Comprehensive Notes on Advanced Medical Biostatistics, Epidemiological Modeling, and Evidence Synthesis

Systematic Reviews and Meta-Analysis

• Evidence Synthesis in Clinical Research:

  • In medical and pharmacological sciences, individual literature databases contain overwhelming volumes of primary research. For example, a search on PubMed for statin[Title] yields over 7,868 publications; restricting this search by clinical trial and randomized controlled trial (RCT) filters still yields 902 individual papers.          

    

        

  • Individual studies often differ in sample size, patient demographics, and geographic regions, leading to conflicting conclusions that require rigorous methods of synthesis.

• Systematic Literature Review (SLR):

  • Definition: A comprehensive, predefined, transparent, and structured method of gathering, critically appraising, and synthesizing all available empirical evidence to answer a clearly formulated clinical research question.

  • Nature: Primarily qualitative in nature, focusing on systematic literature identification, risk-of-bias evaluation, and conceptual appraisal rather than statistical aggregation.

• Sequential Procedure for Systematic Reviews and Meta-Analyses:

  • Step 1: Formulate Research Question: Define clear research objectives using structured frameworks (e.g., PICO: Population, Intervention, Comparator, Outcome).

  • Step 2: Protocol Development: Outline a rigorous protocol in advance detailing inclusion/exclusion criteria, search strategies, and analytical plans.

  • Step 3: Literature Search: Conduct structured literature queries across multiple scientific databases (e.g., PubMed, Embase, Cochrane Library).

  • Step 4: Study Selection: Screen titles, abstracts, and full texts against predefined eligibility criteria.

  • Step 5: Data Extraction: Extract relevant clinical characteristics, sample sizes, and outcome statistics using standardized forms.

  • Step 6: Quality Assessment: Systematically evaluate individual studies for methodological rigor and potential biases.

  • Step 7: Data Synthesis: Summarize the qualitative and quantitative findings.

  • Step 8: Heterogeneity Assessment: Quantify and evaluate clinical and statistical variability across studies.

  • Step 9: Publication Bias Assessment: Analyze potential dissemination biases using graphical and statistical techniques.

  • Step 10: Interpretation and Reporting: Contextualize findings and report them following standardized guidelines (e.g., PRISMA).

  • Step 11: Peer Review and Publication: Submit synthesis findings for formal academic peer review.

• Meta-Analysis:

  • Definition: A formal statistical procedure that mathematically pools the quantitative effect sizes from multiple independent studies to compute an overall, summary effect size and quantify between-study variation.

  • Core Function: An optional, quantitative component of a systematic review. Statistical software tools such as Review Manager (RevMan) are widely utilized for computation (https://www.youtube.com/watch?v=A_p0fBaPIOI).

  • Key Advantages:

    • Integration of Evidence: Provides a single mathematical consensus across a body of literature.

    • Increased Statistical Power: Combines sample sizes from multiple small studies, increasing the likelihood of detecting true effects.

    • Improved Precision: Generates narrower confidence intervals around the estimated effect size.

    • Resolution of Conflicts: Reconciles disparate or contradictory study findings.

    • Investigation of Heterogeneity: Identifies study-level characteristics that explain variation in outcomes.

    • Evidence-Based Clinical Decision Making: Supplies high-level empirical grounding for clinical guidelines.

• Effect Sizes and Standardization:

  • Definition: A standardized metric reflecting the magnitude and direction of the experimental effect or relationship between variables across studies.

  • Common Metrics:

    • Continuous outcomes: Mean Difference ($\bar{X}_1 - \bar{X}_2$) when measurement scales are identical, or Standardized Mean Difference (SMD, Cohen's $d$):           $\text{SMD} = \frac{\bar{X}_1 - \bar{X}_2}{s_p}$

    • Binary/Dichotomous outcomes: Odds Ratio (OR), Risk Ratio (RR), or Risk Difference (RD).

    • Association outcomes: Pearson's correlation coefficient ($r$).

  • Metric Conversion: Before calculating a pooled estimate, disparate statistics across studies (e.g., pooled standard deviation versus per-group standard deviations, or odds ratios from contingency tables versus logistic regression outputs) must be transformed into a single common metric (e.g., log odds ratio).

• Forest Plots:

  • Visual Layout: A standardized graphical representation of meta-analytic data.          

    

        

  • Components:

    • Square Markers: Depict the point estimate of each individual study. The size of the square represents the statistical weight assigned to the study.

    • Horizontal Lines: Represent the $95\%$ confidence interval ($95\%\,\text{CI}$) for each study's effect estimate.

    • Vertical Line of No Effect: Set at $0.00$ for difference metrics (e.g., mean differences) and at $1.00$ for ratio metrics (e.g., odds ratios, hazard ratios).

    • Diamond (Summary Measure): The horizontal endpoints of the diamond represent the $95\%\,\text{CI}$ of the pooled effect, and the vertical center represents the summary point estimate.

• Assessment of Heterogeneity:

  • Underlying Causes: Diversity in study designs, differing patient eligibility criteria, variation in drug dosages/durations, and disparate sample sizes.

  • Cochran's $Q$ Statistic: Quantifies the total weighted squared deviation of study effect sizes ($E_i$) around the overall summary effect ($\bar{E}$):       $Q = \sum_{i} W_i (E_i - \bar{E})^2$       where $W_i$ is the weight of study $i$. The statistic follows a Chi-square distribution with $\text{df} = k - 1$ degrees of freedom, where $k$ is the number of studies.

  • $I^2$ Inconsistency Index: Quantifies the proportion of total variation across studies due to true heterogeneity rather than sampling error (chance):       $I^2 = \frac{Q - \text{df}}{Q} \times 100\%$

  • Cochrane Interpretive Thresholds for $I^2$:

    • $0\%$ to $40\%$: Might not be important.

