panelcond: refreshment-sample designs for panel conditioning

panelcond estimates the panel conditioning effect, the causal effect of prior interviews on a wave-specific survey answer, from a panel that has added a refreshment sample. It implements the naive fresh-versus-continuing contrast, survival matching, symmetric survival matching, the entry-wave correction of Das, Toepoel and van Soest (2011), reinterpreted as a survivor-effect estimator (with a regression-standardised version), attrition IPW and the incremental dose contrast between two continuing cohorts; the decomposition of the naive contrast into conditioning and attrition components; two diagnostic tests that distinguish state-dependent from non-stationary attrition; a joint person bootstrap; worst-case bounds and breakdown values; and a simulation engine for design planning. There are no dependencies beyond base R.

remotes::install_github("sokubo/panelcond")
library(panelcond)
set.seed(1)
sim <- pc_simulate(n_old = 4800, n_new = 960, k = 4, regime = "MNAR_trait")
fit <- pc_estimate(sim$old, sim$new, k = 4, adjust = "x", nboot = 200)
fit
pc_bounds(fit, y_range = c(-4, 4), sd_new = sd(sim$new$y))
mc <- pc_montecarlo(R = 200, k = 4, regime = "MNAR_state")
mc
attr(mc, "tests")

From a wide panel with one row per person, per-wave response indicators r_w1, ..., r_w8, an entry-wave column and outcome columns for waves 1 and 4:

set.seed(2)
n <- 600; R <- matrix(rbinom(n * 8, 1, 0.9), n, 8); R[1:300, 1:3] <- 0L
wide <- data.frame(id = seq_len(n), entry_wave = c(rep(4L, 300), rep(1L, 300)))
for (w in 1:8) wide[[paste0("r_w", w)]] <- R[, w]
wide$y_w1 <- rnorm(n); wide$y_w4 <- rnorm(n)
inp <- pc_from_wide(wide, cohort = "entry_wave", resp = paste0("r_w", 1:8),
                    y_cols = c("1" = "y_w1", "4" = "y_w4"), t = 4, c_old = 1, id = "id")
pc_estimate(inp$old, inp$new, k = inp$k, m = inp$m)

Two continuing cohorts from the same population (an original sample and a cohort-matched refreshment) with doses 12 and 8 at the same wave, each corrected by its own entry wave:

set.seed(3)
a <- pc_simulate(n_old = 4800, n_new = 10, k = 12, regime = "MNAR_trait")$old
b <- pc_simulate(n_old = 960,  n_new = 10, k = 8,  regime = "MNAR_trait")$old
pc_increment(data.frame(y = a$y, y_entry = a$y_entry, s = a$s_prior),
             data.frame(y = b$y, y_entry = b$y_entry, s = b$s_prior), kA = 12, kB = 8, nboot = 200)

The vignette (vignette("panelcond")) walks through the designs on simulated data. pc_design_rank() / pc_is_identified() implement the identification results of Okubo (2026), Panel Conditioning in Fixed-Effects Models: Identification and Bias Propagation, arXiv:2609.28871. The estimators, the diagnostics and the variance formulas are those of Okubo (2026), Identifying Panel Conditioning with Refreshment Samples: Sharp Bounds and Design Assumptions, arXiv:2610.01654, whose Section 10 was computed with version 0.1.6 of this package (replication archive: paper-refreshment-designs-replication, tag paper-v1.5). citation("panelcond") gives the package and both papers.

Design planning

pc_design_rank(entries, t_max) builds the cell-mean design of a staggered panel and reports what the schedule identifies about the conditioning path before any data are collected: the stride d (greatest common divisor of the spacings between entry waves), the dimension of the identified set, whether the last cohort is followed long enough, and a basis of the unidentified directions; pc_is_identified() tests a specific functional. A schedule entering cohorts at waves 1, 5 and 13 has stride 4, so its identified set is four-dimensional: the centred lag-4 second differences tau(s+4) - 2 tau(s) + tau(s-4) are identified and ordinary second differences are not. Those are examples rather than the whole identified space, which is larger; pc_is_identified() is the test. Adding a fourth entry at an even wave (20 or 22) restores stride 1.

pc_design_rank() reports the dimension of the whole kernel and the dimension of its tenure projection separately. They differ when two cohorts never share a period: the extra directions move the period and cohort effects against each other and the conditioning path not at all, so on such a support even exact knowledge of the cohort effects does not pin the path down.

d <- pc_design_rank(c(1, 5, 13), 19); d
lam <- numeric(19); lam[c(1, 5, 9)] <- c(1, -2, 1); pc_is_identified(d, lam)   # TRUE
lam <- numeric(19); lam[2:4] <- c(1, -2, 1);        pc_is_identified(d, lam)   # FALSE