Building Tables for Baseline Characteristics and Balance

What a Baseline Table Should Communicate

A baseline characteristics table lets readers see who was compared and how the treatment groups differed before and after adjustment. In an observational study, it is especially important to separate description from evidence of balance: the table describes observed characteristics, while balance measures summarize the size of observed differences. Neither a table nor a balance measure establishes that unmeasured confounding is absent.

A useful table makes its groups, samples, summaries, and denominators explicit. It should distinguish the original treatment groups from the matched or weighted analysis sample, and state which adjustment method produced each post-adjustment column. For a matched sample, report the number of treated and comparison units retained. For weighting, report the original unit counts and explain that the displayed summaries use weights; if relevant, also report a measure of the effective sample size.

The baseline time point and the analysis population need clear definitions. A variable measured after treatment begins may be affected by treatment, so presenting it as a baseline covariate can be misleading. Explain which participants are included in each column, including relevant exclusions or missing-data rules. The earlier material on missing-data reporting and analysis samples provides the broader reporting context; here the goal is to make those facts visible in the table itself.

Choose Summaries That Fit the Variable

For a roughly symmetric continuous variable, report the mean and standard deviation with units, such as age in years. For a strongly skewed variable, a median and interquartile range may communicate its distribution more faithfully. State which summaries are used and apply the choice consistently across the treatment groups. For a binary variable, report the count and percentage in a clearly identified category, such as “current smoker, n (%).” For a categorical variable with several levels, show each level, usually as its own row, with counts and percentages.

Always identify the denominator behind a percentage. If some values are missing, a percentage among observed values answers a different question from a percentage among everyone in that group. A compact table can include a separate missing row and a note that percentages for the categories use nonmissing observations. Where missingness is substantial or differs across groups, make the missing counts easy to find rather than hiding them in a footnote.

The same rule applies after adjustment: calculate each summary from the adjusted analysis sample and use the relevant weights when the design is weighted. A weighted percentage is the sum of weights for units in the category divided by the sum of weights for units with observed values for that characteristic. Thus, the displayed denominator may be the sum of weights, not the number of people. Label the column “weighted” and explain the denominator or convention in a note.

Show Pre-Adjustment and Post-Adjustment Results Separately

A practical layout places each treatment group in a separate column, with a clear grouping of columns for “Before adjustment” and “After adjustment.” Include the sample size in each group header. For example, a matched table might show the treated and comparison counts retained after matching; a weighted table might show both the number of observed units and a label indicating that post-adjustment summaries are weighted. Avoid a single ambiguous “adjusted” column when readers cannot tell which group or method it represents.

Before adjustment, summaries use the observed treatment groups in the specified starting cohort. After adjustment, summaries correspond to the sample and weights actually used for the analysis. In matching with replacement, for instance, a comparison person can contribute more than once to the matched analysis. The table note should clarify whether counts refer to unique people or matched uses, and the summaries must follow the method used to form the analysis sample. For weighting, the table should not imply that a weighted count is a literal count of people.

For continuous variables, report a mean (SD) or median (IQR) in every relevant column. For categorical variables, report category-specific counts and percentages. If a missing category is shown, clarify whether it is included in the percentage denominator for other categories. Use one consistent number of decimal places, include measurement units in row labels, and define abbreviations and any non-obvious summary conventions in a table note.

Balance Measures Belong Alongside the Descriptions

A baseline table can pair group summaries with an absolute standardized mean difference (SMD) for each covariate, before and after adjustment. The earlier material on covariate balance diagnostics covers the interpretation and calculation of balance measures. For reporting, the key point is to make the comparison scale clear: an SMD is a summary of the magnitude of an observed difference, while a count, percentage, mean, or median describes the covariate itself. Report the chosen balance measure consistently and explain how it was calculated if the convention is not standard for the displayed variable.

Do not use hypothesis-test p-values as a measure of balance. A p-value depends on sample size as well as the observed difference: a small difference may be statistically significant in a very large sample, while a potentially important difference may not be significant in a small one. A non-significant result does not demonstrate that groups are balanced. For this reason, omit routine significance tests from baseline tables and use descriptive summaries plus appropriate balance measures instead.

A short justification clarifies why standardized differences are more useful for this purpose than p-values. A significance test evaluates a null hypothesis about a population difference relative to its sampling uncertainty. Its test statistic therefore changes with sample size, even if the estimated difference stays fixed. A balance summary intended to describe the observed discrepancy should foreground the magnitude of that discrepancy, not whether a sample-size-dependent test crosses a significance threshold. This does not make an SMD a test of causal validity or a guarantee of balance on unmeasured variables.

Worked Table Design and Common Failure Modes

Several table choices can undermine an otherwise sound analysis. Combining original and matched sample summaries under one unlabeled pair of columns hides which population is being described. Showing weighted percentages without labeling them as weighted obscures the calculation. Omitting missing counts can make groups with different data completeness appear directly comparable when they are not. Finally, reporting a p-value next to every characteristic invites readers to treat “not significant” as proof of balance. The correction in each case is to identify the sample, method, denominator, and descriptive measure explicitly.

Tables also have limits. They summarize selected measured covariates and can conceal distributional differences when a single mean or proportion is used. They cannot replace examination of the full covariate distributions, the study’s overlap and missingness assessments, or appropriate balance diagnostics. If a summary is likely to hide important structure, add a suitable diagnostic elsewhere in the report rather than making the table excessively dense.