Essential Statistics for Business Analytics — Module 2

1. Conceptual: Foundations of Statistical Hypothesis Testing In modern data-driven enterprises, managerial decision-making cannot rely on intuition or subjective conjecture. Hypothesis Testing is the formal mathematical procedure that enables business analysts to evaluate whether empirical claims regarding population parameters are statistically supported by sample evidence or whether observed differences are merely artifacts of random sampling variability. 1.1 The Logic of Hypothesis Testing: Proof by Contradiction
2. The: Null Hypothesis (H_0) The hypothesis of no effect, no difference, or status quo. It assumes that any observed variation between sample statistics and population parameters is purely due to chance. It is presumed true until sample evidence overwhelmingly contradicts it.
3. The: Alternative Hypothesis (H_1 or H_a) The operational claim or research proposition that the analyst seeks to validate. It states that a real effect, systemic difference, or statistical relationship exists in the underlying population. 1.2 Directionality: Two-Tailed vs. One-Tailed Tests Test Characteristic Two-Tailed (Non-Directional) Test One-Tailed (Directional) Test Hypothesis Formulation H_0: μ = μ_0 vs. H_1: μ ≠ μ_0 Right-Tailed: H_0: μ ≤ μ_0 vs. H_1: μ > μ_0

• Left-Tailed: H_0: μ ≥ μ_0 vs. H_1: μ < μ_0 Rejection Region Critical rejection area split equally into both tails (α/2 in left tail, α/2 in right tail).

Entire rejection area (α) concentrated exclusively in one extreme tail.

Critical Value at α = 0.05 Critical Z = ± 1.960 Right-Tailed: Z = +1.645 | Left-Tailed: Z = −1.645 Business Application Testing whether a new manufacturing machine alters bottle fill volumes (could be overfilling or underfilling).

Testing whether a new training program increases sales revenue, or a new process reduces delivery time. 1.3 The Statistical Decision Matrix: Type I vs. Type II Errors Statistical Decision Made True State of Nature in the Population Null Hypothesis (H_0) is TRUE Null Hypothesis (H_0) is FALSE Accept H_0 (Fail to Reject) Correct Decision (Confidence Level = 1 − α) Probability of correctly retaining a true null hypothesis.

Type II Error (β / Beta Error) Consumer's Risk: Accepting a false null hypothesis (failing to detect a real improvement).

Reject H_0 (Support H_1) Type I Error (α / Alpha Error) Producer's Risk: Rejecting a true null hypothesis (false alarm; claiming an effect that does not exist).

• Correct Decision: Statistical Power (1 − β) Probability of correctly rejecting a false null hypothesis. 1.4 The 5-Step Hypothesis Testing Workflow Step 1: State Hypotheses: Formulate mutually exclusive $H_0$ and $H_1$ statements.

Step 2: Set Significance Level (α): Establish tolerable risk of Type I error (typically $alpha = 0.05$ or $0.01$).

Step 3: Select Test Statistic: Choose appropriate test ($Z, t, F, chi^2$) based on sample size, known parameters, and distribution.

Step 4: Compute Test Statistic & p-Value: Calculate empirical score from sample data.

Step 5: Formulate Managerial Conclusion: Reject $H_0$ if Test Statistic falls in rejection region (or if $p ext{-value} le alpha$); otherwise, Fail to Reject $H_0$. 1.5 The Modern p-Value Paradigm & Significance Thresholds p-Value Range Empirical Evidence Against Null Hypothesis (H_0) Standard Business Reporting Nomenclature p > 0.10 Virtually no evidence against $H_0$; observed variance is purely random noise.

Statistically Non-Significant (Fail to Reject $H_0$). 0.05 < p ≤ 0.10 Weak / marginal evidence against $H_0$; suggests directional trend.

Marginally Significant (Often reported as a trend in pilot studies). 0.01 < p ≤ 0.05 Strong evidence against $H_0$; less than 5% probability of false alarm.