    • $30\%$ to $60\%$: May represent moderate heterogeneity.

    • $50\%$ to $90\%$: May represent substantial heterogeneity.

    • $75\%$ to $100\%$: Considerable heterogeneity.

• Statistical Models: Fixed-Effect vs. Random-Effects:

  • Fixed-Effect Model:

    • Assumption: There is one identical true effect size ($\theta$) shared by every study in the analysis. Any observed variation in study outcomes is purely attributable to within-study sampling error ($\epsilon_i$):           $Y_i = \theta + \epsilon_i$

    • Weighting Scheme: Weights are inversely proportional to within-study variance ($V_i$):           $W_i = \frac{1}{V_i}$

    • Pooled Estimate Calculation:           $E = \frac{\sum W_i Y_i}{\sum W_i}$

  • Random-Effects Model:

    • Assumption: True effect sizes vary across studies, following a distribution with mean $\mu$ and between-study variance $\tau^2$ (or $B_i$). Observed study estimates ($Y_i$) deviate from $\mu$ due to both between-study heterogeneity ($\zeta_i$) and within-study sampling error ($\epsilon_i$):           $Y_i = \mu + \zeta_i + \epsilon_i$

    • Weighting Scheme: Weights incorporate both within-study variance ($V_i$) and between-study variance ($B_i$):           $W_i = \frac{1}{V_i + B_i}$

    • Implication: Balances study influence by preventing very large studies from completely dominating the analysis while accounting for genuine population differences.

• Evaluation of Meta-Analytic Data (Bjordal et al., 2007 Study):

  • Example 1: Electroacupuncture for Knee Osteoarthritis Pain Relief (100-mm VAS):          

    

        

    • Yurtkuran: $n = 25$ (Treatment Mean $\pm$ SD: $25.00 \pm 12.00$), $n = 25$ (Control: $5.00 \pm 8.90$); WMD (Fixed) = $20.00\,(95\%\,\text{CI}\,[14.14, 25.86])$; Weight = $22.83\%$.

    • Sangdee: $n = 48$ ($48.00 \pm 24.30$), $n = 47$ ($23.00 \pm 24.20$); WMD = $25.00\,[15.25, 34.75]$; Weight = $8.23\%$.

    • Vas: $n = 48$ ($53.50 \pm 21.40$), $n = 49$ ($28.50 \pm 35.20$); WMD = $25.00\,[13.43, 36.57]$; Weight = $5.85\%$.

    • Pooled Subtotal: Treatment $N = 121$, Control $N = 121$; Summary WMD = $21.91\,[17.30, 26.51]$; Total weight = $36.92\%$.

    • Heterogeneity Statistics: $\chi^2 = 1.07$, $\text{df} = 2$ ($P = 0.59$), $I^2 = 0\%$. Test for overall effect: $z = 9.32$ ($P < 0.00001$). Findings indicate homogeneous study outcomes.

  • Example 2: Transcutaneous Electrical Stimulation (TENS) for Knee Osteoarthritis Pain Relief:          

    

        

    • Fargas Babjak: $n = 19$ ($56.00 \pm 37.50$), $n = 18$ ($10.70 \pm 64.00$); WMD = $45.30\,[11.26, 79.34]$; Weight = $0.68\%$.

    • Yurtkuran: $n = 25$ ($25.00 \pm 12.00$), $n = 25$ ($5.00 \pm 8.90$); WMD = $20.00\,[14.14, 25.86]$; Weight = $22.83\%$.

    • Cheing (Study 1): $n = 16$ ($58.00 \pm 27.00$), $n = 16$ ($4.96 \pm 42.40$); WMD = $53.04\,[28.41, 77.67]$; Weight = $1.29\%$.

    • Adeboyin: $n = 15$ ($68.70 \pm 24.00$), $n = 15$ ($43.00 \pm 24.00$); WMD = $25.70\,[8.52, 42.88]$; Weight = $2.65\%$.

    • Cheing (Study 2): $n = 30$ ($35.20 \pm 41.50$), $n = 10$ ($30.00 \pm 24.10$); WMD = $5.20\,[-15.94, 26.34]$; Weight = $1.75\%$.

    • Defrin: $n = 33$ ($28.40 \pm 11.10$), $n = 21$ ($5.70 \pm 13.40$); WMD = $22.70\,[15.83, 29.57]$; Weight = $16.59\%$.

    • Law: $n = 25$ ($46.70 \pm 24.40$), $n = 9$ ($17.00 \pm 39.70$); WMD = $29.70\,[2.06, 57.34]$; Weight = $1.02\%$.

    • Pooled Subtotal: Treatment $N = 163$, Control $N = 114$; Summary WMD = $22.21\,[18.12, 26.30]$; Total weight = $46.82\%$.

    • Heterogeneity Statistics: $\chi^2 = 11.28$, $\text{df} = 6$ ($P = 0.08$), $I^2 = 46.8\%$. Test for overall effect: $z = 10.65$ ($P < 0.00001$). Findings demonstrate moderate heterogeneity requiring qualitative and quantitative exploration.

• Funnel Plots and Publication Bias:

  • Funnel Plot Design: A scatter plot with treatment effect on the horizontal axis (e.g., log odds ratio) and an indicator of study size or precision (standard error) on the vertical inverted axis.          

    

        

  • Interpretation of Geometry:

    • Large, high-powered studies cluster closely near the top around the summary effect.

    • Small studies with larger standard errors scatter broadly across the base, forming an inverted funnel shape.

