip and his wife Martha, who came to call him 'Uncle Phil.' The case demonstrates both the power of situation and system over compliant behavior and the cultural default that resists that explanation.

• If abuse is blamed entirely on a 'few rogue soldiers,' then examine the situation and the chain of command before accepting the dispositional verdict.
• If someone shows 'mindless compliance' — obeying absurd orders, joining chants — then treat it as evidence of situational power, not evidence of inherent bad character.
• If the target group is being described as enemy, savage, or subhuman, then dehumanization is underway and abuse is being licensed; intervene at that framing.
• If you want durable behavior change, then modify the situational forces or teach avoidance rather than only replacing the individual.
• If you can name the influence tactic (authority, conformity pressure, persuasive appeal), then you have already begun to resist it — labeling breaks the automatic pull.

• Defaulting to dispositional explanation ('knowing better but doing worse is a free choice'). Tell: the analysis stops at the individual and never names situation or system.
• Ignoring the System behind the veil of secrecy. Tell: you fix the 'bad apples' but the same abuses recur because the 'bad barrel' and its rules stay intact.
• The evil of inaction — Zimbardo's own sin of omission, failing to intervene when abuse was extreme. Tell: you contain only the worst physical harm while tolerating humiliation you should have stopped.
• Treating resistance as rare heroic magic rather than learnable tactics. Tell: you admire rebels as a 'special breed' and never teach ordinary people the strategies of resistance.
• Missing dehumanization because it feels like normal us-vs-them talk. Tell: the 'cortical cataract' clouds judgment and the Other's suffering stops registering as real.

• ▸Behavior in question: describe the specific compliant or harmful act observed
• ▸Person factors: strengths, virtues, and vulnerabilities the individual brought in
• ▸Situation factors: which force-field drivers are active (roles, uniforms/deindividuation, authority, group norms/chants, self-justification)
• ▸Dehumanization check: is the target framed as enemy/savage/less-than-human? yes/no + evidence
• ▸System factors: rules, secrecy, and chain-of-command leadership that created and legitimized the situation
• ▸Command complicity: who built the barrel and what did 'absentee landlord' leadership permit
• ▸Intervention (public-health): which situational force to modify or teach avoidance of
• ▸Resistance plan: name the influence tactic and the counter-tactic to deploy
• ▸Hero-of-the-moment trigger: what specific call to action would prompt decisive resistance

1. 1Set the four interlocked quantities before collecting data: significance level (alpha), desired power (target .80), expected effect size, and sample size. Fix three, solve for the fourth. In this text, Section 5.15 treats power analysis as the a priori determination of sample size, and Section 13.8 gives sample size for power = .80 in the single-sample repeated-measures case.
2. 2When testing a true effect, name the two errors explicitly: Type I (rejecting a true null) controlled by alpha, and Type II (failing to detect a real effect) whose complement is power. Chapter 1.2 ties Type I error, Type II error, and power together as one framework, so never quote alpha without also stating the power your n supports.
3. 3Guard subject-to-variable ratios so the solution is stable. For a reliable prediction equation, consult Section 3.16 (sample size determination for regression) before interpreting; for factor analysis, apply Section 11.7 on sample size and reliable factors so retained components are not artifacts of too few subjects.
4. 4Improve power directly rather than only raising n: Section 4.10 lists ways of improving power. Reduce error variance (better measurement, homogeneous samples), increase effect size (stronger manipulation), or relax alpha where justified, then re-estimate.
5. 5For multivariate designs, decide how many dependent variables to include (Section 5.14) because adding correlated DVs dilutes power. Estimate multivariate power (Sections 4.11-4.12) using SPSS MANOVA's power option rather than assuming univariate power carries over.
6. 6Control sampling error inflation from multiple tests: Section 1.3 warns that multiple statistical tests raise the probability of spurious results. Budget alpha across the family of tests so the experiment-wise Type I rate stays near nominal.
7. 7Protect the independence assumption before trusting any power figure: Chapter 6.3 and the Chapter 15 note state that even a small violation of independence pushes the actual alpha several times above the nominal level, which corrupts both your Type I control and your power calculation.
8. 8Separate statistical from practical significance (Section 1.4): a large n can make a trivial effect significant, so report effect size alongside power so an adequately powered study is not mistaken for a meaningful one.

Suppose you plan a two-group MANOVA comparing a treatment and control group on several outcomes, following Chapter 4. Before running the study you go to Section 4.9 (Power Analysis) and fix alpha at .05 and target power at .80. You estimate the effect size from a prior study in a good journal within the last 5 years, per the exercise design the author uses for the winter MANOVA quarter. Using multivariate estimation of power (Section 4.12) you find that with only 15 per group your projected power is well under .80, so detecting the true effect is unlikely. You then apply Section 4.10 (Ways of Improving Power): you tighten measurement to cut error variance and drop one weak dependent variable using the guidance in Section 5.14. Re-estimating power on SPSS MANOVA (Section 4.11) with the reduced error variance shows n = 25 per group reaches power = .80, so you commit to that sample before collecting any data.

• If you are choosing sample size and can specify alpha, effect size, and desired power, then solve for n a priori (Section 5.15) rather than justifying n after the fact.
• If observations are correlated (repeated or nested), then do not use general-linear-model power figures; a small independence violation multiplies actual alpha (Ch. 15), so switch to the hierarchical/repeated-measures framework.
• If projected power falls below .80, then first attempt Section 4.10 remedies (reduce error variance, strengthen effect, cut weak DVs) before simply raising n.
• If you run many statistical tests, then adjust for the inflated probability of spurious results (Section 1.3) and recompute power under the corrected alpha.
• If a result is statistically significant with a large sample, then check the effect size against practical significance (Section 1.4) before claiming the effect matters.

• Reporting alpha but never power: the tell is a Methods section that states significance level and n with no a priori power or effect-size justification (contradicts Section 1.2's paired treatment).
• Ignoring correlated observations and quoting nominal alpha: the tell is nested or repeated data analyzed as independent — actual alpha is several times nominal (Ch. 15), so the 'true effect' detection rate is unknown.
• Overloading a MANOVA with too many dependent variables: the tell is a long DV list with no rationale; power erodes and Section 5.14's question 'how many DVs?' was never asked.
• Confusing significance with importance: the tell is a huge n producing p < .05 for a trivial difference, with no effect size reported (Section 1.4).
• Undersized subject-to-variable ratios in regression or factor analysis: the tell is unstable coefficients or factors that fail to replicate because Sections 3.16 and 11.7 guidelines were skipped.

• ▸Analysis type: name it (MANOVA, regression, repeated measures) and cite the matching chapter section.
• ▸Alpha level: state nominal significance level and any multiple-test adjustment (Section 1.3).
• ▸Target power: set desired power (default .80 per Sections 5.15/13.8).
• ▸Expected effect size: give the value and its source (prior study/journal article).
• ▸Subjects per group / total n: enter planned n and the subject-to-variable ratio.
• ▸Number of dependent/predictor variables: list them and justify each (Section 5.14 / 3.16 / 11.7).
• ▸Independence check: confirm observations are independent or note the correlated-data method used (Ch. 6.3 / 15).
• ▸Power-improvement actions: record error-variance reductions or design changes applied (Section 4.10).
• ▸Practical significance note: state the smallest effect worth detecting (Section 1.4).

1. 1Classify your outcome variable before touching any model: continuous, binary, nominal (unordered categories), or ordered categories. This single choice governs the entire model family — linear, binomial logistic, multinomial logistic, or proportional odds logistic.
2. 2For continuous outcomes, run linear regression: fit the model, read coefficients as change in outcome per unit change in input, then check goodness-of-fit via R-squared and the coefficient confidence intervals/p-values.
3. 3For binary outcomes, use binomial logistic regression: model the log odds, then convert coefficients to odds ratios by exponentiation, and separate odds from probability when reporting to stakeholders.
4. 4For nominal outcomes with more than two unordered categories, run multinomial logistic regression: define a reference level explicitly, interpret each category's coefficients relative to that reference, and re-run with a changed reference if a different comparison is more meaningful.
5. 5For ordered category outcomes, use proportional odds logistic regression under the latent-continuous-variable assumption — do not treat ordered categories as either nominal or continuous.
6. 6Simplify the model deliberately: check relevance of each input variable, eliminate variables gradually and safely (checking fit at each removal) to reach the most parsimonious model that still fits.
7. 7Test the assumptions specific to your chosen model. For linear regression: linearity/additivity, constant error variance, normally distributed errors, and absence of high collinearity/multicollinearity between inputs.
8. 8Because people data sets are often small and the goal is usually to understand why rather than to predict, prioritize interpretability and inference over predictive complexity, and confirm sample size can support the chosen method.

Suppose you model a customer's choice among Product A, Product B and Product C — a nominal outcome with three unordered categories. You could first approach it as stratified binomial models, modeling the choice of Product A versus all other products, then modeling the other choices separately. But the cleaner method is a single multinomial logistic regression: you define a reference level (say Product C), run the model, and interpret each coefficient as the effect on the log odds of choosing Product A or B rather than Product C. If stakeholders care more about the A-versus-B contrast, you change the reference level and re-run. You then simplify through gradual safe elimination of variables and assess model fit and goodness-of-fit before reporting. This shows how outcome structure (nominal, three categories) forces the model family, and how reference choice shapes interpretation.

