3.2 Biostatistics, Clinical Epidemiology & Diagnostic Test Interpretation

Diagnostic Test Performance & The $2 \times 2$ Contingency Table

Interpreting diagnostic investigations in outpatient primary care requires understanding the mathematical relationships governing test accuracy, disease prevalence, and predictive values. All diagnostic parameters derive from the standardized $2 \times 2$ contingency table:

Intrinsic Test Parameters: Sensitivity and Specificity

Sensitivity and specificity are inherent biological characteristics of the test itself; they do not vary with changes in population disease prevalence.

1. Sensitivity (True Positive Rate): $\text{Sensitivity} = \frac{TP}{TP + FN}$

   • The probability that a diseased individual tests positive.
   • Clinical Rule (SnNOut): A test with very high Sensitivity, when Negative, rules OUT the disease (few false negatives). High-sensitivity tests are ideal for screening asymptomatic populations (e.g., fifth-generation HIV antigen/antibody immunoassay, fecal immunochemical testing for colorectal cancer, D-dimer in low-risk pulmonary embolism).

2. Specificity (True Negative Rate): $\text{Specificity} = \frac{TN}{TN + FP}$

   • The probability that a disease-free individual tests negative.
   • Clinical Rule (SpPIn): A test with very high Specificity, when Positive, rules IN the disease (few false positives). High-specificity tests are ideal for confirmatory diagnostic testing following a positive screening result (e.g., Western blot/differentiation immunoassay for HIV, tissue biopsy for suspected malignancy).

Prevalence-Dependent Parameters: Predictive Values

In clinical practice, physicians do not know whether the patient truly has the disease; they only know the test result. Predictive values answer the patient's fundamental question: "Given my test result, what is the actual probability that I have the condition?"

1. Positive Predictive Value (PPV): $\text{PPV} = \frac{TP}{TP + FP}$

   • The proportion of individuals with a positive test result who truly have the disease.

2. Negative Predictive Value (NPV): $\text{NPV} = \frac{TN}{TN + FN}$

   • The proportion of individuals with a negative test result who are truly free of the disease.

The Impact of Disease Prevalence (Bayes' Theorem in Action)

Predictive values are heavily dependent on the pre-test probability (prevalence) of disease in the population tested:

• When Prevalence Decreases: True positives decrease, while false positives accumulate. Consequently, PPV decreases dramatically, while NPV increases toward $100%$.
• When Prevalence Increases: True positives accumulate, while false positives become a smaller proportion. Consequently, PPV increases, while NPV decreases.
• Board Exam Implication: Applying a screening test with $90%$ sensitivity and $90%$ specificity to a low-prevalence population ($1%$) yields a PPV of only $\sim 8.3%$, meaning $>91%$ of positive tests are false positives. Conversely, testing in a high-prevalence population ($50%$) yields a PPV of $90%$.

---

Likelihood Ratios, ROC Curves & The Fagan Nomogram

While predictive values fluctuate with prevalence, Likelihood Ratios (LRs) combine sensitivity and specificity into a single metric that quantifies diagnostic power independent of disease prevalence.

Formulas and Definitions

1. Positive Likelihood Ratio ($\text{LR}^+$): $\text{LR}^+ = \frac{\text{Sensitivity}}{1 - \text{Specificity}} = \frac{\text{True Positive Rate}}{\text{False Positive Rate}}$

   • Indicates how much more likely a positive test result is in a patient with the disease compared to one without the disease.

2. Negative Likelihood Ratio ($\text{LR}^-$): $\text{LR}^- = \frac{1 - \text{Sensitivity}}{\text{Specificity}} = \frac{\text{False Negative Rate}}{\text{True Negative Rate}}$

   • Indicates how much more likely a negative test result is in a patient with the disease compared to one without the disease.