  • Asymmetry: Visual skew or an absence of small studies in one corner of the base points to publication bias (suppression of non-significant findings), citation bias, language bias, or true clinical heterogeneity across subgroups.

• Risk of Bias Assessment:

  • Methodological Evaluation: Systematically appraisal of trial designs using standard criteria.          

    

        

  • Seven Core Domains:

    • Random Sequence Generation: Evaluates selection bias during assignment.

    • Allocation Concealment: Checks whether allocations could be foreseen before enrollment.

    • Blinding of Participants and Personnel: Measures performance bias.

    • Blinding of Outcome Assessment: Measures detection bias.

    • Incomplete Outcome Data: Measures attrition bias due to dropouts or exclusions.

    • Selective Reporting: Assesses reporting bias (suppression of null secondary outcomes).

    • Other Bias: Evaluates domain-specific threats to internal validity.

  • Classification Levels: Categorized as Low Risk of Bias (green +), Unclear Risk of Bias (yellow ?), or High Risk of Bias (red -).

Statistical Hypothesis Testing, P-Values, and Confidence Intervals

• Foundations of Hypothesis Testing:

  • Used to deduce whether observed differences in clinical parameters reflect true underlying population effects or random sampling error, measurement noise, and chance.

  • Frequentist vs. Bayesian Approaches:

    • Frequentist Statistics: Treats population parameters as fixed constants and defines probability strictly as the long-run frequency of outcomes across hypothetical identical replications of an experiment.

    • Bayesian Statistics: Treats parameters as random variables and incorporates prior belief distributions updated by empirical data to obtain posterior probabilities (not covered in classical testing frameworks).

• Hypothesis Formulation:

  • Null Hypothesis ($H_0$): Asserts the absence of any true effect or difference:       $H_0: \mu_A - \mu_B = 0$

  • Alternative Hypothesis ($H_a$): Contradicts the null hypothesis:

    • Two-Sided Formulation: Detects differences in either direction:         $H_a: \mu_A - \mu_B \neq 0$

    • One-Sided Formulation: Directs the alternative to a predefined directional superiority or inferiority claim:         $H_a: \mu_A - \mu_B > 0 \quad \text{or} \quad H_a: \mu_A - \mu_B < 0$

• Decision Thresholds and Error Types:

  • Type I Error ($\alpha$): Rejecting the null hypothesis when it is true (false positive rate). Conventionally set to $\alpha = 0.05$.

  • Type II Error ($\beta$): Failing to reject the null hypothesis when it is false (false negative rate).

  • Critical Value Decision Approach: Establish a rejection region under the reference sampling distribution corresponding to $\alpha$. If the test statistic exceeds this critical cutoff, $H_0$ is rejected.

  • P-Value Decision Approach: Compute the exact probability of obtaining a test statistic at least as extreme as observed, assuming $H_0$ is strictly true. If $p < \alpha$, $H_0$ is rejected.

• The Student's $t$-Distribution:

  • Properties: A symmetric, bell-shaped probability distribution indexed by degrees of freedom ($\text{df}$).          

    

        

  • Heavier Tails: For smaller sample sizes, the $t$-distribution exhibits heavier tails than the standard normal ($Z$) distribution, accommodating added uncertainty from estimating the population variance using sample standard deviation. As $\text{df} \to \infty$, the $t$-distribution converges to the standard normal distribution.

• Independent Two-Sample $t$-Test:

  • Test Statistic Formula:       $t = \frac{\widehat{MD}}{SE} = \frac{\bar{x}_1 - \bar{x}_2}{SE}$

  • Variance Assumptions and Calculations:          

    

        

    • Equal Variances Assumed (Pooled Variance):           $t = \frac{\bar{x}_1 - \bar{x}_2}{s_p \sqrt{\frac{1}{n_1} + \frac{1}{n_2}}}$           $s_p^2 = \frac{(n_1 - 1)s_1^2 + (n_2 - 1)s_2^2}{n_1 + n_2 - 2}$           $\text{df} = n_1 + n_2 - 2$

    • Equal Variances Not Assumed (Welch's $t$-Test):           $t = \frac{\bar{x}_1 - \bar{x}_2}{\sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}}}$           $\text{df} = \frac{\left(\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}\right)^2}{\frac{(s_1^2 / n_1)^2}{n_1 - 1} + \frac{(s_2^2 / n_2)^2}{n_2 - 1}}$

  • Levene's Test for Equality of Variances: Employs an $F$-statistic to test $H_0: \sigma_1^2 = \sigma_2^2$. If significant ($p < 0.05$), the equal variance assumption is rejected, requiring Welch's corrected degrees of freedom.

• Parametric Tests vs. Nonparametric Counterparts:

  • Nonparametric rank tests are required when distributions are severely skewed or sample sizes are too small to rely on the Central Limit Theorem.          

    

        

  • 1 Median compared to a specified value: Wilcoxon Signed Rank test corresponds to the One-Sample $t$-test or $z$-test.

  • 2 Dependent (Paired) Medians: Wilcoxon Signed Rank test corresponds to the Paired Samples $t$-test.

  • 2 Independent Medians: Mann-Whitney U test corresponds to the Two-Sample Independent $t$-test.

  • $\ge 3$ Medians, 1 Variable: Kruskal-Wallis test corresponds to One-Way Analysis of Variance (ANOVA).

  • $\ge 3$ Medians, 2 Variables: Friedman test corresponds to Two-Way ANOVA.

  • 2 Categorical Variables for Independence: Chi-Square ($\chi^2$) Test of Independence has no parametric equivalent.

• Worked Example: LDL Follow-Up Comparison:

  • Group A ($n_A = 100$, $m_A = 199.40$, $s_A = 10.6$) versus Group B ($n_B = 100$, $m_B = 219.00$, $s_B = 10.8$).

