7.2 Measures of Central Tendency & Dispersion

Measures of Central Tendency & Dispersion

Descriptive statistics form the mathematical foundation of healthcare data analytics. When analyzing clinical outcomes, financial performance, or operational workflows, a health data analyst must effectively synthesize large volumes of raw transactional data into meaningful, interpretable summary measures. A complete statistical description of any healthcare variable requires two complementary dimensions: measures of central tendency (which identify the central, typical, or expected value of the distribution) and measures of dispersion or variability (which quantify the degree of spread, scatter, or variation among observations).

Reporting a measure of central tendency without an accompanying measure of dispersion presents an incomplete and potentially dangerous clinical picture. For example, two surgical departments may both exhibit an average operative duration of 120 minutes; however, if Department A has a standard deviation of 10 minutes (predictable schedule) and Department B has a standard deviation of 65 minutes (high volatility and operational delays), their operational profiles are vastly different.

---

1. Measures of Central Tendency

Central tendency measures identify the single value that best summarizes an entire distribution of observations. The choice of metric depends strictly on the scale of measurement and the underlying shape of the distribution.

+---------------------------------------------------------------------------------------------------+
|                             MEASURES OF CENTRAL TENDENCY OVERVIEW                                 |
+---------------------------------------------------------------------------------------------------+
| 1. ARITHMETIC MEAN (x̄)  --> Sum of all values divided by sample size n; sensitive to outliers   |
| 2. WEIGHTED MEAN (x̄_w)  --> Accounts for varying group sizes, case-mix weights, or frequencies   |
| 3. GEOMETRIC MEAN (GM)  --> nth root of product; essential for log-normal viral loads and GMLOS   |
| 4. MEDIAN (Q2 / 50th %) --> Middle value when sorted; robust against extreme outliers and skew    |
| 5. MODE                 --> Most frequently occurring value; only central tendency for nominal    |
+---------------------------------------------------------------------------------------------------+

1. Arithmetic Mean

The arithmetic mean (sample mean $\bar{x}$, population mean $\mu$) is the mathematical average of all values in a dataset, calculated as the sum of all observed values divided by the total number of observations ($n$): $\bar{x} = \frac{\sum_{i=1}^n x_i}{n} = \frac{x_1 + x_2 + \dots + x_n}{n}$

• Mathematical Properties: Incorporates every data value in the calculation. The sum of signed deviations from the mean is always zero: $\sum (x_i - \bar{x}) = 0$.
• Healthcare Application: Ideal for continuous, normally distributed physiological variables (e.g., serum electrolytes, adult heights, vital signs in baseline cohorts).
• Major Limitation: Extreme sensitivity to outliers. A single catastrophic ICU stay of 120 days will dramatically inflate the arithmetic mean length of stay for an entire surgical cohort, misrepresenting the experience of typical patients.

2. Weighted Mean

The weighted mean ($\bar{x}_w$) calculates an average where each data point or sub-group is assigned a weight ($w_i$) proportional to its relative importance, population volume, or resource intensity: $\bar{x}_w = \frac{\sum_{i=1}^k (w_i \cdot x_i)}{\sum_{i=1}^k w_i}$

• Healthcare Application: Essential when aggregating clinical quality indicators, readmission rates, or financial metrics across multiple hospitals or clinical departments with unequal patient volumes.
• Example: Calculating a health system's aggregate Case Mix Index (CMI) across three hospitals: $\text{Health System CMI} = \frac{(\text{Discharges}_A \times \text{CMI}_A) + (\text{Discharges}_B \times \text{CMI}_B) + (\text{Discharges}_C \times \text{CMI}_C)}{\text{Total Discharges}_{A+B+C}}$

3. Geometric Mean

The geometric mean ($GM$) is calculated by multiplying $n$ positive numbers together and taking the $n$-th root, or equivalently, by taking the antilogarithm of the arithmetic mean of the log-transformed data: $GM = \sqrt[n]{x_1 \cdot x_2 \dots x_n} = \exp\left( \frac{1}{n} \sum_{i=1}^n \ln(x_i) \right)$

