Practical Guide to Medical Statistics and Data Processing

Many clinicians freeze at the word "statistics": they can't make sense of formulas, don't know how to use the software, and struggle to interpret results. As a result, they either outsource the analysis or copy methods from others' papers—often applying the wrong approach by mistake. In clinical research, however, most data only require a handful of standard methods. This article skips the formulas and derivations. Instead, it focuses on solving three key issues:When to use, how to operate, and how to interpret results。

Before We Begin: You're Not Bad at Math—You Just Never Got the "User Manual"

(1) First, break through three mental barriers

First, statistics is a "multiple-choice question," not a "calculation problem."You don't need to calculate any statistics by hand; the software does that for you. Your real task is to select the correct method from the menu based on your data type and research design.

Second, you only need to know how to "select" and "read."Choosing the right method and understanding key metrics in the output is enough to support most clinical papers.

Third, and most importantly: statistical significance does not equate to clinical meaningfulness.A P<0.05 result may show a difference too small to be clinically meaningful; conversely, P>0.05 may simply reflect insufficient sample size. This principle applies throughout the document—keep it in mind.

(II) The "Three-Question Method" – Always Ask These Three Questions First

Don't launch the software immediately after getting the data. Answer the following three questions in order, and the solution will become clear:

• Question 1: What type of outcome data do I have? Is it continuous (e.g., blood pressure), categorical counts (e.g., disease presence/absence), ordinal ranks (e.g., efficacy grading), or survival data with follow-up times?
• Question 2: How many groups were there? Were they independent or paired? Two groups or more than two? Different sets of people, or the same group before and after treatment?
• Question 3: Do the data meet the assumptions for parametric tests? Check primarily for approximate normality and homogeneity of variances across groups. If assumptions are met, use parametric tests (e.g., t-test, ANOVA); otherwise, use non-parametric tests (e.g., rank-sum test).

II. One Table to Choose the Right Method: A Guide to Selecting Common Medical Statistical Methods

(1) Identify your four data types (30 seconds)

The first step in selecting a method is to determine the data type. Clinical data ultimately falls into four categories:

• Continuous variables (quantitative data): Measurable values with units, such as height, weight, blood pressure, blood glucose, length of hospital stay, and laboratory test results.
• Categorical (nominal) data: Count the number of observations in each category. Categories have no inherent order, such as gender, vital status, presence or absence of complications, and blood type.
• Ordinal Data (Ordered Categories): Categories have a defined order, but the intervals between adjacent levels are not equal. Examples include treatment efficacy (cured/improved/effective/ineffective), pathological grading, cardiac function classification, and TNM staging.
• Survival data: Includes both "follow-up time" and "event occurrence," often containing censored records where the event has not occurred by the end of follow-up.

(II) Core: Comprehensive Table for Selecting Statistical Methods in Medical Research (Most Important Table in This Paper)

The table below lists the preferred methods and their alternatives when conditions are not met, organized by "Data Type × Analysis Purpose × Number of Groups." Simply refer to the table when writing your paper; memorization is unnecessary.

Table 1: Overview of Common Statistical Methods in Medicine

(3) Five Methods Explained: When to Use, Prerequisites, and How to Interpret Results

1. t-test — specifically for comparing "means"

When to use:For measurement data involving only two groups (or one group compared to a known standard value) with approximately normal distribution, there are three types of t-tests. Do not confuse them:

• One-sample t-test: Compares the mean of a single sample to a known standard value (e.g., whether the average blood pressure of patients in a department exceeds the national reference value).
• Independent Samples t-test: Compare the means of two independent groups, such as blood pressure in a new drug group versus a control group.
• Paired t-test: Compares the same group before and after treatment, or two matched individuals, such as a patient's weight before and 8 weeks after medication.

Prerequisites:Data should be approximately normally distributed. The independent samples t-test also assumes equal variances; software automatically checks this using Levene's test.

How to read the results:First, check the significance (Sig) of Levene's test: if P > 0.05, variances are equal—refer to the "Equal variances assumed" row; if P < 0.05, variances are unequal—use the "Equal variances not assumed" row (i.e., Welch's correction). Next, examine the t-value and two-tailed Sig: if P < 0.05, the difference between the two means is statistically significant.Note: In addition to reporting P, also report the two group means, the mean difference, and its 95% confidence interval.

2. Chi-Square Test — Specifically for Comparing "Rates" and "Proportions"

When to use:Categorical data: tests whether two categorical variables are independent, i.e., whether the rates or proportions differ across groups (e.g., whether complication rates differ between two groups).