  • Test statistic: $t = -13.006$.

  • At $\alpha = 0.05$, the critical threshold is approximately $\pm 1.96$.

  • Conclusion: Since $|t| > 1.96$, reject $H_0$. The estimated difference in means has a $95\%\,\text{CI}$ of $[-22.6, -16.7]\,\text{mg/dL}$.

• SPSS Output Interpretation (Medication Use by Gender):

  • Descriptives: Female ($N = 6,031$, Mean = $0.9517$, SD = $1.64185$, SE Mean = $0.02114$); Male ($N = 4,318$, Mean = $1.0919$, SD = $1.77678$, SE Mean = $0.02704$).

  • Levene's Test: $F = 26.608$, Sig. = $0.000$, confirming unequal variances.

  • Equal Variances Not Assumed: $t = -4.084$, $\text{df} = 8,842.811$, Sig. (2-tailed) = $0.000$, Mean Difference = $-0.14019$, SE Difference = $0.03432$, $95\%\,\text{CI}\,[-0.20747, -0.07291]$.

• Confidence Intervals (CI):

  • Formula:       $\text{CI} = \text{Point Estimate} \pm (\text{Critical Value} \times SE)$       $\widehat{MD} \pm 1.96 \times SE$

  • Frequentist Meaning: If the same sampling procedure is repeated across an infinite number of identical experiments, $95\%$ of the resulting intervals will capture the true population parameter.

  • Hypothesis Testing via CI: If the null value ($0$ for differences, $1$ for ratios) lies within the $95\%\,\text{CI}$, the findings are compatible with $H_0$ ($p \ge 0.05$). If the null value falls outside the interval, reject $H_0$ ($p < 0.05$).

  • Reporting Standards: Always report exact point estimates, $95\%\,\text{CI}$, and precise $p$-values. Never report non-significant results as p = n.s.. If software lists Sig. = .000, report as p < 0.001.

Linear Regression Modeling

• Clinical Purpose of Regression Analysis:

  • Allows researchers to either predict a continuous outcome ($Y$) from one or more explanatory variables ($X$), or assess the strength of an association between a primary predictor and an outcome while controlling for confounders.

• Baseline Characteristic Balance (Sabatine et al., 2017 NEJM Trial):

  • Baseline comparability in trials (e.g., Evolocumab, $N=13,784$ vs Placebo, $N=13,780$) is verified using specific statistical tests based on variable distribution:          

    

        

  • Continuous normal variables (e.g., Age: $62.5 \pm 9.1$ vs $62.5 \pm 8.9$; Weight: $85.0 \pm 17.3$ vs $85.5 \pm 17.4\,\text{kg}$) are evaluated using independent $t$-tests ($H_0: \mu_1 = \mu_2$).

  • Categorical variables (e.g., Male sex: $75.4\%$ vs $75.5\%$; Hypertension: $80.1\%$ vs $80.1\%$) are evaluated using $\chi^2$ tests ($H_0: \text{frequency}_1 = \text{frequency}_2$).

  • Skewed continuous variables (e.g., Median time from previous myocardial infarction: $3.4$ vs $3.3\,\text{yr}$; Median LDL: $92\,\text{mg/dL}$) are evaluated using Mann-Whitney U tests ($H_0: \text{Distribution}_1 = \text{Distribution}_2$).

  • Statistically significant baseline differences were documented for patient weight ($P = 0.01$) and aspirin/P2Y12 inhibitor use ($P = 0.03$).

• Mathematical vs. Statistical Formulations:

  • Mathematical Model: Deterministic relationship:       $Y = \beta_0 + \beta_1 X$

  • Statistical Model: Probabilistic formulation incorporating random residual error ($\epsilon$):       $Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + \dots + \beta_p X_p + \epsilon$       where $Y$ is the dependent variable, $X_1, \dots, X_p$ are predictors, $\beta_0$ is the intercept, $\beta_1, \dots, \beta_p$ are regression slopes, and $\epsilon \sim N(0, \sigma^2)$.

• Visualizing the Model:

  • A simple linear regression with one predictor describes a line in two-dimensional space.          

    

        

  • A multiple linear regression with two independent predictors describes a flat plane in three-dimensional space.

• Parameter Estimation (Ordinary Least Squares - OLS):

  • Calculates model parameters by minimizing the sum of squared vertical differences (residuals) between observed values ($y_i$) and fitted values ($\hat{y}_i$):       $\text{Minimize} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2 = \sum_{i=1}^{n} e_i^2$

• Four Classical Model Assumptions:

  • 1. Linearity: The relationship between predictors and the mean of the dependent variable is linear. Assessed via a Residuals vs. Fitted plot; residuals should scatter evenly around zero without non-linear curves.

  • 2. Independent Observations: Data points are sampled independently without clustering or auto-correlation.

  • 3. Normality of Residuals: Residual errors ($\epsilon_i$) follow a normal distribution ($\epsilon_i \sim N(0, \sigma^2)$). Assessed using histograms and Normal Q-Q plots.          

    

        

  • 4. Homoscedasticity: Error terms have constant variance across all fitted values.          

    

        

  • Violations of homoscedasticity produce distinct "fan-shaped" or "bow-tie" patterns, signaling heteroscedasticity that biases standard errors and invalidates inference.

• Wald Hypothesis Testing for Predictor Significance:

  • Assesses whether a specific regression slope differs from zero ($H_0: \beta_i = 0$ vs. $H_a: \beta_i \neq 0$):       $W = \frac{\hat{\beta}_i - \beta_i}{SE(\hat{\beta}_i)}$

  • The Wald statistic follows a $t$-distribution with $\text{df} = n - p - 1$.

