Compensation analytics turns pay records into defensible decisions about cost, employee positioning and program outcomes. This guide develops the skills needed to clean and align compensation data, interpret distributions, calculate meaningful metrics, evaluate regression results and communicate recommendations without overstating evidence. Its 20 original multiple-choice questions and 60 concepts cover the single Compensation Analytics and Insights domain. This representative preview follows the current eight-course CCP structure as of October 16, 2026; it is not a full-length mock or official exam material. Source s2 supports exam identity and format; s3 supports subject relevance, rather than verifying every technical explanation or establishing detailed subtopic weights.
What this guide covers
Build compensation analytics skills with 20 original questions and 60 concepts on pay data and statistics.
Compensation Analytics and Insights
1. Evidence-based compensation decisions
Evidence-based compensation practice connects a defined business question to relevant data, appropriate analysis and an explicit decision. Start with the decision and competing explanations, then select measures that can distinguish those explanations. Combine organizational evidence with relevant research and practical constraints rather than treating one spreadsheet result as decisive. The strength of the recommendation should match the strength of the design. A descriptive pay difference can identify where to investigate, while a credible intervention comparison can support a stronger claim about program effects.
Worked example: Recruiting leaders attribute offer rejections to salary. The analyst compares rejection reasons, offered pay relative to the relevant market and time to decision. A concentration of delays suggests testing approval speed alongside proposed pay changes.
Common mistake: Treating the first measurable association as the complete explanation can direct compensation spending toward a symptom rather than its cause.
Official syllabus reference: [2]
2. Operational definitions make measures reproducible
An operational definition specifies exactly how a concept becomes a measurable variable. For compensation analysis, define pay components, employee eligibility, effective dates, numerator, denominator and treatment of exceptions. For turnover, distinguish voluntary departures from all separations and identify the workforce exposure used in the denominator. Two analysts should obtain the same result from the same records using the definition. A label such as competitive pay is insufficient until the reference market, pay basis and comparison calculation are specified for the decision.
Worked example: One dashboard counts every payment as base pay, while another excludes overtime and allowances. The team defines annual contractual base salary at full-time hours, then rebuilds both comparisons using that same pay field and date.
Common mistake: Assuming familiar labels guarantee comparable measures overlooks differences in pay components, employee coverage and the dates represented by each record.
Official syllabus reference: [2]
3. KPIs need a decision link and denominator
A key performance indicator should measure progress toward a specific objective and support an identifiable action. Its formula needs a stable numerator, denominator, population and reporting period. Distinguish program activity, such as completed pay reviews, from outcomes, such as reduced offer rejection in a target group. Leading indicators may signal emerging conditions, while lagging indicators summarize results already realized. Pair a KPI with relevant balancing measures so improvement in one dimension does not hide rising cost or deterioration in another.
Worked example: A hiring initiative tracks accepted offers divided by decided offers, rather than all open offers. Leaders review that rate alongside offered salary and time to approval, allowing them to distinguish cost changes from process improvements.
Common mistake: Reporting a rising count without its exposure denominator can make workforce growth look like improved program performance when the underlying rate declined.
Official syllabus reference: [2]
4. Data grain determines what a row means
Data grain is the unit represented by one row, such as an employee, assignment, pay transaction or employee-month. Establish it before calculating counts or sums. Multiple assignments can legitimately produce several rows for one employee, while a monthly panel repeats employees across periods. A measure is valid only when aggregation respects that structure. Employee counts may require distinct identifiers, whereas transaction totals require all valid payments. Changing grain without redefining measures can duplicate salaries, distort averages and inflate apparent sample sizes.
Worked example: An employee holds two assignments and receives twelve monthly payments for each. The analyst uses assignment-month rows for payment reconciliation, then constructs a separate employee snapshot before calculating the distribution of annual contractual salaries.
Common mistake: Counting rows as employees silently overstates headcount whenever the dataset contains repeated periods, multiple assignments or several transactions for one person.
Official syllabus reference: [2]
5. Data dictionaries preserve analytical meaning
A data dictionary records each variable's meaning, type, unit, allowed values, source and missing-value conventions. It distinguishes fields that look similar but represent different constructs, such as target incentive opportunity and actual incentive paid. Include derivations and version changes so later users can reproduce calculations. A dictionary does not replace validation: records must still be tested against its definitions. Its analytical value is preventing silent changes in meaning when data move between payroll, human resources systems, surveys and reporting tools.
Worked example: A field named bonus changes from target opportunity to paid awards after a system migration. The dictionary records the change, allowing analysts to split the series and avoid interpreting the resulting jump as improved incentive performance.
Common mistake: Relying on column names alone can merge conceptually different pay measures and create trends that reflect system changes rather than employee compensation.
Official syllabus reference: [2]
6. Join cardinality prevents payroll inflation
Join cardinality describes how many records on one side can match records on the other. A one-to-many join repeats values from the one side, which can inflate subsequent sums. Before joining, test key uniqueness and establish the intended output grain. Aggregate transaction records when the output requires employee-level totals, or retain the transaction grain and avoid summing repeated employee attributes. Reconcile row counts, unmatched keys and control totals afterward. A technically successful join can still produce analytically incorrect compensation results.
Worked example: A salary file has one row per employee, while an award file has three awards for one employee. Summing awards first permits a one-to-one join that preserves the salary once and retains the full award total.
Common mistake: Removing duplicate employee IDs after a transaction join can discard legitimate payments instead of correcting the mismatch between the datasets' grains.
Official syllabus reference: [2]
7. Missingness can change the population analyzed
Missing values can arise independently of the analysis variables, depend on observed characteristics or depend on unobserved values themselves. These mechanisms affect whether complete-record analysis is credible. First distinguish missing data from genuine zero amounts and structural inapplicability. Examine missingness by job, location, period and source system. Deleting incomplete records can systematically exclude particular groups. Imputation may be appropriate under explicit assumptions, but it introduces uncertainty and cannot recover information without assumptions about how the missing values relate to observed data.
Worked example: Incentive payments are absent for many new hires because they were ineligible, while several eligible employees have failed imports. The analyst codes ineligibility separately and resolves import failures instead of converting every blank to zero.
Common mistake: Replacing every blank with zero confuses absent information with a valid absence of pay and can materially depress averages.
