Essential Statistics for Business Analytics — Module 3

1. Foundations of: Correlation Analysis & Bivariate Association In quantitative business analytics, variables rarely operate in isolation. Pricing strategies interact with sales volume, promotional investments drive customer acquisition, and capital expenditure influences operating revenue. Correlation Analysis is the statistical science that quantifies the strength, direction, and nature of the mutual association between two or more random variables. 1.1 Definitional Architecture & The Concept of Co-Variation
2. Definitional: Meaning of Correlation According to Simpson and Kafka: "Correlation analysis deals with the association between two or more variables." When the movement of one variable is accompanied by a systematic, corresponding movement in another variable, the two series are said to be correlated.

2. Covariance: The Unstandardized Foundation Covariance measures the joint variability of two random variables:

Cov(X, Y) = ∑ [ (X − X̄)(Y − Ȳ) ] ÷ n. While Covariance indicates the direction of relationship (+ or −), its numerical magnitude depends on units of measurement (e.g., kilograms vs. grams), making standard comparison impossible without normalization. 1.2 Master Typology of Correlation Classification Basis Correlation Category Mathematical & Behavioral Manifestation By Direction of Movement

1. Positive: Correlation (Direct) Both variables move in the identical direction: as $X$ increases, $Y$ increases (e.g., Advertising expenditure and Sales revenue;

Supply and Price).

1. Negative: Correlation (Inverse) Variables move in opposite directions: as $X$ increases, $Y$ decreases (e.g., Product price and Consumer demand; Interest rates and Bond prices).

By Ratio of Variation

1. Linear: Correlation The ratio of change between variables is strictly constant. When plotted on graph paper, all points cluster along a straight line ($Y = a + bX$).
2. Non-Linear (Curvilinear): The ratio of change is not constant. Plotted data points form a smooth curve (e.g., Agricultural yields versus fertilizer application obeying diminishing returns).

By Number of Variables

1. Simple: Correlation Evaluates the relationship between exactly two variables (e.g.,

Height and Weight).

1. Partial: Correlation Analyzes the relationship between two variables while mathematically holding other confounding variables constant.
2. Multiple: Correlation Studies the simultaneous joint relationship between three or more variables (e.g., Crop yield vs. Rainfall, Temperature, and Soil Quality). 1.3 The Scatter Diagram Method & Spurious Correlation
3. Scatter: Diagram (Visual Inspection) A Cartesian coordinate graph plotting pairs of bivariate values $(X_i, Y_i)$:

Perfect Positive ($r = +1$): All points lie on a straight line sloping upwards from lowerleft to upper-right.

Perfect Negative ($r = -1$): All points lie on a line sloping downwards from upper-left to lower-right.

Zero Correlation ($r = 0$): Points form a chaotic, circular, or horizontal band showing no pattern.

1. Spurious (Nonsense): Correlation A statistical correlation between two variables that is meaningless and lacks causal foundation, driven by sheer coincidence or an unseen lurking/confounding third variable (e.g., high correlation between ice cream sales and drowning incidents, both driven by the lurking variable of summer heat).

• Foundational Maxim: "Correlation does not imply Causation."

1. Karl: Pearson's Coefficient of Correlation & Statistical Significance Formulated by the pioneer of modern statistics, Karl Pearson, the Product-Moment Correlation Coefficient ($r$) is the definitive mathematical measure of linear association between two continuous, normally distributed variables. 2.1 Assumptions Underlying Pearson's Coefficient
2. Bivariate: Normal Distribution: Both variables $X$ and $Y$ are continuous and jointly follow a normal distribution.
3. Strict: Linearity: The fundamental relationship connecting $X$ and $Y$ is linear; Pearson's $r$ drastically underestimates non-linear associations.
4. Absence of: Severe Outliers: Extreme rogue observations severely distort the numerical value of $r$.

4. Homoscedasticity: The dispersion of points around the regression line remains constant across all values. 2.2 Mathematical Formulations of Pearson's r

• MATHEMATICAL FORMULATIONS: KARL PEARSON'S PRODUCT-MOMENT COEFFICIENT Parametric Association Metric D irec t P ro duc t-Mo ment Fo rmula: r = ∑ [ (X − X )(Y − Ȳ ) ] ÷ √ [ ∑ (X − X )^2 × ∑ (Y − Ȳ )^2 ] Raw D ata Co mputatio nal Fo rmula: r = [ n ∑ X Y − (∑ X )(∑ Y ) ] ÷ √ [ [ n ∑ X ^2 − (∑ X )^2 ] × [ n ∑ Y ^2 − (∑ Y )^2 ] ] Core Mathematical Invariants:
• Bounded Limits: Always lies strictly in the interval −1.0 ≤ r ≤ +1.0.
• Origin & Scale Invariance: Pure dimensionless number; unaffected by adding/subtracting constants (origin change) or multiplying/dividing by positive constants (scale change). 2.3 The Coefficient of Determination (r^2) & Probable Error (PE)

1. Coefficient of: Determination (r^2) Represents the percentage of total variation in Y explained by variation in X: r^2 = Explained Variation ÷ Total Variation.