  • Worked Example (Migraine Triptan Utilization by Age):

    • Estimated model: $\text{Total DDD} = 523.828 + 6.599 \times \text{age}$.

    • Constant ($\beta_0$): $B = 523.828$, $SE = 186.158$, $t = 2.814$, $p = 0.005$, $95\%\,\text{CI}\,[157.441, 890.215]$.

    • Age ($\beta_1$): $B = 6.599$, $SE = 4.107$, $\text{Beta} = 0.094$, $t = 1.607$, $p = 0.109$, $95\%\,\text{CI}\,[-1.484, 14.683]$.

    • Conclusion: Fail to reject $H_0$ because $p = 0.109 > 0.05$ and the $95\%\,\text{CI}$ crosses zero. Age does not show a statistically significant linear association with triptan consumption.

• Multiple Regression with Categorical Covariates:

  • Dummy coding converts categorical variables into binary indicators. For instance, coding gender as $1 = \text{Male}$ and $0 = \text{Female}$:       $\text{Number of Medications} = 0.273 + 0.010 \times \text{age} + 0.165 \times \text{gender}$

  • Coefficient Interpretation: At any given age, the model predicts that males take an average of $0.165$ more medications than females ($p < 0.001$).

• Confounding and Multivariable Adjustment:

  • Confounder Definition: A variable that causally influences the exposure and the outcome, without being an intermediate step on the causal pathway.          

    

        

  • Crude vs. Adjusted Models:

    • Crude: $\text{LDL-c} = a_0 + a_1 \times \text{adherence} + \epsilon$

    • Adjusted: $\text{LDL-c} = a_0^* + a_1^* \times \text{adherence} + a_2 \times \text{smoking} + a_3 \times \text{SES} + \epsilon$

  • Including confounders in multivariable regression removes spurious correlation and isolates the independent effect of the primary exposure.

• Statistical Interactions (Effect Modification):

  • An interaction occurs when the effect of an exposure on an outcome varies across levels of a third variable.

  • Model with interaction term:       $\text{LDL-c} = \beta_0 + \beta_1 \times \text{gender} + \beta_2 \times \text{age} + \beta_3 \times (\text{gender} \times \text{age}) + \epsilon$          

    

        

  • In the presence of interaction, slopes are not parallel; testing $H_0: \beta_3 = 0$ determines if the age-associated trajectory of LDL-c significantly differs between men and women.

• Model Fit Assessment:

  • Coefficient of Determination ($R^2$): Quantifies the proportion of total outcome variance explained by the model:       $R^2 = 1 - \frac{SS_{\text{residual}}}{SS_{\text{total}}}$

  • Adjusted $R^2$ ($R^2_{\text{adj}}$): Penalizes the score for each additional predictor, preventing artificial inflation from irrelevant covariates:       $R^2_{\text{adj}} = 1 - \left[\frac{(1 - R^2)(n - 1)}{n - p - 1}\right]$

  • ANOVA for Model Significance: Tests whether the fitted regression explains significantly more variance than a null intercept-only model ($H_0: \beta_1 = \beta_2 = \dots = \beta_p = 0$ or population $R^2 = 0$) using the $F$-ratio:       $F = \frac{MS_{\text{regression}}}{MS_{\text{residual}}}$

  • In the triptan example: Regression $SS = 1,770,366.214$ ($\text{df} = 1$), Residual $SS = 199,545,347.6$ ($\text{df} = 291$), $F = 2.582$, Sig. = $0.109$. The model does not fit the data significantly better than a constant model.

• Non-linear Transformation of Variables:

  • Linear models require linearity in the parameters ($\beta$), not in the predictors ($X$). Curvilinear relationships can be modeled via linear estimation by transforming the predictors:       $Y = a_0 + a_1 \sqrt{X_1} + a_2 \log(X_1) + a_3 X_2^2 + \epsilon$

  • Setting $W_1 = \sqrt{X_1}$, $W_2 = \log(X_1)$, and $W_3 = X_2^2$ yields a standard linear form:       $Y = a_0 + a_1 W_1 + a_2 W_2 + a_3 W_3 + \epsilon$

Logistic Regression for Binary Outcomes

• Limitations of Linear Regression for Binary Outcomes ($Y \in \{0, 1\}$):

  • Ordinary linear regression ($Y = a_0 + a_1 X + \epsilon$) yields unbounded predictions outside the valid probability bounds of $[0, 1]$.

  • Violates homoscedasticity, as the variance of a binary variable depends on the probability: $\text{Var}(Y) = p(1 - p)$.

  • Residuals can only take two discrete values at any given $X$, violating the assumption of normal error distribution.

• Probability vs. Odds:

  • Probability ($p$): Ratio of event occurrences to total opportunities:       $p = \frac{\text{Events}}{\text{Total Trials}} \quad (0 \le p \le 1)$

  • Odds: Ratio of event occurrences to non-occurrences:       $\text{Odds} = \frac{p}{1 - p} = \frac{\text{Events}}{\text{Non-Events}} \quad (0 \le \text{Odds} < \infty)$

  • Odds Ratio (OR): Ratio of two independent odds:       $\text{OR} = \frac{\text{Odds}_1}{\text{Odds}_0}$

  • Worked Example (Bronchitis and Childhood Cough - Holland et al., 1978):

    • In the cough group ($n = 70$): 26 bronchitis, 44 no bronchitis. Risk = $26/70 = 37.1\%$, Odds = $26/44 = 0.591$.

    • In the no-cough group ($n = 1,249$): 247 bronchitis, 1,002 no bronchitis. Risk = $247/1249 = 19.8\%$, Odds = $247/1002 = 0.247$.