Official syllabus reference: [2]
8. Duplicate detection requires a business key
A duplicate is a repeated record that should occur only once under the intended data definition. Identify the business key before deleting rows: employee ID may be insufficient when employees have multiple assignments, dates or award types. Exact duplicates can result from repeated imports, while conflicting duplicates require source reconciliation. Compare amounts and timestamps, document retention rules and verify resulting control totals. Legitimate repeated transactions must remain. Deduplication is therefore a substantive data decision rather than a blanket command to retain one row per employee.
Worked example: Two records share an employee ID and payment date but have different award IDs. They represent separate awards. A third record repeats an existing award ID and amount after a second import, so only that import duplicate is removed.
Common mistake: Using employee ID as the sole duplicate key can erase valid assignments or awards and understate compensation totals.
Official syllabus reference: [2]
9. Pay basis and units must be normalized
Comparable pay analysis requires a common unit and pay basis. Hourly rates, monthly salaries and annual earnings describe different quantities unless converted using appropriate hours or periods. Distinguish contractual rates from realized earnings, which can include overtime, absence or partial-year service. State conversion assumptions explicitly and retain original values for checking. A standard annual-hours assumption may suit one scenario but not another workforce. Unit normalization removes measurement incompatibility; it does not make different jobs, locations or pay components substantively equivalent.
Worked example: One file reports $30 per hour and another reports $62,400 annually. Under the exercise's stated 2,080 paid hours per year, both represent the same annualized rate. Actual earnings may still differ if overtime or unpaid leave occurs.
Common mistake: Applying one annual-hours conversion to every schedule can introduce artificial pay differences when contractual hours or paid periods differ.
Official syllabus reference: [2]
10. Effective dates separate snapshots from flows
A salary snapshot records a rate at a particular date, while payroll earnings accumulate over an interval. Comparing them without aligning time can produce misleading differences. Use effective-dated records to determine which salary, job and location applied when an event occurred. For period costs, prorate rates across the applicable intervals using the organization's specified convention. Avoid attaching today's salary to historical turnover or performance without justification. Temporal alignment is essential because compensation decisions and workforce changes often occur at different points within the reporting period.
Worked example: An employee receives a raise in October. The December salary snapshot shows the new annual rate, but the year's earnings include nine months at the old rate. The analyst uses effective dates to reconcile the two measures.
Common mistake: Multiplying a year-end salary rate by headcount and calling it actual annual payroll ignores timing, vacancies and partial-year employment.
Official syllabus reference: [2]
11. FTE salary and actual earnings answer different questions
Full-time-equivalent salary standardizes a pay rate to a full schedule, while actual earnings describe what the employee receives for the schedule worked. When pay is proportional to hours, divide actual annual scheduled pay by the FTE fraction to obtain the full-time-equivalent rate. Use standardized rates for comparable rate analysis and actual scheduled amounts for payroll cost. Confirm what the source field already represents before converting it. An amount already recorded at full-time equivalent must not be annualized or schedule-adjusted a second time.
Worked example: A 0.75-FTE employee receives $45,000 annually for the scheduled role. The full-time-equivalent rate is $45,000 divided by 0.75, or $60,000. The payroll budget still includes $45,000 unless the employee's schedule changes.
Common mistake: Using full-time-equivalent salary as actual payroll cost overstates spending for part-time employees and confuses pay rates with earnings.
Official syllabus reference: [2]
12. Currency conversion and constant-currency analysis
Currency conversion expresses amounts in a common reporting currency using a defined exchange-rate convention. Specify the rate's date or averaging period and quote direction before calculating. Exchange movements can change reported payroll even when local salaries are unchanged. Constant-currency analysis applies a common rate across comparison periods to isolate that translation effect. It does not adjust for purchasing power, inflation or differences in local labor markets. Keep local-currency results available so leaders can distinguish operating compensation changes from movements caused by financial translation.
Worked example: Local annual payroll remains €1 million. At $1.10 per euro it reports as $1.10 million; at $1.05 it reports as $1.05 million. A constant $1.10 rate shows no local-payroll change despite the reported dollar decline.
Common mistake: Interpreting a reporting-currency decline as a local pay reduction overlooks exchange-rate movements and can misstate employees' compensation experience.
Official syllabus reference: [2]
13. Selection bias limits generalization
Selection bias occurs when inclusion in the analyzed data relates to characteristics relevant to the result. A voluntary survey may overrepresent larger employers or unusually engaged respondents. A sample of accepted offers excludes candidates who rejected the employer's terms. Larger sample size reduces random sampling variability but does not automatically correct systematic selection. Define the target population, compare included and excluded groups where possible and qualify generalizations. Weighting can address known representation differences, but it cannot reliably correct unmeasured selection without additional assumptions.
Worked example: An analyst estimates competitive hiring pay using only accepted offers. Because rejected offers may have been less attractive, the observed dataset cannot fully describe the salaries needed to attract every candidate in the intended market.
Common mistake: Assuming thousands of records guarantee representativeness confuses sample size with whether the right employers, employees or decisions were observed.
Official syllabus reference: [2]
14. Weights define whose experience an average represents
Weights determine how much each observation contributes to an estimate. An employer-weighted salary average gives each participating employer equal influence, whereas an incumbent-weighted average gives larger employee groups more influence. Neither automatically answers every question. Identify whether the intended estimand is the typical employer's reported pay or the average observed employee's pay. Sampling weights serve another purpose by correcting known inclusion probabilities or representation. Document weight construction and avoid applying the same weighting adjustment twice when a supplied statistic is already weighted.
Worked example: One employer reports 10 employees averaging $50,000; another reports 90 averaging $70,000. The employer-weighted average is $60,000, while the incumbent-weighted average is $68,000. Each answers a different question about the same underlying reports.
Common mistake: Calling one weighting method universally more accurate ignores the population and decision the average is supposed to represent.
Official syllabus reference: [2]
15. Arithmetic means reconcile to total pay
The arithmetic mean equals the sum of observations divided by their count. Its connection to totals makes it useful for payroll costing: mean pay multiplied by the corresponding employee count reproduces total pay when definitions align. Every value influences the mean, so high salaries can pull it away from what most employees earn. Report population and pay basis alongside the result. The mean is not necessarily an observed salary, and it should not be interpreted as the middle employee's pay in an asymmetric distribution.
Worked example: Four comparable employees earn $40,000, $45,000, $50,000 and $105,000. Their mean is $60,000, and four times that mean reconciles to $240,000 in total salary. Three employees nevertheless earn less than the reported average.
