HKSM

A statistical method that estimates how a result (dependent variable) changes with one or more factors (independent variables). It quantifies drivers, tests significance, and supports forecasting to produce evidence-based lessons and recommendations at project or phase close.

Key Points

• Used to measure the strength and direction of relationships between outcomes (e.g., schedule or cost variance) and potential drivers (e.g., changes, defects, staffing levels).
• Supports closeout by turning performance data into actionable lessons, benefits insights, and governance recommendations.
• Common forms include simple linear, multiple linear, and logistic regression; choose based on the outcome type.
• Relies on clean, sufficient historical data and checks for assumptions such as linearity, independence, and constant variance.
• Results are documented with coefficients, significance tests, and goodness-of-fit metrics to inform future plans and baselines.
• Complements other analyses (Pareto charts, control charts) by quantifying impact rather than just describing frequency or dispersion.

Purpose of Analysis

• Quantify which factors most influenced cost, schedule, quality, or benefits realization.
• Estimate the expected change in the outcome for a one-unit change in each driver.
• Create defensible recommendations for process improvements and standards.
• Provide predictive insight for future projects and portfolio planning.
• Strengthen closeout reporting with statistically supported findings.

Method Steps

• Frame the question: define the outcome to explain (e.g., final schedule slippage in days) and the decision to support.
• Select candidate drivers based on logic and availability (e.g., number of change requests, team turnover, defect rework hours).
• Collect and clean data: handle missing values, remove obvious outliers, and align units and timeframes.
• Explore data visually with scatterplots and correlations to check direction and rough linearity.
• Fit an appropriate model (e.g., multiple linear regression for continuous outcomes; logistic for pass/fail outcomes).
• Evaluate fit and assumptions using R²/adjusted R², RMSE, p-values, residual plots, and multicollinearity diagnostics (e.g., VIF).
• Refine the model: remove weak or collinear predictors, consider transformations or interaction terms where justified.
• Interpret results in business terms and validate with subject matter experts.
• Document findings, limitations, and practical recommendations in closing reports and lessons learned.

Inputs Needed

• Project performance data: CPI, SPI, cost variance, schedule variance, milestone dates, rework hours, defect counts.
• Change log details: number, size, and timing of approved changes.
• Resource metrics: staffing levels, turnover, overtime, skill mix.
• Quality metrics: test coverage, defect severity distribution, escape rate.
• Risk outcomes: realized risks, mitigation actions, residual impacts.
• External/context data: vendor lead times, market events, tool availability.
• Baseline plans and acceptance criteria to anchor comparisons.
• Data dictionary and measurement definitions to ensure consistency.

Outputs Produced

• Model equation with coefficients and interpretation of each driver.
• Statistical significance indicators (p-values, confidence intervals).
• Goodness-of-fit metrics (R²/adjusted R², RMSE, AIC/BIC as applicable).
• Residual and sensitivity analyses, including prediction intervals.
• Visuals: scatterplots with fitted lines, coefficient charts, residual plots.
• Closeout report content and lessons learned entries that specify what to repeat, avoid, or control better.
• Recommendations for standards, estimating factors, and risk thresholds in future work.

Interpretation Tips

• Translate coefficients into practical units (e.g., each approved change added 2.3 days on average).
• Use adjusted R² to compare models with different numbers of predictors.
• Check multicollinearity; high VIF values indicate unstable coefficient estimates.
• Focus on effect sizes and confidence intervals, not just p-values.
• Use prediction intervals to convey uncertainty for future projects.
• Do not infer causation solely from statistical association; apply domain logic and timing evidence.
• Avoid extrapolating beyond the data range used to fit the model.

Example

A software project finished 18 days late. The team modeled schedule slippage (days) using predictors: approved change requests, defect rework hours, and average team size.

• Model summary: Slippage = 1.6 + 2.1×(Changes) + 0.04×(ReworkHours) − 0.8×(AvgTeamSize).
• Fit: adjusted R² = 0.71; all coefficients significant at p < 0.05.
• Interpretation: each additional approved change added ~2.1 days, while adding one team member reduced slippage by ~0.8 days, within the studied range.
• Action: tighten change control late in the project and plan buffer for high-defect modules.

Pitfalls

• Poor data quality or inconsistent measurement undermining results.
• Overfitting with too many predictors for a small number of observations.
• Omitted variable bias from leaving out relevant drivers.
• Multicollinearity causing unstable or counterintuitive coefficients.
• Assumption violations (nonlinearity, heteroscedasticity, autocorrelation) leading to misleading inferences.
• Misinterpreting correlation as causation or cherry-picking results to fit a narrative.
• Ignoring timing and context, such as late-stage changes having different impacts than early ones.

PMP Example Question

During project closure, the PM wants to determine which factors most contributed to a 12% schedule overrun and to create evidence-based lessons learned. Which action best applies regression analysis?

1. Create a Pareto chart of delay causes from the issue log.
2. Model schedule overrun as the dependent variable with number of approved changes, defect rework hours, and staffing levels as predictors.
3. Compute earned schedule to recalculate SPI(t) at completion.
4. Run a Monte Carlo simulation of remaining schedule risk.

Correct Answer: B — Model schedule overrun with relevant predictors.

Explanation: Regression quantifies how specific factors relate to the overrun and tests significance. Pareto ranks frequency, earned schedule measures performance, and Monte Carlo simulates uncertainty rather than estimating driver impacts.