Conditions for using a 2x2 contingency table (must memorize):

• Total sample size n ≥ 40, and expected frequency T ≥ 5 in each cell: use standard Pearson chi-square.
• If n ≥ 40 but some cells satisfy 1 ≤ T < 5: apply continuity-corrected chi-square test.
• If n<40 or any cell T<1, use Fisher's exact test.

Paired design:For the same set of samples tested with two methods to determine positive/negative results, use McNemar's test instead of the standard chi-square test.

How to read the results:Check the two-sided significance of Pearson's chi-square. Report the actual rates, the rate difference between groups, and the odds ratio (OR) with its 95% confidence interval (CI). Footnotes below the SPSS output table indicate how many cells have expected counts less than 5; use this to decide whether to switch to Fisher's exact test.

Special Notice:Do not use standard chi-square tests for ordinal categorical data. Chi-square only examines frequency distributions across categories, treating "cured," "improved," and "not improved" as equivalent and discarding valuable order information. For ordinal data, use rank sum tests or ordered logistic regression instead.

3. Analysis of Variance (ANOVA) — Comparing means across three or more groups

When to use:Continuous data, three or more groups, approximately normal distribution, and homogeneity of variance.

Why repeated t-tests are problematic:Each additional t-test increases the accumulation of false positives (Type I error: incorrectly detecting a difference when none exists). For example, with three groups requiring 3 pairwise comparisons, the overall false positive rate significantly exceeds 0.05; the more groups involved, the more severe the inflation. This is one of the most common statistical errors in Chinese medical journals.

Correct process (two steps):First, perform an overall ANOVA (F-test). If P < 0.05, it only indicates that differences exist among the groups overall, but does not identify which specific pairs differ. Next, conduct post hoc pairwise comparisons: if variances are homogeneous, choose LSD, Tukey, SNK, or Bonferroni; if variances are heterogeneous, use Tamhane T2 or Games-Howell.

Repeated Measures:When the same subject is measured at multiple time points, it constitutes a repeated-measures design and requires repeated-measures ANOVA. Splitting the data into multiple independent samples t-tests is inappropriate.

How to read the results:First, check the F value and overall significance to determine if there are between-group differences. Then, examine post-hoc comparisons to identify which specific pairs of groups show statistically significant differences.

4. Regression Analysis — Identify Influencing Factors and Make Predictions

The core of regression is selecting a model based on the outcome type, then using it to analyze which factors have independent effects or to build predictive models.

• Outcome is a continuous variable: Use multiple linear regression. The regression coefficient β indicates how much the outcome changes, on average, for each 1 unit increase in that independent variable; R² reflects the proportion of outcome variance explained by the model.
• Outcome is a binary variable (e.g., disease vs. no disease, effective vs. ineffective): Use binary logistic regression. The odds ratio (OR) is calculated as the exponent of the coefficient: OR > 1 indicates a risk factor; OR < 1 indicates a protective factor.
• Outcome is a censored time-to-event: analyzed using Cox proportional hazards regression; results reported as hazard ratios (HR).

The value of multi-factor authentication:Including multiple variables in a model simultaneously allows you to isolate the independent effect of each factor by adjusting for confounders—an approach that single-variable analysis cannot achieve.

How to read the results:Focus on each variable's β or OR/HR, significance (Sig), and the 95% confidence interval. For OR/HR, if the 95% CI includes 1, the factor is not statistically significant at the 0.05 level. Also ensure the sample size is adequate (typically requiring sufficient events relative to the number of variables) and check for multicollinearity among variables.

5. Survival Analysis – Handling "Time + Event + Censoring"

Why standard methods fall short:In follow-up studies, some participants do not experience the event of interest (e.g., death or recurrence) by the end of the study. This censored data is incomplete; using a t-test directly can lead to information loss or bias.

The three pillars of survival analysis:

• Kaplan-Meier (KM) Estimator: Used to plot survival curves and estimate survival probabilities at various time points and median survival time.
• Log-rank test: Used to compare whether two or more survival curves differ; it is an unadjusted, single-factor crude comparison.
• Cox Proportional Hazards Regression: Adjusts for multiple confounders simultaneously to estimate hazard ratios (HR) for each factor. An HR > 1 indicates a higher risk of the event; also check whether the 95% CI includes 1.

How to read the results:Report median survival time, P values from log-rank tests, and HRs with 95% CI from multivariate analysis.

3. Zero-Basis Tool Operation: Enter data and let the results generate automatically

(1) Step 1: How to enter data (easiest to overlook, yet most critical)

Statistical software data tables follow one golden rule:One research subject per row, one variable per column.This is called wide format.