Common mistake: Describing the mean as what the typical middle employee receives can mislead readers when a few high salaries raise the average.
Official syllabus reference: [2]
16. Weighted means require the correct exposure base
A weighted mean equals the sum of each value multiplied by its weight, divided by total weight. The correct weight depends on the quantity being combined. Use employee counts to combine group mean salaries and payroll amounts to combine percentage payroll increases. Averaging percentages without their exposure bases can materially misstate the organization-wide result. Check that groups are mutually exclusive and that weights correspond to the same period and population. A weighted result should fall between the component values when all weights are positive.
Worked example: A department with $1 million payroll receives 5%, while a department with $4 million receives 2%. The combined increase is $130,000 divided by $5 million, or 2.6%, rather than the unweighted average of 3.5%.
Common mistake: Using headcount to weight payroll increase percentages can misstate cost when the groups have different average salaries and payroll sizes.
Official syllabus reference: [2]
17. Medians locate the distribution's midpoint
The median divides ordered observations into lower and upper halves. For an odd count, select the middle observation; for an even count, the conventional sample median averages the two central observations. Because it depends on order rather than every value's magnitude, the median is resistant to extreme pay values. It is useful for describing central salary position in skewed data. However, median pay multiplied by headcount does not generally reproduce payroll, and the median of subgroup medians does not generally equal the overall employee median.
Worked example: Six annual salaries are $42,000, $46,000, $50,000, $54,000, $58,000 and $130,000. The median averages $50,000 and $54,000, giving $52,000. Raising the highest salary further would not change that median.
Common mistake: Averaging departmental medians and labeling the result the organization-wide median ignores department sizes and the ordering of individual salaries.
Official syllabus reference: [2]
18. Percentiles require a stated calculation convention
A percentile identifies a location within an ordered distribution. Sample percentile methods differ in how they assign positions and interpolate between observations, especially in small datasets. Use the method specified by the exercise or reporting system and document it when comparing outputs. Percentile rank and percentile value are different: one describes where an observation sits, while the other gives the amount at a selected distribution location. Percentiles summarize the observed reference population and do not automatically prescribe what every employee should receive.
Worked example: Using position p times n plus one, the 75th percentile of eight salaries lies at position 6.75. If positions six and seven are $68,000 and $80,000, interpolation yields $77,000 under that specified convention.
Common mistake: Treating different software percentile results as data errors before checking interpolation methods can trigger unnecessary corrections to otherwise valid calculations.
Official syllabus reference: [2]
19. Range and interquartile range describe different spread
The statistical range is maximum minus minimum and therefore depends entirely on two observations. The interquartile range, or IQR, is the third quartile minus the first quartile and describes the spread of the central half. Use range to describe the full observed span and IQR to summarize central dispersion with less influence from extremes. Both retain the original measurement units. Neither identifies why pay varies. Distinguish these distribution statistics from an administrative salary range, which is a designed compensation structure rather than an observed-data summary.
Worked example: A salary dataset has a minimum of $40,000, first quartile of $52,000, third quartile of $68,000 and maximum of $180,000. Its full range is $140,000, while the central-half spread is only $16,000.
Common mistake: Interpreting a large maximum-minus-minimum span as evidence that most salaries are widely dispersed overlooks the concentration of the central observations.
Official syllabus reference: [2]
20. Variance and standard deviation quantify dispersion
Variance averages squared deviations from the mean, using the denominator appropriate to a population or sample calculation. Sample variance commonly divides the sum of squared deviations by n minus one when estimating population variance. Standard deviation is its square root and returns to the original units, making salary spread easier to interpret. Both are sensitive to extreme observations because deviations are squared. Specify which formula the task requires. Standard deviation describes variability among observations; it is not the uncertainty of the estimated mean.
Worked example: Three salaries are $40,000, $50,000 and $60,000, with mean $50,000. Their squared deviations total 200 million dollars squared. Dividing by two gives sample variance of 100 million, and its square root gives a $10,000 sample standard deviation.
Common mistake: Using standard deviation as the margin of error for average pay confuses employee-to-employee dispersion with uncertainty about the population mean.
Official syllabus reference: [2]
21. Coefficient of variation compares relative spread
The coefficient of variation divides standard deviation by the mean, often expressing the result as a percentage. It compares dispersion relative to the scale of positive-valued data, such as salaries in groups with different average pay. A larger standard deviation does not necessarily mean greater relative variation if the mean is also much larger. Use the coefficient cautiously when means are near zero or when variables lack a meaningful ratio scale. It describes dispersion and does not establish fairness, consistency of job content or an appropriate compensation policy.
Worked example: A group averaging $50,000 with a $5,000 standard deviation has a 10% coefficient of variation. Another averaging $100,000 with an $8,000 standard deviation has 8%, despite having the larger absolute dollar spread.
Common mistake: Ranking relative salary variability using standard deviations alone ignores differences in average pay and can reverse the comparison.
Official syllabus reference: [2]
22. Skewness changes the meaning of average pay
Skewness describes asymmetry in a distribution. A right-skewed salary distribution has an extended upper tail, often producing a mean above the median because high values influence the mean strongly. A left-skewed distribution has an extended lower tail. Inspect the actual distribution rather than inferring shape from one summary statistic alone. Skewness can affect how useful normal approximations and symmetric summaries are. It is not automatically a data defect: a valid mixture of job levels or a small executive population can produce substantial asymmetry.
Worked example: Most employees earn between $45,000 and $75,000, while a small leadership group earns above $200,000. The analyst reports both median and mean, explaining that the upper tail raises payroll average above the middle employee's salary.
Common mistake: Removing every high salary to make the distribution symmetric changes the population and can erase legitimate compensation differences.
Official syllabus reference: [2]
23. Histograms reveal shape through binning
A histogram groups a quantitative variable into intervals and displays how observations are distributed across them. It can reveal skew, multiple peaks, gaps and unusual concentrations. Bin width affects the picture: very broad bins hide structure, while very narrow bins emphasize random irregularities. Compare relevant views rather than selecting the version that best supports a preferred claim. Keep interval definitions and units clear. For unequal bin widths, area rather than raw bar height must represent frequency correctly, requiring an appropriate density scale.
Worked example: A single broad salary histogram appears smooth, but narrower, sensible intervals reveal two peaks. Splitting the data by job family shows that the peaks represent distinct employee populations rather than an unexplained irregularity in pay administration.
