8.3 Parametric vs. Non-Parametric Tests

Parametric vs. Non-Parametric Tests

Health data analysts operate across diverse healthcare datasets ranging from physiological continuous monitors and billing dollar amounts to ordinal symptom rating scales and discrete categorical diagnostic codes. Selecting the correct statistical hypothesis test is critical: applying a test whose underlying mathematical assumptions are violated produces invalid p-values, misleading clinical conclusions, and flawed organizational policy. For the Certified Health Data Analyst (CHDA), mastering the dichotomy between parametric and non-parametric methods—and executing the appropriate test based on data scale, distribution shape, sample pairing, and group counts—is a core domain competency.

---

1. Core Theoretical Foundations: Parametric vs. Non-Parametric Paradigms

+---------------------------------------------------------------------------------------------------+
|                             THE PARAMETRIC VS. NON-PARAMETRIC DICHOTOMY                           |
+-------------------------------------------------+-------------------------------------------------+
| PARAMETRIC TESTS                                | NON-PARAMETRIC TESTS                            |
| - Evaluates specific population PARAMETERS (\mu)| - Distribution-free; evaluates MEDIANS / RANKS  |
| - Assumes Normal distribution of residuals      | - No assumption of normality                    |
| - Assumes Homogeneity of Variance               | - Robust against extreme skewness and outliers  |
| - Requires Continuous Interval / Ratio data     | - Operates on Ordinal, Rank, or Nominal data    |
| - Higher statistical power when assumptions met | - Slightly lower power (~95% asymptotic eff.)   |
+-------------------------------------------------+-------------------------------------------------+

Core Assumptions of Parametric Methods

1. Normality of Sampling Distribution: The continuous dependent variable (or residuals) is approximately normally distributed. Tested statistically via the Shapiro-Wilk test (for $n < 50$) or Kolmogorov-Smirnov test (for $n \ge 50$), or visually via Normal Q-Q Plots and histograms. (Note: Under the Central Limit Theorem, parametric tests are robust to moderate non-normality when $n \ge 30$).
2. Homogeneity of Variance (Homoscedasticity): The variance of the dependent variable across comparison groups is equal ($\sigma_1^2 = \sigma_2^2 = \dots = \sigma_k^2$). Tested via Levene's Test for Equality of Variances or Bartlett's Test.
3. Scale of Measurement: The outcome variable is measured on a continuous Interval (e.g., temperature in Celsius) or Ratio scale (e.g., blood pressure in mmHg, financial costs in dollars, length of stay in days).
4. Independence of Observations: Individual data points must be independent of one another (unless using dedicated paired/repeated measures designs).

When Non-Parametric or Exact Tests Are Appropriate

1. Severely Skewed Healthcare Variables: Financial expenditures, ICU length of stay, and medication acquisition costs exhibit extreme right skewness with multi-million-dollar outliers that distort the sample mean $\bar{x}$.
2. Ordinal Clinical Staging and Survey Scales: Data captured on discrete ranked scales where numerical distances between categories are unequal or undefined:
   • Visual Analog Pain Scale ($0$ to $10$)
   • Patient Satisfaction Likert Scales ($1 = \text{Very Dissatisfied}$ to $5 = \text{Very Satisfied}$)
   • Glasgow Coma Scale ($3$ to $15$)
   • NYHA Heart Failure Functional Classes ($I, II, III, IV$)
   • Cancer Staging ($0, I, II, III, IV$)
3. Small Sample Sizes with Non-Normal Data ($n < 15$ per group): Where the Central Limit Theorem cannot be invoked and normality cannot be demonstrated.

---

2. Parametric Statistical Tests in Healthcare Analytics

+---------------------------------------------------------------------------------------------------+
|                                 PARAMETRIC TEST ARCHITECTURE                                      |
+-------------------+--------------------+--------------------+-------------------------------------+
| INDEPENDENT t-TEST| PAIRED t-TEST      | ONE-WAY ANOVA      | TWO-WAY ANOVA                       |
| - 2 Unrelated Grps| - 2 Matched / Pre- | - 3+ Independent   | - 2 Simultaneous Independent Factors|
| - df = n1 + n2 - 2|   Post Measures    |   Groups (Omnibus) | - Evaluates Main Effects +          |
| - Compares Means  | - df = n - 1       | - F = MS_B / MS_W  |   Interaction Effect (A x B)        |
+-------------------+--------------------+--------------------+-------------------------------------+

