Biostatistics & Epidemiology 158

Take-Home Points: Basics of Biostatistics

1. Significance of Statistics in Medicine

o Statistics, introduced to biomedical research ~150 years ago, is now indispensable in
understanding medical literature and conducting research.
2. Understanding Data
o Data are the raw material for statistics, consisting of measurements, observations, or counts.
o A variable is a characteristic for which data are collected, and data represent the values of this
variable.
3. Types and Representation of Data
o Categorical Data: Described using percentages or proportions.
o Numerical Data: Observations cluster around a central value, summarized by measures of
central tendency (mean, median, mode) and dispersion (range, variance, standard deviation).
4. Confidence Intervals (CI)
o The CI provides a range likely to contain the true population value, accounting for uncertainty.
o The standard convention is the 95% CI, derived from the standard error.
5. Patterns and Distributions
o Data distributions, such as the normal (bell-shaped Gaussian curve), binomial, and Poisson
distributions, reveal patterns crucial for descriptive statistics.
6. Visual Summarization
o Graphs like histograms, scatter plots, and less common tools like box-and-whisker plots or stem-
and-leaf plots offer insights into data trends and summaries.

Familiarity with these foundational concepts ensures a stronger grasp of biostatistics and its application in medical
research.

Take-Home Points: Hypothesis Testing in Biostatistics

1. Role of Hypothesis Testing

o Hypothesis testing is central to biostatistics, beginning with a null hypothesis (no
change/difference) and an alternative hypothesis (expected effect/difference).
o Researchers study samples to gather evidence to reject the null hypothesis in favor of the
alternative hypothesis.
2. Understanding the P Value
o The P value represents the probability of observing a result as extreme as the one found if the
null hypothesis is true.
o A P value < 0.05 typically signifies statistical significance, warranting rejection of the null
hypothesis.
3. Advances in Statistical Analysis
o Statistical software has simplified calculations, enabling researchers to focus on study design,
hypothesis selection, and accurate execution.
4. Choosing the Right Test
o The choice of hypothesis test depends on:
 Research question: One of five generic types.
 Variables: Numerical vs. categorical; parametric vs. nonparametric.
 Groups: Two vs. more than two.
o Misusing multiple tests to find a P < 0.05 is statistically incorrect.
5. Interpreting Results
o Effect size with a 95% confidence interval is vital for meaningful conclusions.
 o A small P value in large studies may indicate statistical significance but does not guarantee a
meaningful effect. Overlapping confidence intervals between groups suggest inconclusive
differences, even if P < 0.05.
6. Complementary Approaches
o P values and confidence intervals complement each other, offering a comprehensive view of
statistical significance and practical relevance.

Mastering these concepts ensures accurate and ethical application of hypothesis testing in medical research.

Take-Home Points: Analyzing Numerical Data

1. Parametric Tests and Normality

o Parametric tests analyze numerical data assuming a normal distribution defined by specific
parameters.
o Normality can be evaluated visually using a normal probability plot or tested using goodness-of-
fit tests like the Kolmogorov-Smirnov test.
2. Student's t-Test
o Used for comparing means, with three main variants:
 One-sample t-test: Compares a sample mean against a known population mean.
 Independent (unpaired) t-test: Compares means of two independent samples.
 Paired (dependent) t-test: Compares means of related samples.
o The t-test is unsuitable for comparing more than two groups due to an increased risk of Type I
error when multiple comparisons are made.
3. One-Way ANOVA
o Compares means of three or more independent, normally distributed groups.
o For repeated measures (multiple observations from the same subjects), repeated measures
ANOVA is required.
o While ANOVA identifies significant differences, it doesn’t specify which groups differ. Post hoc
tests, like Tukey's HSD test, clarify group-specific differences.
4. Nonparametric Alternatives
o If parametric assumptions are violated, nonparametric tests offer robust alternatives:
 Mann-Whitney U-test: Nonparametric counterpart to the independent t-test.
 Wilcoxon signed-rank test: Counterpart to the paired t-test.
 Kruskal-Wallis test: Nonparametric equivalent of one-way ANOVA.
 Friedman's test: Counterpart to repeated measures ANOVA.
5. Key Considerations
o Misusing parametric tests on non-normally distributed data can lead to erroneous conclusions.
o Post hoc tests are essential after multi-group comparisons to pinpoint specific differences.