• Healthcare Applications:
  • Clinical Laboratory & Immunology: Used when laboratory measurements span multiple orders of magnitude and exhibit log-normal distributions, such as HIV-1 RNA plasma viral loads (copies/mL), COVID-19 antibody neutralization titers, and serial dilution assays.
  • Medicare Inpatient Prospective Payment System (IPPS): CMS calculates the Geometric Mean Length of Stay (GMLOS) for every MS-DRG. GMLOS dampens the distortion of long-stay outlier cases and serves as the statutory baseline for transfer payment policies and post-acute care transfer adjustments.

4. Median (50th Percentile / $Q_2$)

The median is the physical midpoint of a dataset when all observations are arranged in ascending or descending order. At least half of observations are at or below the median and at least half are at or above it; ties can prevent an exact 50/50 split.

• Calculation:
  • If $n$ is odd, the median is the single value at position $\frac{n + 1}{2}$.
  • If $n$ is even, the median is the arithmetic average of the two central values at positions $\frac{n}{2}$ and $\frac{n}{2} + 1$.
• Robustness: The median is resistant (robust) to outliers and skewness. Replacing an extreme value with a number ten times larger has zero effect on the median.
• Healthcare Application: A robust, commonly preferred measure of central tendency for skewed clinical and financial distributions—including Inpatient Length of Stay, Total Billed Hospital Charges, Emergency Department Wait Times, and Pharmacy Costs.

5. Mode

The mode is the data value or category that occurs with the highest frequency in a dataset.

• Modality Classifications:
  • Unimodal: Exactly one peak / most frequent value.
  • Bimodal: Two distinct values share the highest frequency.
  • Multimodal: Three or more values tie for peak frequency.
  • No Mode: All values occur with equal frequency (e.g., in a uniform distribution).
• Healthcare Application: The only permissible measure of central tendency for nominal data (e.g., determining the most common primary admitting ICD-10 diagnosis, the most frequently prescribed antibiotic in the ICU, or the most common blood type in a trauma registry).

---

2. Measures of Dispersion & Variability

Measures of dispersion quantify the extent to which data points spread out around their central tendency. Without dispersion metrics, analysts cannot assess risk, clinical consistency, or process reliability.

+---------------------------------------------------------------------------------------------------+
|                                 MEASURES OF DISPERSION OVERVIEW                                   |
+---------------------------------------------------------------------------------------------------+
| 1. RANGE                  --> Maximum - Minimum; highly vulnerable to sample size and extremes    |
| 2. INTERQUARTILE RANGE    --> IQR = Q3 - Q1 (Middle 50% spread); robust, paired with median       |
| 3. SAMPLE VARIANCE (s²)   --> Sum of squared deviations divided by (n - 1); units are squared     |
| 4. STANDARD DEVIATION (s) --> Square root of variance; restores original units; paired with mean  |
| 5. COEFFICIENT OF VARIATION-> CV = (s / x̄) * 100%; dimensionless relative variability comparison |
+---------------------------------------------------------------------------------------------------+

1. Range

The range is the simplest measure of dispersion, representing the arithmetic difference between the maximum and minimum observed values: $\text{Range} = x_{\max} - x_{\min}$

• Limitation: Highly sensitive to sample size and extreme outliers; ignores the distribution of the remaining $n - 2$ data points.