    • Odds Ratio:           $\text{OR} = \frac{26 / 44}{247 / 1002} = \frac{0.591}{0.247} \approx 2.40$

    • Interpretation: Children with a cough have $2.40$ times higher odds of a bronchitis history than children without a cough.

• Mathematical Structure of Logistic Regression:

  • Logit Link Function: Models the natural logarithm of the odds as a linear function of predictors:       $z = \text{logit}(p) = \log\left(\frac{p}{1 - p}\right) = \beta_0 + \beta_1 X$       where $z \in (-\infty, +\infty)$.

  • Sigmoid (Logistic) Transformation: Maps the unbounded log-odds ($z$) back to a valid probability ($p$):       $p = \frac{1}{1 + e^{-z}} = \frac{1}{1 + e^{-(\beta_0 + \beta_1 X)}}$       where $p \in (0, 1)$.          

    

        

  • Parameter Estimation via Maximum Likelihood: Parameters are estimated by minimizing the cross-entropy loss (negative log-likelihood):       $\text{Log Loss} = -\sum_{i} \left[ y_i \log(p_i) + (1 - y_i) \log(1 - p_i) \right]$       This heavily penalizes confident, incorrect probability predictions.

• Coefficient Interpretation and Exponentiation:

  • In logistic models, slope coefficient $a_1$ represents the change in log-odds of the outcome per 1-unit increase in $X$.

  • Exponentiating the coefficient yields the Odds Ratio:       $\text{OR} = e^{a_1}$

  • Parameter Interpretation:

    • $a_1 > 0 \implies e^{a_1} > 1$: Positive association (increased odds of the event).

    • $a_1 = 0 \implies e^{a_1} = 1$: Null association (no change in odds).

    • $a_1 < 0 \implies e^{a_1} < 1$: Inverse/protective association (decreased odds).

  • Statin and Heart Attack Example:       $\log\left(\frac{P(\text{Heart Attack})}{1 - P(\text{Heart Attack})}\right) = 1.10 - 0.25 \times \text{statin use} + 0.02 \times \text{age}$

    • Statin Effect: $a_1 = -0.25 \implies \text{OR} = e^{-0.25} = 0.78$. Statin therapy is associated with a $22\%$ reduction in the odds of having a heart attack.

    • Age Effect: $a_2 = 0.02 \implies \text{OR} = e^{0.02} = 1.02$. Each one-year increase in age is associated with a $2\%$ increase in the odds of having a heart attack.

• Empirical Application: Varenicline vs. NRT Neuropsychiatric Safety (Wang et al., 2021 BMJ Open):

  • Retrospective inception cohort study evaluating neuropsychiatric adverse events (NPAEs) within 24 weeks.          

    

        

  • Psychiatric Cohort (Varenicline $n=1,427$ vs. NRT $n=1,200$):

    • Overall NPAEs: Events $1,074\,(75.3\%)$ vs. $942\,(78.5\%)$; Crude OR = $0.83\,[0.69, 1.00]$; Adjusted OR = $0.82\,[0.68, 0.99]$.

    • Depression: Events $629\,(44.1\%)$ vs. $548\,(45.7\%)$; Adjusted OR = $0.90\,[0.77, 1.05]$.

    • Anxiety: Events $456\,(32.0\%)$ vs. $472\,(39.3\%)$; Adjusted OR = $0.71\,[0.61, 0.84]$.

    • Insomnia: Events $372\,(26.1\%)$ vs. $352\,(29.3\%)$; Adjusted OR = $0.89\,[0.75, 1.06]$.

  • Non-Psychiatric Cohort (Varenicline $n=4,480$ vs. NRT $n=1,970$):

    • Overall NPAEs: Events $469\,(10.5\%)$ vs. $248\,(12.6\%)$; Adjusted OR = $0.85\,[0.72, 1.00]$.

    • Depression: Events $148\,(3.3\%)$ vs. $57\,(2.9\%)$; Adjusted OR = $1.15\,[0.84, 1.57]$.

    • Anxiety: Events $215\,(4.8\%)$ vs. $110\,(5.6\%)$; Adjusted OR = $0.90\,[0.71, 1.15]$.

    • Insomnia: Events $147\,(3.3\%)$ vs. $105\,(5.3\%)$; Adjusted OR = $0.63\,[0.49, 0.82]$.

Survival Analysis and Time-to-Event Modeling

• Nature of Time-to-Event Data:

  • Analyzes the time duration ($T$) elapsing from a defined baseline until a target endpoint occurs (e.g., mortality, graft failure, relapse).

  • Linear regression cannot be used because duration is strictly positive ($T \ge 0$) and clinical follow-up periods are limited, resulting in censored observations.

• Censoring:

  • Definition: Occurs when an individual does not experience the target event during their observed follow-up window.

  • Causes:

    • The study reaches its calendar conclusion while the patient remains event-free ("withdrawn alive").

    • The patient is lost to follow-up (e.g., moves away, withdraws consent).

    • The patient experiences an independent, competing terminal event.

  • Core Assumption: Non-informative censoring, meaning censored individuals must have the same probability of experiencing the event as uncensored individuals who remain under active observation.

• Hemodialysis Cohort Longitudinal Tracking Example:          

  

      

  • Patient A: Start Jan 2001; End Mar 2004 (Death $\dagger$); Total days = $1,155$.

  • Patient B: Start Sep 2005; End Jan 2008 (Censored/Alive $+$); Total days = $850+$.

  • Patient C: Start Jan 2003; End Jul 2003 (Death $\dagger$); Total days = $180$.