• If the outcome is continuous, then use linear regression and report coefficients as unit-change effects with R-squared for fit.
• If the outcome is binary, then use binomial logistic regression and report exponentiated coefficients as odds ratios, keeping odds distinct from probability.
• If the outcome is unordered categorical with 3+ levels, then use multinomial logistic regression with an explicitly chosen reference level.
• If the outcome is ordered categorical, then use proportional odds logistic regression under the proportional-odds/latent-variable assumption rather than forcing linear or multinomial.
• If the employee population is small or the aim is explanation rather than prediction, then favor a parsimonious, interpretable regression over a complex predictive model.

• Running multivariate models without any grasp of their underlying mathematics — the tell is being unable to defend or explain your method under critique or answer questions about coefficient meaning.
• Treating ordered categories as nominal or as continuous — the tell is a multinomial or linear model applied where proportional odds logistic regression fits the ordinal structure.
• Confusing odds with probability when interpreting logistic coefficients — the tell is reporting an odds ratio as if it were a probability to stakeholders.
• Ignoring linear regression assumptions (non-constant error variance, non-normal errors, high multicollinearity) — the tell is unstable or uninterpretable coefficients and misleading fit.
• Overfitting or retaining irrelevant inputs instead of gradual safe elimination — the tell is a bloated model with no improvement in goodness-of-fit and inputs that don't meaningfully contribute.

• ▸Outcome type: mark continuous / binary / nominal(3+) / ordered.
• ▸Chosen model family: derive directly from outcome type above.
• ▸Reference level (if categorical): state which level and why.
• ▸Key input variables: list candidates and note expected relevance.
• ▸Assumption checklist: for linear — linearity, constant variance, normal errors, collinearity.
• ▸Coefficient interpretation plan: state units (raw, odds ratio, log odds) before running.
• ▸Goodness-of-fit metric: name the fit measure you will report.
• ▸Simplification step: record variables eliminated and fit change at each removal.
• ▸Goal check: confirm whether aim is explanation or prediction and that sample size supports the method.

1. 1Run a data audit against four measurable dimensions before any analysis: accuracy across all systems, completeness in a timely manner, consistency and adherence to business rules, and availability with appropriate security. Fix the source of entry, not just the output: move away from HR-only data entry and get employees to enter their own data, because data are cleanest when entered at their source.
2. 2Assign dedicated team members to manage data quality — their standing job is to check completeness, accuracy, currency and non-corruption by cross-checking each field against other systems until you have one source of truth.
3. 3Set accuracy thresholds by data type rather than demanding 100% everywhere. Hold headcount to the highest accuracy because it underpins many downstream measures and causes disputes; accept directional accuracy for softer fields like leaver reason.
4. 4Integrate the three data domains you will need: core HR data (personnel, pay history, positions, disciplinary actions), employee attribute data, and non-HR business data pulled from existing systems — Finance (revenue, profits, costs), Marketing (customer satisfaction, retention), Operations (defect rates, returned goods).
5. 5Choose infrastructure on total cost of ownership over four years, not sticker price: compare capital, operating and indirect costs of cloud versus on-premise, and decide whether SaaS (provider manages hardware, software, availability and security) fits your accessibility needs — data reachable from any device with a browser.
6. 6Stand up the six analytics technology components in sequence so data flows cleanly: HRIS as the core employee database, an HR data warehouse to consolidate, reporting, statistical/machine-learning, visualisation and cognitive technology. Anchor everything on a master-data and cloud-centric infrastructure layer.
7. 7Redefine processes and governance around the data pipeline so that accuracy is a maintained state, not a one-time cleanup: schedule regular audits of all data systems so data are not questioned when used for analytics.

An engineering group operating three businesses — Healthcare, Life Sciences and Performance Materials — built its People Analytics function on a foundation of master data and cloud-centric infrastructure integrating Finance, group-wide sales and non-sales data, business strategy and HR metrics. The PA team refused to start with the data; they began with the business hypothesis, broke it into sub-questions, and only then went into the data, because they knew from experience it is easy to get lost in the data quickly. To keep the data trustworthy, dedicated team members managed data quality, checking completeness, accuracy and currency and cross-checking against other systems, while shifting entry toward employees themselves since data are cleanest at source. With that integrated, accurate base, they tested drivers of innovation and found diversity was a trigger, with a 'sweet spot' for diversity. That insight — surprising for an innovation-driven engineering firm where everyone held their own views — became a strong input to their people strategy.

• If a field underpins other measures or could trigger disputes (e.g. headcount), then hold it to the highest accuracy; if it only needs to show a directional trend (e.g. leaver reason at 20% vs 23%), then accept lower accuracy and move on.
• If data are entered by HR alone, then redesign the process so employees enter their own data at source, because that produces the cleanest records.
• If four-year total cost of ownership favours cloud (in the cited Hurwitz study, ~$697,656 cloud vs ~$1,400,570 on-premise over four years), then default to cloud/SaaS unless a specific constraint overrides it.
• If an impact measure relates HR to business results, then look first in existing Finance, Marketing or Operations systems — most are already available somewhere in the organisation, so build integration rather than new collection.
• If a data field is questioned or contradicted when used for analytics, then stop and run a cross-system audit before proceeding — you do not yet have one source of truth.

• Jumping straight into the data to show analytics prowess. The tell: the team is mining with sophisticated tools before any business hypothesis or sub-question has been framed, and gets 'lost in the data very quickly'.
• Treating data quality as a one-time cleanup. The tell: no dedicated data owners and no regular audit schedule, so records drift out of date and get contradicted mid-analysis.
• Leaving all entry with HR. The tell: stale, incomplete fields far from their source that require repeated correction because no one who owns the fact entered it.
• Chasing 100% accuracy everywhere. The tell: effort and time sunk into precision on directional fields while critical fields like headcount remain unreconciled.
• Buying infrastructure on licence price alone. The tell: on-premise chosen without accounting for customisation, infrastructure, maintenance, IT personnel and upgrade costs that push four-year TCO to nearly double the cloud alternative.

• ▸Data field/source: name the field and system it lives in
• ▸Domain: mark as HR, employee-attribute, or Non-HR (Finance/Marketing/Operations)
• ▸Accuracy across systems: note result of cross-check against other systems
• ▸Completeness (timely): % populated and whether current
• ▸Consistency/business rules: pass or fail against defined rules
• ▸Availability & security: who can access, from where, with what controls
• ▸Required accuracy level: high (e.g. headcount) or directional (e.g. leaver reason)
• ▸Entry point: entered by HR or at source by employee — flag if entry needs to move
• ▸Infrastructure/TCO note: cloud, on-premise or SaaS and 4-year cost basis

1. 1Classify your target concept first: decide whether it is a concept-by-postulation (built from other concepts, e.g. ethnocentrism, subtle and symbolic racism) or a concept-by-intuition (directly askable). Do not skip to writing questions.
2. 2For concepts-by-postulation, explicitly operationalize them into concepts-by-intuition, and decide whether the indicators are reflective or formative — this choice changes how you later assess quality (Chapters 14, 15).
3. 3Convert each concept-by-intuition into an assertion (a statement indicating the requested concept) using the basic assertion structures — e.g. for behavior use structure 2 (rDy): the respondent does/did y, or structure 3 (rD).
4. 4Transform the assertion into a request for an answer, choosing the request type deliberately (WH-request words vs. other types per Table 3.1), so the wording matches the concept you specified.
5. 5Decide the response format: open request vs. closed categorical request, and if closed, decide how many categories are optimal (Chapter 5) and the item structure (Chapter 6).
6. 6Decide whether the item stands alone or sits in a battery, and note that some prerequests/batteries can change the concept-by-intuition being measured — check the item still targets the intended concept.
7. 7Add supplementary text only as needed: introduction, extra information, definitions, instructions, or a motivation to answer.
8. 8Test before fielding: check face validity, control routing, predict question quality with an instrument (SQP), and run a pilot; only then formulate the final questionnaire.

Take 'satisfaction with the economy,' used in the ESS pilot MTMM study. It is a concept-by-intuition, so you can ask it directly, but the wording and response format are controllable design choices — the study varies them as methods 1, 2, and 3. The same underlying satisfaction is asked with different formulations, and the cross tables (Tables 9.3 and 9.4) show that method 1 vs. method 3 and method 1 vs. method 2 do not produce identical response distributions. Table 9.2 gives the response distribution for the nine requests, and the correlations among the nine variables (Table 9.7) feed the quality estimates in Table 9.8. The lesson: holding the trait constant, changing only the instrument (categories, structure, wording) changes the answers. That is why you specify the concept first, then treat every formulation choice as a variable whose effect on quality you estimate rather than assume.

