Vijaya Industries: Saree Defect Rate Analysis

Solution

Comparing Process A (traditional Chirala technique: 100 defects in 10,000 sarees) and Process B (innovative Siddipet technique: 2 defects in 500 sarees) for Vijaya Industries' Pochampally silk sarees presents several statistical challenges:

• 1. Highly Imbalanced Sample Sizes:
  • The sample size for Process A (n_A = 10,000) is much larger than for Process B (n_B = 500).
  • Challenge: While the defect rate for Process A (p_A = 100/10000 = 0.01 or 1%) is based on a large sample and thus relatively stable, the defect rate for Process B (p_B = 2/500 = 0.004 or 0.4%) is estimated from a much smaller sample. This means the estimate for p_B has higher variability and uncertainty. Standard tests might give disproportionate weight to the larger sample or struggle with the precision difference.
• 2. Small Number of Defects (Rare Events) in Process B:
  • With only 2 defects observed in Process B, we are dealing with a rare event in that sample.
  • Challenge: Many standard hypothesis tests, like the Pearson's chi-squared test or the normal approximation-based z-test for two proportions, rely on asymptotic theory. These approximations work well when expected cell counts (for chi-squared) or the number of successes and failures (for z-test, i.e., np and n(1-p)) are sufficiently large (often a rule of thumb is > 5).
    • For Process B, the observed number of defects is 2. If the true defect rate is low, the expected number of defects in a sample of 500 might also be small. For instance, if comparing against Process A's rate, the expected defects in Process B's sample would be 500 * 0.01 = 5. This just meets the threshold, but the observed is lower. If the null was that p_B = p_A, then the expected count is 5 in cell B-defect, and 495 in B-non-defect. For cell A-defect, 100, and A-non-defect 9900. The issue is more pronounced if p_B is truly much smaller.
• 3. Impact on Reliability of Standard Hypothesis Tests:
  • Chi-squared Test: May not be reliable if expected frequencies in any cell of the 2x2 contingency table (Defect/No Defect vs. Process A/Process B) are too small (typically < 5). Here, the cell for "Process B, Defect" has an observed count of 2. The expected count under the null hypothesis of no difference in proportions (pooled p ≈ 102/10500 ≈ 0.0097) would be 0.0097 * 500 ≈ 4.85, which is borderline or below 5 for some stricter rules of thumb. This can lead to an inaccurate p-value.
  • Z-test for Two Proportions: Relies on the normal approximation to the binomial distribution, which is less accurate when 'np' or 'n(1-p)' is small for either group. For Process B, n_Bp_B = 2 and n_B(1-p_B) = 498. The 'np=2' is problematic.
  • Consequences of Unreliability: Using these tests when their assumptions are violated can lead to:
    • Incorrect p-values: Resulting in an increased risk of Type I error (falsely concluding there's a difference when there isn't) or Type II error (failing to detect a real difference).
    • Reduced Statistical Power: Especially with small counts for one group, the ability to detect a true difference if it exists might be compromised.
• 4. Impact on Weavers' Livelihoods and Business Decisions:
  • The quality control manager needs to make a crucial decision for Vijaya Industries, based in Chirala, before the Sankranti shopping season. An incorrect statistical conclusion could:
    • Lead to unnecessarily discarding the promising Siddipet technique if a true improvement is missed (Type II error), thereby continuing with higher defect rates in their Pochampally silk sarees and affecting efficiency and profitability.
    • Lead to prematurely adopting the new Siddipet technique across all handlooms if a difference is falsely detected or its magnitude overestimated (Type I error or misinterpretation), potentially causing unforeseen scaling issues or initial dips in productivity while weavers across Telangana and Andhra Pradesh adapt.
    This decision directly impacts material costs (precious handloom silk), production efficiency, the reputation of these culturally significant sarees, and ultimately the work and income of hundreds of weavers.

Therefore, more robust statistical methods designed for small counts and imbalanced samples are needed to provide a reliable basis for this important decision.

