Writing the Statistical Section for a Clinical Study Report

Introduction

The statistical section of a Clinical Study Report (CSR) is where the reader learns how the clinical trial data were converted into the statistical results presented elsewhere in the report.

It is therefore much more than a list of statistical tests.

A strong statistical section connects the study objectives, estimands, endpoints, analysis populations, statistical methods, handling of missing data, and presentation of results into one coherent analytical story.

The section should allow a knowledgeable reader to understand:

Which patients were included in each analysis
Which endpoints were analyzed
How each endpoint was defined
Which statistical methods were used
How treatment effects were estimated
How uncertainty was quantified
How missing or incomplete observations were handled
How multiplicity was addressed
Which sensitivity and subgroup analyses were performed
Whether any analyses differed from the prespecified plan

Key idea: The statistical section should explain the analytical methods well enough that the reader can understand and interpret the results without having to reverse engineer the analysis from the tables, listings, figures, and programming outputs.

The Statistical Section Is Not the SAP

One of the most important distinctions is between the Statistical Analysis Plan (SAP) and the statistical section of the CSR.

The SAP is the detailed operational specification of how the analyses will be performed.

The CSR statistical section is a concise description of the methods actually used to generate the reported results.

Document	Primary Purpose
Protocol	Defines the study objectives, design, endpoints, and key methodological requirements
Statistical Analysis Plan	Provides detailed prespecified statistical methodology and programming specifications
CSR statistical section	Summarizes the statistical methods used to produce and interpret the reported results
Tables, Listings, Figures	Present the resulting statistical outputs

The CSR should therefore not simply copy the entire SAP.

At the same time, it should not be so abbreviated that important analytical decisions become invisible.

Practical rule: The SAP tells the statistical programmer exactly what to do. The CSR tells the reader what was done, why it matters, and how the reported results should be interpreted.

Where the Statistical Section Fits in the CSR

The exact CSR structure depends on the reporting standard and study, but the statistical material is commonly distributed across sections addressing study design, statistical methods, efficacy, safety, and other results.

For an integrated statistical discussion, the reader should be able to trace a logical path from:

Study objective — What clinical question is the trial answering?

Estimand — What treatment effect is being targeted?

Endpoint — What outcome is used to quantify that effect?

Analysis population — Which patients contribute to the analysis?

Statistical method — How is the treatment effect estimated?

Uncertainty — How are confidence intervals or standard errors calculated?

Results — What treatment effect was observed?

What Makes a Good Statistical Section?

A high-quality statistical section should be:

Accurate: It must agree with the actual analyses.
Traceable: Statements should be supported by the SAP, datasets, and outputs.
Reproducible: A qualified statistician should understand how the analysis was performed.
Concise: The CSR is not a replacement for the full SAP.
Consistent: Terminology should agree throughout the CSR.
Transparent: Important deviations from the planned analysis should be identified.
Clinically interpretable: Statistical terminology should be connected to the clinical question.

The Core Components

A useful statistical section will normally address several recurring components.

Component	Typical Content
Study design	Randomization, stratification, treatment groups, blinding, and design features relevant to analysis
Analysis populations	Intent-to-treat, full analysis set, per-protocol, safety, pharmacokinetic, or other defined populations
Endpoints	Primary, secondary, exploratory, safety, and other outcome definitions
Estimands	Target treatment effects and handling of intercurrent events
Statistical methods	Models, tests, estimators, covariates, stratification factors, and assumptions
Missing data	Handling rules, imputation, censoring, and sensitivity analyses
Multiplicity	Control of Type I error and adjustment strategies
Subgroups	Prespecified subgroup factors and interaction analyses
Sensitivity analyses	Analyses assessing robustness of primary conclusions
Interim analyses	Timing, purpose, statistical adjustment, and impact on inference

Start With the Statistical Objectives

Before writing individual statistical sentences, identify the actual statistical questions the study is designed to answer.

For example:

Is the experimental treatment superior to control?
Does treatment delay disease progression?
Does treatment improve overall survival?
Does treatment increase the probability of response?
Does treatment reduce a biomarker measurement?
Is the treatment noninferior to an active comparator?

The statistical section should make these objectives visible.

Writing principle: Do not begin with "A t-test was performed." Begin by identifying the clinical question and endpoint. The statistical method should appear as the solution to a defined analytical problem.

Primary and Secondary Endpoints

The statistical section should clearly distinguish the primary endpoint from secondary and exploratory endpoints.

Endpoint Class	Example	Typical Statistical Description
Primary	Change from baseline in HbA1c at Week 24	Prespecified primary comparison and confidence interval
Secondary	Proportion achieving HbA1c <7%	Responder analysis and treatment comparison
Time-to-event	Progression-free survival	Kaplan-Meier estimation and treatment hazard ratio
Exploratory	Change in exploratory biomarker	Descriptive or exploratory inferential analysis

Define the Endpoint Precisely

A statistical method cannot be understood without understanding the endpoint.

For a continuous endpoint, specify relevant elements such as:

Measurement
Baseline definition
Assessment time
Change or percentage change
Analysis visit
Relevant transformation

For example, a change-from-baseline endpoint can be expressed as:

CHG_i = Y_i - Y_{i0}

where \(Y_i\) is the patient's post-baseline value and \(Y_{i0}\) is the corresponding baseline value.

A percentage change endpoint can be defined as:

PCHG_i = \frac{Y_i-Y_{i0}}{Y_{i0}}\times100

The CSR should describe the endpoint in words rather than relying exclusively on mathematical notation.

Analysis Populations

The analysis population is one of the most important components of the statistical section.

The reader needs to know which patients contributed to each analysis.

Common populations include:

Randomized or intent-to-treat population
Full analysis set
Per-protocol population
Safety population
Pharmacokinetic population
Pharmacodynamic population
Biomarker-evaluable population

The terminology varies by study and therapeutic area. The CSR should use the definitions established in the protocol and SAP.

Why Population Definitions Matter

Suppose 500 patients are randomized, 490 receive at least one dose, and 450 have a post-baseline efficacy assessment.

The treatment comparison could therefore involve different denominators depending on the endpoint and analysis population.

Population	Patients	Potential Use
Randomized	500	Disposition and randomized-treatment summaries
Safety	490	Safety analyses
Efficacy-evaluable	450	Endpoint-specific efficacy analyses if prespecified

Do not assume the denominator. The statistical section should make clear which analysis population is used for each major analysis, particularly when different populations produce different sample sizes.