• If the concept is built from other concepts (concept-by-postulation), then operationalize it into concepts-by-intuition explicitly before writing any question; if it is directly askable, formulate the assertion directly.
• If the concept-by-postulation uses reflective indicators, then treat quality per Chapter 14.2; if it uses formative indicators, then treat it per Chapter 14.3 — the two are not interchangeable.
• If the assertion concerns behavior (present/past actions of the respondent), then use structure 2 (rDy) or structure 3 (rD) with an activity-indicating predicator.
• If you place items in a battery or add a prerequest, then re-check the concept-by-intuition, because these features can change what is being measured.
• If two formulations of the same trait must be compared for quality, then run an MTMM (or split-ballot MTMM) experiment rather than judging by inspection.

• Jumping straight to questions without stating how a concept-by-postulation maps to concepts-by-intuition — the tell: reviewers cannot say which concept each item measures, and the instrument may not measure what it claims.
• Assuming wording is cosmetic — the tell: choosing method/format casually, then finding different methods give different response distributions (Tables 9.3–9.6) with no way to explain them.
• Ignoring routing control, especially in computer-assisted collection — the tell: respondents or interviewers get guided the wrong way, producing incomplete responses.
• Overloading items into batteries or adding prerequests without rechecking the concept — the tell: the measured construct quietly shifts from the intended one.
• Skipping pre-field quality prediction and pilot testing — the tell: quality problems only surface after data collection, when they can no longer be fixed by rewording.

• ▸Concept type: mark concept-by-postulation or concept-by-intuition.
• ▸Operationalization (if postulation): list the concepts-by-intuition and mark reflective or formative.
• ▸Assertion: write the statement and its basic structure (e.g. rDy for behavior).
• ▸Request type: choose from Table 3.1 (WH-request word or other) and record it.
• ▸Response format: open or closed; if closed, state number of categories and item structure.
• ▸Battery/prerequest check: note any battery or prerequest and confirm the concept is unchanged.
• ▸Additional text: list introduction, definitions, instructions, or motivation added.
• ▸Quality test plan: mark face-validity check, routing check, SQP prediction, and pilot.

1. 1Define turnover precisely as voluntary severance of employment ties (Hom & Griffeth, 1995) and separate it from involuntary exits before you measure or model anything—conflating the two corrupts every downstream analysis.
2. 2Adopt the 'standard research design' (Hulin, 1966, 1968): collect reliable predictors at Time 1, then collect individual turnover data at a later point. Never rely on retrospective predictors or recalled leaving, which crippled early weighted-application-blank studies.
3. 3Distinguish who leaves from how many leave. Following Porter's functional turnover concept (Dalton, Krackhardt, & Porter, 1981), stratify departures by performer value—loss of surplus, low-quality, or costly labor can help; loss of high talent or high social-capital employees harms most.
4. 4Model both leaving and staying. Incorporate 21st-century embeddedness (Mitchell et al., 2001): assess on-the-job and off-the-job forces—links, fit, sacrifice—across both job and community that hold incumbents in place.
5. 5Add shocks to the model. Per the unfolding model, identify critical events that prompt thoughts of leaving and recognize multiple paths (script-based, job-offer, affect-based); shock-driven leavers quit faster than dissatisfied leavers.
6. 6Track attitudes over time, not once. Use attitudinal trajectories rather than single-point satisfaction snapshots to predict who will quit, since decline patterns carry predictive signal a cross-section misses.
7. 7For unit- or firm-level questions, measure collective turnover as a rate and link it to organizational performance and productivity (Heavey et al., 2013), accounting for context like group cohesiveness, newcomer concentration, and unit size (Hausknecht, Trevor, & Howard, 2009).
8. 8Cost and prioritize interventions using the estimated replacement expense of 90%–200% of annual salary (Allen, Bryant, & Vardaman, 2010); concentrate retention spend where departures are dysfunctional and expensive.

Weitz (1956) tackled attrition among new insurance agents by giving new hires a booklet describing the realistic nature of the work—a realistic job preview (RJP). Retention improved, establishing that setting accurate expectations before entry reduces later quits. Farr, O'Leary, and Bartlett (1973) replicated the principle in a different setting, using work samples to reduce quits among sewing machine operators. The mechanism aligns with Porter and Steers' (1973) met-expectations theory: satisfaction and retention hinge on how closely the job fulfills the expectations formed at entry. Practically, this means the intervention is applied at recruitment and assimilation, not after dissatisfaction sets in. The lasting payoff was a large downstream literature on RJP media, mechanisms, and moderators (Earnest, Allen, & Landis, 2011; Wanous, 1973). The lesson: shrinking the gap between expected and actual work is a documented, replicated retention lever.

• If you are validating a predictor of turnover, then measure predictors at Time 1 and actual leaving later—if either is retrospective, discard the design.
• If a departing employee is a surplus, low-quality, or high-cost performer, then treat the exit as potentially functional rather than a loss to prevent (Dalton, Krackhardt, & Porter, 1981).
• If a quit followed a discrete critical event, then classify it as shock-driven and expect faster exit; if it followed gradual dissatisfaction, expect slower, attitude-driven exit and different remedies.
• If an employee shows strong job or community embeddedness (links, fit, sacrifice), then predict staying even amid dissatisfaction—target retention resources elsewhere.
• If the question is about firm performance impact, then analyze collective turnover rate and its context (cohesion, newcomer concentration, size), not individual intentions alone.

• Treating all turnover as dysfunctional and chasing a lower rate—tell: leadership celebrates rate reduction while high performers keep leaving, the very form Porter warned harms firms most.
• Using retrospective or recalled data for predictors or criteria—tell: predictors and turnover are measured at the same time or from memory, reproducing the flawed early WAB designs.
• Relying on a single satisfaction snapshot—tell: your model misses employees whose attitudes were declining, so predictions fail for people who looked 'fine' at one point.
• Modeling only reasons to leave while ignoring embeddedness—tell: dissatisfied employees who never quit look like model errors because you never accounted for job/community forces holding them.
• Ignoring contagion and diversity side effects—tell: exits cluster (turnover contagion; Felps et al., 2009) or women of color disproportionately leave (Hom, Roberson, & Ellis, 2008), yet your analysis treats each departure as independent and neutral.

• ▸Departure type: mark voluntary vs. involuntary (only voluntary is turnover per Hom & Griffeth, 1995)
• ▸Performer value: rate leaver as high-talent/high-social-capital vs. surplus/low-quality/costly (functional?)
• ▸Replacement cost estimate: apply 90%–200% of the role's annual salary
• ▸Trigger classification: shock-driven (name the event) vs. gradual dissatisfaction
• ▸Unfolding path: script-based / job-offer / affect-based
• ▸Embeddedness score: rate job and community links, fit, and sacrifice (low=at risk)
• ▸Attitude trajectory: rising / stable / declining over last measurement points
• ▸Collective context: unit turnover rate, cohesiveness, newcomer concentration, size
• ▸Recommended lever: RJP/onboarding, embeddedness build, none (functional exit)—with priority rank

1. 1Start with job analysis, not tests: rate how often incumbents perform each task, then factor-analyse to extract the underlying themes of the work (Krzystofiak et al. extracted 60 factors from 754 items across 1,900 employees).
2. 2Map each identified theme to a measurable individual attribute — e.g. 'ability to influence' to PQ dominance, 'attention to detail' to PQ detail-consciousness, 'numeracy' to a numeracy test.
3. 3Validate empirically where you have numbers: correlate test scores with work performance scores (criterion validation), preferring predictive over concurrent designs; this needs large samples.
4. 4For jobs with too few incumbents to compute a local validity, use synthetic validity — pool the predictor and criterion data of everyone across the workforce whose work loads on that same attribute, and infer validity for the small job from the pooled coefficient.
5. 5Judge the size of what you find using the right statistic: use correlation for continuous predictor-performance links and the d statistic (SDs separating group means) for group comparisons.
6. 6Set realistic expectations: treat 0.50–0.60 as near the practical ceiling (a quarter to a third of variance), because performance is also driven by management, climate, co-workers and economy.
7. 7Correct for restriction of range when interpreting validity in pre-selected samples — but ask whether the corrected figure is actionable given that education and social background pre-select applicants in real hiring.

A city authority employs 1,500 people across 300 jobs. Local tax clerk (200 people) has enough incumbents for a conventional validity study, but refuse collection supervisor (25) has too few and crematorium attendant (1) makes statistics impossible. Job analysis identifies underlying themes; 'ability to influence others' runs through 430 people across many jobs, 'attention to detail' through 520, 'numeracy' through 350. Rather than test each small job alone, the selector pools all 430 influence-relevant staff to establish PQ dominance validity (0.30), then infers it applies to the 25 refuse supervisors. The single crematorium supervisor's detail-consciousness validity is proven by pooling that one person's data with the other 520 for whom detail matters. When the city later hires 10 diversity facilitators, the same theme-and-test mapping plans their selection before anyone is in post.