Solution

Given the challenges of imbalanced sample sizes and small defect counts in Process B (the innovative Siddipet technique for Pochampally silk sarees), I would recommend the following alternative statistical approaches to the quality manager at Vijaya Industries in Chirala:

• 1. Fisher's Exact Test:
  • Description: This test is specifically designed for analyzing contingency tables (like a 2x2 table of Process A/B vs. Defect/No Defect) when sample sizes are small or expected frequencies are low, violating assumptions of the chi-squared test. It calculates the exact probability of observing the collected data, or data more extreme, assuming the null hypothesis (that the proportions of defects are the same for both processes) is true.
  • Tradeoffs:
    • Pros: It is an "exact" test, meaning it doesn't rely on large-sample approximations, making it highly suitable for the 2 defects observed in the Siddipet technique. It's very reliable in such situations.
    • Cons: Can be computationally intensive for very large tables (though not an issue here). It can also be somewhat conservative (i.e., have lower power to detect a difference than an approximate test if the approximation were valid), meaning it might require stronger evidence to declare a significant difference.
• 2. Confidence Intervals for Proportions and Their Difference:
  • Description: Instead of just a p-value, calculate confidence intervals (CIs) for:
    • The defect rate of Process A (p_A).
    • The defect rate of Process B (p_B). For Process B with 2 defects in 500 sarees, an "exact" method like the Clopper-Pearson interval or a less conservative but still good option like the Wilson score interval should be used.
    • The difference in proportions (p_A - p_B) or the rate ratio (p_B / p_A). Methods like Newcombe's method for the difference or approaches for rate ratios with small counts can be used.
  • Tradeoffs:
    • Pros: CIs provide a range of plausible values for the true defect rates and their difference/ratio, offering insight into the magnitude and uncertainty of the effect. This is often more informative for business decisions than a p-value alone. The quality manager can assess if the entire CI for the difference lies in a region that indicates a practically meaningful improvement for these valuable Pochampally silk sarees.
    • Cons: Calculating accurate CIs for differences or ratios with very small counts and imbalanced samples can still be complex, and different methods can yield slightly different intervals. Interpretation of overlapping CIs needs care.
• 3. Exact Binomial Test (for Process B against a benchmark):
  • Description: If Process A's defect rate (1%) is considered a stable benchmark, one could perform a one-sample exact binomial test on Process B's data. The null hypothesis would be that Process B's defect rate is greater than or equal to 0.01, versus the alternative that it's less than 0.01. We observed 2 defects in 500 trials for Process B.
  • Tradeoffs:
    • Pros: Directly tests if Process B is better than the established rate. Exact for binomial data.
    • Cons: Doesn't directly use the variability from Process A's sample in the same test, treating p_A as fixed. This might not be ideal if p_A itself has some uncertainty, though with n=10,000, it's fairly stable.
• 4. Bayesian Approaches:
  • Description: Model the defect rates using binomial distributions and assign prior distributions to the unknown defect proportions (p_A and p_B). The experience of the master weaver from Siddipet or historical data from Chirala could inform these priors (e.g., a Beta distribution). After observing the data, calculate the posterior distributions for p_A, p_B, and their difference or ratio. This allows for probabilistic statements like "There is an X% chance that Process B's defect rate is lower than Process A's."
  • Tradeoffs:
    • Pros: Formally incorporates prior knowledge/expertise, which is valuable for traditional Telugu weaving. Handles small counts naturally. Provides intuitive probabilistic outputs (e.g., probability of improvement).
    • Cons: The choice of prior can be subjective and influence results, which might be a concern for regulatory or highly critical decisions. Can be computationally more intensive and conceptually more complex for stakeholders less familiar with Bayesian statistics.

Recommendation Context: Given the cultural significance of Pochampally sarees (often family heirlooms) and the impact on weavers' livelihoods across Telangana and Andhra Pradesh, a cautious yet informative approach is needed. I would recommend using Fisher's Exact Test for a formal significance test, supplemented heavily by exact confidence intervals for the individual defect rates and their difference/ratio. This combination provides both a statistical test robust to small counts and an estimate of the potential magnitude of improvement (or lack thereof) along with its uncertainty. This allows the quality manager at Vijaya Industries to assess not just statistical significance but also practical significance before the crucial Sankranti shopping season.