Common mistake: Changing histogram bins until a preferred pattern appears mistakes a display choice for evidence that the underlying distribution changed.
Official syllabus reference: [2]
24. Boxplots compare center, spread and flagged points
A boxplot typically shows the median, first and third quartiles, whiskers and separately plotted observations beyond a chosen whisker rule. It is useful for comparing salary distributions across groups without displaying every record. Whiskers do not have one universal definition, so identify the plotting convention. Under a common rule, points beyond 1.5 IQR from the quartiles are flagged for attention. Such points are not automatically errors. Boxplots also hide sample detail, so provide counts and inspect underlying data when group sizes or shapes differ substantially.
Worked example: Two departments have similar medians, but one has a much wider box and several flagged high salaries. The analyst investigates its job-level mix before interpreting the display as inconsistent treatment of employees doing comparable work.
Common mistake: Calling every point beyond a boxplot whisker an invalid salary turns a screening convention into an unsupported data-deletion rule.
Official syllabus reference: [2]
25. Outlier treatment follows verification and purpose
An outlier is unusual relative to a defined comparison, but unusual does not mean incorrect. Investigate units, employee identity, pay components, effective dates and source records before deciding treatment. Correct verified errors; retain valid observations when they belong to the target population. Robust summaries, separate subgroup reporting or sensitivity analyses can address influential valid values. Trimming removes observations, while winsorizing replaces extremes with specified boundary values. Both alter the analysis and require a defensible, documented rule rather than adjustment to obtain a preferred result.
Worked example: A $900,000 entry could be a valid executive salary or a monthly-to-annual conversion error. Source verification resolves the distinction. The analyst then reports how including a valid executive affects the mean without silently capping the amount.
Common mistake: Automatically winsorizing unusual salaries before checking their source can conceal extraction errors and distort valid high-pay observations at the same time.
Official syllabus reference: [2]
26. Logarithms and geometric means capture multiplicative scale
A logarithmic transformation expresses proportional differences additively and can make strongly right-skewed positive pay data easier to analyze. The geometric mean equals the exponential of the mean natural logarithm. It summarizes multiplicative central tendency and usually differs from the arithmetic mean used for payroll totals. Logs require positive values, so zeros and negative adjustments need separate consideration. Transforming data changes the interpretation of summaries and model coefficients. It does not guarantee a well-fitting model or remove the need to examine the original dollar scale.
Worked example: Two salaries are $40,000 and $90,000. Their arithmetic mean is $65,000, while their geometric mean is the square root of their product, or $60,000. The geometric mean lies midway between them on a multiplicative scale.
Common mistake: Multiplying a geometric mean by headcount to estimate payroll generally understates the total because payroll reconciles through the arithmetic mean.
Official syllabus reference: [2]
27. Standard scores express distance in standard deviations
A standard score, or z-score, subtracts a distribution's mean from an observation and divides by its standard deviation. It expresses relative position in standard-deviation units, allowing comparisons across differently scaled groups. A positive score is above the mean and a negative score is below it. Calculating a z-score does not require normality, but converting it into a normal-distribution percentile does. Use a meaningful reference group: standardizing within unlike job populations can describe relative position without making their underlying pay levels equivalent.
Worked example: An employee earns $84,000 in a comparable group with mean $72,000 and standard deviation $6,000. The z-score is two. This establishes distance above that group's mean, but a normal percentile requires an additional distributional assumption.
Common mistake: Reading every z-score as a normal percentile ignores distribution shape and can misstate how unusual a salary actually is.
Official syllabus reference: [2]
28. Normal models are assumptions, not salary facts
A normal distribution is symmetric and characterized by its mean and standard deviation. Under that model, about 68% of observations lie within one standard deviation and about 95% within two. These proportions are model properties, not universal salary thresholds. Actual compensation data may be skewed, bounded or mixtures of different job populations. Check whether a normal approximation suits the specific variable and purpose. Averages can have approximately normal sampling distributions under suitable conditions even when individual salaries are not normally distributed.
Worked example: A homogeneous group's salaries are modeled as approximately normal with mean $70,000 and standard deviation $5,000. The model places about 95% between $60,000 and $80,000; the analyst checks the observed distribution before using that approximation.
Common mistake: Applying normal-distribution percentages to a mixed executive and hourly workforce can produce misleading expectations about both tails of the salary distribution.
Official syllabus reference: [2]
29. Standard error measures estimation precision
The standard error describes variability of an estimator across repeated samples, rather than variation among individual employees. For an independent simple sample, the estimated standard error of the mean is the sample standard deviation divided by the square root of sample size. Larger samples generally improve precision under that design. Clustering, unequal weights or repeated employee observations can require a different calculation. A small standard error does not correct selection bias, inaccurate pay records or an estimator that answers the wrong compensation question.
Worked example: A sample of 100 independent observations has salary standard deviation $12,000. Its mean's estimated standard error is $1,200. Increasing the sample to 400 with similar dispersion reduces that standard error to $600, not to one-quarter of its original value.
Common mistake: Treating repeated monthly records from the same employees as independent observations can make estimated precision appear substantially better than it is.
Official syllabus reference: [2]
30. Confidence intervals communicate compatible parameter values
A confidence interval combines an estimate with uncertainty calculated under a specified statistical procedure and assumptions. Its confidence level describes the procedure's long-run coverage across repeated samples. The resulting interval identifies parameter values compatible with the data under that approach. It is not an interval containing a stated percentage of employee salaries. Wider intervals indicate less precise estimation. When analyzing a complete current workforce, distinguish the exact descriptive difference for that workforce from model-based uncertainty about a broader process or future population.
Worked example: A sampled market mean is estimated at $82,000 with a 95% interval from $79,000 to $85,000. The interval describes uncertainty about the market mean under the sampling assumptions, not the range containing 95% of individual market salaries.
Common mistake: Presenting a confidence interval for the mean as the expected spread of individual salaries understates employee-level variation and mislabels statistical uncertainty.
Official syllabus reference: [2]
31. Hypothesis tests assess incompatibility with a null model
A hypothesis test compares observed evidence with what a specified null model would predict. A p-value is the probability, assuming that model and its assumptions, of a result at least as extreme as the observed statistic. It is not the probability that the null is true. Choose the question, direction and decision criterion before examining outcomes. Failure to reject does not prove equality. Test validity depends on the design, data quality and model; a numerical p-value cannot repair biased selection or poorly defined comparison groups.
