Small-sample inference

When is a sample “small”?

Survey research often works with few observations: pilot studies, specialised populations, classroom studies, and early-stage organisational diagnostics may land at n = 10 to 25. Adequacy depends jointly on the design, group balance, outcome distribution, model, effect size, and estimand, so a single word count answers little on its own. With limited data, estimates are often imprecise and assumptions are harder to assess, so an isolated p-value carries limited information.

surveyframe uses n = 30 as a visible advisory heuristic that flags where assumptions deserve a closer look. Its small-sample tools surface exact or alternative procedures where available and pair estimates with uncertainty. Choosing one changes the question asked: a rank-based test may target a different quantity from a test about means, and “distribution-free” carries its own assumptions. Choose the method from the study design and estimand, and report the limited precision.

Planning a small-sample study

sample_size_plan() estimates a required sample size for a design and attaches the same small-sample advisory used elsewhere in the package when the estimate itself falls below 30.

sample_size_plan(type = "t_test", groups = 2)
#> Sample-Size Plan
#>   Target:      t_test
#>   Estimated n: 128
#>   Method:      power
#>   Alpha:       0.050   Power: 0.80
#>   Effect size: d = 0.5
#> 
#> Warning: Assumed a medium effect, d = 0.5. Supply d for the effect you expect.

A worked instrument in this vignette targets n = 20 complete responses, well below that planning estimate. The instrument below has a two-level grouping question, a numeric outcome, and a binary outcome for the logistic example later in this vignette.

group_cs     <- sf_choices(id = "grp_cs", values = c("control", "treatment"),
                            labels = c("Control", "Treatment"))
converted_cs <- sf_choices(id = "conv_cs", values = c("no", "yes"),
                            labels = c("No", "Yes"))

items <- list(
  sf_item(id = "group", label = "Study arm", type = "single_choice", choice_set = "grp_cs"),
  sf_item(id = "outcome", label = "Outcome score", type = "numeric"),
  sf_item(id = "converted", label = "Converted (yes/no)", type = "single_choice",
          choice_set = "conv_cs")
)

study <- sf_instrument(
  title      = "Small-sample pilot",
  version    = "0.1.0",
  authors    = "surveyframe",
  components = c(list(group_cs, converted_cs), items)
)

study
#> <sframe>
#>   Title:      Small-sample pilot
#>   Version:    0.1.0
#>   Items:      3
#>   Scales:     0
#>   Analysis:   0 block(s)
#>   Status:     not validated

Reading the advisory

assumption_report() and sample_size_plan() both attach an advisory element once a relevant n drops below 30. It prints automatically and is also available as a field for programmatic checks.

n20 <- data.frame(score = rnorm(20, mean = 50, sd = 10))
n50 <- data.frame(score = rnorm(50, mean = 50, sd = 10))

ar_small <- assumption_report(n20, variables = "score")
ar_small$advisory
#> [1] "Small sample detected (n = 20). For these assumption checks, consider non-parametric alternatives. Asymptotic p-values may be unreliable. Bootstrap or exact confidence intervals are provided where available."

ar_large <- assumption_report(n50, variables = "score")
is.null(ar_large$advisory)
#> [1] TRUE

Mann-Whitney with the Hodges-Lehmann estimate

The Mann-Whitney runner now reports the Hodges-Lehmann shift estimate alongside the rank-biserial effect size, with a confidence interval on the shift, which a point estimate alone leaves out.

pilot <- data.frame(
  group   = rep(c("control", "treatment"), each = 10),
  outcome = c(rnorm(10, 48, 8), rnorm(10, 55, 8)),
  converted = sample(c("no", "yes"), 20, replace = TRUE, prob = c(0.6, 0.4)),
  covariate = rnorm(20, 0, 1)
)

sf_plan(study) <- list(
  list(id = "RQ1",
       research_question = "Does the treatment arm score higher than control?",
       family = "group_comparison", method = "mann_whitney",
       roles = list(group = "group", outcome = "outcome"),
       options = list(alpha = 0.05))
)

mw_results <- run_analysis_plan(pilot, study)
results_table(mw_results)
RQ Research question Method Result (APA)
RQ1 Does the treatment arm score higher than control? U = 34, z = -1.21, p = .226, r = 0.27, 95% CI [0.02, 0.64], Hodges-Lehmann shift = -3.99, 95% CI [-13.75, 4.58]
mw_results[[1]]$hl_shift
#> [1] -3.987846
mw_results[[1]]$hl_conf_int
#> [1] -13.750068   4.584132

Paired Wilcoxon with the pseudomedian CI

The paired runner reports a Hodges-Lehmann pseudomedian for the within-pair shift, again with an interval in place of a point value.