Handling Treatment Assignment

The statistical section should explain how treatment groups are defined for the analysis.

This can be particularly important when treatment received differs from treatment assigned.

Examples include:

Intent-to-treat analyses based on randomized treatment
As-treated safety analyses
Actual treatment received in pharmacokinetic analyses
Treatment switching
Rescue medication

The relevant rule should be stated clearly and consistently.

Estimands and the Statistical Section

Modern clinical-trial reporting increasingly requires explicit consideration of the estimand.

An estimand describes the precise treatment effect that the trial is intended to estimate.

It can be useful to think of an estimand as answering:

\text{What treatment effect are we actually trying to estimate?}

An estimand is typically characterized by elements such as:

Population
Treatment conditions
Variable or endpoint
Intercurrent-event strategy
Population-level summary

Intercurrent Events

An intercurrent event is an event occurring after treatment initiation that affects either the interpretation or existence of the endpoint.

Examples include:

Treatment discontinuation
Rescue medication
Treatment switching
Death before endpoint assessment
Protocol-defined treatment failure

The statistical section should make clear how such events were handled when they materially affect the endpoint analysis.

Example: Treatment Discontinuation

Suppose a patient discontinues treatment at Week 12 because of an adverse event but would otherwise have had a Week 24 outcome.

Several analytical strategies could be considered depending on the estimand, study design, and SAP.

Strategy	Conceptual Question
Treatment policy	What would the outcome be regardless of treatment discontinuation?
Hypothetical	What would the outcome have been if the intercurrent event had not occurred?
Composite	Is the intercurrent event incorporated into the endpoint itself?
While-on-treatment	What is the treatment effect before the intercurrent event occurs?

The CSR does not need to reproduce the entire estimand framework if it is already fully documented elsewhere, but the reported statistical analysis should be consistent with the estimand being described.

Choosing the Statistical Method

The statistical method should be appropriate for:

The endpoint type
The study design
The estimand
The analysis population
The distribution of the data
The timing of measurements
The randomization structure

Common methods include:

Endpoint	Potential Method
Continuous outcome	ANCOVA, MMRM, analysis of covariance, t-test, regression
Binary outcome	Risk difference, risk ratio, odds ratio, logistic regression
Time-to-event	Kaplan-Meier, log-rank test, Cox proportional hazards model
Count outcome	Poisson or negative binomial model
Repeated measures	MMRM or other longitudinal model
Ordinal outcome	Ordinal logistic or rank-based methods

Do Not Just Name the Test

A common weak CSR statement is:

"The primary endpoint was analyzed using ANCOVA."

This is usually insufficient.

A better description identifies the endpoint, treatment effect, baseline adjustment, relevant covariates, and inferential quantity.

Example

"The primary endpoint was analyzed using an analysis of covariance model with change from baseline at Week 24 as the dependent variable, treatment as a fixed effect, and baseline HbA1c as a covariate. The adjusted treatment difference, two-sided 95% confidence interval, and corresponding p-value were reported."

Continuous Endpoints

For continuous endpoints, the statistical section should identify:

The response variable
Baseline adjustment
Treatment effect
Important stratification factors or covariates
Model assumptions where relevant
Estimated treatment contrast
Confidence interval
Hypothesis test, if applicable

ANCOVA Example

A simple ANCOVA model can be represented as:

Y_i = \beta_0+ \beta_1 Treatment_i+ \beta_2 Baseline_i+ \varepsilon_i

The treatment effect is represented by the appropriate treatment contrast, after adjustment for baseline.

In a two-arm trial, this may be expressed as:

\Delta = \bar{Y}_{\text{Treatment, adjusted}} - \bar{Y}_{\text{Control, adjusted}}

The CSR should generally report the estimated treatment difference rather than only the p-value.

Why Confidence Intervals Matter

A p-value addresses evidence against a specified null hypothesis.

A confidence interval provides information about the magnitude and uncertainty of the estimated treatment effect.

For a treatment difference, the CSR might report:

\[ \Delta=-1.8 \] 95% CI: \(-3.0,\,-0.6\)

This conveys substantially more information than simply stating:

"Treatment was statistically significant, \(p=0.004\)."

Binary Endpoints

For binary endpoints, the statistical section should identify the treatment effect measure.

Common measures include:

Risk difference
Risk ratio
Odds ratio
Adjusted odds ratio
Response-rate difference

For example, if the response rate is 48% in the experimental group and 35% in the control group, the risk difference is:

\[ RD=0.48-0.35=0.13 \]

or 13 percentage points.

Odds Ratios

The odds of response are:

Odds= \frac{P(\text{response})} {1-P(\text{response})}

The odds ratio compares these odds between treatment groups.

The CSR should specify whether the reported odds ratio is crude or adjusted and identify the model used to estimate it.

Time-to-Event Endpoints

Time-to-event analyses require additional statistical detail because both the event and the time at risk are relevant.

Common endpoints include:

Overall survival
Progression-free survival
Time to progression
Duration of response
Time to treatment failure

Kaplan-Meier Analysis

The Kaplan-Meier estimator estimates the probability of remaining event-free over time.

It can be written as:

\hat{S}(t) = \prod_{t_j\leq t} \left( 1-\frac{d_j}{n_j} \right)

where \(d_j\) is the number of events at time \(t_j\), and \(n_j\) is the number of individuals at risk immediately before that time.

The CSR should identify:

The event definition
The censoring definition
The analysis population
The Kaplan-Meier method
The treatment comparison
The reported summary measure

Median Survival

A common summary is the median time to event.

For example:

"Median progression-free survival was 11.4 months in the experimental treatment group and 7.8 months in the control group."

However, the statistical section should also identify the method used to estimate and compare the survival distributions.

Cox Proportional Hazards Model

A Cox model may be written as:

h(t|X)=h_0(t)\exp(\beta X)

The hazard ratio is:

HR=e^\beta

The CSR should explain which covariates or stratification factors were included and how the treatment effect was estimated.

Important: A hazard ratio is not the same thing as a ratio of median survival times. The two quantities answer different statistical questions.

Stratified Analyses

Many randomized trials use stratified randomization.

If stratification factors were incorporated into the primary analysis, the CSR should identify them.

For example:

"The primary analysis was stratified by geographic region and baseline disease severity, consistent with the randomization scheme."

The exact implementation should match the SAP and the statistical model.

Repeated Measures and MMRM

When an endpoint is measured repeatedly over time, a mixed model for repeated measures (MMRM) may be used.

