Missing data and denominators • gtstats

library(gtstats)

The policy in one minute

gtstats does not silently treat missing values as observations with a value. Each analysis retains or reports the observations contributing to its result. The usual default is to use observations with the values needed for that calculation and make the resulting denominator available for review.

There is no universal denominator. A categorical percentage needs observed values of that variable; a correlation needs complete pairs; a rate needs valid event-time records. The right denominator depends on the scientific question.

Start with describe_data() to see completeness for every variable.
Choose the percentage denominator deliberately for descriptive tables.
Use denominators_stats() before reporting results when exclusions matter.
Describe the analysis population when missing data are material.

Key terms

Term	Meaning in gtstats
`n_total`	Rows in the relevant dataset or group before the statistic-specific exclusion
`n_nonmissing`	Rows with the values required for the variable
`n_missing`	Rows excluded because a required value is missing
`level`	Categorical level represented by a displayed cell, when relevant
`numerator`	Count of events or selected levels, where relevant
`denominator`	Eligible observations or total person-time used for the estimate
`rule`	Plain-language record of how the denominator was formed

denominators_stats() returns a formatted table by default, or a plain console tibble with format = "tibble". For categorical summaries, it records the level as well as the group, so row and overall percentages can be checked against the exact denominator used in the displayed cell.

Start by describing missingness

describe_data() reports completeness from all rows in the supplied dataset; it never drops rows to make the overview look complete.

dat <- data.frame(
  arm = factor(c("Control", "Control", "Treatment", "Treatment", "Treatment")),
  age = c(45, NA, 51, 62, 57),
  smoker = factor(c("No", "Yes", NA, "Yes", "No")),
  follow_up = c(12, 10, 8, NA, 11)
)

to_flextable(describe_data(dat))

Variable	Type	Complete	Unique	Overview	Range / levels
arm	binary	5/5 (100.0%)	2	Treatment 3 (60.0%); Control 2 (40.0%)	Control, Treatment
age	categorical	4/5 (80.0%)	4	45 1 (25.0%); 51 1 (25.0%); 57 1 (25.0%)	45, 51, 62, 57
smoker	binary	4/5 (80.0%)	2	No 2 (50.0%); Yes 2 (50.0%)	No, Yes
follow_up	categorical	4/5 (80.0%)	4	10 1 (25.0%); 11 1 (25.0%); 12 1 (25.0%)	12, 10, 8, 11
One row is shown per selected variable.
Potential data-quality findings and interpretation prompts are stored in `$issues`.
Ordered factors are identified as ordinal variables.

This reports the amount of missing data, not its cause. Clinical context, the data dictionary, and the study design are needed to decide whether a complete-case analysis is scientifically appropriate.

Descriptive tables: percentage denominators

For categorical variables, summary_table() makes the denominator explicit. The default is percent = "column".

Setting	Percentage denominator	Typical use
`"column"`	Non-missing observations in each displayed group	Standard Table 1
`"row"`	Non-missing observations for that level across groups	Distribution of a category across groups
`"overall"`	All non-missing observations of the variable	Share of the overall sample
`categorical = "n"`	No percentage	Counts only

summary_table(
  dat,
  by = arm,
  include = smoker,
  percent = "column",
  missing = "ifany"
) |> to_flextable()

Characteristic	Control N = 2	Treatment N = 3
smoker
No	1 (50.0%)	1 (50.0%)
Yes	1 (50.0%)	1 (50.0%)
Missing	0 (0.0%)	1 (33.3%)
Categorical data are n (%).


to_flextable(summary_table(dat, by = arm, include = smoker, percent = "row"))

Characteristic	Control N = 2	Treatment N = 3
smoker
No	1 (50.0%)	1 (50.0%)
Yes	1 (50.0%)	1 (50.0%)
Missing	0 (0.0%)	1 (33.3%)
Categorical data are n (%). Percentages use row denominators.

to_flextable(summary_table(dat, by = arm, include = smoker, percent = "overall"))

Characteristic	Control N = 2	Treatment N = 3
smoker
No	1 (25.0%)	1 (25.0%)
Yes	1 (25.0%)	1 (25.0%)
Missing	0 (0.0%)	1 (33.3%)
Categorical data are n (%). Percentages use overall denominators.

to_flextable(summary_table(dat, by = arm, include = smoker, categorical = "n"))

Characteristic	Control N = 2	Treatment N = 3
smoker
No	1	1
Yes	1	1
Missing	0	1
Categorical data are counts.

missing = "ifany" displays a Missing row only when it is needed. Use "always" to show zero-missing rows or "no" to suppress them. These three settings calculate observed-category percentages from non-missing values; suppressing the row does not change that denominator.