Worked example: A pay-difference test returns p = 0.08 against a predefined 0.05 criterion. The analyst reports insufficient evidence to reject the null under that rule, together with the estimated difference and interval, rather than declaring the groups equal.
Common mistake: Interpreting p = 0.03 as a 3% chance that a pay difference is accidental assigns the p-value a meaning it does not have.
Official syllabus reference: [2]
32. Practical significance connects effects to decisions
Statistical significance concerns evidence relative to a null model, while practical significance concerns the size and consequences of an effect. Evaluate dollar magnitude, percentage difference, number of affected employees, cost and decision context. Large samples can make very small differences statistically detectable, while small samples can leave consequential differences uncertain. Set relevant decision criteria explicitly rather than implying a universal materiality threshold. Report effect estimates and uncertainty together so leaders can distinguish a precise but minor difference from an imprecise result with substantial potential impact.
Worked example: A large workforce shows a statistically detectable $120 annual difference, while a small specialist group shows an uncertain $8,000 difference. Leaders assess scale, employee coverage and uncertainty separately rather than ranking importance solely by the smaller p-value.
Common mistake: Treating statistical significance as proof that a pay difference warrants the same intervention everywhere ignores magnitude, coverage and the decision's consequences.
Official syllabus reference: [2]
33. Multiple comparisons increase false-positive opportunities
Testing many departments, outcomes or model variants creates more opportunities for chance findings. Under independent true nulls and a 5% criterion, twenty tests produce about one false positive on average, though any particular set can differ. Define primary questions in advance and distinguish confirmatory analysis from exploration. Depending on the objective, methods can control familywise error or the false discovery rate. Adjustment does not replace effect-size interpretation. Transparently report the search scope so a selected finding is not presented as the only comparison performed.
Worked example: An analyst tests pay differences across thirty small units and highlights the only result below 0.05. The report instead identifies all comparisons, applies the chosen multiplicity approach and treats the isolated finding as a lead for further investigation.
Common mistake: Running many variations and reporting only the most favorable p-value makes exploratory selection look like independent confirmation of a compensation claim.
Official syllabus reference: [2]
34. Correlation measures linear association
Pearson correlation measures the direction and strength of a linear relationship between two quantitative variables. It ranges from negative one to positive one and is unitless. A correlation near zero can coexist with a strong curved relationship, so inspect a scatterplot. Outliers and restricted ranges can substantially change the result. Correlation does not identify direction of causation or remove common causes. In a simple linear regression with an intercept, its square equals R-squared, but that equivalence should not be generalized indiscriminately to other models.
Worked example: Salary and experience correlate at 0.70 within a job family. The corresponding simple linear model has R-squared 0.49. This describes their observed linear relationship, not a claim that experience causes 49% of every employee's salary.
Common mistake: Interpreting a correlation of 0.70 as 70% of salary explained confuses the correlation coefficient with the corresponding squared measure.
Official syllabus reference: [2]
35. Composition effects can reverse aggregate comparisons
An aggregate pay comparison combines within-group pay differences with differences in workforce composition. If one population contains more senior employees, its overall mean can be higher even when it earns less within every level. This reversal is often called Simpson's paradox. Compare meaningful strata or standardize both populations to common weights to isolate the chosen composition effect. Standardization answers a specified comparison and does not establish causation. The selected weights and grouping variables must be explicit because different reference populations can yield different standardized results.
Worked example: Group A averages $50,000 in junior roles and $100,000 in senior roles; Group B averages $48,000 and $96,000. Giving both populations equal junior and senior weights yields $75,000 versus $72,000, regardless of their actual level mixes.
Common mistake: Reading a higher organization-wide average as evidence of higher pay for comparable work overlooks differences in the jobs each population holds.
Official syllabus reference: [2]
36. Simple regression separates intercept and slope
Simple linear regression models an outcome as an intercept plus a slope multiplied by one predictor. The slope describes the predicted outcome difference associated with one predictor-unit difference. The intercept is the predicted outcome at zero, which may lack substantive meaning if zero lies outside the observed range. Ordinary least squares chooses coefficients that minimize the sum of squared residuals. Use units to interpret coefficients and inspect fit. A fitted salary relationship summarizes observed associations and does not by itself identify the effect of changing the predictor.
Worked example: A model predicts annual salary as $48,000 plus $3,000 per year of relevant experience. At six years it predicts $66,000. The $3,000 slope is a dollar difference per experience year, not a 3% annual salary increase.
Common mistake: Interpreting a cross-sectional experience slope as the raise every employee will receive next year confuses between-person association with individual change.
Official syllabus reference: [2]
37. Multiple regression estimates conditional relationships
Multiple regression relates an outcome to several predictors simultaneously. A coefficient describes the modeled outcome difference for one predictor while holding the included others fixed. This conditional interpretation can differ from a raw comparison because workforce characteristics overlap. Controls must reflect the analytical question and a defensible understanding of how variables relate. Adding more variables does not automatically improve validity; inappropriate controls can change the estimand or introduce bias. Model results remain conditional on recorded variables, coding choices, functional form and the population analyzed.
Worked example: A salary model includes job level, location and relevant experience. The location coefficient compares modeled salaries at the same included level and experience, rather than comparing each location's actual workforce average with its different mix of jobs.
Common mistake: Calling an adjusted coefficient fully explained or causal assumes the model captured every relevant factor and represented their relationships correctly.
Official syllabus reference: [2]
38. Categorical coding defines the reference comparison
Categorical predictors require a coding scheme that makes comparisons explicit. With an intercept, a common approach represents a category with k levels using k minus one indicator variables, leaving a reference category. Each coefficient then compares its category with that reference, conditional on the other predictors. Changing the reference changes coefficient labels and values but not fitted predictions when the model is otherwise equivalent. Avoid treating arbitrary category codes as numerical distances. Job families numbered one, two and three do not necessarily form an equally spaced scale.
Worked example: A model uses North as the reference location and includes South and West indicators. A South coefficient of $4,000 means a modeled $4,000 difference from North at the same included characteristics, rather than a difference from the overall mean.
Common mistake: Entering location codes as one continuous predictor imposes an artificial ordering and equal spacing that the categories may not possess.