• Code grouping variables with numbers (e.g., 1=Treatment Group, 2=Control Group) and define value labels in Variable View to avoid confusion later.
• Independent samples follow a "one group column + one outcome column" structure; paired (pre-post) data uses two columns to record values before and after treatment.

Pay special attention:Excel files you have are often pre-summarized "summary tables" (one row per group with totals), but statistical software requires unsummarized "raw data tables" (one row per patient). Convert the format before entering the data.

(2) Before any analysis, perform a normality test.

This step determines whether to proceed with parametric or non-parametric testing and cannot be skipped.

SPSS Path:Analyze → Descriptive Statistics → Explore; move the continuous variable to be tested into "Dependent List" and the grouping variable into "Factor List"; click "Plots", check "Normality plots with tests", then OK.

How to read the results:For small samples (typically n < 50), rely on the Shapiro-Wilk test's p-value: if p > 0.05, data is approximately normal and parametric tests are appropriate; if p < 0.05, consider non-parametric tests. Also, verify using histograms and Q-Q plots.

Large sample reminder:When sample sizes are large, normality tests become overly sensitive and may report P < 0.05 even for minor deviations. In such cases, rely less on the test result and instead evaluate the histogram, Q-Q plot, skewness, and kurtosis together.

(3) Independent Samples t-Test: 5 Steps to Results

Step 1:Analyze → Compare Means → Independent-Samples T Test.

Step 2:Place outcome variables (e.g., blood pressure) in the "Test Variable" box.

Step 3:Drag the grouping variable (e.g., Group) into the "Grouping Variable" box.

Step 4:Click "Define Group", enter code 1 for group 1 and code 2 for group 2, then click Continue.

Step 5:Click OK, then read the results in the order described earlier: first check Levene's test for homogeneity of variance, then review the corresponding t-value and two-tailed significance (Sig).

(4) Chi-Square Test: Get Results in Four Steps

Step 1:Analyze → Descriptive Statistics → Crosstabs.

Step 2:Place grouping variables in "Rows" and outcome variables in "Columns" (you can reverse them, just ensure your interpretation matches the layout).

Step 3:Click "Statistics" and check "Chi-square"; then click "Cells" and check "Observed" along with either "Row percentages" or "Column percentages" to easily read the actual rates.

Step 4:Click OK, review the Pearson Chi-Square results, and check the footnote regarding expected cell counts. Switch to Fisher's Exact Test if necessary.

(5) One-Way ANOVA: Four Steps to Results

Step 1:Analyze → Compare Means → One-Way ANOVA.

Step 2:Move outcome variables to the "Dependent List" and grouping variables to the "Factor" box.

Step 3:Click "Post Hoc" and select Tukey or LSD for equal variances, Tamhane for unequal variances.

Step 4:Click "Options," check "Description" and "Homogeneity of Variance Tests," then click OK. First, review the homogeneity of variance; then examine the overall F-test and post hoc pairwise comparisons.

(VI) Survival Analysis (Kaplan-Meier) Workflow

SPSS Path:Analysis → Survival Analysis → Kaplan-Meier; place the follow-up time variable in "Time", the event variable in "Status", and click "Define Event" to enter the value representing the event (e.g., 1=Death); place the grouping variable in "Factor". Click "Compare Factor" to select the log-rank test, click "Plots" to choose survival curves, then click OK.

(7) Don't want to install software? These free online tools work just as well.

• Jamovi: Free and open-source with a menu-driven interface. Its workflow closely mirrors SPSS, making it ideal for beginners.
• JASP: Also free, with a menu-driven interface. Supports common tests including t-tests, chi-square, ANOVA, and regression.

Additionally, numerous web-based medical statistics calculators are available for quickly obtaining individual test results after data entry. Please note: while online tools are suitable for preliminary exploratory analysis, formal submissions should rely on reproducible and auditable software. Ensure you properly preserve your raw data, analysis steps, and output logs to facilitate verification by reviewers or peers.

(VIII) Software delivers a lot, but focus on these three essentials

SPSS generates many tables at once. You don't need to understand every one, nor should you copy them directly into your paper. Focus on these three key pieces of information:

• Test Statistics: Report values such as t, χ², and F in the Methods and Results sections of your paper.
• Exact P-value: Report the exact value from the "Sig." column in the output table (e.g., P = 0.006) rather than a threshold (e.g., P < 0.05).
• Effect size and 95% confidence interval: e.g., mean difference, OR, HR, and their 95% CI, to assess the magnitude of differences and their clinical relevance.