Selecting the appropriate test ensures accurate and reliable data analysis, whether parametric or nonparametric.

Take-Home Points: Analyzing Categorical Variables

1. Representation and Organization

o Categorical variables are represented as counts or frequencies, often organized into contingency
tables labeled as r × c tables (rows × columns).
2. Chi-Square (χ²) Tests
o Chi-square tests analyze associations between categorical variables:
 Pearson's χ² test: Assesses differences in the distribution of a categorical variable across
independent groups.
 χ² test for trend: Used for ordered categorical groups.
 o Limitations: Large sample sizes are preferred; small numbers may lead to inaccuracies.
3. Fisher's Exact Test
o Suitable for small samples and 2 × 2 tables.
o Tests the independence of two dichotomous variables.
o Not applicable to contingency tables larger than two dimensions.
4. Paired Sample Analysis
o McNemar's χ² test: Compares paired proportions in a 2 × 2 table.
o Cochran's Q test: Generalizes McNemar’s test for comparing more than two related
proportions.
5. Association Strength
o While χ² tests yield a P value to assess significance, they do not measure the strength of
association.
o Relative risk and odds ratio are derived from 2 × 2 tables to quantify dichotomous associations.

Selecting the appropriate test ensures accurate insights into categorical data relationships, particularly when
considering sample size, independence, and data structure.

4o

Take-Home Points: Sample Size in Biomedical Research

1. Importance of Sample Size

o An adequate sample size ensures reliable study outcomes, whether the data indicates significant
differences or not.
2. Key Determinants
o Type 1 error (α): Risk of falsely concluding a difference exists (false positive).
o Type 2 error (β): Risk of failing to detect a true difference (false negative).
o Power (1 - β): Probability of detecting a true difference; conventionally set at 80% or higher.
o Variance: Greater variability in data increases sample size requirements.
o Effect size: Smaller clinically important differences require larger samples to detect.
3. Conventional Thresholds
o Typical values for sample size calculations:
 α ≤ 5% (probability of Type 1 error).
 Power ≥ 80% (to minimize Type 2 error).
4. Practical Considerations
o Advances in statistical software have simplified the complex mathematics of sample size
determination, allowing researchers to focus on study design and achieving sample size targets.
5. Clinical Relevance
o Effect size should reflect the smallest difference considered clinically meaningful, informed by
clinical judgment and past research.
6. Takeaway
o Thoughtful sample size determination avoids underpowered studies (inconclusive results) and
overpowered studies (wasted resources), ensuring findings are both statistically and clinically
meaningful.

Accurate sample size calculation is essential for meaningful research and facilitates acceptance of study
conclusions.

Key Points: Correlation and Linear Regression

1. Purpose
 o Correlation measures the strength and direction of a linear relationship between two numeric
variables, expressed as a correlation coefficient.
o Linear regression models the relationship between variables, enabling prediction of one variable
based on another.
2. Correlation Coefficients
o Pearson’s correlation coefficient (r): Used when both variables are normally distributed.
o Spearman’s rank correlation coefficient (ρ): A nonparametric alternative when normality is not
assumed.
o Coefficient of determination (r²): Represents the proportion of variability in the dependent
variable explained by the independent variable.
3. Hypothesis Testing in Correlation
o Tests whether a linear relationship exists in the population.
o P < 0.05 suggests a statistically significant correlation.
o A 95% confidence interval provides insight into the population correlation.
4. Linear Regression
o Defines the relationship as y=a+bxy = a + bxy=a+bx, where bbb is the slope and aaa the intercept.
o The least squares method is commonly used to fit the regression line.
5. Assumptions and Limitations
o Linear relationship: Verify with a scatter plot before analysis.
o Outliers or clustered data can distort results.
o Misinterpretation of correlation: Strong correlation ≠ causation.
6. Practical Insight
o Use correlation to describe relationships and regression to predict values.
o Ensure assumptions are met to avoid misleading conclusions.