If $r = 0.80$, then $r^2 = 0.64$ (64% of variance is explained; 36% is unexplained residual variance $k^2 = 1 - r^2$).

1. Probable: Error (PE_r) & Significance Rule Measures reliability of correlation across random samples:

PE_r = 0.6745 × [ (1 − r^2) ÷ √n ]. If r < PE_r: Correlation is completely insignificant.

If r > 6 × PE_r: Correlation is definitely statistically significant. ∑ Worked Illustration 1: Karl Pearson's Correlation & Probable Error Test (Ad Spend vs.

Revenue) Quarterly Operating Data (n = 5 Quarters):

• Ad Spend X (₹ Lakhs): 2, 4, 6, 8, 10 → $sum X = 30$, Mean X̄ = 6.0, $sum (X - X̄)^2 = 40$.
• Revenue Y (₹ Lakhs): 10, 20, 25, 35, 40 → $sum Y = 130$, Mean Ȳ = 26.0, $sum (Y - Ȳ)^2 = 530$.
• Cross-Product Deviations: $sum (X - X̄)(Y - Ȳ) = (-4)(-16) + (-2)(-6) + (0)(-1) + (2)(9) + (4)(14) = 64 + 12 + 0 + 18 + 56 = mathbf{150.0}$.

1. Correlation: Coefficient (r) = 150.0 ÷ √[ 40 × 530 ] = 150.0 ÷ √21,200 = 150.0 ÷ 145.602 = +0.9890.
2. Coefficient of: Determination (r^2) = (0.9890)^2 = 0.9781 (97.81% Explained Variance).
3. Probable: Error = 0.6745 × [ (1 − 0.9781) ÷ √5 ] = 0.6745 × [ 0.0219 ÷ 2.236 ] = 0.0066.

• SIGNIFICANCE CONCLUSION: Since r (0.989) is vastly greater than 6 × PE (6 × 0.0066 = 0.0396), there is an exceptionally strong, statistically significant linear relationship between advertising investment and enterprise sales revenue.

3. Spearman's Rank Correlation: Qualitative & NonParametric Association Formulated by Charles Spearman in 1904, the Rank Correlation Coefficient ($R$ or $ ho$) is the nonparametric counterpart to Pearson's $r$. It is deployed when data cannot be measured on a continuous interval scale but can be arranged in ordinal ranks (e.g., job performance ratings, customer satisfaction rankings, beauty pageant scores). 3.1 Case I: Spearman's Rank Correlation for Non-Tied Ranks

• MATHEMATICAL FORMULATION: SPEARMAN'S RANK CORRELATION (DISTINCT RANKS) Ordinal NonParametric Metric R = 1 − [ (6 ∑ d^2) ÷ [ n(n^2 − 1) ] ] Where: d = R_X − R_Y: Difference between the ranks assigned to an individual unit across the two series. n: Total number of paired observations ranked. 3.2 Case II: Handling Tied / Repeated Ranks (Correction Factor) When two or more items possess identical numerical scores, they are assigned the average of the ranks they would have occupied. To prevent mathematical bias, a correction factor is added to $sum d^2$ for every tied group of size $m$:
• MATHEMATICAL FORMULATION: SPEARMAN'S FORMULA WITH TIED RANK CORRECTION Fractional Ranking Adjustment R_ c o rrec ted = 1 − [ 6 [ ∑ d^2 + ∑ [ m_ i(m_ i^2 − 1) ÷ 12 ] ] ÷ [ n(n^2 − 1) ] ] ∑ Worked Illustration 2: Spearman's Rank Correlation with Tied Ranks (Food Critics Sensory Evaluation) Sensory Evaluation of n = 6 New Coffee Blends by Two Master Tasters:
• Critic 1 Scores: 80, 75, 75, 60, 50, 40 → Blends 2 & 3 tie for ranks 2 and 3 → Rank = $(2+3)/2 = mathbf{2.5}$ ($m_1 = 2$).