1. Independent Samples t-Test (Two-Sample t-Test)

• Purpose: Compares the means of a continuous dependent variable between two independent, unrelated cohorts.
• Clinical Example: Comparing the mean Inpatient Length of Stay (days) between patients undergoing laparoscopic cholecystectomy at Hospital A vs. Hospital B.
• Formula (Equal Variances Assumed): $t = \frac{\bar{x}_1 - \bar{x}_2}{s_p \sqrt{\frac{1}{n_1} + \frac{1}{n_2}}}, \quad \text{where } s_p = \sqrt{\frac{(n_1 - 1)s_1^2 + (n_2 - 1)s_2^2}{n_1 + n_2 - 2}}$ $\text{Degrees of Freedom: } df = n_1 + n_2 - 2$
• Welch's Unequal Variances t-Test: If Levene's test reveals unequal group variances ($p < 0.05$), the standard pooled t-test produces inflated Type I errors. Analysts must utilize Welch's t-test, which modifies the standard error formula and uses the Welch-Satterthwaite equation to calculate fractional degrees of freedom.

2. Paired Samples t-Test (Dependent / Repeated Measures t-Test)

• Purpose: Compares the means of two dependent, matched, or repeated measurements taken on the same individual subjects (or matched case-control pairs).
• Clinical Example: Measuring resting Systolic Blood Pressure (mmHg) in $n = 50$ hypertensive patients immediately before initiating an ACE inhibitor (Baseline) and after 12 weeks of daily therapy (Post-treatment).
• Mathematical Mechanics: Reduces paired observations into a single column of pairwise differences ($d_i = x_{i,\text{post}} - x_{i,\text{pre}}$): $t = \frac{\bar{d}}{s_d / \sqrt{n}}$ where $\bar{d}$ is the mean difference, $s_d$ is the standard deviation of differences, and $df = n - 1$.

3. One-Way Analysis of Variance (ANOVA)

• Purpose: Compares continuous means across three or more independent groups simultaneously, avoiding the severe Type I error inflation that would occur if multiple pairwise t-tests were executed.
• Clinical Example: Comparing mean 6-month HbA1c reduction across $k = 4$ diabetic pharmacotherapy regimens: (1) Lifestyle/Diet alone, (2) Metformin monotherapy, (3) GLP-1 Receptor Agonist, and (4) Basal Insulin.
• Hypotheses:
  • $H_0: \mu_1 = \mu_2 = \mu_3 = \dots = \mu_k$
  • $H_1: \text{At least one group mean } \mu_j \text{ differs from the others}$
• The $F$-Statistic: Compares variance between the group means to variance within the groups: $F = \frac{\text{Mean Square Between } (MS_{\text{between}})}{\text{Mean Square Within } (MS_{\text{within}})} = \frac{SS_{\text{between}} / (k - 1)}{SS_{\text{within}} / (N - k)}$ $\text{Degrees of Freedom: } df_{\text{between}} = k - 1, \quad df_{\text{within}} = N - k$
• Omnibus Nature & Post-Hoc Pairwise Testing: A statistically significant ANOVA $F$-test ($p < 0.05$) is omnibus—it indicates that the groups are not all equal, but does not specify which particular pairs differ. To identify specific pairwise differences while maintaining overall $\alpha = 0.05$, analysts execute post-hoc tests:
  • Tukey's Honestly Significant Difference (HSD): Optimal for all pairwise comparisons when group sample sizes are equal.
  • Bonferroni / Sidak Correction: Conservative post-hoc adjustment for planned pairwise contrasts.
  • Dunnett's Test: Specifically compares multiple active treatment arms against a single reference control group.

4. Two-Way ANOVA (Factorial ANOVA)

• Purpose: Evaluates the simultaneous effects of two categorical independent variables (factors) on a single continuous dependent variable, allowing evaluation of Main Effects and the Interaction Effect.
• Clinical Example: Evaluating the effect of Medication Regimen (Factor A: Drug X vs. Drug Y vs. Placebo) and Cardiac Rehabilitation Intensity (Factor B: Standard vs. Intensive) on 12-month Left Ventricular Ejection Fraction (LVEF %).
• The Interaction Effect ($A \times B$): Evaluates whether the effect of Factor A depends on the specific level of Factor B (e.g., Drug X only improves LVEF when paired with Intensive Rehabilitation).