Both techniques are foundational in statistical analysis but require careful interpretation to differentiate
relationships from causation.

Key Points: Evaluation of Diagnostic Tests

1. Diagnostic Tests in Medicine

o Diagnostic tests, including blood tests, imaging, clinical exams, scoring systems, and
questionnaires, are crucial for making therapeutic decisions.
o Tests categorize patients as either positive (disease likely) or negative (disease unlikely).
2. Test Accuracy and Evaluation
o Sensitivity: The ability of a test to correctly identify those with the disease.
1. Sensitivity = true positives (TP) / (true positives: TP + false negatives: FN)
2. A true positive (TP) occurs when a patient who has the disease obtains a positive test
result on a diagnostic test.
3. A false positive (FP) occurs when a patient who does not have the disease obtains a
positive test result.
o Specificity: The ability of a test to correctly identify those without the disease.
1. Specificity = true negatives / (true negatives + false positives)
2. A true negative (TN) occurs when a patient who does not have the disease obtains a
negative test result.
3. A false negative (FN) occurs when a patient who has the disease obtains a negative test
result.
o Positive Predictive Value (PPV): The probability that a positive test result indicates the presence
of disease.
1. PPV = true positives / (true positives + false positives)
 o Negative Predictive Value (NPV): The probability that a negative test result indicates the absence
of disease.
1. NPV = true negatives / (true negatives + false negatives)
3. Likelihood Ratios
o Combine sensitivity and specificity to assess the likelihood that a test result reflects the presence
or absence of disease.
o Likelihood ratio (LR) helps quantify the likelihood of disease given a test result.
o Positive likelihood ratio (LR+)
1. The probability of an individual with the disease testing positive divided by the
probability of an individual without the disease testing positive
2. LR+ = sensitivity / (1 − specificity)
o Negative likelihood ratio (LR−) The probability of an individual with the disease testing negative
divided by the probability of an individual without the disease testing negative
1. LR− = (1 − sensitivity) / specificity
4. Choosing Cutoff Points
o Tests with numerical results require judgment in selecting a cutoff to distinguish normal from
abnormal.
o There is a trade-off between sensitivity and specificity—a lower cutoff increases sensitivity but
decreases specificity, and vice versa.
5. Receiver Operating Characteristic (ROC) Curve
o The ROC curve plots sensitivity against (1 - specificity), helping to determine the optimal cutoff
point (the "elbow" of the curve).
6. Cohen's Kappa (κ)
o Cohen’s kappa measures the agreement between two tests or raters for categorical variables.
o It is considered more robust than simple percent agreement, as it accounts for chance
agreement.
7. Conclusion
o Accurate diagnostic tests are essential for clinical decisions, and understanding their metrics
(sensitivity, specificity, likelihood ratios, ROC curves, and kappa) ensures proper use and
interpretation in practice.

Sensitivity is an intrinsic test characteristic that does not depend on disease prevalence.

In this case, the physician explained to the patient that triple screening had a sensitivity of 60% in detecting chromosome abnormalities, while amniocentesis had a sensitivity of 90%. A
higher sensitivity typically means more TPs and fewer FNs (Choices B and C).