before <- rnorm(8, 50, 6)
after  <- before + rnorm(8, 4, 5)

sf_plan(study) <- list(
  list(id = "RQ2",
       research_question = "Did scores change from before to after?",
       family = "group_comparison", method = "wilcoxon_pair",
       roles = list(before = "before", after = "after"),
       options = list(alpha = 0.05))
)

wp_results <- run_analysis_plan(data.frame(before = before, after = after), study)
results_table(wp_results)
RQ Research question Method Result (APA)
RQ2 Did scores change from before to after? V = 0, z = -2.52, p = .012, r = 0.89, 95% CI [0.89, 0.90], pseudomedian = -6.46, 95% CI [-11.70, -3.63]
wp_results[[1]]$pseudomedian
#> [1] -6.45986
wp_results[[1]]$pseudomedian_conf_int
#> [1] -11.700678  -3.633836

Fisher’s exact test with the exact odds-ratio CI

For a 2x2 table, fisher_exact now attaches an exact odds-ratio confidence interval alongside the test’s p-value. The interval is only defined for a 2x2 table, and is dropped (not approximated) when simulate_p_value is requested, since base R’s fisher.test() cannot return both together.

sf_plan(study) <- list(
  list(id = "RQ3",
       research_question = "Is conversion associated with study arm?",
       family = "association", method = "fisher_exact",
       roles = list(row = "group", column = "converted"),
       options = list(alpha = 0.05))
)

fe_results <- run_analysis_plan(pilot, study)
results_table(fe_results)
RQ Research question Method Result (APA)
RQ3 Is conversion associated with study arm? Fisher’s exact test, p = 1.000, phi = 0.12
fe_results[[1]]$odds_ratio
#> [1] 0.599308
fe_results[[1]]$odds_ratio_conf_int
#> [1] 0.0390558 6.9366524

Bootstrap confidence interval on a median

bootstrap_ci() computes a bootstrap interval for an arbitrary statistic, useful when no exact interval exists for the estimator you need. It is already exported and used elsewhere in the package, applied here to the median of the pilot outcome scores.

bootstrap_ci(pilot$outcome, FUN = stats::median, R = 999)
#> estimate    lower    upper 
#> 52.82878 47.33676 56.66709 
#> attr(,"resamples")
#> [1] 999
#> attr(,"valid_resamples")
#> [1] 999

Firth logistic regression for a rare or small binary outcome

Ordinary logistic regression can fail to converge or produce severely biased estimates with small samples or separated data. Firth’s penalised-likelihood correction (Firth, 1993) addresses both. surveyframe’s firth_logistic runner needs the optional logistf package.

sf_plan(study) <- list(
  list(id = "RQ4",
       research_question = "Does the covariate predict conversion?",
       family = "regression", method = "firth_logistic",
       roles = list(dependent = "converted", predictors = "covariate"),
       options = list(conf.level = 0.95))
)

firth_results <- run_analysis_plan(pilot, study)
results_table(firth_results)
RQ Research question Method Result (APA)
RQ4 Does the covariate predict conversion? Firth logistic regression (n = 20), likelihood ratio = 0.10
firth_results[[1]]$coefficients
#>               Estimate     ci_low     ci_high         p odds_ratio or_ci_low
#> (Intercept) -0.9722906 -1.9991045 -0.07984311 0.0321831  0.3782157 0.1354565
#> covariate    0.1336834 -0.7614982  0.99469674 0.7555259  1.1430308 0.4669663
#>             or_ci_high
#> (Intercept)  0.9232612
#> covariate    2.7039042

Cohen’s d with a bootstrap interval

cohens_d_ci() pairs Cohen’s d with a bootstrap interval, which is informative precisely where it matters most: interval width grows visibly as n falls, making the extra uncertainty at small n explicit, where a single point estimate leaves it implicit.

cohens_d_ci(pilot$outcome[pilot$group == "treatment"],
            pilot$outcome[pilot$group == "control"], R = 999)
#>   estimate      lower      upper 
#> 0.54201883 0.02928852 1.54044282 
#> attr(,"resamples")
#> [1] 999
#> attr(,"valid_resamples")
#> [1] 999

Citation

citation("surveyframe")

The small-sample methods demonstrated in this vignette are described in full, with guidance on when to prefer each one, in:

Sharafuddin, M. A., Jaleel, A. A., and Madhavan, M. (2026). Quantitative Analysis with Small Samples: A Practical Guide for Students and Early-Career Researchers (Version 0.1.0) [Book]. Zenodo. https://doi.org/10.5281/zenodo.20221929

A companion preprint validating these method choices by simulation is in preparation and will be added here once posted.