Use missing = "as_category" when missing values should form part of the displayed categorical distribution. The Missing row is then included in the percentage denominator, so the observed categories plus Missing sum to 100%.

catheter_data <- data.frame(
  catheter = factor(
    c(rep("Yes", 32), rep(NA_character_, 68)),
    levels = c("No", "Yes")
  )
)

summary_table(
  catheter_data,
  include = catheter,
  missing = "as_category"
) |> to_flextable()

Characteristic	Overall N = 100
catheter
No	0 (0.0%)
Yes	32 (32.0%)
Missing	68 (68.0%)
Categorical data are n (%).

This reports 32/100 (32%) as documented Yes and 68/100 (68%) as Missing. It does not assume that missing records are No. For continuous variables, missing values cannot become a numeric category and therefore remain a separate missingness row.

Continuous summaries use observed finite values of that variable. A displayed Missing row is calculated against all rows in the relevant group.

Use categorical = "n_over_N_percent" when the table should show the non-missing denominator inside every cell, for example 30/59 (50.8%). For a continuous descriptive estimate with precision but no p-value, use statistic = "mean_ci"; this reports a t confidence interval for the observed mean using conf.level.

Proportions and confidence intervals

proportion_stats() and add_proportion() use non-missing observations of the selected variable as the denominator. The numerator is the chosen level, and the confidence interval describes that observed-data proportion.

smoking <- proportion_stats(dat, smoker, by = arm, level = "Yes")
to_flextable(smoking)

	Control		Treatment
Event	n (%)	95% CI	n (%)	95% CI
Yes	1 (50.0%)	9.5–90.5%	1 (50.0%)	9.5–90.5%
Selected event: smoker = Yes. Estimates use Wilson score 95% confidence intervals.
This function estimates a proportion; it does not test differences between groups.
The denominator must define the population at risk, and observations should be independent.

denominators_stats(smoking)

Variable	Level	Group	Eligible observations	Used in analysis	Missing / excluded	Numerator	Denominator	Rule
Denominator audit
smoker	NA	arm = Control	2	2	0	NA	2	Non-missing observations; event level = Yes
smoker	NA	arm = Treatment	3	2	1	NA	2	Non-missing observations; event level = Yes

If unknown status should be included in a categorical descriptive distribution, use missing = "as_category". Do not assume that an unknown value is a non-event; recode it to No only when that meaning is justified by the data definition.

An all-missing categorical variable remains visible in summary_table() as an em dash, with a Missing row by default. It is not silently dropped from the table.

Rates use complete event-time records

rate_stats() and add_rate() require finite, non-negative event counts and person-time values. Rows missing either value are excluded. The denominator is the sum of valid person-time, not the number of patients. A group with zero accumulated time is retained but its rate is reported as not estimable.

rate_dat <- data.frame(
  arm = c("Control", "Control", "Treatment", "Treatment"),
  events = c(1, 0, 2, 1),
  person_years = c(1.2, NA, 0.8, 1.0)
)

rate <- rate_stats(rate_dat, event = events, time = person_years, by = arm)
denominators_stats(rate)

Variable	Level	Group	Eligible observations	Used in analysis	Missing / excluded	Numerator	Denominator	Rule
Denominator audit
events	NA	Control	2	1	1	1	1.2	Accumulated person-time; rate per 1000
events	NA	Treatment	2	2	0	3	1.8	Accumulated person-time; rate per 1000

The audit records both valid event-time records and the person-time denominator.

Cross-tabs, comparisons, and correlations

crosstabs() uses complete row-column pairs: a row is excluded when either cross-tabulated variable is missing. Its margins, percentages, test, and 2×2 measures therefore refer to that same complete-pair population.

cross <- crosstabs(dat, row = arm, col = smoker)
denominators_stats(cross)

Variable	Level	Group	Eligible observations	Used in analysis	Missing / excluded	Numerator	Denominator	Rule
Denominator audit
smoker	Yes	Treatment	3	2	1	1	2	Complete row/outcome observations; event level = Yes
smoker	Yes	Control	2	2	0	1	2	Complete row/outcome observations; event level = Yes

For independent compare_groups(), an observation requires both variable and group. For paired analysis, an identifier must be observed once in each group with the required outcome; the denominator becomes the number of complete, aligned pairs.