4o mini

Key Points: Estimating and Interpreting Risk in Studies

1. Risk Assessment in Studies

o In both observational and interventional studies, estimating risk is crucial for understanding the
association between outcomes and exposure to certain factors.
2. Incidence and Prevalence
o Incidence rate uses person-time as the denominator, which provides a more accurate measure
than simple counts.
o Prevalence refers to the proportion of individuals with a disease at a particular point in time.
3. Comparing Risks Between Groups
o To assess the importance of a risk factor, compare the outcome risk in the exposed group with
that in the nonexposed group.
o Comparison can be made using ratios (e.g., odds ratio, relative risk) or the difference in risks.
4. Key Ratios and Measures
 o Odds Ratio (OR): The ratio of the odds of an event in the exposed group to the odds in the
nonexposed group.
 OR = 1 indicates no difference in odds.
 OR > 1 suggests exposure increases risk.
 OR < 1 suggests exposure protects against risk.
 The 95% confidence interval (CI) of OR should be reported—if it includes 1, the OR's
significance is diminished.
o Relative Risk (RR): The ratio of risk (probability) of an event in the exposed group to the risk in
the nonexposed group.
 Interpretation is similar to OR but is often more useful when the event is common.
 For uncommon events, OR and RR are usually close in value.
5. Absolute Risk Reduction (ARR)
o ARR measures the effectiveness of an intervention, representing the difference in the event
occurrence between the control and treatment groups.
 ARR = Proportion of events in control group - Proportion in treated group.
 A higher ARR suggests a more effective intervention.
6. Number Needed to Treat (NNT)
o NNT is the number of individuals that need to be treated to achieve one additional successful
outcome compared to the control group.
 NNT = 1 / ARR.
 It can also reflect the number needed to prevent one adverse outcome.
7. Number Needed to Harm (NNH)
o NNH measures the number of individuals that need to be exposed to a treatment to cause one
additional harmful outcome.
o NNT and NNH are important for understanding the balance between benefit and harm and are
crucial for policy makers when evaluating treatment effectiveness.
8. Conclusion
o When evaluating risk in studies, it is essential to use metrics like OR, RR, ARR, NNT, and NNH to
interpret the effectiveness and potential harm of interventions. These measures provide a
clearer understanding of outcomes in both clinical and policy contexts.

Key Points: Survival Analysis in Time-to-Event Data

1. Purpose of Survival Analysis

o Survival analysis focuses on "time to event" data, where the event may not always be adverse. It
is concerned with determining how long it takes for a specific event to occur (e.g., death, relapse,
recovery, etc.).
2. Censored Observations
o In survival studies, it is common for subjects to leave the study prematurely or not experience
the event by the end of the study period. These are called censored observations because
complete information is unavailable for these subjects.
o Survival analysis methods are designed to handle censored data without treating it as missing,
which distinguishes it from other statistical techniques.
3. Descriptive Methods for Survival Times
o To explore survival times in a sample, common methods include:
 Life table method
 Kaplan-Meier estimator (which provides a cumulative survival probability over time)
 Distribution fitting (for advanced modeling)
4. Kaplan-Meier Plot
o The Kaplan-Meier plot is the most widely used method for visualizing survival analysis. It shows
the cumulative survival probability over time and is a central tool in biomedical survival analysis.
5. Comparing Survival Experiences Between Groups
 o Log-rank test is commonly used to compare survival experiences between two or more groups. It
tests whether there is a statistically significant difference in survival between groups.
o The hazard ratio (HR) is an estimate of the relative risk of an event between groups, which can
also be derived from the log-rank test. It represents how much more or less likely the event is to
occur in the test group compared to the control group.
6. Cox’s Proportional Hazards Model
o Cox’s proportional hazards model is a regression method used to estimate the impact of
multiple predictors (covariates) on survival, without making assumptions about the underlying
survival distribution.
 It accommodates both categorical and continuous variables.
 The model produces adjusted hazard ratios (HRs), similar to how adjusted odds ratios
are used in logistic regression, allowing for the assessment of multiple factors
influencing survival.
7. Limitations and Enhancements
o Although powerful, the traditional log-rank test has limitations in some scenarios, leading to the
development of various modifications and enhancements to improve its accuracy and
applicability.

Summary

Survival analysis is crucial for analyzing time to event data, especially in cases where the event of interest is not
guaranteed to occur for all participants during the study period. It handles censored observations effectively and
offers methods like Kaplan-Meier plots and the log-rank test for comparing survival between groups. Advanced
techniques, such as Cox’s proportional hazards model, allow researchers to account for multiple covariates,
providing a deeper understanding of the factors that affect survival outcomes.