2. Interquartile Range (IQR)

The interquartile range (IQR) measures the spread of the middle 50% of ordered observations, calculated as the difference between the third quartile ($Q_3$, 75th percentile) and the first quartile ($Q_1$, 25th percentile): $\text{IQR} = Q_3 - Q_1$

• Quartile Definitions:
  • $Q_1$ (25th Percentile / Lower Quartile): Value below which 25% of data lie.
  • $Q_2$ (50th Percentile / Median): Value below which 50% of data lie.
  • $Q_3$ (75th Percentile / Upper Quartile): Value below which 75% of data lie.
• Outlier Detection Fences (Tukey Method):
  • $\text{Lower Inner Fence} = Q_1 - 1.5 \times \text{IQR}$
  • $\text{Upper Inner Fence} = Q_3 + 1.5 \times \text{IQR}$
  • Values falling outside these fences are classified as statistical outliers.
• Healthcare Application: The standard companion dispersion measure reported alongside the median for all non-normal, skewed healthcare metrics (LOS, costs, cycle times).

3. Variance

Variance measures the average squared deviation of each data point from the arithmetic mean.

• Sample Variance ($s^2$): Divides by $n - 1$ (Bessel's correction) to provide an unbiased estimator of the population variance: $s^2 = \frac{\sum_{i=1}^n (x_i - \bar{x})^2}{n - 1}$
• Population Variance ($\sigma^2$): Used when the dataset comprises the entire population ($N$): $\sigma^2 = \frac{\sum_{i=1}^N (x_i - \mu)^2}{N}$
• Limitation: Expressed in squared units of measurement (e.g., $\text{days}^2$, $\text{dollars}^2$, $(\text{mg/dL})^2$), which lacks intuitive clinical interpretability.

4. Standard Deviation

The standard deviation (sample $s$, population $\sigma$) is the positive square root of the variance, restoring the dispersion metric to the original units of measurement: $s = \sqrt{s^2} = \sqrt{\frac{\sum_{i=1}^n (x_i - \bar{x})^2}{n - 1}}$

• Healthcare Application: The primary dispersion metric paired with the arithmetic mean for normally distributed clinical datasets. It quantifies the expected clinical fluctuation or biological variation within a patient cohort.

5. Coefficient of Variation (CV)

The coefficient of variation ($CV$), also called relative standard deviation (RSD), expresses the standard deviation as a percentage of the arithmetic mean: $CV = \left( \frac{s}{\bar{x}} \right) \times 100\%$

• Key Properties: Dimensionless (unitless) metric that adjusts for differences in scale or magnitude.
• Healthcare Applications:
  • Clinical Laboratory Quality Control: Used to benchmark the analytical precision and repeatability of laboratory analyzers across different clinical chemistry assays. A blood glucose analyzer with a $CV < 3%$ demonstrates high precision.
  • Comparing Disparate Clinical Measures: Allows an analyst to determine whether systolic blood pressure ($s = 15\text{ mmHg}, \bar{x} = 120\text{ mmHg} \rightarrow CV = 12.5%$) exhibits greater or lesser relative variability than serum potassium ($s = 0.5\text{ mmol/L}, \bar{x} = 4.0\text{ mmol/L} \rightarrow CV = 12.5%$).

---

3. Shape of Distributions: Skewness & Kurtosis

Beyond center and spread, health data analysts must evaluate distribution shape via skewness (asymmetry) and kurtosis (tailedness).

+---------------------------------------------------------------------------------------------------+
|                                 DISTRIBUTION SHAPE CHARACTERISTICS                                |
+-----------------------------------+-----------------------------------+---------------------------+
| SYMMETRIC (Normal)                | POSITIVE SKEW (Right-Tailed)      | NEGATIVE SKEW (Left-Tailed|
| - Skewness = 0                    | - Skewness > 0                    | - Skewness < 0            |
| - Mean = Median = Mode            | - Mode < Median < Mean            | - Mean < Median < Mode    |
| - Mesokurtic (Kurtosis = 3)       | - High-cost outliers, LOS         | - Gestational age         |
+-----------------------------------+-----------------------------------+---------------------------+

Skewness (Asymmetry)

Skewness measures the degree and direction of asymmetry in a distribution relative to a normal distribution.