A typical model can include:

Treatment
Visit
Treatment-by-visit interaction
Baseline measurement
Other prespecified covariates

The treatment comparison may then be made at the primary assessment visit.

The CSR should not simply state "MMRM was used."

Instead, identify the principal fixed effects and covariance approach where important for understanding the analysis.

Missing Data

Missing data are one of the most important areas of statistical reporting.

The statistical section should distinguish between:

Patients with no post-baseline assessment
Patients with intermittent missing visits
Patients who discontinue early
Missing outcome values
Missing covariates
Censored observations in time-to-event analyses

Missing Data Should Not Be Described Vaguely

A statement such as:

"Missing data were handled appropriately."

provides almost no useful information.

The reader needs to know what "appropriately" means for the endpoint.

For example:

Better reporting

"Repeated measurements were analyzed using a mixed model for repeated measures without explicit imputation of missing outcome values, under the model-based assumptions specified in the SAP."

The exact wording should reflect the actual analysis.

Last Observation Carried Forward

Last observation carried forward (LOCF) has historically been used in clinical trials, but it should not be presented as a generic solution to missing data.

If LOCF is used, the CSR should state:

Which endpoint used LOCF
Why it was specified
Which observations were carried forward
Whether it was a primary or sensitivity analysis

Modern analyses often use model-based or multiple-imputation approaches where appropriate.

Multiple Imputation

Multiple imputation creates multiple plausible completed datasets rather than a single deterministic replacement.

Conceptually:

D_1,D_2,\ldots,D_M

are analyzed separately and then combined using an appropriate pooling method.

The CSR should identify the imputation model and important assumptions when multiple imputation materially affects the primary inference.

Sensitivity Analyses

Sensitivity analyses investigate whether the primary conclusion depends strongly on a particular analytical assumption.

Examples include:

Alternative missing-data assumptions
Alternative covariance structures
Alternative censoring rules
Alternative analysis populations
Per-protocol analyses
Pattern-mixture models
Tipping-point analyses

Key distinction: A sensitivity analysis is not simply another analysis. It should address a specific uncertainty or assumption underlying the primary analysis.

Multiplicity

Clinical trials frequently involve multiple statistical comparisons.

For example:

Multiple primary endpoints
Sequential testing
Multiple doses
Multiple treatment comparisons
Multiple time points
Hierarchical secondary endpoints

If multiple testing is not addressed appropriately, the probability of at least one false-positive conclusion can exceed the nominal significance level.

Family-Wise Type I Error

If \(m\) independent tests are each performed at level \(\alpha\), the probability of at least one false positive can be approximated as:

1-(1-\alpha)^m

For example, five independent tests at \(\alpha=0.05\) produce:

1-(1-0.05)^5 \approx0.226

or approximately 22.6%.

This illustrates why the statistical section should describe the multiplicity strategy when multiple confirmatory hypotheses are tested.

Common Multiplicity Strategies

Approach	Concept
Bonferroni	Adjusts the significance level or p-values across comparisons
Holm	Sequentially rejects hypotheses using ordered p-values
Hierarchical testing	Tests hypotheses in a prespecified sequence
Gatekeeping	Controls testing across families of hypotheses
Alpha spending	Allocates Type I error across sequential analyses
Graphical approaches	Represent alpha allocation and recycling across hypotheses

Interim Analyses

If an interim analysis was conducted, the statistical section should explain its role.

Relevant details can include:

Timing of the interim analysis
Purpose
Data cutoff
Statistical method
Stopping boundaries
Alpha-spending approach
Whether the analysis was blinded or unblinded
Impact on the final analysis

For a group-sequential design, for example, the final significance threshold may differ from the nominal 0.05 level because the Type I error has been allocated across interim and final analyses.

Subgroup Analyses

Subgroup analyses are commonly performed for demographic and clinical factors such as:

Age
Sex
Race or ethnicity where appropriate and permitted
Geographic region
Baseline disease severity
Prior treatment
Biomarker status
Baseline risk category

The statistical section should distinguish between:

Prespecified subgroup analyses
Exploratory subgroup analyses
Post hoc analyses

Interaction Tests

A difference in statistical significance between two subgroups does not demonstrate that treatment effects differ between those subgroups.

For example:

"Treatment was statistically significant in patients aged <65 years but not in patients aged ≥65 years."

This statement alone does not establish treatment-effect heterogeneity.

A treatment-by-subgroup interaction should generally be evaluated when a formal assessment of effect modification is required.

The model may contain an interaction term:

Treatment \times Subgroup

Important: "Significant in one subgroup but not significant in another" is not equivalent to "significantly different between subgroups."

Protocol Deviations

Protocol deviations can affect which patients are included in particular analyses.

The statistical section should describe the relationship between important protocol deviations and analysis populations.

Examples include:

Major eligibility violations
Prohibited concomitant medication
Incorrect treatment assignment
Important assessment-window deviations
Unblinding
Noncompliance

The statistical section should not independently invent a new protocol deviation classification.

Instead, it should use the definitions established by the protocol, SAP, and relevant blinded review processes.

Baseline Characteristics

Baseline characteristics are usually summarized descriptively.

Continuous variables may be summarized using:

Number of observations
Mean
Standard deviation
Median
Minimum
Maximum

Categorical variables may be summarized using:

Number of patients
Percentage of patients

The statistical section should explain the summary conventions sufficiently for the reader to interpret the tables.

Descriptive Statistics

For a continuous variable \(X\), the arithmetic mean is:

\bar{X} = \frac{1}{n} \sum_{i=1}^{n}X_i

The sample standard deviation is:

SD = \sqrt{ \frac{ \sum_{i=1}^{n}(X_i-\bar{X})^2 }{ n-1 } }

The CSR does not normally need to reproduce these formulas, but understanding them helps ensure that statistical terminology is used correctly.

Reporting Percentages

For a categorical variable, the percentage is typically:

Percentage = \frac{n}{N}\times100

The denominator \(N\) should be clearly understood.

A percentage calculated among randomized patients may differ substantially from one calculated among treated patients or patients with available assessments.

Rounding Rules

Rounding should be consistent throughout the CSR.

The statistical section may describe conventions such as:

Percentages displayed to one decimal place
Means and standard deviations to one decimal place
Confidence intervals to two decimal places
P-values displayed to three decimal places
Very small p-values reported as \(p<0.001\)

Do not over-round intermediate calculations. Statistical computations should generally use appropriate underlying precision, with rounding applied for presentation rather than before the analysis unless specifically required by the analysis method.