Official syllabus reference: [2]
39. Residuals reveal patterns the model misses
A residual is observed outcome minus fitted outcome. Positive salary residuals indicate pay above the model's prediction; negative residuals indicate pay below it. Residual plots can reveal curvature, changing spread, clusters and unusual observations that a single fit statistic hides. A large residual is not automatically an error or an unjustified salary: it may reflect omitted job information, valid individual circumstances or model misspecification. Examine residuals alongside source records and model assumptions. Their interpretation depends on whether the fitted model is appropriate for the analytical purpose.
Worked example: A model predicts $80,000 for an employee paid $86,000, giving a $6,000 residual. The analyst checks whether a specialized responsibility was omitted before interpreting the employee as unusually highly paid relative to genuinely comparable work.
Common mistake: Using every negative residual as an automatic adjustment amount treats the model as a complete pay policy rather than an imperfect analytical representation.
Official syllabus reference: [2]
40. Heteroscedasticity affects conventional uncertainty estimates
Heteroscedasticity means the variance of model errors changes across predictor values or fitted outcomes. Salary models may show wider dollar variation at higher pay levels, producing a fan-shaped residual plot. Under otherwise suitable conditions, this need not bias ordinary least squares coefficients, but conventional constant-variance standard errors can be unreliable. Appropriate responses may include heteroscedasticity-robust standard errors, a justified transformation or a variance model. Robust standard errors address inference under relevant assumptions; they do not repair omitted variables, nonlinearity or inaccurate compensation records.
Worked example: Residuals cluster within about $3,000 at low predicted salaries but spread much more widely at high predictions. The analyst examines a log-salary specification and robust standard errors rather than assuming one constant dollar variance fits every level.
Common mistake: Assuming robust standard errors fix the entire model overlooks possible curvature, omitted predictors and systematic errors in the salary data.
Official syllabus reference: [2]
41. Multicollinearity separates prediction from attribution
Multicollinearity occurs when predictors contain overlapping information, making their separate coefficients difficult to estimate precisely. Experience and years since career entry may predict salary jointly while providing little information to distinguish their individual associations. Coefficients can become unstable and standard errors large even when predictions remain useful. Examine predictor relationships and diagnostics such as variance inflation factors, interpreted in context rather than through an unsupported universal cutoff. Resolve redundancy according to the model's purpose and variable meaning, not by mechanically deleting whichever predictor has the largest p-value.
Worked example: Replacing one sample with another barely changes predicted salaries, but experience and career-length coefficients swap magnitude. Because the variables nearly coincide, the analyst selects the measure matching the intended construct instead of interpreting both coefficients as independent effects.
Common mistake: Treating stable overall predictions as proof that every individual coefficient is stable overlooks the distinct estimation problem created by redundant predictors.
Official syllabus reference: [2]
42. Log-salary coefficients require back-transformation
When natural log salary is the outcome, a coefficient represents an additive log difference and a multiplicative salary relationship. For a one-unit predictor difference, the exact percentage difference is 100 times exp(coefficient) minus one, holding included predictors fixed. Multiplying the coefficient by 100 is a small-value approximation. Exponentiating a fitted log value does not generally recover the conditional arithmetic mean salary without considering retransformation. Distinguish proportional modeled differences from dollar differences and keep the reference category clear when interpreting group indicators.
Worked example: A group coefficient is −0.08. With exp(−0.08) approximately 0.9231, the exact modeled difference is about 7.7% lower than the reference, rather than exactly 8%. Reversing the comparison produces a different percentage because the denominator changes.
Common mistake: Reporting every log coefficient as an exact percentage can introduce avoidable errors, especially for larger coefficients or reversed reference comparisons.
Official syllabus reference: [2]
43. Prediction intervals include individual variation
A confidence interval for a fitted mean describes uncertainty about the average outcome at specified predictor values. A prediction interval for a new individual outcome also includes residual individual variation, so it is generally wider under the same model and confidence level. The distinction matters when communicating expected employee salaries versus uncertainty about average pay relationships. Both depend on model assumptions and where the predictor values lie. A narrow mean interval should not be presented as evidence that nearly every employee with those characteristics will earn a similar amount.
Worked example: At eight years of experience, a model gives a mean salary interval of $70,000 to $74,000 but an individual prediction interval of $58,000 to $86,000. The wider interval reflects employee-level variation beyond uncertainty about the mean.
Common mistake: Using a narrow confidence interval for average salary as an individual employee's expected pay band omits the model's residual variation.
Official syllabus reference: [2]
44. Extrapolation extends beyond the evidence
Interpolation predicts within the region represented by the data, while extrapolation predicts beyond it. A relationship fitted among employees with two to ten years of experience may behave differently at twenty years. Extrapolation can also arise from combinations of predictors absent from the sample, even when each variable individually lies within its observed range. Identify the model's support before applying it. If a decision requires unsupported predictions, obtain relevant data or present the extension explicitly as an assumption-driven scenario with appropriately limited confidence.
Worked example: A salary model is built entirely from entry and intermediate analysts. Applying it to a senior specialist with twenty years of experience assumes the same relationship continues into an unobserved role and experience range, requiring separate validation.
Common mistake: Checking only whether each predictor lies between its minimum and maximum can miss unsupported combinations of job level, location and experience.
Official syllabus reference: [2]
45. Unadjusted pay gaps combine several mechanisms
An unadjusted pay gap compares aggregate pay between defined groups without controlling for differences in recorded characteristics. Specify whether it uses means or medians and identify the reference denominator. The result can reflect job distribution, location, working hours, tenure and within-job differences together. It is valuable as a descriptive outcome but cannot isolate a single explanation. Report population and pay basis clearly. An unadjusted gap neither proves nor disproves unlawful treatment, and legal interpretation depends on the applicable jurisdiction and relevant facts.
Worked example: Group A averages $80,000 and Group B $72,000. Using Group A as reference gives a 10% lower mean for Group B. The analyst then examines job and schedule composition before attributing that aggregate difference to any specific mechanism.
Common mistake: Calling an aggregate gap a direct measure of unequal pay for identical work skips the comparability analysis needed to support that interpretation.
Official syllabus reference: [2]
46. Adjusted pay gaps depend on model choices
An adjusted pay gap estimates a group difference conditional on included predictors and the model's specification. It can help distinguish recorded workforce differences from a remaining modeled association, but it is not a universal fairness score. Job level, performance and prior salary may themselves reflect earlier decisions, so their inclusion changes the question being answered. Compare defensible specifications and explain their purposes. Neither a remaining coefficient nor its disappearance establishes legal compliance or causation. Statistical findings require substantive investigation of pay practices and applicable jurisdictional requirements.