comparison <- compare_groups(dat, variable = age, group = arm)
denominators_stats(comparison)

Variable	Level	Group	Eligible observations	Used in analysis	Missing / excluded	Numerator	Denominator	Rule
Denominator audit
age	NA	arm = Control	2	1	1	NA	1	Non-missing outcome observations within group
age	NA	arm = Treatment	3	3	0	NA	3	Non-missing outcome observations within group

correlation() similarly uses complete pairs: both variables must be observed in the same row.

cor_dat <- data.frame(
  age = c(34, 41, 45, 49, 53, 57, 62, 68),
  follow_up = c(12, 11, NA, 9, 8, 7, 6, 5)
)
to_flextable(correlation(cor_dat, x = age, y = follow_up))

Variables	n	Correlation (95% CI)	p-value
age and follow_up	7	-1.00 (-1.00 to -0.98)	<0.01
Pearson when both absolute sample skewness values are below 1; otherwise Spearman in auto mode.
Inspect a scatterplot: Pearson requires an approximately linear relationship; Spearman requires a monotonic relationship.
Correlation inference assumes independent observation pairs and no dominating influential observations.
Analysis used 7 complete finite pairs; 1 pairs were excluded.

Distribution assessment

assess_distribution() keeps data-quality information visible. It reports n, Missing, and Non-finite; skewness and Shapiro-Wilk use usable finite values only. Missing or non-finite values are flagged in the guidance rather than being interpreted as evidence of non-normality.

dist_dat <- data.frame(
  arm = factor(rep(c("Control", "Treatment"), each = 8)),
  age = c(34, 39, 44, 48, 52, 57, NA, 63,
          36, 41, 46, 51, 56, 61, 66, Inf)
)
to_flextable(assess_distribution(dist_dat, vars = age, by = arm))

Variable	Group	n	Missing	Non-finite	Skewness	Shape	Suggested presentation	Shapiro p
age	Control	7	1	0	0.05	Little/no asymmetry	Mean (SD) reasonable	0.990
	Treatment	7	0	1	0.00	Little/no asymmetry		0.949
Shape categories use absolute sample skewness: little/no asymmetry < 0.50; some asymmetry 0.50 to < 1.00; marked skew >= 1.00. They are descriptive guidance, not formal classifications.
Shapiro-Wilk is sensitive to sample size. Interpretation should consider skewness, graphical assessment, sample size and subject-matter knowledge.
Suggested summaries are intended for descriptive reporting only and should not be used alone to determine inferential statistical methods.
For grouped data, the suggested presentation applies to all groups of each variable.

For a grouped continuous variable, assess_variance() also excludes missing and non-finite values from its SD and variance calculations, while retaining their counts in the output.

to_flextable(assess_variance(dist_dat, vars = age, by = arm))

Variable	Control	Treatment	Observed SD ratio	Observed variance ratio	Levene p	Interpretation
age	n = 7; SD = 10.12; variance = 102.48 (1 excluded)	n = 7; SD = 10.80; variance = 116.67 (1 excluded)	1.07	1.14	0.814	Descriptive spread shown; interpret it in the context of the study design. Missing or non-finite values are excluded from SD and variance. Some values were excluded.
SD and variance ratios are the largest group value divided by the smallest group value. They describe observed spread; they are not pass/fail tests.
For independent groups, Welch t-tests and Welch ANOVA do not require equal variances. `assess_variance()` does not select an inferential test and does not assess pairing or repeated-measures sphericity.
The displayed Levene test is median-centred (Brown-Forsythe). It is supporting information only: its p-value neither proves equal variances nor selects an inferential test.
Interpret spread alongside sample size, distributional shape, outliers, missingness, and the study design.

Reporting checklist

Does the displayed percentage use the denominator that answers the question?
Are missing values shown, suppressed, or described elsewhere?
Is the analysis population the same for every variable? If not, is this clear?
For rates, is the denominator person-time rather than participants?
For paired analysis, have you reported the number of complete pairs?
For repeated-measures ANOVA, have you reviewed sphericity rather than using a cross-occasion variance test as a substitute?
Have you reviewed denominators_stats() where exclusions affect interpretation?

What gtstats does not do automatically

gtstats reports the observed-data calculation. It does not impute missing values, diagnose the missing-data mechanism, apply weighting, or decide whether complete-case analysis is scientifically valid. Prepare the appropriate analysis dataset first when those methods are required, then use gtstats to report the result transparently.