P-Values

P-values should be reported consistently.

For example:

Actual P-Value	Recommended Display
0.4372	0.437
0.0127	0.013
0.0004	<0.001

The exact convention should follow the study's reporting standards.

Statistical Significance vs. Clinical Relevance

A statistical section should avoid implying that statistical significance automatically means clinical importance.

For example, with a very large sample size, a small treatment difference may produce a very small p-value.

Conversely, an important treatment difference may have a wide confidence interval in a small study.

The statistical section should therefore report the estimated effect and its uncertainty rather than relying solely on significance labels.

Effect Sizes

Examples of effect measures include:

Treatment difference
Risk difference
Risk ratio
Odds ratio
Hazard ratio
Mean ratio
Median difference

The appropriate effect measure depends on the endpoint and estimand.

Confidence Intervals

A two-sided 95% confidence interval for an estimated treatment effect can conceptually be written as:

Estimate \pm Critical\ Value \times Standard\ Error

For a simple normally distributed estimator, the critical value is commonly approximately 1.96.

Estimate \pm 1.96\times SE

The exact method used in the trial should be described when it is important to interpret the reported interval.

Noninferiority Trials

Noninferiority trials require particularly careful statistical reporting.

The statistical section should identify:

The noninferiority margin
The direction of the treatment effect
The analysis populations
The confidence interval used for the assessment
The primary noninferiority criterion
Any subsequent superiority assessment

Suppose the treatment difference is:

\Delta=1.2

with a 95% confidence interval:

[-0.8,\;3.2]

Whether this establishes noninferiority depends on the endpoint direction and the prespecified noninferiority margin.

Never interpret a noninferiority confidence interval without stating the margin and direction of benefit.

Superiority After Noninferiority

Some trials are designed to assess noninferiority first and superiority second.

The statistical section should describe the prespecified testing hierarchy.

A conclusion such as:

"The treatment was noninferior and subsequently demonstrated superiority."

should only be made when the prespecified statistical criteria support both conclusions.

Bayesian Analyses

Some clinical studies use Bayesian methods rather than, or in addition to, frequentist inference.

The statistical section should identify:

The prior distribution
The likelihood/model
The posterior quantity of interest
The decision criterion
The credible interval or posterior probability

For example, the posterior distribution can be represented as:

p(\theta|Data) \propto p(Data|\theta)p(\theta)

The CSR should use appropriate Bayesian terminology rather than describing a credible interval as though it were a frequentist confidence interval.

Transformation of Data

Some endpoints require transformation before analysis.

Common transformations include:

Log transformation
Square-root transformation
Logit transformation

For a log-transformed endpoint:

Y^*=\log(Y)

The statistical section should explain the transformation and, where relevant, how results were transformed back for clinical interpretation.

Geometric Means

Geometric means are commonly used for pharmacokinetic and other approximately log-normal measurements.

For positive observations \(X_1,\ldots,X_n\), the geometric mean is:

GM = \left( \prod_{i=1}^{n}X_i \right)^{1/n}

Equivalently:

GM = \exp \left( \frac{1}{n} \sum_{i=1}^{n}\log X_i \right)

Analysis of Adverse Events

The statistical section should explain key conventions used for safety analyses.

These may include:

Treatment-emergent adverse events
Multiple occurrences within a patient
Severity grading
Relationship to study treatment
Serious adverse events
Adverse events leading to discontinuation
Deaths
Exposure-adjusted incidence rates where applicable

Multiple Occurrences of the Same Adverse Event

A patient may experience the same adverse event more than once.

The CSR should specify whether the patient is counted once or multiple times for the relevant summary.

For many patient-incidence tables, the denominator is the number of patients in the safety population and a patient is counted once for the relevant event category.

Important: Patient incidence and event counts are not the same thing. The statistical section should make the counting convention clear.

Exposure-Adjusted Analyses

When treatment exposure differs substantially between groups, exposure-adjusted incidence rates may be useful.

A simple incidence rate can be expressed as:

Incidence\ Rate = \frac{Number\ of\ Events} {Total\ Person\text{-}Time}

The statistical section should identify the time unit and exposure definition if such analyses are presented.

Pharmacokinetic and Pharmacodynamic Analyses

If pharmacokinetic or pharmacodynamic analyses are included, the statistical section may need to describe:

Analysis populations
Derivation of PK parameters
Log transformation
Geometric mean ratios
Confidence intervals
Dose proportionality analyses
Exposure-response analyses

The exact statistical methods should be consistent with the SAP and PK analysis plan.

Multiplicity Across the CSR

Multiplicity should not be described only in one isolated paragraph if the testing strategy affects interpretation throughout the results.

The reader should be able to understand:

Which hypotheses were confirmatory
Which hypotheses were tested formally
Which hypotheses were exploratory
How alpha was allocated
Which p-values are adjusted
Which confidence intervals correspond to confirmatory inference

Post Hoc Analyses

A post hoc analysis is one that was not prespecified in the applicable statistical planning documentation.

Post hoc analyses can be useful, but they should be identified as such.

For example:

"An exploratory post hoc analysis was conducted to evaluate treatment effects according to baseline biomarker quartile."

This is preferable to presenting a post hoc analysis as though it were part of the original confirmatory testing strategy.

Analysis Cutoff Dates

The statistical section should identify important data cutoff dates when they matter to interpretation.

Examples include:

Database lock
Primary efficacy data cutoff
Safety data cutoff
Interim analysis cutoff
Long-term follow-up cutoff

This is particularly important when different sections of the CSR use different data snapshots.

Handling Unscheduled Assessments

Clinical trials frequently contain assessments outside the nominal visit schedule.

The SAP may specify whether these assessments are:

Included in longitudinal analyses
Assigned to the nearest nominal visit
Used only for selected endpoints
Excluded from specific visit-based summaries

The statistical section should summarize the convention when it affects the interpretation of the analysis.

Baseline Definition

Baseline must be defined consistently.

Depending on the endpoint, baseline may be:

The last nonmissing assessment before first dose
The last assessment before randomization
The average of repeated pretreatment assessments
A protocol-defined baseline measurement

The exact definition should follow the SAP.

Common mistake: Using "baseline" as though it automatically means the first measurement in the database. In clinical-trial analysis, baseline is a defined analytical concept.

Analysis Windows

Visit-based analyses may use predefined windows around nominal assessment times.