Ranks Critic 1 ($R_1$): 1, 2.5, 2.5, 4, 5, 6.

• Critic 2 Scores: 90, 85, 70, 70, 70, 50 → Blends 3, 4, 5 tie for ranks 3, 4, 5 → Rank = $(3+4+5)/3 = mathbf{4.0}$ ($m_2 = 3$).

Ranks Critic 2 ($R_2$): 1, 2, 4, 4, 4, 6.

• Rank Differences $d = R_1 - R_2$: 0, +0.5, −1.5, 0, +1.0, 0.
• Squared Differences $d^2$: 0, 0.25, 2.25, 0, 1.0, 0 → $sum d^2 = mathbf{3.50}$.

1. Correction for: Critic 1 ($m_1 = 2$) = [ 2(4 − 1) ] ÷ 12 = 6 ÷ 12 = 0.50.
2. Correction for: Critic 2 ($m_2 = 3$) = [ 3(9 − 1) ] ÷ 12 = 24 ÷ 12 = 2.00.

3. Adjusted ∑ d^2 = 3.50 + 0.50 + 2.00 = 6.00. Denominator = $6(36 - 1) = 6 imes 35 = mathbf{210}$.

1. Rank: Correlation R = 1 − [ (6 × 6.00) ÷ 210 ] = 1 − [ 36 ÷ 210 ] = 1 − 0.1714 = +0.8286.

• EVALUATION VERDICT: R = +0.8286 indicates strong consensus between the two independent tasters regarding coffee flavor profiles.

1. Principles of: Linear Regression & Ordinary Least Squares (OLS) While correlation evaluates co-variation, Regression Analysis establishes functional mathematical dependence. Derived from Sir Francis Galton's hereditary studies on "regression toward mediocrity", modern regression enables analysts to predict the value of an unknown dependent variable ($Y$) from a known independent explanatory variable ($X$). 4.1 Correlation vs. Regression: Master Comparative Framework Analytical Parameter Correlation Analysis Regression Analysis
2. Core: Nature Measures degree and direction of linear association.

Models causal/functional dependence of one variable on another.

1. Variable: Symmetry Symmetric: $r_{xy} = r_{yx}$; variables are interchangeable.

• Asymmetric: $Y$ depends on $X$ ($b_{yx} e b_{xy}$); clear cause-and-effect roles.

1. Origin &: Scale Units Pure dimensionless number bounded between −1 and +1.

Expressed in physical units of measurement (e.g., ₹ revenue per ₹ ad spend).

1. Predictive: Capability Zero forecasting capacity; cannot predict values.

Primary tool for business forecasting and scenario modeling. 4.2 The Ordinary Least Squares (OLS) Principle The Principle of Least Squares fits a mathematical trend line through scatter points such that the sum of the squares of vertical deviations (residuals) is minimized: $sum e_i^2 = sum (Y_i - hat{Y}_i)^2 ightarrow ext{Minimum}$. 4.3 The Two Regression Equations & Normal Equations

• MATHEMATICAL FORMULATIONS: BIVARIATE OLS REGRESSION MODELS Predictive Trend Modeling
• Regressio n E quatio n o f Y o n X : Y − Ȳ = b_ y x (X − X ) Regressio n Co ef f ic ient b_ y x = r × ( σ _ y ÷ σ _ x ) = [ n ∑ X Y − (∑ X )(∑ Y ) ] ÷ [ n ∑ X ^2 − (∑ X )^2 ]
• Regressio n E quatio n o f X o n Y: X − X = b_ xy (Y − Ȳ ) Regressio n Co ef f ic ient b_ xy = r × ( σ _ x ÷ σ _ y ) = [ n ∑ X Y − (∑ X )(∑ Y ) ] ÷ [ n ∑ Y ^2 − (∑ Y )^2 ] ∑ Worked Illustration 3: Fitting OLS Regression Lines for Corporate Revenue Forecasting Dataset (n = 5 Business Units):
• Marketing Spend X: Mean X̄ = ₹6.0 Lakhs | Standard Deviation σ_x = 2.828 Lakhs.
• Sales Revenue Y: Mean Ȳ = ₹26.0 Lakhs | Standard Deviation σ_y = 10.295 Lakhs.
• Correlation Coefficient r = +0.9890.