• If a job has too few incumbents for a local validity study, then use synthetic validity by pooling everyone whose work shares that attribute-theme.
• If an effect is very small (r around 0.11 or d around 0.08), then look for something better — unless the method is so cheap and repeated across many vacancies that small gains accumulate (as with informal recruiting).
• If you are comparing two groups' averages, then report d (SDs between means); if you are linking a predictor to performance, then report a correlation.
• If a validity coefficient exceeds 0.30, then it clears the EEOC's practical barrier for US employers; below that, expect challenges to establishing validity.
• If you have PAQ job-analysis data, then use it to predict the GATB profile of successful incumbents, since PAQ scores correlate well with GATB scores and fairly well with GATB validity.

• Dismissing tests because correlations sit around 0.30 — the tell is treating 10% of variance as 'nothing' while ignoring that 0.50–0.60 is the realistic ceiling given all the other influences on performance.
• Sifting on what managers claim matters (proven ability, prior position) while actually rejecting for untidy or badly written applications — the tell surfaces in policy-capturing analyses where stated and real criteria diverge (Machwirth et al.).
• Misusing predictive information idiosyncratically, e.g. campus recruiters screening OUT high-GPA applicants despite GPA's link to performance and mental ability (McKinney et al.).
• Trusting range-corrected 0.70+ true correlations as directly actionable — the tell is forgetting that pre-selection by education and social background is a fact of life, so the unselected-sample figure may not apply to real hiring.
• Postulating utility savings that 'dwarf the national debt' without demonstrating real organizational gains — the tell is grand estimates with no criterion evidence (Vance and Colella's complaint).

• ▸Job title: name each role in the workforce
• ▸N incumbents: count per job to flag which need synthetic validity
• ▸Theme/attribute: the trait the job demands (e.g. ability to influence, detail, numeracy)
• ▸Mapped test: the specific instrument for that attribute (e.g. PQ dominance, numeracy test)
• ▸Attribute flag (XX/X/–): mark whether this job loads on each attribute
• ▸Pooled N per attribute: total people workforce-wide sharing this attribute
• ▸Validity coefficient: the pooled correlation for that attribute-test link
• ▸Statistic type: note correlation (predictor-performance) or d (group comparison)
• ▸Action threshold: record whether coefficient clears 0.30 and is worth deploying

1. 1Define the outcomes the role actually pays for before assessing anyone — Michael pays people well 'for what they do,' which requires knowing what 'what they do' is. Separate the outcome (satisfy the customer, run the show) from the method.
2. 2Diagnose each person's recurring patterns — their style, motivation, and personal situation — by spending real time with them: listening, meals, holidays. Michael got to know Brad, Susan, Gary and Emma this way, then matched them to what they did best.
3. 3Look for the fit between an existing talent and the demands of the specific role, not for a talent you hope to install. Brad's talent is performing for an audience he respects (customers); that fits waiting, not managing.
4. 4Test fit against the role's real demands: Emma's awareness and responsibility fit team-building; Brad's disrespect for new employees disqualifies him from managing, since those employees would become his audience.
5. 5When fit exists, remove yourself — let people run the show. Michael's best team trained new hires, set the example and ejected people who didn't fit for three years without his interference.
6. 6When fit is absent and performance consistently fails, act fast. Michael fires quickly because waiting on a consistent underperformer 'makes matters worse,' not better.
7. 7Refuse to promote out of fit. Pay people well and make it rewarding to keep doing what they do best rather than moving a great performer into a role that doesn't match.

Michael's strongest case of person-role fit is his decision NOT to promote Brad. Brad is a great waiter because he loves to perform for an audience he respects, and he respects the customers. But Brad would make a terrible manager, because as a manager his audience becomes the new employees — and Brad is 'less respectful' of them, so the very trait that makes him excellent on the floor would fail him one rung up. Michael also reads Brad's temperament precisely: raise your voice and Brad shuts down and is crushed, so Michael reasons through disagreements quietly. Contrast this with Gary, whose fit was real enough that Michael fired and rehired him twice — a firm hand worked with Gary and would have destroyed Brad. Michael's rule follows directly: 'Don't overpromote people. Pay them well for what they do.' The fit is preserved by keeping the person in the role their talent matches, not by advancing them.

• If a person excels because of an audience or condition the current role supplies, then check whether the new role removes that condition before promoting — if it does, don't promote.
• If someone is consistently underperforming, then fire fast; waiting is not a favor and makes matters worse.
• If a person has a talent that fits the role but a behavior that crosses a line (Gary's joking), then correct the behavior firmly without discarding the talent — rehire when the fit still holds.
• If you catch yourself treating everyone the same to avoid favoritism, then stop — treat people differently according to their style and situation, because that is how they feel understood.
• If your best team runs itself — trains, sets the example, ejects misfits — then interfere less, not more.

• Trying to fix people into the role instead of finding people who fit it. The tell: you're spending energy making someone 'more responsible' rather than casting a naturally responsible person like Emma.
• Rewarding excellent performance with a promotion that changes the role's demands. The tell: your best waiter becomes a mediocre, resentful manager and you've lost value on both ends.
• Managing everyone identically. The tell: your firm hand works on one person and crushes another, and you keep applying it uniformly.
• Not knowing your people's style, motivation, or personal situation, so you can't judge fit. The tell: you're relying on 'familiarity breeds contempt' as an excuse for distance.
• Waiting on a bad hire because you think patience is kind. The tell: things are 'falling apart' and you're still deliberating — misfit doesn't self-correct.

• ▸Role's paid-for outcome: state the result this role exists to produce, not the method.
• ▸Person's recurring strength: the pattern they show without prompting.
• ▸Enabling condition: what the current role supplies that makes that strength work (e.g., an audience they respect).
• ▸Fit evidence: specific instances where their pattern already produced the outcome.
• ▸Misfit risk: which role demand their pattern works against.
• ▸Promotion test: does any considered move remove the enabling condition? Yes = do not move.
• ▸Management style match: what response do they give to firmness vs. quiet reasoning?
• ▸Decision: keep-and-reward / correct-behavior-keep / exit-fast — with the trigger you observed.

1. 1Start from a job definition or, better, a competency model — the passages treat this as the foundation. Codify what successful performance actually looks like before selecting any method (echoing Taylor's minute study of how the most skilled work).
2. 2Reject weak predictors as primary screens. Age and educational qualifications are inputs to behaviour, not behavioural outputs, and correlate only slightly with job performance — do not select arbitrarily on them.
3. 3Choose methods for their validity against the defined criteria: psychometric tests (personality, ability/reasoning), structured or competency-based interviews, and assessment centres, joined up rather than treated as separate silos.
4. 4Set the highest defensible cut-off on any predictor. Given a positive correlation, other things equal, the highest possible cut-off reduces false positives and false negatives; but improve the predictor itself rather than just raising an arbitrary bar.
5. 5Combine complementary measures — pair a personality questionnaire with a high-level reasoning test — because a relaxed person with weak reasoning may still fail under load, while a tense person with high reasoning may cope.
6. 6Structure the interview around behaviour and derived interview models (criterion-based, competency-based, critical incident) rather than leaving it unstructured; reserve unstructured questioning for transitory factors like present circumstances, motivation and financial expectations.
7. 7Check every procedure against equal opportunities law: eliminate indirect discrimination where a screen favours one group without relation to job capability (e.g. excluding a geographic area that disproportionately contains an ethnic minority).
8. 8For one-off or shortlist senior hires, bring in psychometrics to surface what CV-sifting and conventional interviews (both experience-focused, both weak predictors) missed — treating it as insurance against costly hiring mistakes.

Consider the common employer instinct to raise academic standards — hiring first-class rather than second-class degree holders. The relationship between degree class and job performance may be slight, but if any positive correlation exists at all, raising the cut-off improves the odds of a successful hire. However, the passages warn this is 'leveraging the weak predictor': the market is dynamic, first-class candidates know their worth and demand higher salaries, complicating cost. The better move is to substitute a validated test procedure for the academic standard entirely, and to fund more research to define competencies and tailor the test off the shelf — this shifts the shape of the test/performance relationship closer to a straight line, shrinking both false positives (hired but fail) and false negatives (rejected but could have succeeded).

• If a screen is an input (age, experience, degree class) rather than a behavioural output, then treat it as a weak predictor and do not use it as a primary selection filter.
• If a predictor has any positive correlation with performance, then set the highest defensible cut-off — but if you can afford it, replace the weak predictor with a validated test rather than just raising its threshold.
• If you need transitory information (current circumstances, motivation, salary expectations), then use an unstructured interview; if you need personality or reasoning evidence, then use psychometrics and structured behavioural interviewing.
• If a procedure favours one group without that basis relating to job capability, then it constitutes indirect discrimination and must be removed.
• If a senior or key hire has passed CV-sift and conventional interview, then add psychometric assessment to catch what those experience-based methods miss before committing to an expensive mistake.