Clinical Interpretation of Likelihood Ratios

Clinicians use the following rules of thumb to interpret likelihood ratios and their corresponding shifts on a Fagan nomogram (converting pre-test probability to post-test probability):

Receiver Operating Characteristic (ROC) Curves & Cutoff Thresholds

An ROC curve graphs diagnostic performance across continuous cutoffs:

• Y-axis: Sensitivity (True Positive Rate).
• X-axis: $1 - \text{Specificity}$ (False Positive Rate).
• Area Under the Curve (AUC): Measures overall test discriminative ability. An $\text{AUC} = 0.5$ represents a worthless test equivalent to a coin toss. An $\text{AUC} = 1.0$ represents a perfect test. An $\text{AUC} \ge 0.80$ to $0.90$ indicates excellent diagnostic accuracy.
• Shifting Cutoff Values:
  • Lowering the Cutoff Value (e.g., lowering the fasting glucose threshold for diabetes from 126 mg/dL to 100 mg/dL): Captures more true positives $\rightarrow$ Sensitivity increases, but false positives also increase $\rightarrow$ Specificity decreases.
  • Raising the Cutoff Value (e.g., raising the troponin threshold for myocardial infarction): Eliminates false positives $\rightarrow$ Specificity increases, but false negatives accumulate $\rightarrow$ Sensitivity decreases.

---

Epidemiological Study Designs: Methodological Hierarchy

Evidence-based clinical guidelines synthesize data across diverse epidemiological study architectures:

1. Randomized Controlled Trials (RCT)

• Design: Participants are randomly assigned to an experimental intervention or control (placebo or standard of care) and followed prospectively for clinical outcomes.
• Strengths: Gold standard for demonstrating causality. Randomization balances both known and unknown confounding variables across study arms.
• Analysis Principles:
  • Intention-to-Treat (ITT) Analysis: All randomized participants are analyzed in the group to which they were originally assigned, regardless of non-compliance, protocol deviations, or dropout. ITT preserves randomization, prevents attrition bias, and provides a pragmatic estimate of real-world effectiveness.
  • Per-Protocol Analysis: Analyzes only participants who completed the full assigned protocol. Evaluates maximal biological efficacy but introduces confounding and attrition bias.

2. Cohort Studies (Prospective and Retrospective)

• Design: Defined by exposure status (exposed vs. unexposed cohorts). Followed over time to assess the incidence of outcome events.
• Primary Measure of Effect: Relative Risk (RR) and Incidence Rate Ratio.
• Strengths: Optimal for studying rare exposures (e.g., occupational chemical exposures); establishes clear temporal sequence (exposure precedes disease); allows assessment of multiple potential outcomes from a single exposure.
• Weaknesses: Inefficient for rare outcomes or diseases with long latency periods; expensive and vulnerable to loss to follow-up (attrition bias).

3. Case-Control Studies

• Design: Defined by outcome status (cases with disease vs. controls without disease). Assessed retrospectively for frequency of past exposures.
• Primary Measure of Effect: Odds Ratio (OR). Cannot directly calculate incidence or Relative Risk because the investigator determines the proportion of cases and controls.
• Strengths: Highly efficient, low cost, and optimal for rare diseases (e.g., mesothelioma) or diseases with long latency periods (e.g., decades-long carcinogenesis).
• Weaknesses: Cannot establish incidence; highly vulnerable to recall bias (cases remember past exposures more vividly than controls) and selection bias (inappropriate control selection).

4. Cross-Sectional Studies (Prevalence Surveys)

• Design: Assesses exposure status and disease status simultaneously at a single point in time in a defined population.
• Primary Measure of Effect: Prevalence and Prevalence Odds Ratio.
• Strengths: Rapid, inexpensive; excellent for public health planning and evaluating disease burden.
• Weaknesses: Cannot establish temporal sequence (cannot determine whether exposure preceded disease or vice versa); susceptible to Neyman (prevalence-incidence) bias, which over-represents indolent, chronic cases while missing rapidly fatal or brief cases.