1. Regression: Coefficient of Y on X: b_yx = r × (σ_y / σ_x) = 0.9890 × (10.295 / 2.828) = 0.9890 × 3.640 = 3.600.

2. Equation of Y on X: Y − 26.0 = 3.60(X − 6.0) → Y = 3.60X − 21.6 + 26.0 → Ŷ = 4.40 + 3.60X.

1. Managerial: Prediction: If Marketing Spend is set to X = ₹15.0 Lakhs: → Expected Revenue Ŷ = 4.40 + 3.60(15.0) = 4.40 + 54.0 = ₹58.40 Lakhs.

• ECONOMIC INTERPRETATION: The slope b_yx = 3.60 indicates that for every incremental ₹1 Lakh invested in marketing, sales revenue increases by an expected ₹3.60 Lakhs.

1. Mathematical: Properties of Regression & Standard Error of Estimate In classical statistical theory, the regression coefficients $b_{yx}$ and $b_{xy}$ possess fundamental algebraic and geometric properties that validate econometric models and guard against computational errors. 5.1 Seven Inviolable Properties of Regression Coefficients Property Rule Mathematical Formulation Analytical & Diagnostic Meaning
2. Geometric: Mean Property r = ± √[ b_yx × b_xy ] The correlation coefficient is the geometric mean of the two regression coefficients.
3. Identical: Signs Rule Sign(r) = Sign(b_yx) = Sign(b_xy) All three coefficients must have identical signs; it is mathematically impossible for one to be positive and another negative.
4. Magnitude: Reciprocal Bound If |b_yx| > 1, then |b_xy| < 1 Both regression coefficients cannot simultaneously be greater than unity in absolute terms because $r^2 le 1$.
5. Arithmetic: Mean Bound [ b_yx + b_xy ] ÷ 2 ≥ r The arithmetic mean of regression coefficients is always greater than or equal to the correlation coefficient.

5. Scale vs. Origin Shift Independent of origin; dependent on scale.

Shifting origin does not alter $b_{yx}$; multiplying by constants alters $b_{yx}$ by the scale ratio.

1. Centroid: Intersection Intersect at ( X̄, Ȳ ) Both regression lines always pass through the point of sample means.
2. Angle: Between Lines (θ) tan θ = [ (1 − r^2) / r ] × [ σ_x σ_y / (σ_x^2 + σ_y^2) ] If $r = pm 1$, $ an heta = 0$ (lines coincide); if $r = 0$, lines are perpendicular ($ heta = 90^circ$). 5.2 Standard Error of the Estimate (S_yx) & Prediction Intervals

• MATHEMATICAL FORMULATION: STANDARD ERROR OF THE ESTIMATE Predictive Dispersion Metric S_ y x = σ _ y × √ [ 1 − r^2 ] = √ [ ∑ (Y − Ŷ )^2 ÷ (n − 2) ] ∑ Worked Illustration 4: Standard Error of the Estimate & Forecast Margin of Error
• Predictive Confidence Parameters: Sales revenue regression model with $sigma_y = ₹10.295$ Lakhs and $r = 0.9890$ ($r^2 = 0.9781$). Forecasted Revenue for $X = ₹15$ Lakhs is $hat{Y} = ₹58.40$ Lakhs.

1. Standard: Error of Estimate = 10.295 × √[ 1 − 0.9781 ] = 10.295 × √0.0219 = 10.295 × 0.14798 = ₹1.523 Lakhs. 2. 95% Prediction Interval (± 1.96 S_yx) = ₹58.40 ± 1.96(1.523) = ₹58.40 ± 2.985 = [ ₹55.415 Lakhs to ₹61.385 Lakhs ].

• DIAGNOSTIC CONCLUSION: The exceptionally low standard error of ₹1.52 Lakhs confirms high model reliability, guaranteeing that true sales will fall within ±₹2.98 Lakhs of our forecast with 95% probability. 5.3 The Gauss-Markov Theorem & BLUE Properties of OLS Estimators BLUE Pillar Econometric Criterion Practical Diagnostic Significance B - Best Minimum variance among all linear unbiased estimators.

Guarantees that OLS produces the tightest confidence intervals and maximum forecasting precision.

L - Linear Estimators are linear combinations of the dependent variable observations: $hat{eta} = sum w_i Y_i$.

Ensures computational simplicity and tractability across enterprise modeling software.

U Unbiased Expected value of the regression coefficient equals the true population parameter: E(b) = β.

The estimated regression slope does not systematically under-predict or over-predict the true structural relationship.

E Estimator Rule or algorithm applied to sample observations to estimate the unknown population parameter.

Under the classical assumptions (zero error mean, no heteroscedasticity, no autocorrelation), OLS is mathematically unbeatable.