• Pearson's Coefficient of Skewness: $\text{Skewness} = \frac{3(\bar{x} - \text{Median})}{s}$
• Interpretation Rules:
  • $\text{Skewness} = 0$: Symmetrical distribution ($\text{Mean} = \text{Median} = \text{Mode}$).
  • $\text{Skewness} > 0$ (Positive / Right Skew): Tail stretches toward positive infinity ($\text{Mean} > \text{Median} > \text{Mode}$).
  • $\text{Skewness} < 0$ (Negative / Left Skew): Tail stretches toward negative infinity ($\text{Mean} < \text{Median} < \text{Mode}$).

Kurtosis (Tailedness & Peakedness)

Kurtosis measures the "tailedness" and outlier propensity of a probability distribution relative to a normal Gaussian curve.

• Mesokurtic: Normal distribution baseline; Kurtosis $= 3$ (Excess Kurtosis $= 0$).
• Leptokurtic: Sharp central peak with heavy, fat tails; Kurtosis $> 3$ (Excess Kurtosis $> 0$). Indicates a higher probability of extreme outlier events (e.g., catastrophic financial claims, severe ICU complications).
• Platykurtic: Flatter peak with thin, light tails; Kurtosis $< 3$ (Excess Kurtosis $< 0$). Indicates data points are more uniformly dispersed with fewer extreme outliers.

---

4. The Empirical Rule vs. Chebyshev's Theorem

When evaluating how observations are distributed across standard deviation intervals from the mean, health data analysts must choose between the Empirical Rule and Chebyshev's Theorem based on distribution normality.

+---------------------------------------------------------------------------------------------------+
|                        EMPIRICAL RULE VS. CHEBYSHEV'S THEOREM COMPARISON                          |
+-------------------------------------------------+-------------------------------------------------+
| EMPIRICAL RULE (68 - 95 - 99.7 RULE)            | CHEBYSHEV'S THEOREM (1 - 1/k²)                  |
| - Applies STRICTLY to NORMAL (Bell-Shaped) Data | - Applies to ANY Distribution (Skewed, Bimodal) |
| - Within 1 SD (μ ± 1σ): ~ 68.27% of data        | - Within 1 SD (k = 1): No guarantee (0%)        |
| - Within 2 SD (μ ± 2σ): ~ 95.45% of data        | - Within 2 SD (k = 2): At least 75.0% of data   |
| - Within 3 SD (μ ± 3σ): ~ 99.73% of data        | - Within 3 SD (k = 3): At least 88.89% of data  |
| - Outside 2 SD: ~ 4.55% (2.28% in each tail)    | - Within 4 SD (k = 4): At least 93.75% of data  |
+-------------------------------------------------+-------------------------------------------------+

The Empirical Rule (68-95-99.7 Rule)

• Prerequisite: The data distribution must be symmetric and bell-shaped (normally distributed).
• Intervals:
  • Approximately 68.27% of all observations lie within $\mu \pm 1\sigma$.
  • Approximately 95.45% (commonly rounded to 95%) of all observations lie within $\mu \pm 2\sigma$.
  • Approximately 99.73% of all observations lie within $\mu \pm 3\sigma$.
• Healthcare Application: Establishing clinical laboratory reference intervals (e.g., normal serum sodium range is defined as the central 95% interval $\mu \pm 2\sigma = 135–145\text{ mmol/L}$). Observations beyond $\pm 2\sigma$ are flagged as abnormal, and beyond $\pm 3\sigma$ as critical/panic values.

Chebyshev's Theorem

• Prerequisite: Applies to ANY distribution, regardless of shape, skewness, or modality, for any constant $k > 1$.
• Mathematical Formula: The minimum proportion of observations falling within $k$ standard deviations of the mean is at least: $\text{Proportion} \ge 1 - \frac{1}{k^2}$
• Standard Multipliers:
  • For $k = 2$: At least $1 - \frac{1}{2^2} = 1 - \frac{1}{4} = 75.0%$ of observations lie within $\bar{x} \pm 2s$.
  • For $k = 3$: At least $1 - \frac{1}{3^2} = 1 - \frac{1}{9} = 88.89%$ of observations lie within $\bar{x} \pm 3s$.
  • For $k = 4$: At least $1 - \frac{1}{4^2} = 1 - \frac{1}{16} = 93.75%$ of observations lie within $\bar{x} \pm 4s$.
• Healthcare Application: Evaluating patient wait times, hospital charges, or surgical length of stay where distributions are heavily skewed and normal distribution assumptions are violated.