5. Ecological Studies

• Design: Units of observation are populations or geographic groups rather than individuals (e.g., comparing country-level per capita wine consumption with national rates of coronary heart disease).
• Major Limitation (Ecological Fallacy): Erroneously inferring that an association observed at the population level holds true for individuals within that population.

---

Measures of Association & Effect Size Calculations

Board examinations frequently test manual calculation of clinical effect sizes using $2 \times 2$ event rates:

• Control Event Rate (CER): $\text{CER} = \frac{c}{c + d}$ (proportion of control group experiencing the outcome).
• Experimental Event Rate (EER): $\text{EER} = \frac{a}{a + b}$ (proportion of experimental group experiencing the outcome).

1. Relative Risk (RR) and Relative Risk Reduction (RRR)

• Relative Risk (Risk Ratio, RR): $\text{RR} = \frac{\text{EER}}{\text{CER}}$
  • $\text{RR} = 1.0$: No association between exposure and outcome.
  • $\text{RR} < 1.0$: Protective effect (reduced risk in experimental group).
  • $\text{RR} > 1.0$: Harmful effect (elevated risk in experimental group).
• Relative Risk Reduction (RRR): $\text{RRR} = \frac{\text{CER} - \text{EER}}{\text{CER}} = 1 - \text{RR}$
  • Caution: Pharmaceutical marketing frequently emphasizes RRR because it presents an impressive percentage reduction even when absolute baseline risk is negligible (e.g., reducing an event from 2 in 10,000 to 1 in 10,000 yields a $50%$ RRR, but an ARR of only $0.01%$).

2. Absolute Risk Reduction (ARR) & Number Needed to Treat (NNT)

• Absolute Risk Reduction (Risk Difference, ARR): $\text{ARR} = \text{CER} - \text{EER}$
  • Reflects the true absolute decrease in clinical event rate attributable to the intervention.
• Number Needed to Treat (NNT): $\text{NNT} = \frac{1}{\text{ARR}} = \frac{1}{\text{CER} - \text{EER}}$
  • The number of patients who must be treated with the intervention for a specified duration to prevent one additional adverse outcome event.
  • Critical Rounding Rule: NNT must always be rounded UP to the next whole integer (e.g., an NNT of 16.2 must be rounded up to 17) to provide a conservative, honest estimate of clinical effort required.

3. Absolute Risk Increase (ARI) & Number Needed to Harm (NNH)

• Absolute Risk Increase (ARI): Applied when experimental therapy increases an adverse event: $\text{ARI} = \text{EER} - \text{CER}$
• Number Needed to Harm (NNH): $\text{NNH} = \frac{1}{\text{ARI}} = \frac{1}{\text{EER} - \text{CER}}$
  • The number of patients exposed to the intervention required for one additional patient to experience a harmful adverse effect.
  • Critical Rounding Rule: In safety reporting, NNH is conventionally rounded DOWN to the nearest whole integer (or reported conservatively) to present the fewest patients exposed before causing harm.

4. Odds Ratio (OR) & The Rare Disease Assumption

• Odds Ratio (OR): $\text{OR} = \frac{\text{Odds of exposure in cases}}{\text{Odds of exposure in controls}} = \frac{a / c}{b / d} = \frac{a \times d}{b \times c}$
• The Rare Disease Assumption: When the baseline prevalence of a disease in the population is low (generally $<10%$ or $<5%$), $a$ is very small compared to $b$, and $c$ is very small compared to $d$. Under this condition, the Odds Ratio closely approximates the Relative Risk ($\text{OR} \approx \text{RR}$).

5. Hazard Ratio (HR)

• Used in survival analysis (time-to-event data modeled via Cox proportional hazards regression).
• Represents the ratio of event rates (hazards) occurring at any specific point in time between two groups. An $\text{HR} = 0.70$ indicates that the treated group experiences events at $70%$ of the rate of the control group at any given time point ($30%$ instantaneous hazard reduction).

