01 · Specify an explanatory question
Specify an explanatory question
Multiple linear regression models a continuous response using several predictors. Warehouse cycle time might depend on order size, distance, congestion, shift, and handling method. Decide whether the goal is predicting time or estimating a particular conditional association. These goals need different variable choices. A predictor available only after the order finishes may help explain historical times but cannot support advance staffing forecasts.
For causal questions, identify confounders, mediators, and consequences using process knowledge before selecting columns. Adjusting for a variable caused by the intervention may remove part of the total effect you wanted to estimate. Adjusting for a common consequence can create bias. Adding every recorded field is not a substitute for an explicit process model, and adjustment cannot resolve unmeasured confounding.
02 · Define the model and coefficients
Define the model and coefficients
Write yi = β0 + Σ βjxij + εi. Here y is the outcome, xij is predictor j for observation i, β0 is the intercept, βj is a conditional coefficient, and ε is error. A coefficient describes the mean change for one predictor unit while the other modeled predictors are held fixed. This condition must be feasible and supported by the data.
In matrix notation, b = (XᵀX)−1Xᵀy when the design matrix X has full column rank. X includes an intercept and the chosen predictor terms. Software normally uses numerically safer decompositions rather than explicitly inverting this matrix. Estimated covariance is s²(XᵀX)−1, where s² = SSE/(n−p), n is sample size, and p counts every fitted coefficient including the intercept. Exact conventional intervals assume independent normal constant variance errors and an appropriate mean model.
03 · Worked two predictor warehouse example
Worked two predictor warehouse example
Consider fictional cycle time with two coded inputs: x = −1 for small orders and +1 for large, z = −1 for uncrowded aisles and +1 for crowded. There are two independent observations for each combination. Times are 6 and 8 at (−1,−1), 12 and 14 at (−1,+1), 10 and 12 at (+1,−1), and 16 and 18 at (+1,+1).
The balanced design gives ŷ = 12 + 2x + 3z. Moving from small to large, a two unit change in x, corresponds to four additional minutes at either congestion level. Moving from uncrowded to crowded corresponds to six additional minutes. Residuals are −1 and +1 in each cell, so SSE = 8. With n = 8 and p = 3, s² = 8/5 = 1.6 and s = 1.2649.
Since the coded columns are orthogonal with squared lengths of eight, each coefficient standard error is √(1.6/8) = 0.4472. The 95% interval for the size coefficient is 2 ± 2.5706 × 0.4472, or approximately 0.850 to 3.150. The large versus small contrast is twice that interval, approximately 1.701 to 6.299 minutes. Total response variation is 112, giving R² = 1−8/112 = 0.9286. The balance makes this example unusually clean; actual observational warehouse data rarely have such separation.
04 · Understand coding and interaction
Understand coding and interaction
Categorical predictors require indicator or other explicit coding. A three level method factor usually needs two independent columns. Treating method labels 1, 2, and 3 as a numeric predictor imposes a linear spacing assumption that may be meaningless. Centering numeric inputs can make the intercept represent a realistic operating point. Standardizing helps compare scales but does not convert associations into causes.
Interactions make one coefficient depend on another predictor. With ŷ = b0 + bxx + bzz + bxzxz, the slope in x is bx + bxzz. Retain relevant lower order terms when modeling interactions. Quadratic terms can represent curvature while remaining linear in the unknown coefficients. Interpret predictions and contrasts at specified process conditions rather than calling a main effect universal.
05 · Diagnose confounding and instability
Diagnose confounding and instability
Correlated predictors can produce multicollinearity: coefficients become imprecise and sensitive to small changes despite good overall prediction. Variance inflation factor for predictor j is VIFj = 1/(1−R²j), where R²j comes from predicting that column using the others. VIF is a diagnostic, not a universal deletion rule. Investigate redundant measurements, narrow operating ranges, and impossible combinations.
If large orders always travel farther, the data may contain little information about a distance effect holding size fixed. Even a mathematically estimable coefficient can rely heavily on a few unusual rows. Inspect leverage, influence, and partial relationships. Removing a predictor because its p value is large can change other coefficients and their meaning. Repeated automated selection exaggerates the apparent certainty of the final model unless selection is accounted for.
06 · Validate fit and predictions
Validate fit and predictions
Plot residuals against fitted values, each important predictor, and run order. Curves indicate omitted structure; changing spread suggests a variance problem; group patterns suggest missing effects; serial patterns challenge independence. Examine unusual orders with process staff and verify measurements. Robust standard errors can address some variance misspecification, but they do not fix a wrong mean function, unmeasured confounding, or cluster structure automatically.
Use suitable held out data or cross validation for prediction. Keep dependent units together and respect time order when predicting future periods. R² always tends to increase when predictors are added; adjusted R² penalizes parameter count but does not replace validation. Report errors in minutes and assess performance for peak periods and important order classes. Mean confidence intervals differ from individual prediction intervals, and both become less credible outside the observed joint predictor region.
07 · Judge strength and DMAIC contribution
Judge strength and DMAIC contribution
A strong application starts with a process question, credible variable definitions, adequate independent observations, and a defensible model. It reports conditional estimates, uncertainty, diagnostic findings, and prediction limits. A weak application feeds every field into stepwise selection and names significant coefficients root causes. High R² can arise from leakage, composition, or deterministic accounting relationships rather than a useful explanation.
Analyze uses the model to compare candidate drivers and identify an intervention worth testing. Improve manipulates controllable factors while measuring unintended effects. Control tracks whether relationships and residual variation persist under new conditions. Save the coding, predictor availability, fitted terms, excluded records, diagnostic rationale, and validation results. The equation alone cannot communicate when a conditional comparison is trustworthy.
08 · Check the joint region of support
Check the joint region of support
In the example, all four size and congestion combinations are observed equally. This permits a size comparison at either congestion level without relying on an unseen combination. In an actual warehouse, large orders may appear almost exclusively during peak congestion. A fitted conditional coefficient can then rely on very little overlap. A prediction for large orders in an empty aisle may be outside the observed joint region even if each individual predictor value falls inside its separate range.
Inspect joint plots and relevant cross tabulations before making conditional claims. If information is weak, obtain deliberate observations in missing feasible combinations or narrow the conclusion. Do not describe an adjusted coefficient as fully isolating a factor when the study lacks comparable units. Good process knowledge and design can provide more useful separation than adding another model term to the same confounded dataset.
Connections
Related methods and phase use
Placed under Analyze because of the outcome it supports. Use the method in other phases when its purpose and assumptions fit the work; the phase list is a guide to selection, not a compulsory sequence.
Sources
Professional foundations and further reading
The explanation and fictional teaching examples are written for LeanTrainingLab. Linked primary sources provide professional foundations; example values and local conventions are stated in the guide. The accepted toolset draws on Six Sigma+Lean Toolset, second edition (Lunau, editor; Springer, 2013). No book text or examples are reproduced here.