Section 4: Standards for Medical Statistical Charts – How to Create and Avoid Errors in Four Chart Types

(I) First, remember the universal laws of charts

• Figures must be self-explanatory: readable without the main text; axes must clearly state "Name / Unit".
• Place figure captions below figures and table captions above tables, numbering them consecutively with Arabic numerals.
• Error bars must be clearly labeled: standard deviation (SD), standard error (SE), and 95% confidence intervals are distinct concepts and must not be left unmarked.
• Avoid 3D effects, flashy color schemes, and unrelated images forced together just to meet image counts.

(II) Bar Chart

Purpose:Display means or rates for each group to facilitate between-group comparisons.

Production GuidelinesThe y-axis should generally start at 0 to avoid visually exaggerating differences between groups; error bars must be labeled as SD, SE, or 95%CI. For continuous data, box plots or scatter plots are preferred over plain bar charts, as they better display the true distribution and variability of the data.

Common Errors:Hiding the original data distribution with bar charts, error bars without labels, and truncated y-axes without explanation.

(3) Survival Curve (Kaplan-Meier Curve)

Essential ElementsY-axis: Survival probability; X-axis: Follow-up time (with unit); Use small vertical bars to indicate censored data. Include a "Number at Risk" table below the figure showing the number of patients still under follow-up at each time point for each group. Report median survival times and log-rank test P-values.

Common Errors:Do not mark censored data or display the number at risk table. A sharp drop in the survival curve during late follow-up often indicates a small remaining sample size and unstable results; interpret such trends cautiously alongside confidence intervals. Also, ensure you do not confuse "cumulative survival rate" with "cumulative event rate."

(IV) Forest Plot

Purpose:Commonly used in meta-analysis and for visually displaying the effects of various factors in multivariate regression.

How to pronounce:The vertical line in the center represents the "line of no effect" (OR/RR/HR = 1 or mean difference = 0). Each square indicates the effect size of a study or variable, with its size reflecting the weight. The horizontal lines extending from each square represent the 95% confidence interval. The diamond at the bottom shows the pooled overall effect.

Production GuidelinesRatio-type effect measures (OR, RR, HR) should be plotted on a logarithmic scale so that 0.5 and 2 are symmetric around 1, while the x-axis labels display the original values. If a confidence interval line crosses the null line, the effect is not statistically significant.

Common Errors:Misusing a linear scale for ratio-based effects causes visual asymmetry. Invalid lines are not drawn, and weights are not labeled.

(5) ROC Curve

Purpose:Evaluate the discriminatory ability of diagnostic tests or prediction models and select the optimal cut-off point accordingly.

Essential Content:The area under the curve (AUC) must be reported. The AUC is interpreted as follows: 0.5 indicates no discrimination; 0.7–0.8, acceptable; 0.8–0.9, good; and >0.9, excellent (these are empirical guidelines, not absolute thresholds). Additionally, report the optimal cutoff value along with its corresponding sensitivity and specificity.

Common Errors (Frequently Occurring in Clinical Studies):

• AUC below 0.5 without action usually indicates the direction was set incorrectly; please verify the test direction.
• When two ROC curves intersect, relying solely on a single AUC value to draw conclusions is insufficient; the overall AUC cannot adequately reflect performance differences.
• When two AUC values are very close, do not compare them directly without statistical testing. For curves obtained from the same group of subjects, use DeLong's test to determine if the difference is statistically significant.
• Do not rely solely on AUC while ignoring model calibration and clinical decision utility. A high AUC does not guarantee the model performs well with real patients; instead, combine it with calibration curves and decision curve analysis.

5. Common Statistical Errors in Theses: Check Your Work Against These Case Studies

(I) Overuse of P-values: The Top Problem in Clinical Research

The American Statistical Association (ASA) has outlined key principles regarding P-values in its official statement. Below, we explain each principle by addressing common clinical misinterpretations.

Error 1: Treating P<0.05> as a binary on/off switch.The ASA explicitly states that scientific conclusions should not depend solely on whether a P-value crosses a specific threshold. The strength of evidence for 0.049 and 0.051 is actually comparable.

Error 2: Misinterpreting the p-value as a measure of effect size or result importance.P-values do not measure effect size. With a sufficiently large sample, even a difference too small to be clinically meaningful can yield a very small P-value.

Error 3: Misinterpretation of the P-value.A p-value is not the probability that the research hypothesis is true, nor is it the probability that the results occurred entirely by chance. It is the probability of observing a result as extreme as, or more extreme than, the one observed, assuming the null hypothesis is true.

Error 4: Specify only P<0.05, P<0.01>; do not provide exact values.Report the exact P value (e.g., P=0.006); if very small, report it as P<0.001.