For example, a Week 12 assessment window might include observations collected within a prespecified range around the target visit.

The CSR should summarize the rule when it materially affects which observation is selected.

Derived Variables

Many statistical analyses depend on derived variables.

Examples include:

Change from baseline
Percentage change
Responder status
Best overall response
Time to event
Censoring indicators
Exposure duration

The statistical section does not need to reproduce every programming derivation, but key derivations should be understandable.

Example: Change From Baseline

Suppose baseline is 100 and the Week 12 measurement is 72.

The change is:

\[ CHG=72-100=-28 \]

The percentage change is:

PCHG= \frac{72-100}{100}\times100 = -28\%

The CSR should distinguish between these two quantities.

Software Used for Statistical Analysis

The statistical section may identify the statistical software used for the analyses.

For example:

"Statistical analyses were performed using SAS® software, Version 9.4 or later."

If R, Python, or other software was used for specific analyses or figures, the reporting should accurately reflect the actual validated analysis environment.

Software versioning should be consistent with the study's validated computing environment and reporting standards.

Statistical Programming and Validation

The statistical section generally does not need to describe every programming quality-control step.

However, the broader CSR should provide appropriate assurance that analyses were conducted according to established procedures.

Statistical outputs should be traceable to:

Source data
Analysis datasets
Derived variables
Statistical programs
Validated outputs

Tables, Listings, and Figures

The statistical section should work together with the TLFs rather than duplicating them.

For example, the text might state:

"The primary efficacy analysis is summarized in Table 14.2.1. The adjusted treatment difference was −1.8 percentage points (95% CI −3.0 to −0.6; p=0.004)."

The table then provides the complete statistical presentation.

Don't Repeat the Entire Table in Prose

A common CSR-writing problem is repeating every number from a table in the narrative.

The statistical narrative should focus on:

The primary conclusion
Important treatment effects
Important uncertainty
Clinically relevant patterns
Unexpected findings

The table remains the authoritative detailed presentation.

Example of Weak vs. Strong Reporting

Weak	Stronger
"The data were analyzed statistically."	"The primary endpoint was analyzed using ANCOVA with treatment as a fixed effect and baseline value as a covariate."
"Missing values were handled appropriately."	"The primary analysis used the prespecified MMRM approach without explicit imputation of missing outcome values."
"There was a significant treatment effect."	"The adjusted treatment difference was −2.1 units (95% CI −3.4 to −0.8; p=0.002)."
"Subgroup analyses were performed."	"Prespecified subgroup analyses were conducted by age group, sex, baseline disease severity, and geographic region."

How to Write the Statistical Methods Paragraph

A useful statistical methods paragraph often follows this sequence:

Identify the analysis population.

Identify the endpoint.

Identify the statistical model or test.

Identify treatment effects and important covariates.

Identify confidence intervals and hypothesis-testing conventions.

Describe missing-data and sensitivity-analysis methods where relevant.

Example Statistical Methods Section

Consider a randomized, double-blind, placebo-controlled trial with a continuous primary endpoint measured at Week 24.

Example CSR language

"The primary efficacy analysis was performed in the full analysis set according to randomized treatment assignment. The primary endpoint was change from baseline in the disease activity score at Week 24. The primary endpoint was analyzed using an analysis of covariance model including treatment as a fixed effect and baseline disease activity score as a covariate. The adjusted treatment difference between the active-treatment and placebo groups was estimated together with its two-sided 95% confidence interval and corresponding p-value. The statistical hypothesis test was performed at the prespecified two-sided significance level. Missing post-baseline observations were handled according to the prespecified primary analysis method described in the SAP. Sensitivity analyses were conducted to evaluate the robustness of the primary result to alternative assumptions regarding missing data."

Notice that this paragraph answers several questions without reproducing the entire SAP.

Reporting a Primary Result

A strong result statement usually includes:

The treatment groups
The estimated effect
The confidence interval
The p-value when relevant
The direction of the effect

For example:

"At Week 24, the adjusted mean change from baseline was −12.4 in the active-treatment group and −9.7 in the placebo group, corresponding to an adjusted treatment difference of −2.7 (95% CI −4.1 to −1.3; p<0.001)."

Do Not Report Only the P-Value

Compare:

"Treatment was statistically significant, \(p=0.003\)."

with:

"The adjusted treatment difference was −3.2 units (95% CI −5.2 to −1.2; p=0.003)."

The second statement communicates both the estimated effect and its uncertainty.

Reporting Non-Significant Results

A non-significant result should not automatically be described as evidence of no treatment effect.

For example, avoid:

"There was no treatment effect, \(p=0.21\)."

unless the study design and inferential framework genuinely support such a conclusion.

Prefer:

"The estimated treatment difference was 1.4 units (95% CI −0.8 to 3.6; p=0.21)."

The confidence interval allows the reader to evaluate the range of effects compatible with the data.

Reporting Subgroup Results

Subgroup results should be presented carefully.

A useful narrative might state:

"Treatment effects were generally consistent across the prespecified subgroups of age, sex, baseline disease severity, and geographic region. There was no strong evidence of treatment-by-subgroup interaction."

The exact wording should reflect the observed estimates and interaction analyses.

Reporting Sensitivity Analyses

A sensitivity-analysis narrative should answer:

What assumption was changed?
What alternative method was used?
Did the conclusion change?

For example:

"The results of the sensitivity analyses using alternative assumptions for missing Week 24 outcomes were consistent with the primary analysis, with the estimated treatment effect remaining in the same direction and of similar magnitude."

When Sensitivity Analyses Disagree

If a sensitivity analysis materially changes the conclusion, do not hide the difference.

For example:

"The primary analysis demonstrated a statistically significant treatment effect; however, the sensitivity analysis under the specified conservative missing-data assumption produced a smaller treatment effect and did not meet the prespecified significance criterion."

This is more transparent and scientifically useful than reporting only the primary analysis.

Statistical Assumptions

Some statistical methods rely on assumptions that may affect interpretation.

Examples include:

Normality of residuals
Homogeneity of variance
Independence
Missing-at-random assumptions
Proportional hazards
Linearity
Correct model specification

The statistical section should mention important assumptions when they are material to the analysis.

Proportional Hazards

For a Cox model, one important assumption is proportional hazards.

Conceptually, this means that the hazard ratio between treatment groups is stable over time under the model.

If the proportional-hazards assumption is questionable, alternative analyses may be appropriate.