Worked example: A raw gap is 12%, while a model including job level and location estimates 3%. The report explains that the adjusted result compares employees at the same included characteristics and separately examines why the groups occupy different levels.
Common mistake: Describing every control variable as a legitimate explanation assumes its measurement and causal role are appropriate without examining how it arose.
Official syllabus reference: [2]
47. Comparable groups balance relevance and information
A comparison group should reflect the pay decision under study, using meaningful job content, level, location and compensation arrangements. Overly broad groups mix unlike work; excessively narrow groups can leave too little information to estimate patterns reliably. Examine whether categories are consistent across systems and whether employees genuinely share the relevant pay-setting process. Statistical grouping is an analytical choice and may differ from legally defined comparators in a particular jurisdiction. State that boundary and avoid treating an internal job label as sufficient proof of comparable work.
Worked example: Two employees share an analyst title, but one manages enterprise systems and another performs routine reporting. The analyst reviews responsibilities and level before pooling them, preventing a common title from masking materially different pay-setting characteristics.
Common mistake: Assuming identical titles establish comparability overlooks differences in responsibility, skill requirements and the actual processes used to determine compensation.
Official syllabus reference: [2]
48. Compa-ratio measures position against midpoint
Compa-ratio commonly divides an employee's base salary by the applicable range midpoint. A value of one places salary at midpoint; values below or above one indicate relative position. Confirm that salary and midpoint share currency, schedule basis and effective date. For group reporting, the average of individual ratios can differ from total salary divided by total midpoint because the latter weights employees by midpoint size. Compa-ratio is a positioning metric, not a complete judgment about appropriate pay, performance or fairness within a job population.
Worked example: An employee earns $76,000 against an $80,000 midpoint, producing a 0.95 compa-ratio. After the midpoint moves to $84,000 with salary unchanged, the ratio falls to about 0.905 even though the employee receives no pay reduction.
Common mistake: Interpreting a falling compa-ratio as proof of a salary cut overlooks changes in the reference midpoint and its effective date.
Official syllabus reference: [2]
49. Range penetration uses the full range width
Range penetration locates salary between an administrative range's minimum and maximum. Under a common definition, subtract minimum from salary and divide by maximum minus minimum. Zero corresponds to minimum and one to maximum; values outside that interval can describe salaries below or above the range. Unlike compa-ratio, the calculation uses range endpoints rather than midpoint. Define treatment of out-of-range values explicitly instead of silently truncating them. Penetration can change after range redesign even when salary does not, so trend comparisons require consistent reference structures.
Worked example: A $70,000 salary in a $50,000 to $90,000 range has 50% penetration. In a redesigned $60,000 to $100,000 range, the same salary has 25% penetration, reflecting the reference change rather than a reduction in pay.
Common mistake: Dividing salary by the range maximum produces a different ratio and does not measure progress from minimum through the range width.
Official syllabus reference: [2]
50. Structure simulations distinguish movement from adjustment cost
A salary-structure simulation applies proposed reference changes to employee records and measures their consequences. Recalculate midpoints, minima, maxima and employee positioning using consistent assumptions. Moving a range does not itself change employee pay, although it may create below-minimum cases or alter policy-driven adjustments. Estimate those adjustments separately using explicit eligibility and timing rules. Compare alternatives through cost, affected counts and distributional outcomes. This analytical exercise evaluates consequences of proposed structures; it does not imply that one universal midpoint relationship determines every employee's proper salary.
Worked example: A proposed 4% range movement shifts an employee's minimum from $50,000 to $52,000 while salary remains $51,000. Under the scenario's rule to bring salaries to minimum, the required adjustment is $1,000, not 4% of salary.
Common mistake: Equating a percentage movement in salary ranges with the same percentage payroll increase confuses reference values with actual employee pay changes.
Official syllabus reference: [2]
51. Annualized increases differ from current-year expense
An annualized salary increase measures the full-year run-rate effect after a pay change. Current-year expense depends on when the change takes effect and how much of the year remains. Apply the scenario's proration convention consistently, and separate base salary from employer costs such as payroll-related charges when those are included. Employee eligibility, vacancies and changing schedules can further affect realized spending. Report both run-rate and period cost when leadership needs to understand ongoing commitment as well as the immediate budget impact.
Worked example: A $4,800 annual salary increase takes effect with three months remaining in the budget year. Under simple monthly proration, current-year base-pay cost is $1,200, while the continuing annualized increase remains $4,800 before other employer costs.
Common mistake: Presenting a prorated first-year cost as the continuing annual commitment understates the next full year's payroll requirement for the same increase.
Official syllabus reference: [2]
52. Distinguishing incentive payouts from goal attainment
Incentive award dollars reflect both the size of an employee's incentive opportunity and the rules converting results into payment. Dividing actual payout by target award normalizes payment against that opportunity, making relative payout levels easier to compare. This ratio does not independently measure goal attainment. Thresholds, payout curves, caps, individual modifiers and combinations of performance measures can produce different payments for similar attainment. To compare performance, obtain the underlying results and goal definitions, then assess whether their measurement bases are comparable. Report payout ratios and attainment measures separately with clear labels.
Worked example: Mira has an $8,000 target award and receives $10,000; Leon has a $16,000 target and receives $14,400. Their payout-to-target ratios are 125% and 90%. Mira has the higher relative payout despite receiving fewer dollars. Without comparable payout rules and goal results, the analyst cannot rank their goal attainment.
Common mistake: Treating payout divided by target award as goal attainment overlooks how payout formulas translate performance into money; normalized payments alone do not establish relative performance.
Official syllabus reference: [2]
53. Payroll bridges separate headcount and average-pay effects
Payroll equals headcount multiplied by average pay when population and pay definitions align. A bridge decomposes a change using a stated convention for assigning the interaction between changing headcount and changing average pay. One exact convention values the headcount change at old average pay and the average-pay change at new headcount. The components then reconcile to total change. Average pay can change through hiring mix, exits, promotions and raises, so the average-pay component should not automatically be interpreted as the cost of individual salary increases.