• Treating assessment methods as separate silos rather than joined-up understanding of capability — the tell: each method run by different owners with no integration of test, interview and reference data.
• Confusing inputs with outputs — selecting on age or degree class as if they were performance itself; the tell is arbitrary age bands 'sometimes defined as narrowly as a per[cent]' with no validation.
• Overconfidence in interviewing skill — 'all interviewers think they are excellent interviewers'; the tell is that almost no manager will ever admit to being a poor interviewer despite the University-of-Life basis of their expertise.
• Building rating factors on an arbitrary basis with no research on feedback or follow-up — the tell is the process becoming 'little more than a ritual', with appraisers reluctant to record negative ratings.
• Spending enormous time and money on early procedures (CV-sifting, conventional interviews) then wondering what was missed — the tell is a late external-consultant referral, which is an implicit admission the prior methods were weak predictors.

• ▸Role & competency model: list behavioural competencies derived from study of successful performers
• ▸Criteria definition: state what 'success in the job' concretely looks like for each competency
• ▸Methods mapped: for each competency, name the method (psychometric / structured interview / assessment centre) that validly measures it
• ▸Weak-predictor audit: list any age/experience/qualification screens in use and mark input vs. output
• ▸Cut-off setting: record each predictor's cut-off and note whether raising it or replacing it improves prediction
• ▸Complementarity check: confirm ability and personality measures are paired, not used in isolation
• ▸Equal-opportunities test: flag any screen that favours a group without job-capability basis (direct or indirect)
• ▸Cost-of-mistake note: estimate the impact of a bad hire on the organization and hiring executive to justify added assessment
• ▸Unstructured-scope limit: list only the transitory items (circumstances, motivation, expectations) reserved for unstructured questioning

1. 1Write company-level OKRs first, before team or employee OKRs, to solidify the corporate vision from top to bottom; make these Objectives applicable and actionable by top-level management.
2. 2State each Objective as a specific, ambitious, time-bound goal — not an ethereal one. Reject 'I want to be rich' or 'I want to grow'; force a destination like 'By next quarter, cut user signup time in half.'
3. 3Attach 2-4 Key Results to each Objective. Each must contain a numerical value, be trackable over time, and be objectively gradable so fulfillment is not open to interpretation.
4. 4Test each Key Result by asking 'What value is this measuring?' If the answer takes a while, you chose a bad Key Result — rewrite it with a blatantly obvious metric (growth, revenue, engagement, performance).
5. 5Build team-level OKRs collaboratively with the whole team participating — not decided by the manager alone — and make sure they complement the company OKRs. Assign one team leader to monitor and track the metrics.
6. 6Let OKRs flow bottom-to-top-and-back (grassroots up to leadership, then back down) rather than shoving targets down through the floors, so employees feel empowered by the process.
7. 7Publish every level's OKRs transparently (e.g., on the company intranet, as Google does) so any team out of alignment automatically stands out and goals are easy to grade and modify.
8. 8Run regular check-ins and weekly updates with face-to-face assessment meetings; use them to grade progress, course-correct, and discard goals that are not working.

An insurance company owner sets the Objective: 'By next quarter, reduce the amount of time it takes users to sign up to my website by half.' This is a destination — specific, ambitious, and time-fixed — unlike his first pass, 'I want to be rich.' Key Result #1: Interview 100 customers and ask them to describe why they found the website too complicated to use. A parallel content example uses KR #2: Contact 20 website owners and ask to interview them for a podcast, and KR #3: Start an email list and sign up 25% of the existing customer base. Each Key Result has an objectively established finish line — interview 100, contact 20, sign up 25% — so progress is a metric, not an opinion. They act as milestones on the road: markers that confirm you're on the right track toward the Objective. As John Doerr frames it: the Objective is what you want accomplished, the Key Result is how you're going to get it done.

• If a stated goal is too vague to measure (e.g., 'grow the company'), then it is not an Objective yet — rewrite it as a specific, time-bound destination.
• If a Key Result's fulfillment is open to interpretation, then it fails — add a numerical target so grading is not a matter of opinion.
• If your teams are tiny, close-knit units of under half a dozen people, then skip individual employee OKRs and let team OKRs cover them — there are no hard rules at the employee level.
• If your company already runs an agile framework with short work segments and frequent reassessment, then adopt OKRs — they complement agility's check-ins and rapid course-correction seamlessly.
• If a goal isn't working during a weekly check-in, then discard or modify it rather than protecting it to term.

• Setting ethereal goals ('make a lot of money') and calling them Objectives — the tell is that no one can point to a finish line or a number.
• Writing Key Results with no quantifiable metric, so they read like a to-do list — the tell is you can't answer 'what value is this measuring?' quickly.
• Manager decides the team OKRs alone — the tell is low buy-in and misalignment between team and employee goals, because alignment only happens when all members participate.
• Keeping goals on a 'need to know'/pay-grade basis instead of publishing them — the tell is that a misaligned team goes unnoticed because nobody can see it against others.
• Building employee OKRs before the company vision exists — the tell is bottom-level effort that doesn't roll up to any corporate destination.

• ▸Level (Company / Team / Individual): state which tier this OKR sits at
• ▸Objective: one specific, ambitious, time-bound destination (avoid vague wishes)
• ▸Alignment link: name the higher-level Objective this complements
• ▸Key Result 1: numerical, gradable target with a clear finish line
• ▸Key Result 2: numerical, gradable target with a clear finish line
• ▸Key Result 3: numerical, gradable target with a clear finish line
• ▸Metric check: for each KR, write what value it measures
• ▸Owner: name the one person monitoring and tracking the metrics
• ▸Cadence & visibility: check-in frequency (e.g., weekly/quarterly) and where it's published for transparency

1. 1Name the two systems as you work: label the fast, automatic response as System 1 and the effortful, sequential response as System 2 so you can talk about each separately. System 1 generates impressions; System 2 endorses or overrules them.
2. 2Run a substitution check on any hard judgment. Write down the actual question you face, then ask, 'What easier question am I answering instead?' The executive faced 'Is Ford stock underpriced?' but answered 'Do I like Ford cars?' — catch that swap.
3. 3Watch for the affect heuristic: when a judgment tracks your feelings of liking or disliking with little reasoning, flag it. Emotion driving a decision is a tell that System 1 has substituted a feeling for an analysis.
4. 4Test whether the intuition comes from skill or from a heuristic. Ask: does this person have prolonged, relevant practice in an environment with reliable feedback (chess master, firefighter)? If yes, trust it more; if no, treat the fast answer as a candidate for error.
5. 5Recognize that knowing the bias does not stop it. Like the Müller-Lyer illusion, you keep seeing the biased impression even after measuring. The only defense is to learn to recognize the pattern in advance and recall what you know about it.
6. 6Force System 2 engagement when stakes are high: slow down, hold the relevant statistical facts in working memory, and check them against the intuitive answer — the way you must retrieve and run the multiplication program for 17 × 24 rather than accepting a guess.
7. 7Protect against depletion: don't attempt hard reflective checks when tired or after prolonged self-control. Ego depletion and low glucose degrade System 2 monitoring, so schedule high-stakes judgments when your capacity to override System 1 is intact.

An executive decides whether to invest in Ford stock. The relevant question an economist would ask is narrow: is Ford stock currently underpriced? Instead of answering it, he listened to his intuition — he liked the cars, liked the company, liked the idea of owning the stock — and bought. This is the affect heuristic: the decision was guided directly by feelings of liking, with little deliberation. Notice the substitution: the difficult question ('Is the stock underpriced?') was silently replaced by an easy one ('Do I like Ford cars?'), and the answer to the easy question determined the choice. He had no relevant expertise in stock picking, so the fast answer was a heuristic, not skill — and from what we know about the accuracy of stock picking, it is reasonable to believe he did not know what he was doing. The fix would have been to write out the original question, notice the swap, and mobilize System 2 to actually assess price.

• If a judgment tracks how much you like or dislike something rather than the evidence, then treat it as the affect heuristic and re-ask the original question.
• If the person judging has prolonged practice with reliable feedback (chess, firefighting), then weight the intuition as skill; if not, then treat the fast answer as heuristic and check it.
• If an answer comes to mind quickly and effortlessly for a hard question, then suspect you answered an easier substitute question and identify what it was.
• If you notice a biased impression persists even after you know the correct fact (like still seeing unequal lines), then rely on a learned recognition rule, not on willpower to 'see' correctly.
• If the task requires holding several facts at once or statistical reasoning, then deliberately engage System 2, because System 1 is not designed to think about many things at once.

• Trusting fluency: the answer feels obvious, so you accept it. Tell — you felt no strain and never engaged the effortful, sequential steps System 2 requires.
• Mistaking a heuristic for expertise. Tell — the fast intuition arrives in a domain (stock picking) with poor feedback and no genuine prolonged practice.
• Availability distortion: judging importance by what comes easily to mind, like dramatic media topics over unexciting ones. Tell — your examples are the ones mentioned often, while equally important, less-available issues never surfaced (Kahneman caught himself doing exactly this).
• Assuming that knowing the bias immunizes you. Tell — you still 'see' the illusory pattern; the biased impression returns despite your correct belief.
• Attempting override while depleted. Tell — after self-control or hard cognitive work your blood glucose drops and you 'don't feel like' making the effort, so System 2 rubber-stamps System 1.

