presence <- liss %>%
dplyr::filter(
t %in% analysis_waves,
!is.na(bmi) | !is.na(fv) | !is.na(pa)
)03 · Sample selection with weasel
The analytic sample keeps every respondent observed on at least three of the seven analysis waves, and estimates under full-information maximum likelihood. This vignette expresses that rule as an explicit, named, auditable scenario with weasel, shows the audit outputs the article draws on, and demonstrates the machinery live on synthetic data.
Semantics first
weasel counts a respondent as observed at a wave when a row with that (id, wave) pair exists in the long data. The study’s notion of observed is substantive: at least one focal measure (BMI, fruit and vegetable consumption, or sport hours) reported at that wave. The bridge is a presence filter applied before planning:
Rows with a wave record but no focal measurement therefore do not count as observed, exactly as in the pipeline’s own rule.
The rule as a scenario
Three defaults ship with weasel_plan() (anchored_strict, anchored_balanced, lenient_info_max); the study’s rule is a fourth, custom one. Within the full seven-wave span, at least three observed waves means at most four missing; endpoints are not required, and interior-gap limits are set non-binding.
ws_plan <- weasel::weasel_plan(
presence[c("id", "t", "female", "age")],
id = "id", wave = "t", span = "full",
scenarios = data.frame(
scenario = "min3_of7",
require_endpoints = FALSE,
max_missing = 4L,
n_gap_max = 5L,
max_gap_max = 5L
)
)
analysis_ids <- unique(weasel::weasel_apply(ws_plan, "min3_of7")$id)The pipeline (chunk cb112 of analysis/reciprocal-health-dynamics.R) runs this plan alongside its own one-liner and asserts setequal() between the two id sets, so the transcript proves at every render that the scenario and the rule are the same selection: 5,676 respondents.
The audits the article uses
Three diagnostics turn the rule from a stipulation into an argued choice.
Tolerance sensitivity. weasel_sensitivity() sweeps the minimum-wave tolerance and reports the retained sample at each level, which is the sample-size trade-off behind choosing three as the threshold:
weasel::weasel_sensitivity(
ws_plan,
require_endpoints = FALSE, max_missing = 0:6,
n_gap_max = 5L, max_gap_max = 5L
)Attrition selectivity. weasel_selectivity() compares retained and excluded respondents on covariates via standardised mean differences, the empirical check on whether completeness-based selection skews the sample that the missing-at-random discussion in the article leans on:
weasel::weasel_selectivity(ws_plan, "min3_of7")Prose justification. weasel_justify_subset() generates methods-section text stating the window, the constraints, the retained count, and the coverage achieved, regenerable from the same inputs:
cat(weasel::weasel_justify_subset(ws_plan, "min3_of7"))scripts/02_select_sample.R runs all of the above against the real panel and writes the id list and the justification to output/.
Live demonstration on synthetic data
Everything above executes here on generated data with the study’s exact scenario, so the outputs can be inspected without any LISS access.
set.seed(42)
d <- do.call(rbind, lapply(1:400, function(i) {
w <- sort(sample(1:7, sample(1:7, 1)))
data.frame(id = i, t = w, female = stats::rbinom(1, 1, 0.5),
age = stats::rnorm(1, 50, 15))
}))
sc <- data.frame(
scenario = "min3_of7", require_endpoints = FALSE,
max_missing = 4L, n_gap_max = 5L, max_gap_max = 5L
)
p <- weasel::weasel_plan(d, id = "id", wave = "t", span = "full", scenarios = sc)✔ plan ready: span 1:7 (full, L = 7)p<weasel_plan>
span: 1:7 (full, L = 7)
respondents observed in span: 400
scenario n_ids mean_prop_present endpoint_rate score recommended
min3_of7 297 0.696 0.549 2.396 TRUE
data: 1604 row(s) attachedweasel::weasel_print_table(
weasel::weasel_sensitivity(p, require_endpoints = FALSE, max_missing = 0:6,
n_gap_max = 5L, max_gap_max = 5L),
title = "sample size by minimum-wave tolerance"
)── sample size by minimum-wave tolerance ─────────────────────────────────────── require_endpoints max_missing n_gap_max max_gap_max n_ids prop_ids
FALSE 0 5 5 56 0.140
FALSE 1 5 5 103 0.258
FALSE 2 5 5 165 0.412
FALSE 3 5 5 233 0.583
FALSE 4 5 5 297 0.743
FALSE 5 5 5 350 0.875
FALSE 6 5 5 400 1.000
mean_prop_present
1.000
0.935
0.852
0.770
0.696
0.634
0.573weasel::weasel_print_table(
weasel::weasel_selectivity(p, "min3_of7"),
title = "retained vs excluded"
)── retained vs excluded ──────────────────────────────────────────────────────── variable n_retained n_excluded mean_retained mean_excluded diff smd
female 297 103 0.488 0.544 -0.055 -0.111
age 297 103 50.252 48.855 1.398 0.090cat(weasel::weasel_justify_subset(p, "min3_of7"))To construct a longitudinal analysis sample, we selected respondents whose wave participation satisfied explicit structural criteria using the WEASEL framework (Wave-based Extraction and Selection for Longitudinal Data) (R package weasel). Specifically, we focused on waves 1 to 7 (L = 7) and did not require observed endpoints, allowing unanchored participation, allowed up to 4 missing waves within the window, and restricted the missingness structure (at most 5 interior missing block(s), each no longer than 5 wave(s)). This strategy retained 297 respondent(s), reflecting an explicit trade-off between sample size and within-window completeness. In the resulting subset, mean within-window coverage was 0.696, endpoint coverage was 0.549. The analysis window was selected using the package's span rule (full), which prioritizes a coherent window with comparatively strong participation. All selection decisions were rule-based and reproducible, and can be regenerated from the same inputs and parameters using the weasel workflow.sub <- weasel::weasel_apply(p, "min3_of7")
identical(
sort(unique(sub$id)),
sort(as.integer(names(which(table(d$id) >= 3))))
)[1] TRUEThe final identical() is the same equivalence the pipeline asserts on the real data: the scenario is the rule.
Interactive exploration
Before committing to any rule, the scope pipeline maps the participation landscape:
weasel::set_weasel_scope(presence, "id", "t", gap = 1)
weasel::evaluate_weasel_scope()
weasel::weasel_summarize_waves()
weasel::weasel_print_table(weasel::weasel_filter_wave_summary(), n = 10)
weasel::weasel_scope_info()
weasel::weasel_clear_scope()Panels on non-consecutive schedules (biennial waves, waves recorded as years) are handled with grid = "observed" in both pipelines.