12.2 Biostatistics and Literature Evaluation

Biostatistics and Clinical Literature Evaluation

A critical care pharmacist must rapidly analyze and apply clinical trial results to bed-side patient care. This requires a sophisticated understanding of biostatistics, data structures, and literature evaluation. Selecting and interpreting statistical analyses ensures that clinical protocols and patient-specific interventions are grounded in solid scientific evidence rather than random variation.

Classification of Data Types

Data are the building blocks of clinical research and are categorized into discrete (categorical) and continuous (scale) data. The mathematical properties of the data dictate the statistical tests that can be legally performed.

Discrete (Categorical) Data

Discrete data are divided into distinct categories where values cannot be subdivided.

• Nominal Data: Categorical variables with no inherent quantitative value or logical order. Examples:
  • Survival status at day 28 (alive vs. dead).
  • Primary pathogen group (Gram-positive vs. Gram-negative vs. fungal).
  • Development of ventilator-associated pneumonia (yes vs. no).
  • Blood type (A, B, AB, O).
• Ordinal Data: Categorical variables that possess a logical order or ranking, but the mathematical distance between the ranks is not uniform or quantifiable. Examples:
  • Richmond Agitation-Sedation Scale (RASS) score, ranging from -5 (unarousable) to +4 (combative).
  • Critical-Care Pain Observation Tool (CPOT) score (0 to 8).
  • Glasgow Coma Scale (GCS) score (3 to 15).
  • New York Heart Association (NYHA) functional class (I, II, III, IV).
  • APACHE II severity score.

Continuous (Scale) Data

Continuous data can take on any value along a specific range and have consistent, measurable intervals between data points.

• Interval Data: Continuous data with consistent intervals, but without a true absolute zero point (zero does not indicate the absence of the property). Example: Temperature in Celsius or Fahrenheit (0 degrees does not mean there is no heat).
• Ratio Data: Continuous data with a true absolute zero point, where zero represents the complete absence of the measured property. This allows for ratio comparisons (e.g., 10 mg is twice as much as 5 mg). Examples:
  • Mean arterial pressure (MAP) in mmHg.
  • Vasoactive infusion rates (e.g., norepinephrine dose in mcg/kg/min).
  • Serum creatinine concentration (mg/dL).
  • Urinary output (mL/kg/hour).
  • Intracranial pressure (ICP) in mmHg.
  • Volume of distribution ($V_d$) and clearance ($Cl$) of an antibiotic.

Parametric vs. Non-parametric Statistical Tests

Statistical tests are selected based on the data type, the number of comparative groups, and the distribution of the data.

Parametric Tests

Parametric tests assume that the data are continuous and follow a normal (Gaussian, bell-shaped) distribution, and that there is homogeneity of variance between the groups. These tests are mathematically powerful because they evaluate the actual values (means and standard deviations).

• Student's t-test (Independent samples): Compares the means of two independent groups (e.g., comparing mean MAP in patients randomized to norepinephrine vs. vasopressin).
• Paired t-test: Compares the means of two related or matched groups (e.g., comparing a patient's arterial pH before and after receiving a sodium bicarbonate infusion).
• Analysis of Variance (ANOVA): Compares the means of three or more independent groups (e.g., comparing mean arterial blood gas values between patients randomized to three different ventilator modes).

Non-parametric Tests

Non-parametric tests do not assume a normal distribution (distribution-free). They are used for ordinal data, categorical data, or continuous data that are skewed (non-normally distributed). Instead of comparing means, these tests analyze the ranks of the data.

• Mann-Whitney U Test (Wilcoxon Rank-Sum Test): Compares two independent groups (e.g., comparing median days of mechanical ventilation between a propofol group and a dexmedetomidine group, where ventilation days are skewed).
• Wilcoxon Signed-Rank Test: Compares two paired or related groups (e.g., comparing median CPOT pain scores in the same patient before and after receiving intravenous fentanyl).
• Kruskal-Wallis Test: Compares three or more independent groups (e.g., comparing median ICU length of stay across three different sedation protocols).
• Chi-square ($\chi^2$) Test: Compares the proportions of nominal data between two or more independent groups. It requires a large sample size (expected cell counts in a contingency table must be $\ge 5$). E.g., comparing 30-day mortality rates between hydrocortisone and placebo groups in septic shock.
• Fisher's Exact Test: Compares proportions of nominal data when sample sizes are small (any cell in the contingency table has an expected count $< 5$). E.g., comparing the rate of a rare adverse event, like propofol infusion syndrome (PRIS), between low-dose and high-dose propofol groups.
• McNemar's Test: Compares proportions of paired nominal data (e.g., checking presence of delirium [yes/no] before and after a sedation vacation in the same cohort).