  • Patient D: Start Mar 2004; End Nov 2005 (Death $\dagger$); Total days = $605$.

  • Patient E: Start Jan 2004; End Jan 2008 (Censored/Alive $+$); Total days = $1,460+$.

  • Patient F: Start Apr 2000; End Nov 2002 (Death $\dagger$); Total days = $940$.

• Life-Table Construction:

  • Computes cumulative survival probabilities over discrete time intervals:          | Time (Days) | $N$ at Risk | Events | Censored | Proportion with Event | Interval Survival | Cumulative Survival |     | :--- | :--- | :--- | :--- | :--- | :--- | :--- |     | $0$ | $6$ | $-$ | $-$ | $-$ | $1.000$ | $1.000$ |     | $180$ | $6$ | $1$ | $0$ | $1/6$ | $5/6$ | $5/6 \approx 0.833$ |     | $605$ | $5$ | $1$ | $0$ | $1/5$ | $4/5$ | $(5/6) \times (4/5) = 4/6 \approx 0.667$ |     | $850$ | $4$ | $0$ | $1$ | $0$ | $1.000$ | $4/6 \approx 0.667$ |     | $940$ | $3$ | $1$ | $0$ | $1/3$ | $2/3$ | $(4/6) \times (2/3) = 4/9 \approx 0.444$ |     | $1,155$ | $2$ | $1$ | $0$ | $1/2$ | $1/2$ | $(4/9) \times (1/2) = 2/9 \approx 0.222$ |     | $1,460$ | $1$ | $0$ | $1$ | $0$ | $1.000$ | $2/9 \approx 0.222$ |

• Kaplan-Meier Curve:

  • Graphic Form: A non-parametric step-function where vertical drops represent event occurrences and tick marks denote censored individuals.          

    

        

  • Median Survival Time: The time point at which the cumulative survival probability reaches exactly $50\%$ ($0.50$). In the hemodialysis example, this occurs at approximately $940$ days.

  • Assumptions: Event criteria remain stable over time; early and late enrollees share identical hazard risks; non-informative censoring holds.

• Log-Rank (Mantel-Haenszel) Test:

  • Hypotheses:       $H_0: \text{The survival curves of the groups are identical}$       $H_a: \text{The survival curves of the groups are not identical}$

  • Limitations: A univariate test that cannot adjust for confounding covariates. It also loses statistical power when survival curves cross, which indicates non-proportional hazards.

• Cox Proportional Hazards Regression:

  • Hazard Rate ($h(t)$): The instantaneous rate of experiencing an event during the interval $[t, t + \Delta t]$, conditional on surviving up to time $t$.

  • Semi-Parametric Model: Models the log hazard without requiring parametric assumptions about baseline hazard shape:       $\log(h(t)) = \log(h_0(t)) + a_1 X_1 + a_2 X_2 + \dots + a_p X_p$       $\text{Hazard Ratio (HR)} = e^{a_1}$

  • Interpretation of Hazard Ratios:

    • $\text{HR} > 1$: Associated with an increased event rate over time.

    • $\text{HR} = 1$: Null effect; identical hazard to the reference.

    • $\text{HR} < 1$: Associated with a reduced event rate over time.

  • Cox Regression SPSS Output Analysis:          

    

        

    • Gender: $B = 0.183$, $SE = 0.309$, Wald = $0.352$, $\text{df} = 1$, Sig. = $0.553$, $\text{Exp}(B) = 1.201\,(95\%\,\text{CI}\,[0.655, 2.203])$. Gender differences are not statistically significant.

    • BMI: $B = -0.071$, $SE = 0.036$, Wald = $3.860$, $\text{df} = 1$, Sig. = $0.049$, $\text{Exp}(B) = 0.931\,(95\%\,\text{CI}\,[0.867, 1.000])$. Each unit increase in BMI reduces the instantaneous hazard rate by $6.9\%$ ($1 - 0.931 = 0.069$).

    • Agegroup (Overall Factor): Wald = $7.333$, $\text{df} = 3$, Sig. = $0.062$.

    • Agegroup(1): $B = 0.103$, $SE = 0.522$, Wald = $0.039$, $\text{df} = 1$, Sig. = $0.844$, $\text{Exp}(B) = 1.108\,[0.399, 3.081]$.

    • Agegroup(2): $B = 0.906$, $SE = 0.457$, Wald = $3.940$, $\text{df} = 1$, Sig. = $0.047$, $\text{Exp}(B) = 2.475\,[1.012, 6.058]$. Group 2 has a $2.475$ times higher hazard rate than the baseline reference category.

    • Agegroup(3): $B = 1.010$, $SE = 0.453$, Wald = $4.978$, $\text{df} = 1$, Sig. = $0.026$, $\text{Exp}(B) = 2.745\,[1.131, 6.665]$. Group 3 has a $2.745$ times higher hazard rate than the baseline reference category.

  • Cox Model Assumptions:

    • Proportional Hazards: The ratio of hazards between groups must remain constant over time.                  

      

              

    • If hazard lines cross over time, the proportionality assumption is violated, requiring time-dependent covariates or stratified Cox models.

    • Linearity: Log-hazard must have a linear relationship with continuous covariates.

Ordinal Outcomes, Statistical Misconceptions, and Multiple Testing Corrections

• Ordinal Outcomes in Medical Research:

  • Definition: Categorical variables that follow a natural hierarchical sequence but lack uniform numeric intervals between levels.

  • Clinical Examples: Medication Adherence Rating Scale (MARS), Drug Attitude Inventory (DAI), and the Clinician Rating Scale (CRS).

  • Clinician Rating Scale (CRS) 7-Level Tier:          

    

        

    • Rating 1: Complete refusal.