Worked example: Headcount rises from 50 to 60 and mean salary from $40,000 to $42,000. The bridge assigns $400,000 to headcount and $120,000 to average pay, reconciling the $520,000 payroll increase under the stated convention.
Common mistake: Using old headcount and old average pay for both components omits their interaction and leaves the payroll bridge short of the actual change.
Official syllabus reference: [2]
54. Difference-in-differences removes a shared time change
Difference-in-differences compares the before-and-after change in an intervention group with the contemporaneous change in a comparison group. Subtracting the latter attempts to remove changes that both groups would have experienced without the intervention. A causal interpretation requires credible parallel trends in the untreated potential outcomes and no other differential shocks or spillovers that undermine the comparison. Examine pre-intervention patterns and changes in group composition. The method improves on a simple before-and-after comparison but does not automatically make a nonrandom compensation pilot equivalent to a randomized experiment.
Worked example: Target attainment rises eight percentage points in pilot locations and five in comparison locations. The difference-in-differences estimate is three points. The analyst checks whether staffing changes or local demand shifts affected only the pilot locations before attributing that estimate to incentives.
Common mistake: Assuming any comparison group validates the method overlooks whether its untreated trend credibly represents what would have happened in the pilot group.
Official syllabus reference: [2]
55. Turnover models need time order and validation
A turnover model relates information available at a defined point to a subsequent departure outcome. Establish the prediction horizon and use predictors known before that outcome to avoid data leakage. Separate voluntary departures when they are the intended target and evaluate performance on data not used to fit the model. Calibration checks whether predicted risks match observed frequencies, while discrimination assesses ranking. An association between low pay position and departure can support investigation, but it does not establish how much turnover a pay adjustment would prevent.
Worked example: A model predicts next-quarter voluntary turnover using salary position and job information recorded at quarter start. It excludes termination codes entered after departure and is validated on a later quarter, preventing future information from inflating apparent predictive accuracy.
Common mistake: Including final payroll or exit-processing fields in a prospective model can reveal the outcome rather than genuinely predict future employee departures.
Official syllabus reference: [2]
56. Sensitivity analysis reveals assumption dependence
Sensitivity analysis recalculates results under plausible changes in assumptions, inputs or analytical choices. For compensation decisions, vary factors such as participation, eligibility, increase rates, turnover or model specification while identifying what changes in each scenario. Distinguish a one-factor sensitivity from a combined scenario that changes several conditions together. The purpose is to find which assumptions materially influence the decision, not to assign probabilities without evidence. A recommendation is more informative when it identifies conditions under which its preferred option would cease to be attractive.
Worked example: A proposed adjustment costs $300,000 at full participation. The analyst shows costs at 80%, 90% and 100% participation, then separately tests alternative eligibility rules. Leaders can see whether uncertainty comes mainly from take-up or the policy definition.
Common mistake: Labeling optimistic, central and pessimistic scenarios as confidence intervals invents probabilistic meaning unless a defensible probability model supports those bounds.
Official syllabus reference: [2]
57. Chart choice follows the analytical relationship
Choose a chart according to the relationship the reader must assess. Scatterplots show paired quantitative relationships, line charts show ordered time trends, bars compare categories and histograms show distributions. A chart that aggregates too early can hide variation essential to the decision. Preserve relevant units, population and period, and include uncertainty when it affects interpretation. Direct labels and restrained emphasis help readers identify the intended comparison. A visually polished display remains misleading if its geometry implies continuity, ordering or relationships that the underlying data do not support.
Worked example: To examine whether salary levels flatten with experience, an analyst uses a scatterplot with employee observations and a fitted curve. Department mean bars would conceal the requested relationship, while a histogram would show salary shape without linking it to experience.
Common mistake: Connecting employees in arbitrary ID order with a line implies a sequence and continuity that have no substantive compensation meaning.
Official syllabus reference: [2]
58. Axes and denominators control visual interpretation
Visual comparisons depend on scale, baseline and denominator. Bar lengths generally need a zero baseline because readers compare their physical lengths, while a clearly labeled restricted axis can be appropriate for a line chart examining small changes over time. Distinguish percentage change from percentage-point change and display rates alongside relevant exposure counts when needed. Avoid three-dimensional effects or unequal intervals that distort magnitude. Comparable panels should use consistent scales unless differences are clearly marked. The goal is faithful comparison rather than maximizing apparent movement.
Worked example: Acceptance rises from 80% to 84%. The change is four percentage points and a 5% relative increase. A truncated bar axis can make the improvement look enormous, so the analyst uses clear labels and an appropriate baseline.
Common mistake: Describing a movement from 80% to 84% as four percent confuses percentage points with relative percentage change and understates the latter.
Official syllabus reference: [2]
59. Privacy-aware reporting prevents indirect disclosure
Compensation records contain sensitive individual information, and aggregated reporting can still reveal identities when groups are small or easily inferred. Limit access to the information needed for the analytical purpose, separate identifiers where appropriate and assess disclosure through combinations of tables. Suppressing one small cell may be ineffective if totals reveal its value by subtraction. Apply the organization's approved privacy rules and relevant jurisdictional obligations without assuming a universal minimum cell size. Useful reporting balances decision needs with protection against both direct and indirect identification.
Worked example: A department report suppresses one executive's salary but displays the department total and every other salary. Readers could recover the amount by subtraction. The analyst changes the aggregation and complementary disclosures rather than relying on a single suppressed cell.
Common mistake: Removing names alone does not anonymize pay data when job titles, locations, totals or other released details allow individuals to be identified.
Official syllabus reference: [2]
60. Decision narratives separate finding, inference and action
A compensation narrative should connect the decision question, the most relevant finding, its limitations and a proportionate recommendation. Separate what the data directly show from explanations inferred through a model and outcomes that remain uncertain. Quantify the affected population, cost and plausible benefit using transparent assumptions. Explain why the proposed action follows from the evidence and how its results will be evaluated. Technical detail belongs where it supports that judgment. A strong narrative makes uncertainty usable rather than hiding it or allowing it to prevent every decision.
Worked example: A report finds higher turnover among lower-positioned employees, estimates a proposed adjustment's cost and recommends a targeted pilot. It states that the relationship is observational and defines comparison outcomes before launch so the next decision can use stronger evidence.
Common mistake: Presenting a dashboard finding as a guaranteed business outcome overstates the evidence and leaves leaders unable to judge the recommendation's assumptions.
Official syllabus reference: [2]