3.5.5 1 Sample Sign

1 Sample Sign Introduction The 1 Sample Sign test is a nonparametric hypothesis test used to evaluate a median when data are: - Not normally distributed - Ordinal (ranked) or continuous but skewed - Possibly affected by outliers It is an alternative to the 1-sample t-test when assumptions for the t-test are not met. Instead of using the raw values, the 1 Sample Sign test uses only the signs (+ or −) of differences from a hypothesized median. --- Purpose of the 1 Sample Sign Test When to Use Use the 1 Sample Sign test when: - The parameter of interest is the population median (not the mean). - Data are at least ordinal (can be ordered). - The distribution may be: - Skewed - Contaminated with outliers - Not well modeled by a normal distribution - There is a specified target or benchmark median value to compare against. Typical situations include: - Comparing process performance to a target when data are strongly skewed. - Evaluating median time, median cost, or median satisfaction level. - Analyzing before/after differences when only direction (better/worse) is reliable. --- Hypotheses and Parameter Parameter of Interest The 1 Sample Sign test focuses on the population median: - θ = true population median You test whether θ differs from a hypothesized value θ₀. Form of Hypotheses - Two-sided test: - H₀: θ = θ₀ - H₁: θ ≠ θ₀ - Upper-tailed test: - H₀: θ = θ₀ - H₁: θ > θ₀ - Lower-tailed test: - H₀: θ = θ₀ - H₁: θ < θ₀ Choose the form based on the practical question: - Two-sided: checking for any difference from the target. - Upper-tailed: checking if the median is too high. - Lower-tailed: checking if the median is too low. --- Data Requirements and Assumptions Data Requirements - A single sample from the population - Independent observations - Measurement scale: - Ordinal, interval, or ratio is acceptable - Hypothesized median θ₀ must be specified in advance. Core Assumptions - Observations are independent and identically distributed. - The probability that a value is exactly equal to θ₀ is small, or any ties are handled correctly. - The distribution is continuous or approximately so (important for the sign behavior). The test does not assume: - Normality - Equal variances - Any specific distributional shape --- Test Logic and Test Statistic Conceptual Logic For each observation ( xi ), compare it to the hypothesized median ( θ0 ): - If ( xi > θ0 ) → record a + sign - If ( xi < θ0 ) → record a − sign - If ( xi = θ0 ) → usually discard the observation from the test If the true median equals θ₀, then: - Values are equally likely to be above or below θ₀. - The signs should behave like fair coin flips: - P(+) = 0.5 - P(−) = 0.5 The 1 Sample Sign test checks whether the observed imbalance between + and − is too large to be attributed to random variation. Test Statistic Definition Let: - n = number of usable observations (excluding ties with θ₀) - S = number of positive signs (xᵢ > θ₀) - Then number of negative signs is n − S Under H₀: θ = θ₀, the test statistic S follows a Binomial(n, 0.5) distribution. The 1 Sample Sign test uses: - The value of S (or sometimes the smaller of S and n − S for a two-sided test). - The binomial distribution to compute the p-value. --- Performing the 1 Sample Sign Test Step 1: State the Question and Hypotheses - Identify the practical question in terms of the median. - Choose θ₀ (the target or benchmark). - Formulate H₀ and H₁ clearly. Example pattern: - H₀: median cycle time = 20 minutes - H₁: median cycle time > 20 minutes Step 2: Prepare the Data For each observation: - Compute the difference ( di = xi − θ_0 ). - Assign a sign: - dᵢ > 0 → + - dᵢ < 0 → − - dᵢ = 0 → exclude from n Count: - n = number of nonzero differences - S = number of positive signs Check that n is not too small. Very small n reduces test power. Step 3: Determine the Test Statistic and Distribution - Under H₀: S ~ Binomial(n, 0.5). - For a given n, the distribution of S is fully defined. For large n, software may approximate this binomial with a normal distribution, but conceptually the test remains binomial. Step 4: Compute the p-Value - Two-sided test: - Evaluate how extreme S is in either direction under Binomial(n, 0.5). - p-value = P(S ≤ observed S) + P(S ≥ opposite tail value), or a symmetric equivalent based on the smaller tail probability. - Upper-tailed test (H₁: θ > θ₀): - p-value = P(S ≥ observed S | Binomial(n, 0.5)) - Lower-tailed test (H₁: θ < θ₀): - p-value = P(S ≤ observed S | Binomial(n, 0.5)) In practice, use statistical software, but know that the underlying calculation is based on the binomial distribution. Step 5: Decision and Interpretation Compare the p-value with the chosen significance level (α), commonly 0.05: - If p-value ≤ α: - Reject H₀. - Conclude that the data provide evidence that the median differs from θ₀ (direction depending on H₁). - If p-value > α: - Do not reject H₀. - Conclude that the data do not provide sufficient evidence to say the median differs from θ₀. Interpret in terms of the process median or performance target, not just in statistical terms. --- One-Sided vs Two-Sided Tests Choosing