---

3. Categorical Data Analysis: Chi-Square & Fisher's Exact Tests

+---------------------------------------------------------------------------------------------------+
|                                 CATEGORICAL ASSOCIATION TESTS                                     |
+-------------------------------------------------+-------------------------------------------------+
| CHI-SQUARE (\chi^2) TEST OF INDEPENDENCE        | FISHER'S EXACT TEST                             |
| - r x c Contingency Tables                      | - 2 x 2 Contingency Tables                      |
| - Compares Observed (O) vs Expected (E) counts  | - Exact hypergeometric probability              |
| - Formula: \chi^2 = \sum (O - E)^2 / E          | - Preferred for sparse 2x2 tables or when asymptotic assumptions fail |
| - df = (r - 1) * (c - 1)                        | - Small pediatric, oncology, or rare cohorts    |
+-------------------------------------------------+-------------------------------------------------+

Chi-Square ($\chi^2$) Test of Independence

• Purpose: Assesses whether a statistically significant association exists between two nominal or categorical variables presented in an $r \times c$ contingency table.
• Clinical Example: Evaluating whether 30-Day Hospital Readmission (Binary: Yes vs. No) is associated with Discharge Destination (Categorical: Home Self-Care vs. Home Health vs. Skilled Nursing Facility).
• Test Statistic Formula: $\chi^2 = \sum_{i=1}^r \sum_{j=1}^c \frac{(O_{ij} - E_{ij})^2}{E_{ij}}$ where $O_{ij}$ is the observed cell count and $E_{ij}$ is the expected cell count under $H_0$.
• Expected Frequency Formula: $E_{ij} = \frac{\text{Row Total}_i \times \text{Column Total}_j}{\text{Grand Total } N}$
• Degrees of Freedom: $df = (r - 1) \times (c - 1)$
• Assumptions of Chi-Square:
  1. Random sampling and independent observations.
  2. Categories must be mutually exclusive and exhaustive.
  3. Sample Size Assumption: All expected cell frequencies must be $\ge 1$, and at least 80% of cells must have expected frequencies $E_{ij} \ge 5$.

Fisher's Exact Test

• Purpose: An exact non-parametric test for $2 \times 2$ contingency tables utilized when sample sizes are small and the Chi-Square assumption of expected cell frequencies $\ge 5$ is violated.
• Mathematical Mechanics: Calculates the exact hypergeometric probability of observing the specific cell frequencies ($a, b, c, d$) given fixed marginal totals: $P = \frac{(a+b)! \, (c+d)! \, (a+c)! \, (b+d)!}{a! \, b! \, c! \, d! \, N!}$
• Clinical Use Case: Evaluating surgical site infections between two rare pediatric neurosurgical approaches where total cases $N = 25$ and observed infections are $0$ and $3$.

---

4. Non-Parametric Equivalents: The Complete Analytic Crosswalk

When continuous healthcare variables violate normality assumptions or when data is ordinal, analysts substitute parametric tests with their corresponding non-parametric rank-based equivalents.

+---------------------------------------------------------------------------------------------------+
|                             PARAMETRIC TO NON-PARAMETRIC MAPPING                                  |
+-------------------------------+-------------------------------+-----------------------------------+
| ANALYTIC SCENARIO             | PARAMETRIC TEST (Normality)   | NON-PARAMETRIC EQUIVALENT (Ranks) |
+-------------------------------+-------------------------------+-----------------------------------+
| 2 Independent Groups          | Independent Samples t-Test    | MANN-WHITNEY U TEST (Wilcoxon)    |
| 2 Paired / Pre-Post Measures  | Paired Samples t-Test         | WILCOXON SIGNED-RANK TEST         |
| 3+ Independent Groups         | One-Way ANOVA (F-Test)        | KRUSKAL-WALLIS H TEST             |
| 3+ Repeated / Paired Measures | Repeated Measures ANOVA       | FRIEDMAN TEST                     |
| 2 Categorical Variables       | Pearson Chi-Square (\chi^2)   | FISHER'S EXACT TEST (Small n)     |
+-------------------------------+-------------------------------+-----------------------------------+

1. Mann-Whitney U Test (Wilcoxon Rank-Sum Test)

• Parametric Equivalent: Independent Samples t-Test.
• Mechanics: Combines all observations from both groups, ranks them in ascending order from $1$ to $N$, and calculates the sum of ranks for each group. Tests whether the distribution medians differ.
• Clinical Application: Comparing Emergency Department Length of Stay (hours, severely right-skewed) between patients presenting during Day Shift vs. Night Shift.

2. Wilcoxon Signed-Rank Test

• Parametric Equivalent: Paired Samples t-Test.
• Mechanics: Calculates pairwise differences ($d_i = x_{\text{post}} - x_{\text{pre}}$), drops zero differences, ranks the absolute values $|d_i|$ from lowest to highest, attaches the original positive/negative signs to the ranks, and sums the ranks.
• Clinical Application: Comparing patient-reported pain scores ($0$ to $10$ Visual Analog Scale) measured on the same cohort of oncology patients immediately before and after palliative acupuncture.