Error 5: P>0.05This is more likely due to insufficient sample size and low statistical power. The correct phrasing should be "no statistically significant difference," rather than concluding "no effect."

Correct approach:Report the P value along with effect sizes (mean difference, OR, HR, etc.) and the 95% confidence interval. For primary outcomes, decisions should be based primarily on the confidence interval and clinical significance.

(2) Missing Confidence Interval

Without confidence intervals, readers cannot gauge the magnitude of the effect or the precision of the estimate.

How to pronounce confidence interval:Using a 95% CI as an example: if the interval includes null values (a mean difference of 0, or an OR/RR/HR of 1), the difference is not statistically significant at the 0.05 level. A narrower interval indicates more precise estimation, while a wide interval often suggests insufficient sample size and high uncertainty in the results. Reiterating, for primary outcomes, confidence intervals are more important than P-values alone.

(III) Misuse of Statistical Methods: Eight Cautionary Cases

Case 1:Comparing three group means by performing 3 t-tests leads to inflated false-positive rates due to accumulation; use one-way ANOVA with post-hoc correction instead.

Case 2:Using the chi-square test for ordered outcomes like "cured/effective/ineffective" discards the ordinal information. Instead, use the rank-sum test or ordinal logistic regression.

Case 3:Paired data from the same patients before and after treatment were incorrectly analyzed using an independent samples t-test. Instead, a paired t-test or Wilcoxon signed-rank test should be used.

Case 4:Data is clearly skewed and variances are unequal across groups. Do not use t-tests or ANOVA; instead, apply Welch's correction or non-parametric tests.

Case 5:The total sample size in the 2x2 table is less than 40. Use Fisher's exact test instead of the standard chi-square test.

Case 6:Repeated measurements of the same object at multiple time points should not be analyzed as independent t-tests; use repeated measures ANOVA or a mixed-effects model instead.

Case 7:Performing multiple tests without correction for multiple comparisons inflates the false positive rate; apply Bonferroni or False Discovery Rate (FDR) correction.

Case 8:Do not include all variables in a multivariate regression based solely on a single-factor P < 0.05. Variable selection should be guided by domain expertise, collinearity among variables, and sample size—not just P values.

(IV) Pre-submission Self-Checklist (Check off each item)

• Are the data types and study design clearly described in the Methods section?
• Does the selected method match the data type, number of groups, and design?
• Have the assumptions for parametric tests (normality, homogeneity of variance) been checked and explained?
• Were the test statistic, exact P value, effect size, and 95% confidence interval all reported?
• Were multiple comparisons or multiple testing corrections applied?
• Are the charts standardized, are error bars labeled, and are figure/table captions complete?
• Is there a clear distinction between "statistical significance" and "clinical significance"?
• Has the software name and version, significance level α, and whether a one-sided or two-sided test is used been specified?
• Are the original data and full analysis process retained for verification?

VI. Conclusion: Three Key Messages for Clinicians

First, choose right. Then, calculate accurately.Choosing the right method is half the battle for the statistical section—far more efficient than fixing it later.

Second, always report the effect size with a 95% confidence interval alongside the p-value.It helps you distinguish between "statistical significance" and "clinically meaningful differences."

Third, consult a statistician or methodologist early if you are unsure.An hour of communication during the design phase is often worth three months of back-and-forth revisions after submission.

References

[1] Wasserstein R L, Lazar N A. The ASA's statement on p-values: context, process, and purpose[J]. The American Statistician, 2016, 70(2): 129-133.

[2] He Z, Zhang Q, Song M, et al. Four overlooked errors in ROC analysis: how to prevent and avoid[J]. BMJ Evidence-Based Medicine, 2025, 30(3): e113078.

[3] Park S H, Han K, Park S Y. Mistakes to avoid for accurate and transparent reporting of survival analysis in imaging research[J]. Korean Journal of Radiology, 2021, 22(10): 1587-1593.

[4] Wu S, Jin Z, Guo X, et al. Misuse of statistical methods in 10 leading Chinese medical journals in 1998 and 2008[J].. The Scientific World Journal, 2011, 2011: 139494.

[5] Nahm FS. Receiver operating characteristic curve: overview and practical use for clinicians[J]. Korean Journal of Anesthesiology, 2022, 75(1): 25-36.

[6] Cochrane Training. Technical supplement to Chapter III: Considerations and recommendations for figures in Cochrane reviews—graphs of statistical data [EB/OL]. (2024) [2026-11-06]. https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/graph-recommendations.

[7] StatPearls Publishing. Survival analysis [EB/OL]. (2023) [2026-11-06]. https://www.ncbi.nlm.nih.gov/books/NBK560604/.