Hypothesis Testing, Type I/II Errors, and Statistical Power

Clinical trials are designed to test a hypothesis. The Null Hypothesis ($H_0$) states that there is no difference between the treatment groups, while the Alternative Hypothesis ($H_1$) states that a difference exists.

Type I Error ($\alpha$)

A Type I error occurs when researchers reject a true null hypothesis, concluding that a difference exists when there is actually no difference (a false positive).

• Alpha ($\alpha$): The maximum acceptable probability of committing a Type I error, set a priori, typically at 0.05. A p-value of $< 0.05$ indicates that the probability of the observed difference occurring by random chance under the null hypothesis is less than 5%.

Type II Error ($\beta$) and Statistical Power

A Type II error occurs when researchers fail to reject a false null hypothesis, concluding that no difference exists when a true therapeutic difference does exist (a false negative).

• Beta ($\beta$): The probability of committing a Type II error, usually set at 0.10 or 0.20.
• Statistical Power ($1 - \beta$): The probability of correctly rejecting the null hypothesis and detecting a statistically significant difference if a true difference exists. A standard power target is 80% ($\beta = 0.20$) or 90% ($\beta = 0.10$).

Factors Influencing Statistical Power

1. Sample Size ($N$): As sample size increases, standard error decreases, which directly increases statistical power.
2. Effect Size: The magnitude of the difference between groups. Larger differences are easier to detect and require a smaller sample size to achieve the same power.
3. Variability (Standard Deviation): Greater variability or noise in the data (common in heterogeneous ICU populations) reduces power. Minimizing variability increases power.
4. Alpha Level ($\alpha$): Decreasing the acceptable Type I error rate (e.g., setting $\alpha = 0.01$) increases the risk of a Type II error, thereby reducing power.

To avoid conducting underpowered trials (which risk missing clinically beneficial treatments), investigators must perform an a priori power analysis to determine the minimum sample size required to detect a clinically meaningful effect size.

P-values and Confidence Intervals

P-values

A p-value is a continuous measure of evidence against the null hypothesis. However, p-values do not indicate the magnitude or direction of the treatment effect, nor do they reflect clinical relevance.

Confidence Intervals (CIs)

CIs provide a range of values within which the true population parameter is expected to lie with a specified probability (typically 95%). CIs are superior to p-values because they show the precision of the estimate, the direction of effect, and the clinical significance.

• For Difference Measures (e.g., Mean Difference, Absolute Risk Reduction): If the 95% CI contains 0 (the null value for differences), there is no statistically significant difference between the groups ($p \ge 0.05$).
• For Ratio Measures (e.g., Relative Risk [RR], Odds Ratio [OR], Hazard Ratio [HR]): If the 95% CI contains 1.0 (the null value for ratios, indicating equal risk), there is no statistically significant difference ($p \ge 0.05$). E.g., an OR for mortality of 0.82 with a 95% CI of [0.65 to 1.04] is not statistically significant.

Clinical vs. Statistical Significance

Statistical significance indicates that the observed difference is unlikely to be due to chance. It is heavily influenced by sample size: in very large studies, tiny differences can achieve a p-value $< 0.05$.

• Clinical Significance: Reflects whether the magnitude of the effect is large enough to impact patient care. For example, a trial enrolling 20,000 patients might find that a new vasoactive drug increases MAP by 0.7 mmHg compared to norepinephrine, with a statistically significant p-value of 0.002. However, a 0.7 mmHg change is clinically meaningless.
• Minimum Clinically Important Difference (MCID): The smallest change or treatment effect that patients or clinicians perceive as beneficial. Clinicians must benchmark trial results against the MCID rather than relying solely on p-values.