3. Kruskal-Wallis H Test

• Parametric Equivalent: One-Way ANOVA.
• Mechanics: Extends the Mann-Whitney U concept to $k \ge 3$ independent groups. Ranks all $N$ pooled observations and computes the $H$-statistic based on rank sums per group ($df = k - 1$).
• Post-Hoc Testing: When Kruskal-Wallis is statistically significant ($p < 0.05$), analysts perform Dunn's Post-Hoc Test with Bonferroni adjustment to identify specific pairwise differences.
• Clinical Application: Comparing Hospital Consumer Assessment of Healthcare Providers and Systems (HCAHPS) patient satisfaction ratings ($1$ to $5$ Likert scale) across 5 distinct regional health system clinics.

4. Friedman Test

• Parametric Equivalent: Repeated Measures ANOVA.
• Mechanics: Ranks observations across $k \ge 3$ repeated time points or conditions within each individual subject.
• Clinical Application: Tracking Functional Independence Measure (FIM) motor scores in stroke rehabilitation patients across four sequential assessment intervals: Admission, Week 2, Week 4, and Discharge.

---

5. Step-by-Step Worked Chi-Square Calculation Example

Healthcare Scenario: A hospital quality analytics team evaluates whether implementing an "Enhanced Recovery After Surgery" (ERAS) multimodal clinical pathway significantly reduces 30-day all-cause readmissions among colorectal surgery patients compared to standard traditional post-operative care. The quality department tracks $N = 400$ total surgical discharges over 12 months.

Step 1: Formulate the Hypotheses

• Null Hypothesis ($H_0$): 30-Day Readmission is independent of the surgical care pathway (Discharge pathway and readmission status are unrelated).
• Alternative Hypothesis ($H_1$): 30-Day Readmission is dependent on the surgical care pathway (Readmission rates differ significantly between ERAS and Standard Care).
• Significance Level: $\alpha = 0.05$.

Step 2: Construct the $2 \times 2$ Contingency Table of Observed Counts ($O_{ij}$)

Step 3: Calculate Expected Frequencies ($E_{ij} = \frac{R_i \times C_j}{N}$)

1. ERAS / Readmitted ($E_{11}$): $E_{11} = \frac{R_1 \times C_1}{N} = \frac{200 \times 70}{400} = \mathbf{35.00}$
2. ERAS / Not Readmitted ($E_{12}$): $E_{12} = \frac{R_1 \times C_2}{N} = \frac{200 \times 330}{400} = \mathbf{165.00}$
3. Standard / Readmitted ($E_{21}$): $E_{21} = \frac{R_2 \times C_1}{N} = \frac{200 \times 70}{400} = \mathbf{35.00}$
4. Standard / Not Readmitted ($E_{22}$): $E_{22} = \frac{R_2 \times C_2}{N} = \frac{200 \times 330}{400} = \mathbf{165.00}$

Assumption Check: All expected frequencies are $35.0$ and $165.0$, all strictly $\ge 5$. The Chi-Square assumption is satisfied.

Step 4: Calculate Cell Chi-Square Components ($\frac{(O - E)^2}{E}$)

Step 5: Sum Components & Determine Critical Value

$\chi^2 = 6.4286 + 1.3636 + 6.4286 + 1.3636 = \mathbf{15.5844} \approx 15.58$ $\text{Degrees of Freedom: } df = (r - 1) \times (c - 1) = (2 - 1) \times (2 - 1) = 1 \times 1 = \mathbf{1}$

• Critical Value for $df = 1, \alpha = 0.05$: $\chi^2_{\text{crit}} = \mathbf{3.841}$.
• Critical Value for $df = 1, \alpha = 0.001$: $\chi^2_{\text{crit}} = \mathbf{10.828}$.

Step 6: Statistical Decision & Clinical Interpretation

• Statistical Decision: Because the calculated $\chi^2 = 15.58$ substantially exceeds the critical threshold of $3.841$ (and exceeds $10.828$, yielding $p < 0.001$), we Reject $H_0$ at $\alpha = 0.05$.
• Clinical Conclusion: There is a statistically significant association between the clinical care pathway and 30-day readmissions. Colorectal surgery patients managed under the ERAS pathway experienced a significantly lower readmission rate ($20/200 = 10.0%$) compared to standard care ($50/200 = 25.0%$).

---

6. Master Decision Guide for Statistical Test Selection