    • Rating 2: Partial refusal or only accepts minimum dose.

    • Rating 3: Accepts only because compulsory, or very reluctant / requires persuasion, or questions the need for medication often (e.g., every 2 days).

    • Rating 4: Occasional reluctance (e.g., questions the need for medication once a week).

    • Rating 5: Passive acceptance.

    • Rating 6: Moderate participation, some knowledge and interest in medication and no prompting required.

    • Rating 7: Active participation, readily accepts, and shows some responsibility for regimen.

  • Analytical Strategies:

    • Dichotomization: Collapses levels into two groups (e.g., Ratings 1–4 coded as $0$ for "some form of refusal"; Ratings 5–7 coded as $1$ for "some level of adherence"), enabling standard binary logistic regression.

    • Methodological Caveat: Dichotomization discards statistical information and reduces statistical power. It should only be used when clinically justified.

    • Ordinal Logistic Regression: Uses proportional odds models to analyze the outcome across all ordinal categories without losing information.

• Misinterpretations of P-Values:

  • Hypothetical Scenario: Drug vs. Placebo ($n = 20$), two-sample $t$-test, $t = 4.7$, $\text{df} = 18$, $p = 0.01$.

  • Common False Interpretations:

    • "We have disproved the null hypothesis": False. Empirical science never mathematically proves or disproves hypotheses; it only rejects or fails to reject them based on probabilistic evidence.

    • "We have found the probability of the null hypothesis being true ($P(H_0) = 0.01$)": False. A frequentist $p$-value is not the posterior probability of the hypothesis ($P(H_0 \mid \text{Data})$). Rather, it is the conditional probability of the data given the null ($P(\text{Data} \ge \text{Observed} \mid H_0)$).

    • "We have found the probability of the alternative hypothesis being true": False. Hypotheses do not possess probability distributions in frequentist inference.

    • "We have found a practically/clinically significant result": False. Statistical significance reflects sample size and variance, not clinical meaningfulness.

  • The Effect of Sample Size on $p$-Values:       $t = \frac{\bar{X}_1 - \bar{X}_2}{s_p \sqrt{\frac{2}{n}}}$       As sample size ($n$) increases toward infinity, the standard error shrinks toward zero, driving the $t$-statistic toward infinity and the $p$-value toward zero. Even clinically trivial differences eventually cross arbitrary significance thresholds ($p < 0.05$) in sufficiently large cohorts.

  • Formal Definition of a P-Value: The probability of obtaining a test statistic at least as extreme as the one calculated from the observed sample data, assuming the null hypothesis is true, the population distribution is identical, and the sampling scheme is repeated with the same sample size.

  • American Statistical Association (ASA) Warning on P-Values:          

    

        

  • The ASA issued guidelines noting that widespread misuse of $p$-values has contributed to the scientific reproducibility crisis. P-values cannot determine whether a scientific hypothesis is true or whether the findings are clinically meaningful.

• Common Methodological Errors in Statistical Analysis:

  • Misinterpreting $p$-values as effect sizes or clinical importance.

  • Confusing correlation with causation.

  • Ignoring distribution and modeling assumptions (e.g., homoscedasticity, normality, independence).

  • Data dredging or $p$-hacking: Conducting multiple post-hoc comparisons until a test achieves $p < 0.05$.

  • Failing to quantify and report uncertainty using confidence intervals.

  • Ignoring influential outliers.

  • Miscommunicating findings to clinical audiences.

• Remedies for Statistical Pitfalls:

  • Distinguish between confirmatory research (testing a predefined hypothesis) and exploratory research (generating hypotheses through data mining).

  • Pre-register study protocols (mandatory in pharmacoepidemiology, such as IADB-related registries) to prevent post-hoc changes to hypotheses.

  • Let substantive scientific theory guide study designs rather than opportunistic statistical mining.

  • Report confidence intervals alongside $p$-values to show the plausible range of values compatible with the data.

• Multiple Comparison Corrections:

  • The Problem of Multiplicity: Performing $m$ independent statistical tests at a fixed significance level $\alpha = 0.05$ inflates the Family-Wise Error Rate (FWER):       $\text{FWER} = 1 - (1 - \alpha)^m$       Running multiple comparisons without adjustment substantially increases the likelihood of false-positive conclusions (Type I errors).

  • Bonferroni Correction:

    • Mechanism: Divides the family-wise error rate equally across all planned hypothesis tests:           $\alpha_{\text{bonferroni}} = \frac{\alpha}{m}$

    • Decision Rule: Reject $H_0$ for test $i$ only if its $p$-value satisfies $p_i \le \alpha_{\text{bonferroni}}$.

    • Limitation: Highly conservative; substantially increases Type II errors (false negatives) when many tests are conducted.

  • Benjamini-Hochberg (BH) False Discovery Rate Procedure:

    • Mechanism: Controls the False Discovery Rate (FDR, $q$, typically $0.05$ or $0.10$), which is the expected proportion of false positives among all rejected null hypotheses.

    • Step-by-Step Procedure:

      1. Run all $m$ planned hypothesis tests and obtain their raw $p$-values ($p_1, p_2, \dots, p_m$).

      2. Rank the $p$-values in ascending order: $p_{(1)} \le p_{(2)} \le \dots \le p_{(m)}$.

      3. Calculate the critical value for each rank $k$ (where $k \in \{1, 2, \dots, m\}$):                $\text{Critical Value} = \frac{k}{m} \times q$

      4. Identify the largest rank $k$ for which the observed $p$-value is less than or equal to its critical value ($p_{(k)} \le \frac{k}{m} q$).

      5. Reject the null hypothesis for that test and all tests with smaller $p$-values (ranks $1$ through $k$).