Statistically Significant at α = 0.05 (*) (Reject $H_0$). p ≤ 0.01 Overwhelming evidence against $H_0$; less than 1% chance result is accidental.

Highly Statistically Significant at α = 0.01 (**) (Decisive rejection).

1. Large: Sample Hypothesis Tests: Single & Difference of Proportions (Z-Tests) When analyzing categorical binary attributes across large enterprise datasets ($n ge 30$, with $np ge 5$ and $nq ge 5$), the sampling distribution of proportions follows the standard normal distribution under the Central Limit Theorem. 2.1 Z-Test for a Single Population Proportion

• MATHEMATICAL FORMULATION: Z-TEST FOR SINGLE PROPORTION Large Sample Parametric Test Z_ c alc ulated = (p − P _ 0) ÷ √ [ P _ 0(1 − P _ 0) ÷ n ] Where: p: Sample proportion ($x / n$); P_0: Hypothesized population proportion under $H_0$.
• Standard Error: $SE_P = sqrt{P_0(1 - P_0) / n}$. ∑ Worked Illustration 1: Testing Telecom Customer Churn Rate Against Benchmark
• Industry Benchmark: Telecom standard historical monthly churn proportion $P_0 = 10%$ (0.10). A new retention plan is tested across a random sample of n = 400 subscribers, resulting in x = 28 churned users (Sample proportion $p = 28 / 400 = 0.07$ or 7.0%).
• Hypotheses: $H_0: P ge 0.10$ vs. $H_1: P < 0.10$ (One-Tailed Left-Tailed Test at $alpha = 0.05$). Critical $Z_{0.05} = -1.645$.

1. Standard: Error = √[ (0.10 × 0.90) ÷ 400 ] = √[ 0.09 ÷ 400 ] = √0.000225 = 0.015.

2. Z-Calculated = (0.07 − 0.10) ÷ 0.015 = −0.03 ÷ 0.015 = −2.00.

• DECISION & CONCLUSION: Since Z_calc (−2.00) < Z_critical (−1.645), we REJECT H_0. The new retention program significantly reduces customer churn from 10% to 7% at the 5% significance level ($p = 0.0228$). 2.2 Z-Test for Difference Between Two Population Proportions
• MATHEMATICAL FORMULATION: Z-TEST FOR TWO INDEPENDENT PROPORTIONS Comparative A/B Analytics Z_ c alc ulated = (p_ 1 − p_ 2) ÷ √ [ p̂ (1 − p̂ ) × (1/n_ 1 + 1/n_ 2) ] P o o led P ro po rtio n E stimate: p̂ = (x_ 1 + x_ 2) ÷ (n_ 1 + n_ 2) ∑ Worked Illustration 2: Digital A/B Testing Landing Page Conversion Comparison E-Commerce Experiment:
• Landing Page A: $n_1 = 1,000$ visitors → $x_1 = 80$ purchases ($p_1 = 0.080$ or 8.0%).
• Landing Page B: $n_2 = 1,000$ visitors → $x_2 = 120$ purchases ($p_2 = 0.120$ or 12.0%).
• Hypotheses: $H_0: P_1 = P_2$ vs. $H_1: P_1 e P_2$ (Two-Tailed Test at $alpha = 0.05$; Critical $Z = pm 1.960$).

1. Pooled: Proportion (p̂) = (80 + 120) ÷ (1,000 + 1,000) = 200 ÷ 2,000 = 0.100 (10%).
2. Standard: Error = √[ 0.10 × 0.90 × (1/1000 + 1/1000) ] = √[ 0.09 × 0.002 ] = √0.00018 = 0.013416.

3. Z-Calculated = (0.080 − 0.120) ÷ 0.013416 = −0.040 ÷ 0.013416 = −2.981.

• A/B TEST CONCLUSION: |Z_calc| (2.981) > 1.960 → REJECT H_0 ($p = 0.0028$).