The CSR should accurately describe any diagnostics or alternative methods that were prespecified or used.

Multiplicity and Confidence Intervals

If multiplicity adjustments are applied to hypothesis tests, consider whether the corresponding confidence intervals are also adjusted or otherwise aligned with the inferential framework.

The statistical section should avoid presenting adjusted p-values alongside unadjusted confidence intervals without explaining the distinction when it could confuse interpretation.

Statistical Analysis of Safety Data

Safety analyses are often primarily descriptive.

The statistical section may describe:

Patient incidence
Event counts
Exposure-adjusted incidence
Severity
Relationship to treatment
Seriousness
Deaths
Laboratory abnormalities
Vital-sign abnormalities
ECG abnormalities

Formal hypothesis testing of safety endpoints should be described when it is part of the prespecified analysis.

Laboratory Data

Laboratory analyses may include:

Observed values
Change from baseline
Shift tables
CTCAE grading
Potentially clinically significant abnormalities

The statistical section should identify the conventions used for clinically important laboratory analyses.

Analysis of Vital Signs and ECGs

Similar principles apply to:

Blood pressure
Heart rate
Temperature
Respiratory rate
QT interval
Other ECG parameters

If derived measures such as corrected QT intervals are used, the relevant method should be identified.

Statistical Section and Clinical Interpretation

The statistical section should describe methods rather than overstate clinical conclusions.

For example, it is better to write:

"The adjusted mean treatment difference favored the active-treatment group."

than:

"The treatment was definitively better."

The latter is a clinical conclusion that may require consideration of magnitude, uncertainty, multiplicity, safety, and other evidence.

What Belongs in the Statistical Section vs. Results?

Statistical Section	Results Section
Which analysis population was used	How many patients were analyzed
Which model was used	Estimated treatment effect
How missing data were handled	Observed missingness and its consequences
Multiplicity strategy	Adjusted p-values or conclusions
Subgroup analysis methodology	Observed subgroup treatment effects
Censoring rules	Observed median survival and hazard ratio

In practice, the boundary is not always absolute, but this distinction is a useful writing framework.

Statistical Section Writing Workflow

Read the final protocol and current SAP.

Confirm the final analysis populations.

Identify the primary and key secondary estimands and endpoints.

List the statistical models and treatment contrasts.

Document missing-data and censoring conventions.

Document multiplicity and interim-analysis methods.

Identify subgroup and sensitivity analyses.

Confirm actual analyses against the SAP.

Draft the statistical methods in logical order.

Perform statistical and medical-writing review.

Start With a Statistical Inventory

Before writing prose, create an inventory of the analyses.

Analysis	Population	Endpoint	Method	Effect Measure
Primary efficacy	FAS	Week 24 change	ANCOVA	Adjusted mean difference
Secondary efficacy	FAS	Responder rate	Logistic regression	Odds ratio
Survival	ITT	PFS	Kaplan-Meier/Cox	Hazard ratio
Safety	Safety set	TEAEs	Descriptive	Patient incidence

This inventory becomes the backbone of the statistical section.

Traceability to the SAP

Every major statistical statement in the CSR should be traceable to the applicable SAP or other approved analysis documentation.

Useful traceability checks include:

Endpoint definition matches the SAP.
Analysis population matches the SAP.
Model specification matches the SAP.
Covariates match the SAP.
Stratification factors match the SAP.
Multiplicity strategy matches the SAP.
Missing-data strategy matches the SAP.
Subgroup analyses match the SAP.
Confidence-interval conventions match the SAP.

Best practice: If the CSR describes an analysis differently from the SAP, determine whether the difference is merely editorial or represents an actual methodological change. Methodological differences should be documented appropriately.

Changes From the SAP

Sometimes the final analysis differs from the prespecified plan.

Examples include:

A different covariance structure
A revised analysis population
A changed censoring rule
A different statistical model
An additional sensitivity analysis
A changed endpoint derivation

Such changes should not be silently incorporated into the CSR.

The statistical reporting should identify important deviations and explain their rationale where required.

Example of a Statistical-Method Change

Example language

"The SAP specified an analysis of covariance model using the baseline value as a covariate. Following review of the final analysis dataset, the prespecified model was retained for the primary analysis. An additional sensitivity analysis using a mixed model for repeated measures was conducted to evaluate the robustness of the treatment effect."

The exact wording should reflect the actual circumstances and documentation.

Blinded vs. Unblinded Information

Statistical writing must respect the study's blinding procedures.

Before database lock, analysts may have access to blinded information, while certain personnel may have access to treatment-unblinded results only under defined procedures.

The CSR should accurately describe the final analysis process without exposing information that was not appropriately available during study conduct.

Database Lock and Final Analysis

The statistical section should make clear which data snapshot supports the final results.

The distinction between:

Data cutoff
Database lock
Analysis dataset creation
Final TLF generation

can be important, particularly when follow-up data are updated after an initial analysis.

Statistical Section Quality Control

The statistical section should undergo a dedicated statistical QC review.

At minimum, check:

All endpoint names
All analysis populations
All model names
All covariates
All stratification factors
All alpha levels
All multiplicity methods
All confidence-interval levels
All missing-data methods
All censoring rules
All subgroup definitions
All sensitivity analyses
All software/version statements

Cross-Check the CSR Against the TLFs

The statistical narrative should agree with the actual output.

For example, if the statistical section says:

"The adjusted treatment difference was estimated using ANCOVA."

the corresponding efficacy table should contain the treatment contrast generated by that model.

If the text states that a confidence interval is 95%, the table should not display a 90% interval unless the difference is intentional and explained.

Common CSR Statistical Writing Errors

Copying the SAP verbatim. The CSR should summarize the analysis rather than reproduce every programming instruction.
Being too vague. "The data were analyzed using appropriate methods" tells the reader almost nothing.
Reporting only p-values. Effect estimates and confidence intervals are usually more informative.
Ignoring analysis populations. Readers need to know who contributed to each analysis.
Calling post hoc analyses prespecified. The distinction should remain transparent.
Failing to describe important deviations from the SAP. Material methodological changes should not disappear from the record.
Confusing statistical significance with clinical relevance. A small p-value does not automatically establish a clinically important benefit.
Confusing subgroup significance with interaction. A significant result in one subgroup and a nonsignificant result in another does not prove heterogeneity.
Using inconsistent terminology. "ITT," "FAS," "full analysis set," and other population names should be used consistently.
Failing to describe missing-data methods. The reader needs to understand how incomplete observations entered the analysis.
Overstating non-significant results. A non-significant p-value does not necessarily prove equivalence or absence of an effect.
Confusing hazard ratios and median survival. These are different effect measures.