the Test Direction - Use a two-sided test when: - Any deviation from the target is important (higher or lower). - There is no specific directional expectation prior to the analysis. - Use a one-sided test when: - Only one direction is critical (for example, cycle time being too long). - There is a justified directional practical concern. Effect on p-Value - For the same data, a one-sided test: - Has a smaller p-value than a two-sided test if the effect is in the hypothesized direction. - The decision about one-sided vs two-sided must be made before looking at the data to maintain integrity. --- Handling Ties and Zero Differences Ties at the Hypothesized Median If some observations equal θ₀ exactly: - They do not favor + or −. - Standard practice: - Exclude them from the analysis. - Reduce n accordingly (n = total observations − ties). Implications: - Many ties reduce effective sample size and power. - Extreme numbers of ties may suggest: - Data rounding - Insufficient measurement resolution - A true median very close to θ₀ Practical Treatment When ties are present: - Report: - Total sample size - Number of ties - Effective n used in the test - Consider whether the measurement system is adequate to support median-based decisions. --- Comparison with the 1 Sample Wilcoxon Test The 1 Sample Sign test is related to, but simpler than, the 1 Sample Wilcoxon (Wilcoxon signed-rank) test. Key Differences - Information used: - 1 Sample Sign: - Uses only the sign of each difference (+ or −). - 1 Sample Wilcoxon: - Uses both sign and magnitude (ranks of absolute differences). - Efficiency: - 1 Sample Sign: - Simpler, more robust but usually less powerful (requires larger n to detect the same effect). - 1 Sample Wilcoxon: - More powerful under standard conditions but slightly more sensitive to distribution shape. - When Sign is preferable: - When magnitude is untrustworthy (e.g., rough or coarsely measured data). - When outliers or heavy-tailed distributions might overly influence rank-based magnitudes. - When only direction of difference can be meaningfully interpreted. Understanding this comparison helps justify the choice of the 1 Sample Sign test in appropriate conditions. --- Strengths and Limitations Strengths - Distribution-free: - No normality assumption. - Robust to outliers: - Outliers affect only a sign, not magnitude. - Simple interpretation: - Counts of “above target” vs “below target.” - Applicable to ordinal data: - Can be used when distances between categories are not meaningful but order is. Limitations - Lower power: - Less sensitive than tests that use magnitudes (such as the Wilcoxon test or t-test when assumptions hold). - Loss of information: - Ignores size of differences. - Difficult with many ties: - Large number of values equal to θ₀ reduces usable data. Recognizing these trade-offs guides appropriate and efficient use of the test. --- Practical Considerations and Common Pitfalls Sample Size and Power - Small n: - Limited resolution of p-values. - Harder to detect practical differences. - Larger n: - Gives more reliable inferences about the median. When planning analysis, consider whether the sample size is sufficient to detect a practically meaningful shift in the median. Misinterpretation of Non-Significance A non-significant result (p-value > α): - Does not prove that the median equals θ₀. - Indicates insufficient evidence to conclude a difference. - May reflect: - Small sample size - High data variability - A difference that is smaller than the test can reliably detect Always interpret results in the context of: - Practical requirements - Data quality - Sample size Confusion Between Mean and Median The test focuses on the median, not the mean: - Passing or failing the test does not directly speak about the mean. - In skewed distributions, mean and median can differ substantially. - When the center of interest is truly the mean, and normality is approximately satisfied, a parametric test (like the t-test) may be more appropriate. Staying clear about which measure of central tendency is under analysis avoids incorrect conclusions. --- Summary The 1 Sample Sign test is a simple, robust nonparametric method for testing a population median against a specified value. It: - Converts each observation into a sign based on whether it is above or below the hypothesized median. - Uses the number of positive signs as a binomial test statistic. - Requires minimal assumptions about the underlying distribution. - Is especially useful with skewed data, ordinal data, or data contaminated by outliers. Mastery of the 1 Sample Sign test includes: - Formulating correct hypotheses about the median. - Correctly assigning signs and handling ties. - Using the binomial distribution to interpret the test statistic and p-value. - Choosing one-sided or two-sided tests based on the practical question. - Understanding strengths and limitations relative to other tests, such as the 1 Sample Wilcoxon test. Used appropriately, the 1 Sample Sign test provides a clear, distribution-free way to evaluate whether a process median aligns with a target or requirement.