Landing Page B generates a statistically superior conversion rate, justifying permanent deployment. 2.3 Valid Mathematical Conditions for Proportion Z-Tests

1. The: Success-Failure Condition The normal approximation to the binomial distribution is valid only when both expected successes and expected failures are at least 5 (or 10 in conservative analytics): n × P_0 ≥ 5 and n × (1 − P_0) ≥ 5.
2. Independence &: Randomness Condition Sampled observations must be independent of one another. When sampling without replacement from finite customer databases, the sample must not exceed 10% of the total population ($n le 0.10 N$) to preserve independence.
3. Small: Sample Tests: Student's t-Distribution & Practical Applications Formulated by William Sealy Gosset in 1908 under the pen name "Student", the Student's t-Distribution is deployed when the sample size is small ($n < 30$) and the population standard deviation (σ) is unknown. The t-distribution is symmetric and bell-shaped like the Z-distribution but exhibits heavier tails (higher kurtosis), reflecting greater estimation uncertainty. 3.1 Master Selection Matrix: When to Use Z-Test vs. t-Test Sample Size Condition Population Variance (σ^2) Status Population Underlying Distribution Appropriate Statistical Test Large Sample ($n ge 30$) Known or Unknown (Use sample $s$) Normal or Non-Normal (CLT applies) Standard Normal ZTest Small Sample ($n < 30$) Known (σ available) Normal Standard Normal ZTest Small Sample ($n < 30$) Unknown (Only $s$ available) Normal (or approx. normal) Student's t-Test ($df = n - 1$) 3.2 Typology of Student's t-Tests
4. One-Sample t-Test: Tests whether a single sample mean differs significantly from a known population mean: $t = (ar{X} - mu) / (s / sqrt{n})$, with degrees of freedom $df = n - 1$.
5. Independent: Two-Sample t-Test Compares means of two unrelated groups (e.g., salaries of male vs. female executives): uses pooled variance $s_p^2$ with degrees of freedom $df = n_1 + n_2 - 2$. 3.3 The Paired-Samples t-Test (Dependent t-Test)

• MATHEMATICAL FORMULATION: PAIRED-SAMPLES TTEST Pre-Post Experimental Design t_ c alc ulated = d ÷ [ s_ d ÷ √ n ] D egrees o f Freedo m: df = n − 1 Where: d = X_post − X_pre: Difference score for each paired subject. d̄: Mean of difference scores; s_d: Standard deviation of difference scores. ∑ Worked Illustration 3: Evaluating Corporate Sales Training Effectiveness (Paired t-Test)
• Corporate HR Analytics: A sales executive training workshop tested on n = 10 sales executives. Weekly sales revenue measured Before ($X_1$) and After ($X_2$) training (in ₹'000):
• Difference Scores ($d = X_2 - X_1$): +8, +6, +10, +4, +7, +9, +5, +8, +11, +2.
• Sum of Differences $sum d = 70$ → Mean Difference $d̄ = 70 / 10 = mathbf{+7.00}$. Sample variance of differences $s_d^2 = 7.333 ightarrow s_d = mathbf{2.708}$.
• Hypotheses: $H_0: mu_d le 0$ vs. $H_1: mu_d > 0$ (Right-Tailed at $alpha = 0.05$; $df = 10 - 1 = 9$; Critical $t_{0.05, 9} = mathbf{1.833}$).

1. Standard: Error of Mean Difference = $s_d divide adic{n} = 2.708 divide adic{10} = 2.708 divide 3.1623 = mathbf{0.8563}$.

2. t-Calculated = $7.00 divide 0.8563 = mathbf{+8.175}$.

• HR ANALYTICS DECISION: t_calc (8.175) is overwhelmingly greater than critical t (1.833) → REJECT H_0 ($p < 0.0001$). The sales training generates a highly statistically significant average weekly sales lift of ₹7,000 per executive.