Common Statistical Language Problems

Avoid	Prefer
"There was a significant improvement."	"The adjusted treatment difference was X (95% CI X to X; p=X)."
"No difference was observed."	"The estimated treatment difference was X, with a 95% CI of X to X."
"The treatment was better."	"The estimated treatment effect favored the active-treatment group."
"The subgroup analysis was positive."	"The estimated treatment effect in the subgroup was X."
"Missing data were handled appropriately."	Describe the actual missing-data method.
"The Cox model showed improvement."	"The estimated hazard ratio was X (95% CI X to X)."

How Detailed Should the Statistical Section Be?

The correct level of detail depends on the complexity of the study.

A small Phase 1 study may require a relatively concise statistical description.

A large Phase 3 trial with multiple estimands, hierarchical testing, time-to-event endpoints, interim analyses, and extensive sensitivity analyses requires considerably more detail.

A useful test is:

Could a qualified reader understand exactly what statistical approach produced the major results without opening the programming code?

If not, the section may be too vague.

What Should Not Be Included?

The statistical section should not become a dumping ground for every technical detail from the analysis process.

Avoid unnecessary inclusion of:

Programming syntax
Dataset variable names unless useful for traceability
Internal programmer comments
Temporary debugging decisions
Every intermediate calculation
Implementation details that do not affect interpretation

The goal is statistical transparency, not code documentation.

Statistical Section Template

The following structure works well as a starting point for many CSRs.

General statistical considerations
Statistical principles, significance level, confidence intervals, and general conventions.

Analysis populations
Definitions and use of each analysis set.

Demographic and baseline analyses
Descriptive summaries and relevant comparisons.

Efficacy analyses
Primary, secondary, and exploratory endpoints.

Safety analyses
Adverse events, laboratory data, vital signs, ECGs, and other safety endpoints.

Missing data
Imputation, model-based handling, censoring, and sensitivity analyses.

Multiplicity
Adjustment procedures, hierarchical testing, and alpha allocation.

Subgroup analyses
Prespecified factors and interaction analyses.

Sensitivity analyses
Robustness analyses supporting the primary conclusions.

Interim analyses and deviations
Important interim procedures and departures from the planned analysis.

Example: Complete Statistical Overview

The following is an illustrative example of how the opening of a statistical section might read.

Illustrative CSR language

"The statistical analyses were performed according to the prespecified Statistical Analysis Plan. Unless otherwise specified, efficacy analyses were conducted according to randomized treatment assignment in the full analysis set, while safety analyses were conducted in the safety population according to treatment received. Continuous variables were summarized using the number of observations, mean, standard deviation, median, minimum, and maximum. Categorical variables were summarized using counts and percentages. The primary efficacy endpoint was analyzed using the prespecified statistical model, with treatment as the principal comparison and adjustment for the specified baseline covariate and stratification factors. The treatment effect was summarized using the prespecified effect measure together with a two-sided 95% confidence interval and corresponding p-value. Secondary endpoints were analyzed using the methods specified in the SAP. Multiplicity was controlled using the prespecified testing strategy. Missing data were handled according to the primary analysis method specified in the SAP, with sensitivity analyses performed to assess the robustness of the primary conclusion. Prespecified subgroup analyses were conducted for the demographic and baseline clinical factors identified in the SAP. Additional exploratory analyses are identified as such in the CSR."

Writing the Statistical Section From a Blank Page

If you are staring at a blank CSR document, do not start by writing prose.

Instead, create five working lists.

List	Questions
Endpoints	What are the primary, secondary, and exploratory endpoints?
Populations	Which patients are included in each analysis?
Methods	Which model or test is used for each endpoint?
Exceptions	Were there any deviations from the SAP?
Interpretation	Which statistical features are important for understanding the results?

Once these lists are complete, the prose becomes much easier to construct.

A Practical Statistical Writing Matrix

For complex studies, a statistical writing matrix can be particularly useful.

Endpoint	Population	Method	Effect	CI	P-Value	Multiplicity
Primary	FAS	ANCOVA	Mean difference	95%	Yes	Yes
Secondary 1	FAS	Logistic regression	Odds ratio	95%	Yes	Hierarchical
Secondary 2	ITT	Cox model	Hazard ratio	95%	Yes	Hierarchical
Exploratory	Biomarker set	Regression	Regression coefficient	95%	Exploratory	No

This matrix can also be used as a QC checklist.

Statistical Section and Tables

Every major statistical method should have an obvious corresponding output.

For example:

ANCOVA → adjusted treatment-difference table
Kaplan-Meier → survival curve and summary table
Cox model → hazard-ratio table
Responder analysis → response-rate table
MMRM → longitudinal estimated means and treatment contrasts
Subgroup analysis → forest plot

This one-to-one relationship helps reviewers trace the statistical narrative through the CSR.

Statistical Section and Figures

The statistical section should also provide enough information to understand important figures.

For example, if a CSR contains a Kaplan-Meier plot, the statistical methods should identify the event definition and censoring convention.

If the CSR contains a forest plot, the statistical section should identify the subgroup-analysis method and treatment-effect measure.

If the CSR contains a waterfall plot, the analysis population and tumor-response derivation should be clear.

When a Statistical Section Becomes Too Long

A common problem is attempting to place every detail from the SAP into the CSR.

When this happens, the section can become difficult to read.

A useful hierarchy is:

Essential: methods needed to understand the major analyses.

Important: assumptions, missing-data methods, multiplicity, and sensitivity analyses affecting interpretation.

Supporting: technical details already fully documented in the SAP.

The CSR should emphasize A and B while avoiding unnecessary duplication of C.

When a Statistical Section Is Too Short

The opposite problem is also common.

Warning signs include statements such as:

"Standard statistical methods were used."
"Appropriate methods were used for the data."
"Missing data were handled appropriately."
"Subgroup analyses were performed."
"Sensitivity analyses supported the primary analysis."

These statements may be true but do not provide enough information.

Replace generic language with specific methodological descriptions.

Reviewer Questions Your Statistical Section Should Answer

A good statistical section should allow a reviewer to answer:

What was the primary estimand?
What was the primary endpoint?
Who was included?
How was baseline defined?
What statistical model was used?
Which covariates were included?
How was treatment compared?
How was missing data handled?
How was multiplicity controlled?
Were interim analyses performed?
Were subgroup analyses prespecified?
Which analyses were exploratory?
Were there deviations from the SAP?
How should the reported confidence intervals be interpreted?