---

5. Step-by-Step Worked Calculation Example: Patient Length of Stay

A health data analyst extracts the Inpatient Length of Stay (LOS in days) for a sample of $n = 10$ patients admitted to a post-surgical step-down unit: $\text{Raw LOS Data (days)}: \{4, 3, 12, 5, 2, 8, 3, 23, 4, 6\}$

Step 1: Sort Data in Ascending Order

$\text{Sorted Array}: \{2, 3, 3, 4, 4, 5, 6, 8, 12, 23\}$

Step 2: Calculate Central Tendency Measures

1. Arithmetic Mean ($\bar{x}$): $\sum x = 2 + 3 + 3 + 4 + 4 + 5 + 6 + 8 + 12 + 23 = 70$ $\bar{x} = \frac{70}{10} = 7.0\text{ days}$
2. Median ($Q_2$): Since $n = 10$ is even, average the 5th and 6th values (positions $\frac{10}{2} = 5$ and $5 + 1 = 6$): $\text{5th value} = 4, \quad \text{6th value} = 5$ $\text{Median} = \frac{4 + 5}{2} = 4.5\text{ days}$
3. Mode: The values 3 and 4 each appear twice. The distribution is bimodal (Modes = $3$ and $4$).

Step 3: Calculate Quartiles and Interquartile Range (IQR)

• Lower half of sorted array: ${2, 3, 3, 4, 4} \rightarrow Q_1 = 3.0\text{ days}$
• Upper half of sorted array: ${5, 6, 8, 12, 23} \rightarrow Q_3 = 8.0\text{ days}$
• Interquartile Range (IQR): $\text{IQR} = Q_3 - Q_1 = 8.0 - 3.0 = 5.0\text{ days}$
• Tukey Outlier Fences:
  • $\text{Lower Fence} = Q_1 - 1.5 \times \text{IQR} = 3.0 - (1.5 \times 5.0) = 3.0 - 7.5 = -4.5\text{ days}$
  • $\text{Upper Fence} = Q_3 + 1.5 \times \text{IQR} = 8.0 + (1.5 \times 5.0) = 8.0 + 7.5 = 15.5\text{ days}$
  • Outlier Identification: Patient with $\text{LOS} = 23\text{ days}$ exceeds $15.5$ and is identified as a statistical outlier.

Step 4: Calculate Sample Variance and Standard Deviation

• Sample Variance ($s^2$): $s^2 = \frac{\sum (x_i - \bar{x})^2}{n - 1} = \frac{358.0}{10 - 1} = \frac{358.0}{9} \approx 39.78\text{ days}^2$
• Sample Standard Deviation ($s$): $s = \sqrt{39.78} \approx 6.31\text{ days}$
• Coefficient of Variation ($CV$): $CV = \left( \frac{6.31}{7.0} \right) \times 100\% \approx 90.14\%$

Step 5: Analytical Interpretation for Health Data Analysts

• Notice the relationship: $\text{Mode } (3, 4) < \text{Median } (4.5) < \text{Mean } (7.0)$. The data is positively skewed.
• The single outlier patient ($23\text{ days}$) inflated the arithmetic mean to $7.0\text{ days}$, even though 80% of patients (8 out of 10) were discharged in 8 days or fewer.
• Reporting $\bar{x} = 7.0\text{ days}$ and $s = 6.31\text{ days}$ would lead hospital administrators to overestimate typical post-surgical recovery times. The appropriate summary is Median LOS = 4.5 days (IQR = 5.0 days).