1. Analysis of: Variance (ANOVA), Two-Way Factorial ANOVA & MANOVA When an experiment involves comparing means across three or more groups, conducting multiple pairwise t-tests results in severe Type I Error inflation ($alpha_{family-wise} = 1 - (1 - alpha)^k$). Analysis of Variance (ANOVA), developed by Sir Ronald A. Fisher, overcomes this by testing whether all group means are simultaneously equal using the $F$-distribution. 4.1 Partitioning Total Variation in One-Way ANOVA Source of Variation Sum of Squares (SS) Degrees of Freedom (df) Mean Square (MS) F-Ratio Statistic Between Groups (Treatment) SSB = ∑ n_j (X̄_j − X̄_grand)^2 k − 1 MSB = SSB ÷ (k − 1) F = MSB ÷ MSW Within Groups (Error) SSW = ∑ ∑ (X_ij − X̄_j)^2 N − k MSW = SSW ÷ (N − k) Total Variation SST = SSB + SSW N − 1 — — 4.2 Two-Way Factorial ANOVA & Interaction Effects
2. Main: Effects vs. Interaction Effects In a $2 imes 2$ factorial design (e.g., Testing 2 Advertising Creatives across 2 Price Points), TwoWay ANOVA isolates:

• Main Effect of Factor A: Independent effect of Creative.
• Main Effect of Factor B: Independent effect of Price.

Interaction Effect (A × B): Whether the effect of Creative depends on the specific Price tier.

1. Multivariate ANOVA (MANOVA): Extends ANOVA to evaluate multiple correlated dependent variables simultaneously (e.g., evaluating how marketing strategies affect Brand Awareness, Brand Loyalty, and Customer Retention concurrently), using test statistics such as Wilks' Lambda (Λ), Pillai's Trace, and Hotelling's Trace. ∑ Worked Illustration 4: One-Way ANOVA Comparing Three Retail Promotional Formats

• Retail Marketing Experiment: Testing daily sales (₹ Lakhs) across k = 3 store formats with n = 5 stores each (Total N = 15):
• Format 1 (Discount Display): Mean = 12.0 | Format 2 (Influencer
• Endorsement): Mean = 16.0 | Format 3 (Contest/Gamification): Mean = 20.0.
• Grand Mean (X̄_grand) = (12 + 16 + 20) / 3 = 16.0 Lakhs.
• Calculated Variations: $SSB = 5 imes [(12-16)^2 + (16-16)^2 + (20-16)^2] = 5 imes [16 + 0 + 16] = mathbf{160.00}$. $SSW = mathbf{48.00}$.

1. Mean: Square Between (MSB) = SSB / (k − 1) = 160.00 / (3 − 1) = 160.00 / 2 = 80.00.
2. Mean: Square Within (MSW) = SSW / (N − k) = 48.00 / (15 − 3) = 48.00 / 12 = 4.00.

3. Computed F-Statistic = MSB ÷ MSW = 80.00 ÷ 4.00 = 20.00. (Critical $F_{0.05, 2, 12} = mathbf{3.89}$).

• ANOVA DECISION: F_calc (20.00) >> F_critical (3.89) → REJECT H_0 ($p < 0.001$). There is a highly significant difference in sales revenue across promotional formats; post-hoc Tukey tests confirm Format 3 significantly outperforms Formats 1 and 2. 4.3 Core Assumptions of ANOVA & Post-Hoc Multiple Comparisons Analytical Pillar Statistical Assumption / Procedure Diagnostic Test & Remedial Action

1. Normality: Residual errors within each treatment group must be normally distributed.

Assessed via Shapiro-Wilk test or Q-Q plots; robust to mild violations if sample sizes are equal.

1. Homogeneity of: Variances (Homoscedasticity) Variances across all $k$ treatment groups must be equal ($sigma_1^2 = sigma_2^2 = dots = sigma_k^2$).

Tested using Levene's Test or Bartlett's Test; if violated, deploy Welch's ANOVA.

1. Post-Hoc: Analysis (Tukey's HSD) When the omnibus $F$-test rejects $H_0$, post-hoc tests identify precisely which pairs of means differ significantly.

Tukey's Honestly Significant Difference (HSD): Controls the Family-Wise Error Rate ($alpha_{FW} le 0.05$) across all pairwise group comparisons.