Solution

To recommend a minimum sample size for Process B (the innovative Siddipet technique) for Vijaya Industries, we need to perform a sample size calculation for comparing two proportions. This requires several inputs to be decided in discussion with the quality manager, balancing statistical rigor with the practical costs of testing precious Pochampally handloom silk before the wedding season:

Inputs for Sample Size Calculation:

• 1. Baseline Defect Rate (Process A - Chirala): p_A = 100/10,000 = 0.01 (1%).
• 2. Expected Defect Rate for Process B (or Minimum Detectable Effect - MDE): This is the crucial part. What level of improvement would make it worthwhile for Vijaya Industries to switch?
  • The pilot showed p_{B_observed} = 2/500 = 0.004 (0.4%).
  • The quality manager might decide that if Process B can reliably achieve a defect rate of, say, p_{B_target} = 0.005 (0.5%), this 50% reduction from 1% would be a significant improvement worth detecting. This means an absolute difference of 0.01 - 0.005 = 0.005.
• 3. Significance Level (Alpha, α): The probability of a Type I error (falsely concluding Process B is better). Commonly set at 0.05. Since they are looking for fewer defects, a one-sided test is appropriate (H₀: p_B ≥ p_A vs H₁: p_B < p_A).
• 4. Statistical Power (1-β): The probability of correctly detecting a true difference if Process B is indeed better by the MDE. Commonly set at 0.80 (80%) or 0.90 (90%). Higher power requires larger sample sizes. Let's use 80% for this example.

Sample Size Calculation (Illustrative):

Using these parameters (p_A=0.01, p_{B_target}=0.005, α=0.05 one-sided, Power=0.80), we can use statistical software or an online calculator for comparing two proportions. (Note: The sample size for Process A is very large, so we are mainly calculating the required n_B).

For these values, a typical sample size calculation would yield a required sample size for Process B (n_B) in the range of approximately 2,500 to 3,000 sarees.

(Calculation detail for context: Using a formula or calculator, if n_A is considered very large or infinite, and we want to detect a shift from p_A=0.01 to p_B=0.005 with 80% power and one-sided alpha=0.05, n_B ≈ (Z_α√(p_A(1-p_A)) + Z_β√(p_B(1-p_B)))² / (p_A-p_B)². With Z_α=1.645 and Z_β=0.84, this would be roughly n_B ≈ (1.645*√(0.01*0.99) + 0.84*√(0.005*0.995))² / (0.005)² ≈ (0.1637 + 0.0592)² / 0.000025 ≈ (0.2229)² / 0.000025 ≈ 0.04968 / 0.000025 ≈ 1987. This is a simplified formula; actual software might give slightly different numbers, often higher when accounting for continuity corrections or specific test properties, e.g., around 2600-2800 if using tools that account for unequal sample sizes or specific test statistics for proportions.)

Recommendation to the Quality Manager:

"To make a data-driven decision with reasonable statistical confidence before the peak wedding and Sankranti seasons, and to truly verify if the innovative Siddipet technique reduces defects in our Pochampally silk sarees from the current 1% (Chirala process) to around 0.5% or better, I would recommend increasing the sample size for Process B to approximately 2,800 sarees.

Balancing Costs and Confidence:

• This sample size aims for 80% power to detect a reduction in defect rate to 0.5% at a 5% significance level (one-sided test). This means if the Siddipet technique is truly this good, we'd have an 80% chance of statistically confirming it.
• The current sample of 500 sarees is likely insufficient to provide this level of confidence for such a critical improvement. With only 500 sarees, we might miss a real benefit (commit a Type II error).
• Testing an additional ~2,300 sarees with the new Siddipet technique involves costs related to precious handloom silk and weavers' time. However, this must be weighed against:
  • The potential long-term savings from reduced defects if the new technique is adopted (less wasted silk, fewer reworks).
  • The enhanced reputation of Vijaya Industries for producing high-quality, culturally significant Pochampally sarees.
  • The risk of making a wrong decision based on insufficient data, which could impact resources and the livelihoods of weavers across Telangana and Andhra Pradesh.
• This recommended sample size seeks a balance. We could aim for higher power (e.g., 90%), which would require an even larger sample, or accept detecting only a larger difference, which would require a smaller sample. However, reducing defects by half (from 1% to 0.5%) seems like a practically significant goal for these valuable garments."

This approach ensures that the decision to potentially overhaul a traditional Telugu weaving technique is based on robust evidence, respecting both the heritage and the need for modern quality improvements, crucial for the upcoming peak seasons.