4o mini

Key Points: Multivariate Analysis

1. Definition of Multivariate Analysis

o Multivariate analysis involves statistical techniques that examine the relationships between
three or more variables at the same time, aiming to uncover underlying connections and
patterns.
2. Types of Multivariate Analysis
o Dependence techniques: Focus on understanding the relationship between one or more
dependent variables and their independent predictors.
o Interdependence techniques: Treat all variables equally and search for underlying relationships
without differentiating between dependent and independent variables.
3. Common Multivariate Techniques
o Multiple Linear Regression: Predicts a single numerical dependent variable from multiple
numerical independent variables.
o Logistic Regression: Used when the outcome variable is dichotomous (i.e., binary, such as
yes/no, success/failure).
o Log-Linear Models: Analyze count data and can be used for examining cross-tabulations with
more than two variables.
o Analysis of Covariance (ANCOVA): An extension of ANOVA that includes an additional
independent variable (the covariate) to account for its effect on the dependent variable.
o Multivariate Analysis of Variance (MANOVA): Used when there are multiple numerical
dependent variables and you want to test for differences across groups.
4. Interdependence Techniques
 o Exploratory Factor Analysis (EFA): Seeks to identify latent factors that explain the correlations
among observed variables.
o Principal Component Analysis (PCA): Reduces a large number of variables to a smaller set of
components that capture the most variance in the data.
o Cluster Analysis: Identifies homogeneous groups or clusters in a dataset without prior
knowledge of group membership.
5. Applications of Multivariate Analysis
o Interdependence techniques are widely used in fields such as psychometrics, social sciences, and
market research to explore complex relationships within data.
6. Barriers and Advancements
o Historically, the computational complexity of multivariate analysis limited its use. However, the
increasing availability and sophistication of statistical software have made these techniques
more accessible, encouraging researchers to incorporate them into real-world data analysis.

Summary

Multivariate analysis allows researchers to examine complex relationships between multiple variables, using
techniques like multiple regression, logistic regression, and cluster analysis. Dependence techniques focus on
predicting outcomes based on other variables, while interdependence techniques explore how variables are
related to each other without distinguishing between dependent and independent factors. The advent of advanced
statistical software has made these powerful methods more accessible, enabling their widespread application in
diverse research fields.

A group of researchers conducted a randomized, double-blind, placebo-controlled trial of amitriptyline (1 mg per

kg of body weight per day), topiramate (2 mg per kg per day), and placebo in children and adolescents age 8-17
with migraine.

 The patients were randomly assigned in a 2:2:1 ratio to receive one of the medications or placebo.
 The primary outcome was the number of headache days in the 24-week trial.
 Secondary outcomes were headache-related disability, number of trial completers, and serious adverse
events that emerged during treatment.
 The study had 80% power to detect a relative reduction ≥50% in the number of headache days between
the first 28 days of the trial (baseline period) and the last 28 days of the trial.
 Statistical significance was established at 0.01.

Which of the following is closest to the probability of incorrectly finding a relative reduction ≥50% in the number
of headache days between the 28-day baseline period and the last 28 days of this trial?

Research studies generally compare a null hypothesis (H0) (eg, no relative reduction ≥50%) against an alternative
hypothesis (Ha) (eg, relative reduction ≥50%). Hypothesis testing may result in 1 of 4 possible outcomes:

 Two correct decisions:

o Determining there is no relative reduction ≥50% when one truly does not exist (correctly failing
to reject a true H0)
o Determining there is a relative reduction ≥50% when one truly does exist (correctly rejecting a
false H0); this reflects the power of a study
 Two incorrect decisions:
o Determining there is a relative reduction ≥50% when one truly does not exist (incorrectly
rejecting a true H0); this reflects a type I error
 o Determining there is no relative reduction ≥50% when one truly does exist (incorrectly failing to
reject a false H0); this reflects a type II error
 The probability of a type I error is known as the significance level (α); it is used to establish statistical
significance in hypothesis testing. This means that α is the probability of incorrectly rejecting a true H0.
 In this study, H0 represents no relative reduction ≥50% in the number of headache days between the first
28 days (baseline period) and the last 28 days of the trial. The statistical significance was established at
0.01 in the study, so α was set to 0.01. Therefore, the probability of incorrectly finding a relative
reduction ≥50% in the number of headache days in this trial (ie, type I error) is 0.01.

The significance level (α) is the probability of a type I error (ie, probability of incorrectly rejecting a true null
hypothesis H0).