• ▸Original question: state the exact hard question you must answer
• ▸Answer that came to mind: write the fast intuitive response
• ▸Easier question actually answered: name the substitute (e.g., 'Do I like it?')
• ▸Feeling flag: is this driven by liking/disliking? (affect heuristic yes/no)
• ▸Skill or heuristic: does the judge have prolonged practice + reliable feedback? (yes/no)
• ▸Relevant facts ignored: list statistical or base-rate facts System 1 skipped
• ▸Recognized bias pattern: name the known illusion at play (availability, anchoring, etc.)
• ▸Capacity check: are you depleted/tired/rushed right now? (yes/no)
• ▸System 2 verdict: after deliberate check, endorse or overrule the intuition

1. 1Name the theoretical construct you claim to measure before touching data, then state which operational indicators stand in for it — construct validity fails when this link is left implicit, so make it explicit in your research design (the logic linking data to conclusions to initial questions).
2. 2Assemble multiple sources of evidence for the same construct rather than a single measure; case study research inherently draws on a wide variety of evidence (documents, direct observation, interviews) where experiments limit to recorded behavior and surveys to verbal information.
3. 3Converge those independent sources on the same construct — integrate them so behavioral and verbal evidence corroborate one another, as Zigler and Muenchow combined personal experience, historical/field research, and 200+ interviews into one coherent case.
4. 4Specify each plausible rival explanation for what your indicator might actually be measuring, then test its other implications against other data sets (ramification extinction) to reduce its plausibility — do not rely on unnamed controls.
5. 5Maintain a chain of evidence so a reader can trace each conclusion back through the indicators to the original construct and questions, protecting against threats to validity.
6. 6Review your emerging design against the four standard tests — construct validity, internal validity, external validity, reliability — treating construct validity as the gate on whether your operational measures represent the concept at all.
7. 7Where quantification is not available, still pursue validity: apply the plausible-rival-hypotheses strategy from the experimental-isolation paradigm, specifying and controlling each named rival rather than abandoning the goal of getting it right.

Zigler and Muenchow's study of Head Start shows construct validity built through convergence. The construct — the program's early development and its standing as one of the most successful programs — could have rested only on Zigler's personal experience as its first director, which alone would invite the rival that his account is self-serving. Instead the book calls on independent sources: the coauthor's historical and field research, plus interviews of more than 200 persons associated with Head Start. These multiple sources are integrated into a coherent and well-documented case rather than left as separate anecdotes. Because behavioral, documentary, and verbal evidence point to the same conclusions, the plausibility of the self-report rival is reduced. The result — readable but well-documented — is what convergent evidence on a single construct produces.

• If your construct is defined only implicitly, then stop and articulate the theory of what is being studied before collecting data, because the data-to-question link is where construct validity is won or lost.
• If you have a single source measuring the construct, then add at least one independent source of a different type (behavioral, documentary, verbal) so the sources can converge.
• If a rival explanation for your indicator remains plausible, then trace its other implications on other data sets and rule it out by ramification extinction rather than by assertion.
• If you cannot name which rival hypotheses your design controls, then you are relying on randomization-style blanket control you do not actually have — specify each rival explicitly as the experimental-isolation paradigm requires.
• If you lack quantification or significance tests, then still hold to validity-seeking: the plausible-rivals strategy works for a historian's conjecture as much as for a causal law.

• Treating research as mechanistic data collection without theory — the tell is a study with rich data but no stated construct or proposition the data are meant to represent.
• Relying on one source for a construct: an all-interview or all-observation base — the tell is that no independent source could corroborate or contradict the central claim.
• Confusing this research use of the case with a teaching case, so evidence is chosen to illustrate rather than to test the construct — the tell is that no rival explanation is ever entertained.
• Using the term validity-seeking language while quietly abandoning the goal of getting it right — the tell is invoking hermeneutics to excuse not disputing who has it correct.
• Assuming randomization or scale mechanics 'control everything' — the tell is that no specific rival hypothesis is named, so an unexamined confound can later prove the indicator measured something else.

• ▸Construct name: the theoretical concept you intend to represent
• ▸Operational indicators: the specific measures/data standing in for it
• ▸Source types used: list each (documents, direct observation, interviews, archival)
• ▸Convergence check: which conclusions are corroborated by two or more independent sources
• ▸Named rival explanations: alternative constructs the indicator might actually reflect
• ▸Ramification-extinction test: other data sets each rival implies, and the fit result
• ▸Chain of evidence: how a reader traces conclusion → indicator → construct → question
• ▸Design tests review: notes on construct, internal, external validity, and reliability
• ▸Residual threat: the strongest rival not yet ruled out

1. 1Reframe reliability operationally: your goal is that a later investigator following your documented procedures on the *same* case reaches the *same* findings and conclusions — this is repeatability of procedure, not replication of results across cases.
2. 2Write a case study protocol before data collection that records every procedural step (field questions, data sources, sequence of contacts, coding conventions). Without this documentation you cannot even repeat your own work, let alone allow an external reviewer to.
3. 3Make each step as operational as possible: convert vague instructions ('assess neighborhood change') into specific, followable operations (which records, which interviews, which observable events count).
4. 4Build a case study database that keeps your raw evidence (notes, documents, interview records) separate from your final report, so a later investigator can trace and re-run your steps against the same base of evidence.
5. 5Conduct the research 'as if someone were looking over your shoulder' — assume an external auditor will retrace every decision, and structure your notes accordingly.
6. 6When multiple researchers are involved, curb individual intuition in favor of consistency across investigators; standardize procedures through training so the same case handled by different people yields the same result.
7. 7Treat reliability as ongoing design work: apply the protocol and database tactics throughout the conduct of the study, not only at the design stage.

Take 'neighborhood change,' a common case study topic cited in the passages. A researcher claims a neighborhood 'changed its character,' but with no prior specification of what operational events constitute 'change,' a reader cannot tell whether the claimed changes were measured consistently or reflect the researcher's subjective, preconceived notions. To make the study repeatable, you would specify in a protocol the exact significant events counted as 'change' (e.g., defined shifts in named records or observable indicators), document the data collection procedures for gathering them, and store the underlying evidence in a case study database. A later investigator, following that same protocol on the same neighborhood, should then arrive at the same finding of change. This documentation both minimizes errors and biases and answers the critics who found the case study method suspect precisely because its procedures were poorly documented.

• If a later researcher redoing the *same* case would get different findings, then your procedures are underspecified — document them further before proceeding.
• If you cannot repeat your own analysis from your own notes, then your documentation fails the reliability prerequisite; build the case study protocol and database.
• If multiple researchers are collecting data, then standardize and train for consistency rather than relying on individual intuitive adaptation, to minimize introduced bias.
• If a key concept is stated in vague terms, then define the operational events or measures for it before data collection so its recording is consistent.
• If the goal is generalizing findings to other situations, that is external validity, not reliability — do not conflate 'repeating the same case' with 'replicating results in another case.'

• Confusing reliability with replication: doing a second, different case to 'confirm' results, when reliability means redoing the *same* case with the same procedures. Tell: the researcher describes selecting a new case as their reliability check.
• Poorly documented procedures — the historical shortcoming that made reviewers suspicious of the whole method. Tell: no protocol exists and the researcher cannot reconstruct how a specific finding was reached.
• Leaving concepts subjective and unspecified, so data collection drifts toward confirming preconceived notions. Tell: claimed 'changes' or effects with no prior operational definition of what counts.
• Letting a lone brilliant researcher intuitively adapt without recording those adaptations, so no one — including the researcher — can retrace the steps. Tell: findings depend on undocumented on-the-fly judgment calls.
• Treating reliability as a one-time design checkbox rather than ongoing work throughout data collection and analysis. Tell: the protocol was written once and never used to govern actual fieldwork.

• ▸Case defined: state the single case being studied and its bounds
• ▸Operational definitions: list each key concept and the specific observable events/measures that count
• ▸Data collection procedures: step-by-step sequence a later investigator would follow
• ▸Data sources: name each source and how it is accessed
• ▸Field questions: the mental line of inquiry driving each data contact
• ▸Coding/analysis conventions: rules for turning raw evidence into findings
• ▸Case study database location: where raw notes, documents, and records are stored separately from the report
• ▸Chain of evidence: how each finding traces back to its source data
• ▸Multi-researcher consistency: training/standardization steps if more than one investigator

1. 1State the inferential goal and choose a design that isolates it before touching software: decide whether you need experimental control (random assignment), covariate adjustment (ANCOVA), or repeated observations, because each carries distinct validity threats.
2. 2Set your Type I/Type II error budget up front: fix alpha, target power at .80, and calculate required sample size using the guidelines the text provides for each analysis so results can generalize.
3. 3Count your statistical tests and control the family-wise error rate: multiple tests inflate the probability of spurious results, so plan Bonferroni-type or stepdown control on overall Type I error rather than running many uncorrected comparisons.
4. 4Screen the data before modeling: check for outliers, verify assumptions (independence, homogeneity, normality as required), and treat independence as non-negotiable — even small violations push actual alpha to several times the nominal level.
5. 5When observations are correlated (repeated measures, nested/clustered data), abandon the general linear model's independence assumption and move to the Hierarchical Linear Model or a correlated-observations analysis.
6. 6Guard against capitalization on chance in any maximization procedure (multiple regression, discriminant analysis, SEM): expect shrinkage, report it, and estimate it as a function of R-squared and whether predictors were culled from a larger set.
7. 7Cross-validate any data-driven modification: hold out a sample or use a validation approach for predictor selection and SEM specification searches, and refuse modifications that lack substantive theoretical justification.
8. 8Disclose threats explicitly: report internal versus external validity limits, nonresponse in surveys, and any financial conflict of interest that could bias the design or reporting.

