Market Penetration & Regression Modeling

Regression-Based Decision Modeling

• Left-hand side (LHS) = outcome we care about; right-hand side (RHS) = predictors we choose.

  • This structure ( $Y = f(X<em>1, X</em>2, \ldots, X_k)$ ) is fixed; only the nature of the $X$’s changes with context.

• Goal of modeling = unlock a concrete decision.

  • Begin by defining the decision space (menu of actions) → ask what variables inform the choice.

  • Ex: "Should we enter a new geographic market?"; "How much should we spend on advertising?".

Market Entry & Market Penetration

• Context used in class to illustrate modeling logic.

• Key notions

  • Market size = total attainable customers/sales in the geography.

  • Penetration rate: $\text{Penetration} = \dfrac{\text{Our Sales}}{\text{Total Market}}$ (can be defined in units, households, or dollars).

  • Share of wallet = customer-level analogue (proportion of a single customer’s spend captured by the firm).

• Practical motivations

  • New product launch, new geography, or expansion of existing chain (restaurants, ISPs, etc.).

  • Benchmark vs competitors, optimise ad spend, evaluate promotions.

  • Special case: highly concentrated markets (>80 % share in hands of one player) → defensive vs aggressive entry strategies differ.

Examples & Anecdotes

• Restaurants

  • Numerous student capstones on identifying best city/site for a new location.

  • Variables explored: population, income, competition, parking, weather, local excitement (survey-based).

• Expansion failures: Krispy Kreme over-expanded; Subway’s franchising issues.

• Chipotle–McDonald’s partnership: McD bought ≈20–25 % pre-IPO for real-estate & supply-chain know-how; illustrates data-driven market penetration analysis.

• ISP scenarios

  • High-penetration/high-price states attractive for premium high-speed entrants (e.g., Google Fiber).

  • Low-penetration/rural states attractive for low-price or satellite offerings (e.g., Starlink).

Data Set Used in Class (50 U.S. States)

• Outcome $Y$: Internet access (% households with Internet, 2013).

• Predictors $X$

  • CollegeGrad = % adults with bachelor’s or higher.

  • UrbanRate = % population living in urban areas.

  • RedBlue (dummy) = 1 if state voted Democratic in 2012, 0 if Republican.

  • Coastal (dummy) = 1 if state is on East/West coast; Gulf not included.

• Hypotheses brainstormed

  • More college grads → higher Internet access (knowledge-worker effect).

  • Higher urbanisation → cheaper infrastructure per capita → higher access.

  • Red states may be more rural/lower grad rate, so RedBlue could proxy for the two variables (multicollinearity concern).

  • Political climate may also predict regulatory friendliness, taxation, etc.

Modeling Strategy Options

• Bottom-up: start with one predictor, add variables gradually; good for exploratory, theory-light situations.

• Top-down: dump in all customary predictors, prune; useful when standard frameworks exist.

• Instructor recommendation: anchor on the decision & theoretical expectations, then choose variables accordingly.

• Machine-learning alternative: optimise predictive accuracy regardless of interpretability → great for black-box tasks (recommendation systems) but weak for theory-based managerial insight.

R Workflow Highlights

• Inspect data

  • summary(data) → min/mean/max for each variable.

  • Scatter plots (plot(), ggplot2) to eyeball linearity & clusters.

• Create dummy: data$RedBlueD <- ifelse(data$Winner2012=="Obama",1,0).

• Combine columns: cbind() to build quick summary tables.

• Run models: lm(Y ~ X1 + X2, data=data); interactions via X1*D or X1:D.

Core Regression Models & Findings

• Model 1: $\text{Internet} = \beta<em>0 + \beta</em>1(CollegeGrad) + \varepsilon$

  • Intercept ≈ 47.9 % (predicted access if no college grads).

  • Slope $\beta_1 ≈ 0.886$ → each 1 ppt rise in CollegeGrad lifts Internet access by ≈0.886 ppt.

  • Re-expressible: 10 ppt Grad increase ⇒ ≈8.86 ppt Internet gain.

• Model 2: $\text{Internet} = \beta<em>0 + \beta</em>1(RedBlue) + \varepsilon$

  • Red (0) mean ≈69.9 %; Blue gain ≈+5.6 ppt.

  • Note unequal state counts & populations → interpret carefully.

• Model 3: $\text{Internet} = \beta<em>0 + \beta</em>1(CollegeGrad) + \beta_2(RedBlue) + \varepsilon$

  • CollegeGrad remains significant; RedBlue loses significance → redundancy/multicollinearity evident.

• Interaction Model (Model 4): $Y = \beta<em>0 + \beta</em>1(CollegeGrad) + \beta<em>2(RedBlue) + \beta</em>3(CollegeGrad\times RedBlue) + \varepsilon$

  • Two intercepts (Red vs Blue) + two slopes.

  • Red slope = $\beta_1$.

  • Blue slope = $\beta<em>1 + \beta</em>3$ (shallower by ≈0.78 ppt in class results).

  • Visual: distinct lines; proves relationship strength depends on political grouping.

• Quadratic / Curvilinear Model (Model 5): $Y = \beta<em>0 + \beta</em>1X + \beta_2X^2 + \varepsilon$

  • \beta_2 < 0 in example → diminishing returns: Internet adoption rises quickly with early grad-rate gains, then tapers.

  • Plot shows upward curve flattening at high CollegeGrad.

• Key rules

  • Always include lower-order terms when specifying interactions or polynomials.

  • Interpretation: coefficients represent marginal effect holding included terms constant.

Formulas & Statistical Reminders

• Generic multiple regression: $Y<em>i = \beta</em>0 + \sum<em>{k=1}^K \beta</em>k X<em>{ki} + \varepsilon</em>i$.

• Dummy-coded intercept shift: if $D\in{0,1}$ → $\beta<em>0$ = group 0 mean (adjusted), $\beta</em>0+\beta_D$ = group 1.

• Interaction interpretation: effect of $X$ on $Y$ depends on $D$; difference-in-slopes captured by $\beta_3$.

• Quadratic turning point: $X^* = -\beta<em>1 / (2\beta</em>2)$ (maximum or minimum).

Decision Implications of the Internet-Access Case

• Penetration strategy differs by state cluster

  • High-grad, high-urban → potential for premium, high-speed offerings; intensive competition likely.

  • Low-grad, rural → lower price sensitivity? Could test price elasticity; satellite tech may offer unique advantage.

• Redundant predictors waste degrees of freedom & may obscure interpretation; choose parsimonious, theory-relevant variables.

• Interaction findings alert managers that one playbook may not fit all political/ demographic segments.

• Diminishing returns suggest spending to boost college-grad segments in already highly educated states offers limited bang-for-buck.

Broader Ethical & Practical Considerations

• Over-expansion risk (Krispy Kreme case) underscores need for rigorous, data-backed forecast rather than hype.

• Surveys capture intangible factors (awareness, excitement) absent from transactional data; integrating qualitative inputs can prevent model blind spots.

• Machine-learning black boxes excel at prediction but can fail to justify investment to stakeholders; transparency matters for strategic decisions.

• Political/regulatory environment must be treated carefully—dummy variables may proxy sensitive socio-economic attributes; guard against mis-use or discriminatory inferences.

Study Checklist

✓ Understand difference between main effects, interaction effects, and quadratic terms.
✓ Practice writing and interpreting regression equations with dummy variables.
✓ Be able to compute & explain market penetration and share-of-wallet metrics.
✓ Rehearse R commands: ifelse, summary, lm, interaction syntax, plotting.
✓ Relate statistical outcomes back to managerial decisions (enter, expand, target, price).