Final Statistical QC Checklist

Confirm the protocol version and SAP version used for the analysis.

Confirm every analysis population definition.

Confirm every endpoint definition.

Confirm baseline and analysis-visit conventions.

Confirm statistical models and covariates.

Confirm treatment contrasts and effect measures.

Confirm confidence-interval levels.

Confirm significance levels and multiplicity procedures.

Confirm missing-data and censoring rules.

Confirm subgroup and sensitivity analyses.

Confirm interim-analysis procedures where applicable.

Confirm all methodological deviations are appropriately documented.

Cross-check statistical language against the actual TLFs.

Perform an independent statistical review before finalization.

Final Checklist for the Statistical Writer

Question	Check
Does the section identify the major analysis populations?	✓
Are primary and secondary endpoints clearly defined?	✓
Are the major estimands and intercurrent-event strategies understood?	✓
Are the statistical models clearly identified?	✓
Are treatment effects and confidence intervals described?	✓
Are missing-data methods explained?	✓
Is multiplicity addressed?	✓
Are subgroup analyses appropriately characterized?	✓
Are sensitivity analyses identified?	✓
Are post hoc analyses clearly labeled?	✓
Are important deviations from the SAP transparent?	✓
Does the text agree with the actual TLFs?	✓

The Most Important Writing Principle

The statistical section should tell the reader enough to understand how the reported evidence was generated.

It should not be a copy of the SAP.

It should not be a list of statistical buzzwords.

And it should not simply say that "appropriate statistical methods were used."

Instead, it should establish a clear chain:

\text{Clinical Question} \rightarrow \text{Estimand} \rightarrow \text{Endpoint} \rightarrow \text{Population} \rightarrow \text{Statistical Method} \rightarrow \text{Treatment Effect} \rightarrow \text{Uncertainty}

When that chain is clear, the statistical section becomes much easier to write and much easier for clinicians, reviewers, regulators, and other statisticians to understand.

Bottom Line

Bottom line: The statistical section of a Clinical Study Report should provide a concise, accurate, and reproducible description of how the study data were analyzed. It should identify the relevant analysis populations, endpoints and estimands, statistical models, treatment-effect measures, confidence intervals, hypothesis-testing procedures, missing-data methods, multiplicity strategy, subgroup analyses, sensitivity analyses, interim analyses, and important deviations from the prespecified plan. The strongest statistical sections do not merely name statistical tests. They explain the connection between the clinical question and the statistical method used to answer it. They also report treatment effects with appropriate measures of uncertainty rather than relying exclusively on p-values. Most importantly, the statistical section should remain consistent with the protocol, SAP, analysis datasets, statistical programs, and final tables, listings, and figures. A reader should be able to follow the analytical path from the study objective to the final reported treatment effect without having to reconstruct the methodology independently.

References

International Council for Harmonisation (ICH). (1995). E3: Structure and Content of Clinical Study Reports.

International Council for Harmonisation (ICH). (1998). E9: Statistical Principles for Clinical Trials.

International Council for Harmonisation (ICH). (2019). E9(R1): Addendum on Estimands and Sensitivity Analysis in Clinical Trials to the Guideline on Statistical Principles for Clinical Trials.

International Council for Harmonisation (ICH). (2022). E8(R1): General Considerations for Clinical Studies.

European Medicines Agency. Guideline on Missing Data in Confirmatory Clinical Trials.

U.S. Food and Drug Administration. Guidance for Industry: Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products.

U.S. Food and Drug Administration. Multiple Endpoints in Clinical Trials: Guidance for Industry.

Little, R.J.A., D'Agostino, R., Cohen, M.L., et al. (2012). The Prevention and Treatment of Missing Data in Clinical Trials. New England Journal of Medicine, 367, 1355–1360.

Rubin, D.B. (1987). Multiple Imputation for Nonresponse in Surveys. Wiley.

ICH E9 Working Group. Estimands and Sensitivity Analysis in Clinical Trials.

Writing the Statistical Section for a Clinical Study Report

What You'll Learn

Introduction

The Statistical Section Is Not the SAP

Where the Statistical Section Fits in the CSR

What Makes a Good Statistical Section?

The Core Components

Start With the Statistical Objectives

Primary and Secondary Endpoints

Define the Endpoint Precisely

Analysis Populations

Why Population Definitions Matter

Handling Treatment Assignment

Estimands and the Statistical Section

Intercurrent Events

Example: Treatment Discontinuation

Choosing the Statistical Method

Do Not Just Name the Test

Continuous Endpoints

ANCOVA Example

Why Confidence Intervals Matter

Binary Endpoints

Odds Ratios

Time-to-Event Endpoints

Kaplan-Meier Analysis

Median Survival

Cox Proportional Hazards Model

Stratified Analyses

Repeated Measures and MMRM

Missing Data

Missing Data Should Not Be Described Vaguely

Last Observation Carried Forward

Multiple Imputation

Sensitivity Analyses

Multiplicity

Family-Wise Type I Error

Common Multiplicity Strategies

Interim Analyses

Subgroup Analyses

Interaction Tests

Protocol Deviations

Baseline Characteristics

Descriptive Statistics

Reporting Percentages

Rounding Rules

P-Values

Statistical Significance vs. Clinical Relevance

Effect Sizes

Confidence Intervals

Noninferiority Trials

Superiority After Noninferiority

Bayesian Analyses

Transformation of Data

Geometric Means

Analysis of Adverse Events

Multiple Occurrences of the Same Adverse Event

Exposure-Adjusted Analyses

Pharmacokinetic and Pharmacodynamic Analyses

Multiplicity Across the CSR

Post Hoc Analyses

Analysis Cutoff Dates

Handling Unscheduled Assessments

Baseline Definition

Analysis Windows

Derived Variables

Example: Change From Baseline

Software Used for Statistical Analysis

Statistical Programming and Validation

Tables, Listings, and Figures

Don't Repeat the Entire Table in Prose

Example of Weak vs. Strong Reporting

How to Write the Statistical Methods Paragraph

Example Statistical Methods Section

Reporting a Primary Result

Do Not Report Only the P-Value

Reporting Non-Significant Results

Reporting Subgroup Results

Reporting Sensitivity Analyses

When Sensitivity Analyses Disagree

Statistical Assumptions