---

Hypothesis Testing, Error Types, Power & Confidence Intervals

Null Hypothesis ($H_0$) and Error Matrix

• Null Hypothesis ($H_0$): States that there is no true difference between study groups.
• Alternative Hypothesis ($H_1$): States that a true difference exists.

Statistical Parameters Defined

• Type I Error ($\alpha$): Conventionally set at $0.05$. Represents the probability of finding a statistically significant result purely by chance when no true difference exists.
• Type II Error ($\beta$): Conventionally set at $0.10$ to $0.20$. Represents the probability of failing to detect a true difference.
• Statistical Power ($1 - \beta$): Conventionally target $\ge 80%$ ($0.80$). The probability of correctly rejecting the null hypothesis when a true effect exists. Statistical power increases when:
  1. Sample size ($N$) increases.
  2. Effect size (difference between groups) increases.
  3. Study population variance/noise decreases.
  4. The alpha threshold is relaxed (e.g., from $0.01$ to $0.05$).

Interpreting $p$-Values vs. $95%$ Confidence Intervals (CI)

• A $p$-value $< 0.05$ indicates statistical significance, meaning the probability of obtaining the observed result (or more extreme) by random chance under the null hypothesis is $< 5%$. However, $p$-values provide zero information regarding the magnitude of clinical benefit or precision.
• A $95%$ Confidence Interval (CI) indicates that if the study were repeated 100 times, 95 of the calculated confidence intervals would contain the true population parameter.
• Rules for Evaluating Statistical Significance via Confidence Intervals:
  • Difference Measures (Mean difference, Absolute Risk Reduction):
    • If the $95%$ CI includes $0.0$ (e.g., $95%\text{ CI: } -2.1\text{ to } +4.5$), the finding is NOT statistically significant ($p \ge 0.05$).
    • If the $95%$ CI does not cross zero, the finding is statistically significant.
  • Ratio Measures (Relative Risk, Odds Ratio, Hazard Ratio):
    • If the $95%$ CI includes $1.0$ (e.g., $\text{RR } 0.82; 95%\text{ CI: } 0.65\text{ to } 1.04$), the finding is NOT statistically significant ($p \ge 0.05$).
    • If the $95%$ CI does not cross 1.0 (e.g., $\text{HR } 0.72; 95%\text{ CI: } 0.58\text{ to } 0.89$), the finding is statistically significant ($p < 0.05$).

---

Biases, Confounding & Effect Modification

Systematic Epidemiological Biases

Bias is a systematic error in study design, conduct, or analysis that distorts results:

• Selection Bias: Flawed participant selection or retention.
  • Berkson's Bias: Selecting controls exclusively from hospitalized inpatients who have higher comorbidity rates than the general population.
  • Healthy Worker Effect: Comparing an employed occupational cohort to the general population (which includes disabled/chronically ill individuals), falsely making the occupational exposure appear healthier.
  • Attrition Bias: Differential loss to follow-up between study groups.
• Information (Measurement) Bias: Inaccurate classification of exposure or outcome.
  • Recall Bias: Cases with poor outcomes recall past exposures more vividly than healthy controls.
  • Hawthorne Effect: Participants alter their behaviors simply because they know they are being actively observed.
• Screening Biases:
  • Lead-Time Bias: Early detection of disease through screening falsely creates the appearance of prolonged survival, when in reality only the point of diagnosis was shifted earlier without altering the natural history or date of death. (Can be recognized when survival time from diagnosis increases, but overall age-adjusted population mortality rates remain unchanged).
  • Length-Time Bias: Screening tests preferentially detect slow-growing, indolent, less aggressive cases with long preclinical phases, while aggressive, rapidly fatal cases present clinically between screening intervals. This falsely inflates apparent survival in screened cohorts.

Confounding vs. Effect Modification

A core distinction frequently tested on board examinations: