16.2 Biostatistical Concepts: Hazard Ratios, Survival Analysis & Precision Metrics

16.2 Biostatistical Concepts: Hazard Ratios, Survival Analysis & Precision Metrics

Biostatistical literacy is paramount for oncology pharmacists appraising medical literature, designing clinical trials, evaluating formulary requests, and translating survival metrics into clinical practice. Oncology studies rely heavily on time-to-event survival models, multivariable regression, and precision risk statistics that differ fundamentally from standard continuous or categorical biostatistical tests.

---

1. Survival Analysis Mechanics & Kaplan-Meier Estimation

Unlike traditional clinical trials where outcomes are measured at a static single time point (e.g., blood pressure reduction at 12 weeks), oncology trials analyze time-to-event outcomes (time from study entry until disease progression, relapse, or death).

The Nature of Censoring in Oncology Trials

Survival data are rarely normally distributed and are invariably subject to censoring:

• Right-Censoring (Most Common): Occurs when the study ends before the patient experiences the event, the patient withdraws consent, or the patient is lost to follow-up. The exact event time is unknown, but is known to be at least as long as the last documented follow-up time.
• Left-Censoring: Occurs when the event occurred prior to study enrollment at an unobserved time.
• Interval-Censoring: The event is known to have occurred between two specific observation timepoints (e.g., between two surveillance CT scans at 8 and 16 weeks).

+---------------------------------------------------------------------------------------------------+
|                         KAPLAN-MEIER PRODUCT-LIMIT ESTIMATOR                                      |
|                                                                                                   |
|   SURVIVAL FUNCTION S(t):                                                                         |
|   - Probability that an individual survives longer than time t: S(t) = P(T > t)                   |
|   - Non-parametric step-function estimated as:                                                    |
|                                                                                                   |
|                 S_hat(t) = PRODUCT_{t_i <= t} [ 1 - (d_i / n_i) ]                                 |
|                                                                                                   |
|   Where:                                                                                          |
|   - t_i = Time of event occurrence                                                                |
|   - d_i = Number of events (deaths or progressions) occurring at time t_i                         |
|   - n_i = Number of patients alive, at risk, and uncensored immediately prior to time t_i         |
|                                                                                                   |
|   ANATOMY OF A KAPLAN-MEIER SURVIVAL CURVE:                                                       |
|   - Vertical Drops: Represent occurrence of an event (death or progression).                      |
|   - Tick Marks (+): Represent censored patients (last known event-free follow-up).                |
|   - Median Survival: The time point corresponding to S(t) = 0.50 (50% cumulative survival).       |
|   - Why Median over Mean? Mean survival cannot be accurately computed unless ALL patients die;    |
|     furthermore, survival data are highly right-skewed. Median is the universal standard.         |
+---------------------------------------------------------------------------------------------------+

graph TD
    A["Kaplan-Meier Curve Analysis"] --> B["Event Occurs (Death/Progression)"]
    A --> C["Patient Censored (Loss to Follow-up/Study End)"]
    B --> D["Step Curve Drops Vertically by (d_i / n_i)"]
    C --> E["Tick Mark (+) Added; Denominator n_i Decreases for Next Event"]
    D --> F["Median Survival: S(t) = 0.50"]
    E --> F

---

2. Comparative Survival Statistics & Cox Proportional Hazards Modeling

The Log-Rank Test

The log-rank test (Mantel-Cox test) is the standard non-parametric hypothesis test used to compare the entire survival distributions of two or more independent groups over time:

• Null Hypothesis ($H_0$): There is no difference in the probability of an event between groups at any time point ($S_1(t) = S_2(t)$).
• Mechanics: Calculates the observed number of events ($O_i$) versus expected events ($E_i$) under the null hypothesis at each discrete event time across groups.
• Weighting: The log-rank test places equal weight on events regardless of whether they occur early, midway, or late in the trial.
• Alternative Tests:
  • Wilcoxon (Breslow-Gehan) Test: Weights events by the number of patients at risk ($n_i$), placing greater emphasis on early events (sensitive when curves separate early, but loses power with late differences).
  • Tarone-Ware Test: Uses intermediate square-root weighting ($\sqrt{n_i}$) between log-rank and Wilcoxon.

Cox Proportional Hazards Regression Model

The Cox proportional hazards model is a semi-parametric multivariable regression model that assesses the relationship between survival time and one or more predictor variables (covariates):

$h(t \mid X) = h_0(t) \exp(\beta_1 X_1 + \beta_2 X_2 + \dots + \beta_k X_k)$

Where:

• $h(t \mid X)$ = Hazard of event at time $t$ given covariates $X$
• $h_0(t)$ = Baseline hazard function at time $t$
• $\beta$ = Regression coefficient (log-hazard ratio)
• $\exp(\beta)$ = Hazard Ratio (HR)

The Hazard Ratio (HR)

The Hazard Ratio measures the relative risk of an event occurring in the experimental group compared to the control group per unit time across the entire study period:

$HR = \frac{\text{Hazard in Experimental Group}}{\text{Hazard in Control Group}}$

• Interpretation:
  • $HR = 1.00$: No difference in hazard between treatment arms.
  • $HR = 0.70$: At any given time point, patients in the experimental arm have a 30% lower hazard (risk) of the event compared to the control arm (Relative Hazard Reduction $= 1 - HR = 30%$).
  • $HR = 1.35$: Patients in the experimental arm have a 35% higher hazard of the event per unit time.
• 95% Confidence Interval for HR: If the 95% CI for superiority excludes 1.00 (e.g., $HR = 0.68, 95%\text{ CI: } 0.52\text{--}0.88$), the difference is statistically significant at $\alpha = 0.05$.

The Proportional Hazards (PH) Assumption & Immuno-Oncology Anomalies

The Cox model assumes that the ratio of hazards between treatment groups remains constant over time (the proportional hazards assumption):

+---------------------------------------------------------------------------------------------------+
|                    NON-PROPORTIONAL HAZARDS IN IMMUNO-ONCOLOGY (IO)                               |
|                                                                                                   |
|   [CLASSICAL CYTOTOXIC CHEMOTHERAPY]                                                              |
|   - Curves separate immediately (early benefit).                                                  |
|   - Hazards remain parallel over time ==> Proportional Hazards Assumption holds (HR valid).       |
|                                                                                                   |
|   [IMMUNE CHECKPOINT INHIBITORS (PD-1 / PD-L1 / CTLA-4)]                                          |
|   - Delayed Separation of Curves: No difference in survival for first 3–6 months while immune    |
|     activation occurs; curves separate substantially only at 6–12+ months.                        |
|   - Crossing Survival Curves: Early toxic deaths/rapid progression in IO vs delayed durable rescue.|
|   - Long Survival Tail (Plateau): Subgroup of long-term survivors/cures producing flat asymptote. |
|                                                                                                   |
|   ==> VIOLATION OF PROPORTIONAL HAZARDS ASSUMPTION:                                               |
|   - Standard Cox HR produces an artificial, time-averaged artifact that underestimates late       |
|     durable benefit. Standard log-rank test loses statistical power!                              |
|                                                                                                   |
|   ==> RECOMMENDED STATISTICAL ALTERNATIVES:                                                       |
|   1. Restricted Mean Survival Time (RMST): Area under the KM curve up to a fixed time horizon tau.|
|   2. Milestone Survival Analysis: Pre-specified survival rate at fixed time (e.g., 2- or 5-year OS).|
|   3. Fleming-Harrington Weighted Log-Rank Test: G(rho, gamma) weighting favoring late events.    |
+---------------------------------------------------------------------------------------------------+

---

3. Precision, Risk & Association Metrics

+---------------------------------------------------------------------------------------------------+
|                         EPIDEMIOLOGICAL & CLINICAL TRIAL RISK EQUATIONS                           |
|                                                                                                   |
|   Control Event Rate (CER) = c / (c + d)                                                          |
|   Experimental Event Rate (EER) = a / (a + b)                                                     |
|                                                                                                   |
|   [RELATIVE RISK (RR)]:                                                                           |
|   RR = EER / CER                                                                                  |
|                                                                                                   |
|   [RELATIVE RISK REDUCTION (RRR)]:                                                                |
|   RRR = (CER - EER) / CER = 1 - RR                                                                |
|                                                                                                   |
|   [ABSOLUTE RISK REDUCTION (ARR)]:                                                                |
|   ARR = | CER - EER |                                                                             |
|                                                                                                   |
|   [NUMBER NEEDED TO TREAT (NNT)]:                                                                 |
|   NNT = 1 / ARR  (MANDATORY: Always round UP to the next whole integer!)                         |
|                                                                                                   |
|   [ABSOLUTE RISK INCREASE (ARI) & NUMBER NEEDED TO HARM (NNH)]:                                   |
|   ARI = EER_toxicity - CER_toxicity                                                               |
|   NNH = 1 / ARI  (MANDATORY: Always round DOWN for safety/conservative harm or UP per convention)|
+---------------------------------------------------------------------------------------------------+

Odds Ratio (OR) vs Relative Risk (RR)

• Relative Risk (RR): Evaluates prospective cohort studies and randomized controlled trials ($RR = \frac{a/(a+b)}{c/(c+d)}$).
• Odds Ratio (OR): Evaluates retrospective case-control studies ($OR = \frac{a \times d}{b \times c}$).
• The Rare Disease Assumption: When the baseline incidence of the outcome is low ($<5\text{--}10%$ in the population), the odds ratio closely approximates the relative risk ($OR \approx RR$). When the outcome is common ($>10%$), the odds ratio significantly overstates the relative effect compared to relative risk.

---

4. Diagnostic Biomarker Accuracy & Prevalence Dynamics

In precision oncology, companion diagnostic biomarkers (e.g., NGS panels, IHC for PD-L1/HER2, PCR for BCR-ABL1) dictate targeted therapy eligibility.

+---------------------------------------------------------------------------------------------------+
|                         DIAGNOSTIC CONTINGENCY MATRIX (2 x 2 TABLE)                               |
|                                                                                                   |
|                                       DISEASE PRESENT (Gold Standard)                             |
|   TEST RESULT                         Positive (+)                    Negative (-)                |
|   ---------------------------------------------------------------------------------------------   |
|   Test Positive (+)                   True Positive (TP)              False Positive (FP)         |
|   Test Negative (-)                   False Negative (FN)             True Negative (TN)          |
|   ---------------------------------------------------------------------------------------------   |
|                                       Total Disease = TP + FN         Total Disease-Free = FP + TN|
+---------------------------------------------------------------------------------------------------+

Core Diagnostic Formulas

• Sensitivity (True Positive Rate): $\text{Sensitivity} = \frac{TP}{TP + FN}$ (Ability of assay to correctly identify patients with the biomarker/disease; SnNOut: High Sensitivity rules OUT disease when negative).
• Specificity (True Negative Rate): $\text{Specificity} = \frac{TN}{TN + FP}$ (Ability of assay to correctly identify patients without the biomarker/disease; SpPIn: High Specificity rules IN disease when positive).
• Positive Predictive Value (PPV): $\text{PPV} = \frac{TP}{TP + FP}$ (Probability that a patient who tests positive truly has the disease).
• Negative Predictive Value (NPV): $\text{NPV} = \frac{TN}{TN + FN}$ (Probability that a patient who tests negative is truly disease-free).
• Positive Likelihood Ratio ($LR+$): $LR+ = \frac{\text{Sensitivity}}{1 - \text{Specificity}}$ ($LR+ > 10$ provides strong evidence to rule in disease).
• Negative Likelihood Ratio ($LR-$): $LR- = \frac{1 - \text{Sensitivity}}{\text{Specificity}}$ ($LR- < 0.1$ provides strong evidence to rule out disease).

Impact of Disease Prevalence on Predictive Values

• Prevalence Independence: Sensitivity, Specificity, and Likelihood Ratios are intrinsic properties of the test and do not change with disease prevalence.
• Prevalence Dependence: PPV and NPV are directly dependent on disease prevalence:
  • As disease prevalence decreases (e.g., cancer screening in an asymptomatic population), PPV drops dramatically (more false positives than true positives), while NPV increases.
  • As disease prevalence increases (e.g., testing high-risk refractory patients), PPV rises, while NPV drops.

---

5. Statistical Hypothesis Testing & Variable Selection Matrix

Selecting the appropriate statistical test depends on the number of groups, study design (paired vs unpaired), and variable data type (continuous parametric vs non-parametric vs categorical).