5. Chi-Square (χ^2) Tests: Goodness of Fit & Independence of Attributes The Chi-Square (χ^2) Test, formulated by Karl Pearson, is the premier non-parametric test used to evaluate categorical and nominal frequencies without requiring underlying normal distribution parameters. 5.1 Chi-Square Test of Goodness of Fit

• MATHEMATICAL FORMULATION: PEARSON'S CHI-SQUARE TEST STATISTIC Non-Parametric Analysis χ^2_ c alc ulated = ∑ [ (O _ i − E _ i)^2 ÷ E _ i ] E xpec ted Frequenc y (Go o dness o f Fit): E _ i = N × P _ i | df = k − 1 − c Where:

O_i: Observed sample frequency; E_i: Expected theoretical frequency under $H_0$.

• Fundamental Condition: Every expected cell frequency must be at least 5 ($E_i ge 5$). 5.2 Chi-Square Test of Independence of Attributes (Contingency Tables) Analytical Dimension Standard r × c Contingency Table 2 × 2 Table with Yates' Continuity Correction Expected Cell Frequency E_ij = (Row Total_i × Column Total_j) ÷ Grand Total N Same expected formula applied to all 4 cells.

Degrees of Freedom df = (r − 1)(c − 1) df = (2 − 1)(2 − 1) = 1 Yates' Correction Rule Not required for tables with $df > 1$. Mandatory when $df = 1$ and sample size is small to prevent Type I error: χ^2_Yates = ∑ [ (|O_i − E_i| − 0.5)^2 ÷ E_i ] ∑ Worked Illustration 5: Chi-Square Test of Independence (Customer Brand Preference vs.

Age Group) Market Research Contingency Data (N = 200 Consumers):

• Age < 30: Brand Alpha = 60 (Observed), Brand Beta = 40 (Observed) → Row Total = 100.
• Age ≥ 30: Brand Alpha = 30 (Observed), Brand Beta = 70 (Observed) → Row Total = 100.
• Column Totals: Brand Alpha = 90 | Brand Beta = 110 | Grand Total N = 200.
• Expected Frequencies:
• E(Young, Alpha) = (100 × 90) / 200 = 45.0 | E(Young, Beta) = (100 × 110) / 200 = 55.0.
• E(Older, Alpha) = (100 × 90) / 200 = 45.0 | E(Older, Beta) = (100 × 110) / 200 = 55.0.

1. Cell 1: (60 − 45)^2 ÷ 45 = 225 ÷ 45 = 5.000.

2. Cell 2: (40 − 55)^2 ÷ 55 = 225 ÷ 55 = 4.091.

3. Cell 3: (30 − 45)^2 ÷ 45 = 225 ÷ 45 = 5.000.

4. Cell 4: (70 − 55)^2 ÷ 55 = 225 ÷ 55 = 4.091. Total χ^2_Calculated = 5.000 + 4.091 + 5.000 + 4.091 = 18.182. (Critical χ^2_0.05, df=1 = 3.841).

• CONCLUSION: χ^2_calc (18.182) >> 3.841 → REJECT H_0 ($p < 0.0001$). Brand preference is strongly dependent on customer age group; younger buyers show a significant preference for Brand Alpha while older demographics favor Brand Beta. 5.3 Measuring Association Strength: Phi Coefficient & Cramér's V

1. Phi (φ) Coefficient (for 2 × 2 Tables) Measures the correlation strength between two binary variables: φ = √[ χ^2 ÷ N ]. Ranges from 0 (no relationship) to 1.0 (perfect association). For our worked example: φ = √[18.182 / 200] = 0.3015 (Moderate-to-Strong association).

2. Cramér's V (for r × c Tables) Generalized effect size metric for contingency tables larger than $2 imes 2$: V = √[ χ^2 ÷ (N × min(r − 1, c − 1)) ]. Protects against sample size inflation in large enterprise cross-tabulation datasets.