Take the SEM specification-search problem the text documents. A researcher fits a CFA, gets poor fit, and lets the program suggest post hoc modifications — a data-driven process highly susceptible to capitalization on chance. MacCallum et al. (1992) showed such searches usually proceed with no substantive justification, no cross-validation, many modifications, and samples too small, calling it a dangerous methodological trend. Their study found that no sequence of four modifications produced the same modified model when sample size was 250 or less; consistency emerged only at N=400. Because most published studies run between 100 and 350, most reported modified models are unstable and unlikely to cross-validate. The rigorous move: refuse untheorized modifications, hold sample size to at least 400 before trusting a search, and cross-validate the final model on an independent sample.

• If observations are correlated (nested, clustered, or repeated), then use HLM or a correlated-observations analysis — do not use the general linear model, because even small independence violations multiply actual alpha.
• If you run multiple statistical tests, then control overall Type I error (e.g., stepdown or Bonferroni-type adjustment); otherwise expect spurious 'significant' results.
• If predictors are selected from a much larger set, then anticipate heavy shrinkage in R-squared and cross-validate before generalizing.
• If an SEM specification search is needed with N ≤ 250, then treat the result as unstable and unreproducible; require N around 400 and independent cross-validation.
• If a modification or finding lacks substantive theoretical justification, then reject it regardless of how much it improves fit — you are likely finding a nonexistent pattern you wanted to find.

• Treating statistical significance as practical significance — the tell is a tiny effect declared important solely because p < .05 with a large N.
• Ignoring the independence assumption on clustered/repeated data — the tell is an inflated significance rate; the actual alpha is several times the nominal level.
• Running many uncorrected tests and reporting whichever came out significant — the tell is a scattershot results section with no stated error-rate control.
• Accepting a maximization procedure's sample-optimized solution as generalizable — the tell is no reported shrinkage estimate and no cross-validation for regression, discriminant, or SEM models.
• Post hoc SEM tweaking on a small sample without theory — the tell is a long list of program-suggested modifications and a final model that cannot be reproduced at N=400.

• ▸Inferential goal: state the single question and the design chosen to isolate it.
• ▸Error budget: record alpha, target power (default .80), and required N from the guidelines.
• ▸Test count and control: list all planned tests and the family-wise correction method.
• ▸Independence status: note whether observations are correlated and the model chosen if so (HLM/correlated-observations).
• ▸Assumption checks: list assumptions to verify and screening steps for outliers.
• ▸Capitalization-on-chance guard: name any maximization procedure and the cross-validation or shrinkage plan.
• ▸SEM/search discipline: if a specification search is possible, state minimum N and cross-validation requirement.
• ▸Validity threats: document internal/external validity limits and survey nonresponse plan.
• ▸Conflict of interest: disclose any financial or other conflict affecting design or reporting.

1. 1Separate your model into the two parts LISREL requires: a measurement model that maps latent factors to observed indicators, and a structural model that specifies causal relations among the factors. Only after the measurement model holds should you interpret structural coefficients.
2. 2Name each hypothetical construct (intelligence, anxiety, self-esteem, poverty) and list the multiple observed indicators you claim measure it. Spearman's original logic applies: each measurement shares a common component tied to the factor plus a unique component holding measurement error and irrelevant variance.
3. 3Specify the factor structure explicitly — how many factors, which indicators load on which factor, and whether factors are orthogonal (uncorrelated) or oblique (correlated). Default to Joreskog's oblique specification because correlated factors are more realistic than the algebraic convenience of orthogonality.
4. 4Estimate reliabilities of the indicators as part of the measurement model. You need good reliability estimates to correct for the attenuation inherent in research using errorful measures.
5. 5Decide whether you are in an exploratory stage (using factor analysis to identify which constructs underlie a set of observed variables) or a confirmatory stage (testing whether items designed to measure a construct actually reveal the hypothesized factor structure). Use confirmatory maximum likelihood factor analysis when you have an a priori structure to test.
6. 6Fit the model and check whether the data reveal the hypothesized factor structure. If items do not load as designed, treat this as evidence to modify the theoretical model, combine or separate measures, or delete variables before proceeding to structural interpretation.
7. 7Guard against measurement-side misspecification: if you preconceived a variable to perfectly measure a factor but it does not, its coefficient is biased by unspecified measurement error. Model that error rather than assuming perfect measurement.

Take Spearman's original problem: measuring the construct 'general intelligence.' He proposed that various measurements of mental abilities each share a common component associated with the general intelligence factor, while each also carries a unique component representing measurement error plus an irrelevant component specific to that measurement. In LISREL terms this is a single latent factor with multiple observed indicators, where each indicator's variance decomposes into the common (factor) part and the unique (error) part. You would specify the ability tests as indicators loading on one intelligence factor, estimate the loadings and the reliabilities, and inspect whether the single-factor structure reproduces the observed covariances. Spearman's model was later extended to multiple factors, each measured by two or more variables, so if a single factor fails you can test a multi-factor oblique structure. The reliability estimates you obtain then let you correct downstream causal estimates for attenuation from errorful measurement.

• If a construct is not directly measurable, then estimate it from multiple indicators as a latent factor rather than treating any single measured variable as the construct.
• If you are identifying unknown constructs, then use factor analysis as exploratory; if you have a hypothesized factor structure to verify, then use confirmatory maximum likelihood factor analysis.
• If factors are theoretically related, then specify them as oblique (correlated); reserve orthogonal specification only when the algebraic simplification is justified and independence is defensible.
• If a variable was assumed to perfectly measure a factor but carries error, then model that measurement error explicitly, because ignoring it biases the coefficient.
• If the measurement model does not reproduce the hypothesized factor structure, then revise the theoretical model before interpreting any structural (causal) parameters.

• Interpreting regression coefficients on observed variables as structural causal effects without first establishing a valid measurement model — the tell is unexpected signs (e.g., a significant negative coefficient where you predicted positive) among correlated predictors.
• Treating a single observed variable as if it perfectly measures a latent construct; the tell is a coefficient that shifts or biases once measurement error is admitted into the model.
• Assuming factors are uncorrelated for algebraic convenience when they are substantively related — the tell is a poorly fitting orthogonal model that fits better once oblique factors are allowed.
• Failing to estimate reliabilities, so effects stay attenuated by errorful measurement — the tell is systematically weakened structural coefficients that don't correct for measurement error.
• Omitting a major causative factor correlated with an included predictor, so the predictor acts as a proxy and produces a spurious effect — the tell is a coefficient reflecting the omitted factor's effect weighted by its correlation with the included variable.

• ▸Construct name: state the hypothetical construct (e.g., self-esteem, intelligence)
• ▸Observed indicators: list the multiple measured variables claimed to indicate this construct
• ▸Factor count: specify how many latent factors the indicators are hypothesized to represent
• ▸Loading pattern: note which indicators load on which factor
• ▸Factor correlation choice: mark orthogonal or oblique and justify
• ▸Error components: identify measurement-error and irrelevant components per indicator
• ▸Reliability estimate: record the estimated reliability of each indicator for attenuation correction
• ▸Analysis stage: mark exploratory (identify constructs) or confirmatory (test hypothesized structure)
• ▸Misspecification check: note any omitted correlated causal factor that could make an indicator a proxy

1. 1Begin analysis only after you name your goal: description or theory. If theory, commit to integrating findings into an overarching explanatory scheme rather than stopping at a summary of themes.
2. 2Break the data apart. Take a passage and perform microanalysis — detailed coding around a word or phrase to surface its varied meanings, then examine each component's properties and dimensions.
3. 3Ask the grouping question of disparate data instances: 'what do these share in common?' Name that shared property as a concept (as with bird, plane, kite sharing 'flight'), then explore each instance in depth for similarities and differences along that concept.
4. 4Categorize and abstract: group related concepts into higher-order categories, specifying each category's properties and their dimensional ranges.
5. 5Write memos continuously as you code — capture conceptual reasoning about what participants convey, and separately capture your emotional/reflexive reactions so you can distinguish the two.
6. 6Reread coded interviews for tension, contradictions, and conflicting codes; treat those friction passages as sites to test and refine interpretations rather than smooth over.
7. 7Reread your memos and rewrite any that reflect your emotional response rather than a conceptualization of what respondents are telling you, correcting where you slant the data.
8. 8Iterate between data and concepts until you reach 'feeling right' — the point where, after immersion, the findings express what participants convey through word, action, and emotion as seen through the analyst's eyes.