A 38-year-old primigravid woman comes to the physician's office at 20 weeks gestation for prenatal counseling.
She is concerned about her baby's risk for Down syndrome and asks about measures to diagnose it early. The
physician explains that triple screening may detect up to 60% of cases of chromosomal abnormalities and that
amniocentesis may detect approximately 90% of cases. The patient decides not to undergo any testing. When
explaining that amniocentesis detected a higher percentage of cases compared to triple screening, the physician
was referring to which of the following values?

A cohort study of 4,000 patients examines the role of vitamin D supplements on the incidence of colon cancer. Relative risk (RR) calculations with their corresponding 95% confidence intervals
(CI) are reported for subjects taking vitamin D versus controls after a 5-year follow-up period:

RR (95% CI)

Colon cancer adjusted for age and gender 0.71 (0.61-0.81)

Colon cancer mortality adjusted for age and gender 0.75 (0.60-0.90)

A separate double-blind randomized clinical trial assigns 3,900 subjects to vitamin D supplementation or placebo. The following results are reported for subjects taking vitamin D versus
placebo after a 5-year follow-up period:

Hazard ratio (95% CI)

Colon cancer 0.98 (0.92-1.07)

Colon cancer mortality 1.01 (0.94-1.10)

Which of the following best describes the results of both studies?

Unknown residual confounders may impact the statistical analysis of study results.

Randomization removes the effects of known and unknown confounders.

A study is conducted to assess the role of different treatment regimens on cardiovascular outcomes. Two
treatment arms are evaluated: high-dose hydrochlorothiazide (100 mg/day) and low-dose hydrochlorothiazide (25 mg/day). After a defined follow-up period, the mean
systolic blood pressure (BP) in the high-dose group is 139 mm Hg, with a mean diastolic BP of 88 mm Hg. In the low-dose group, the mean values are 143 mm Hg
and 92 mm Hg, respectively. A 2-sample t-test gives p-values of 0.03 for the systolic BP and 0.04 for the diastolic BP differences between the high-
dose and low-dose groups. The relative risk of sudden cardiac death in the low-dose as compared to the high-dose group is 0.4 (95% confidence interval 0.25-0.55).

When assessing measures of effect, it is important to recognize which groups are being compared as the interpretation of the measures can differ.

Another subset of patients from this sample was given a placebo, and the risk of sudden cardiac death was assessed. When compared to the low-dose hydrochlorothiazide group,
the placebo group's relative risk of sudden cardiac death was very close to 1.0. The mean systolic and diastolic blood pressures in the placebo group were higher when compared to the low-
dose hydrochlorothiazide group. Which of the following is the best statement concerning the risk of sudden cardiac death in the overall sample of patients?

The relative risk (RR) of sudden cardiac death (SCD) in the placebo group compared to the low-dose hydrochlorothiazide (HCTZ) group is ~1.0. This means that patients taking low-dose
HCTZ have a risk of SCD that is approximately equal to that of patients taking placebo.
Summarizing the data provided in these 2 questions:

 RR of SCD in the low-dose HCTZ group as compared to the placebo group = ~1.0 (Equation i)
 RR of SCD in the low-dose HCTZ group as compared to the high-dose HCTZ group = 0.4 (Equation ii)

Because RR is a ratio obtained by dividing 2 values, the numerator and the denominator can be inverted (ie, x → 1/x) to change the reference. Therefore, by inverting the values in (Equation
ii):

 RR of SCD in the high-dose HCTZ group as compared to the low-dose HCTZ group = 1.0/0.4 = 2.5 (Equation iii)

Finally, taking (Equation iii) and (Equation i):

 RR of SCD in the high-dose HCTZ group as compared to the low-dose HCTZ group = 2.5
 RR of SCD in the low-dose HCTZ group as compared to the placebo group = ~1.0

Therefore, RR of SCD in the high-dose HCTZ group as compared to the placebo group is ~2.5. In other words, high-dose HCTZ treatment seems to increase the risk of SCD by a factor of 2.5
(not 50% [Choice B]) when compared to both low-dose treatment and placebo.