Corbin ran a mini-research project on Vietnam veterans, coding interview data (provided as the MAXQDA project 'JC-BasicsQR.mx3'). She coded segments in the Text Browser, wrote memos, and attached them to codes as she worked. Practicing self-reflection, she noticed on rereading that some memos captured her emotional response to disturbing war stories rather than a conceptualization of what respondents were telling her — she identified strongly with the stories. She rewrote those memos so the analysis stayed grounded in participants' meanings while still acknowledging her presence in it. She followed Hamberg and Johansson's practice: rereading coded interviews to scrutinize passages with tension, contradictions, or conflicting codes, and rereading memos to recall reactions and prior coding discussions. Participants read chapters involving them and gave input, making the result partly a co-construction that presents their words alongside hers.

• If the research question can only be answered by getting at participants' inner experience and discovering rather than testing variables, then choose qualitative analysis over a structured quantitative format.
• If the goal is a few qualitative questions inside a larger study, then a summary of major themes suffices; if the goal is theory, then integrate concepts into an overarching explanatory scheme.
• If a memo reflects your emotional reaction more than a conceptualization of what respondents said, then rewrite it — but keep it, because seeing yourself in the analysis is legitimate self-reflection.
• If two coders' codes conflict or a passage carries tension and contradiction, then treat it as a priority site for finding a more reasonable, legitimate interpretation.
• If you find yourself obsessed with following coding procedures step-by-step, then loosen up — the analytic process should stay relaxed, flexible, and driven by insight from interaction with data.

• Doing description and calling it theory — the tell is a report that never integrates findings into an explanatory scheme yet claims to be theoretical, leaving readers unsure what is theory and what is not.
• Overdoing analysis into minutiae — the tell is a report so descriptive and detailed it becomes boring; you failed to know when to let an idea go.
• Treating procedures as directives rather than tools — the tell is rigid, mechanical coding where the fluid, dynamic nature of analysis is lost.
• Letting emotional identification masquerade as coding — the tell is memos that record your feelings about the data instead of conceptualizing what participants convey.
• Chasing 'objectivity' as if it were achievable — the tell is denial that your paradigms, training, and biases are woven into the process, which forfeits the sensitivity that comes from putting yourself into the research.

• ▸Source segment: paste the exact passage/quote being coded
• ▸Microanalysis notes: varied meanings of the key word or phrase
• ▸Concept name: the label for the shared property that groups instances
• ▸Properties: the defining characteristics of this concept
• ▸Dimensions: the range along which each property varies
• ▸Comparison instances: other data segments sharing this property, with similarities and differences
• ▸Reflexive note: your emotional/personal reaction — flag if it slants the data
• ▸Conflict/tension check: contradictions or conflicting codes to revisit
• ▸Feeling-right status: does this express participants' meaning? (yes / not yet / revise)

1. 1State your goal before collecting data: description, concept analysis, theme extraction, or full theory construction. Corbin insists you 'be very clear at the beginning of a study what it is you are setting out to do' — if it is description, do quality description; if theory, integrate findings into an overarching explanatory scheme.
2. 2Make your epistemological assumptions explicit rather than importing procedures blindly. Most researchers 'have not reflected upon the assumptions that underlie the method.' Write down where you stand on objectivity, multiple perspectives, and whether reality can be 'captured' in data.
3. 3Position yourself inside the study, not outside it. Accept that objectivity as total detachment is not the aim; the postmodern shift 'put the researcher right into the center of the study.' Treat data as co-constructed by researcher and participant.
4. 4Run reflexivity as a continuous practice during both collection and analysis, not a one-time confession. Track your own sadness, anger, approval, and disapproval, because 'such emotions are conveyed to participants' who adjust their stances in response.
5. 5Decide the dose of reflexivity deliberately — 'how much, when, and how' — matched to your philosophical orientation. Do not assume complete self-accounting is possible; Cutcliffe notes much transpires 'within the deeper levels of consciousness.'
6. 6Use the autobiographical analysis of self as a 'nurturing bed' for findings, per Chesney — placed alongside participants' words, not competing with them, to keep the research process transparent.
7. 7Keep analysis a researcher-driven thinking-and-feeling process even when using computer programs. Let the evolving analysis determine how the tool is used; the program should 'support' analysis, never 'take over' or 'direct' it.
8. 8Balance competing commitments explicitly: honor the value of theory (Strauss's vision) while accepting that thick description, case analysis, changing a difficult situation, and telling a story are all valid research.

Corbin's own trajectory is the worked case. Trained as a grounded theorist under Strauss when there was supposedly 'one grounded theory approach,' she went 'about business as usual' writing interpretive research and one day found herself classified and labeled a 'post-positivist' — the very act of categorizing she performed on her own data. The Qualitative Revolution had dismissed objectivity as impossible, declared capturing 'reality' in data a fantasy, and insisted on multiple perspectives. Her initial reaction was exasperation and fear that researchers would end up 'examining their own navels' and 'telling nice stories,' losing sight of building an empirical body of knowledge. Rather than reject the shift, she reflexively examined it, concluded the deconstructionist and constructionist schools made 'some valid points,' and reconstructed her stance: she still believes generating concepts is useful, yet no longer believes theory is the only route to knowledge. Teaching analysis to students across many countries fed this revision — the interactions 'helped shape my new understanding.' The result is a stance that keeps Strauss's belief in theory while holding her own conviction that description, case analysis, and storytelling are each powerful.

• If your research question can only be answered by getting at participants' inner experience and how meanings form in culture, then commit to qualitative work and frame the question so quantitative methods cannot answer it.
• If your goal is description, then do quality description and label it as such; if your goal is theory, then integrate findings into an overarching explanatory scheme — never call description theory.
• If you feel a strong emotional reaction during an interview or observation, then log it as data about co-construction, because that reaction is being read and reciprocated by the participant.
• If a subsequent observation yields no data on a previously identified concept, then ask about that concept directly at the end of the day rather than assuming it is absent.
• If you cannot fully account for yourself because processes occur below consciousness, then still treat reflexivity as a valuable tool for examining impact — partial transparency beats none.

• Treating the book or method as a recipe. The tell: procedures applied step-by-step with no adaptation to the study — Corbin says she 'would be offended if it is viewed as such.'
• Calling description theory. The tell: findings presented as a theory but never integrated into an explanatory scheme, leaving the reader confused about what is theory and what is not.
• Letting the computer program direct analysis. The tell: a rigidified analytic process driven by the software's structure rather than by the evolving analysis and the researcher's thinking.
• Running research with unexamined assumptions. The tell: a researcher who uses only the procedures and has 'not reflected upon the assumptions that underlie the method' — objectivity and reality treated as unproblematic givens.
• Reflexivity that collapses into navel-gazing. The tell: self-examination and 'nice stories' that displace the purpose of generating a professional body of knowledge, or self-analysis that competes with participants' words instead of grounding them.

• ▸Study goal: name one — description / concept analysis / theme extraction / theory construction.
• ▸Epistemological position: state your stance on objectivity and whether reality can be captured in data.
• ▸Ontological assumptions: note what you assume about the world (change, contingency, complexity, symbolic meaning).
• ▸Researcher-as-instrument acknowledgment: describe how you enter the center of the study.
• ▸Emotional log entry: record felt reactions (sad/angry/approval/disapproval) during this collection session.
• ▸Co-construction note: how did the participant adjust to your responses today?
• ▸Reflexivity dose: decide how much, when, and how you will self-examine for this phase.
• ▸Below-consciousness caveat: name what you likely cannot fully account for.
• ▸Tool use check: confirm the computer program is supporting, not directing, the analysis.

1. 1Place your current practice on the maturity continuum before choosing tools: descriptive (anecdote, reactive check, ongoing report, segment/comparison) → diagnostic (correlation) → predictive (causation, simulation and forecasting). Name where you sit so you stop over-claiming.
2. 2Start from a high-impact business problem, not the data. Ask whether solving it lifts business performance; if not, drop it. This protects against the rookie move of mining data with fancy tools for its own sake.
3. 3Act as architect before analyst: draw the 'design of analyses' blueprint (what must be analysed to solve the problem) before any data gathering, mirroring an architect starting with a blueprint, not bricks.
4. 4Build the data foundation: use the HRIS (PeopleSoft, SuccessFactors, Workday) as the system of record, then take regular periodic exports into a purpose-built HR data warehouse where data are cleansed and validated in staging areas before becoming the master record.
5. 5Classify your metrics as hard vs soft data (volume, quality, cost, time on the hard side; service, satisfaction, engagement, retention on the soft side) and confront the conversion problem — turning soft measures into monetary values.
6. 6Match analysis type to the question: descriptive for 'what happened', diagnostic for 'why/what drives it' (correlation), predictive for 'what is likely to happen' (causation, ROI forecasts).
7. 7Confirm an engaged stakeholder/sponsor and check that intended actors will understand and believe the insight, because analysis that isn't acted on produces no organisational value.
8. 8Overlay analytics maturity on HR-function maturity — as HR moves from Basic Personnel Administration through People Development, Line Management Empowerment, Value Addition to Anticipating the Future, raise the analytics tier to match.