Grid exploration for svyplan objects

Evaluate a svyplan result at new parameter combinations. Returns a data frame with the varied parameters and resulting quantities, suitable for sensitivity analysis or plotting.

Usage

# S3 method for class 'svyplan_n'
predict(object, newdata, ...)

# S3 method for class 'svyplan_cluster'
predict(object, newdata, ...)

# S3 method for class 'svyplan_power'
predict(object, newdata, ...)

# S3 method for class 'svyplan_prec'
predict(object, newdata, ...)

Arguments

object: A svyplan object (svyplan_n, svyplan_cluster, svyplan_power, or svyplan_prec). For svyplan_prec, only types "proportion", "mean", and "ratio" are supported. Types "cluster" and "multi" are not supported.
newdata: A data frame of parameter combinations to evaluate. Column names must be valid parameters for the object type (see Details). Parameters not in newdata stay at their original values from the object.
...: Additional arguments are not supported and produce an error.

Value

A data frame with newdata columns followed by result columns. The result columns depend on the object type:

svyplan_n: n, se, moe, cv, rmoe, except for a type = "change" result, which reports one size per occasion as n1 and n2 in place of n, and a joint constrained allocation, which reports the design and cost columns described under Details
svyplan_cluster: n_psu, n_per_psu, (opt. n_per_ssu), total_n, cv, cost
svyplan_power: n, power, effect
svyplan_prec: se, moe, cv, rmoe

Details

Valid parameters for newdata by object type:

n_prop: p, moe, rmoe, cv, alpha, N, deff, resp_rate, df, min_cases
n_mean: var, mu, moe, rmoe, cv, alpha, N, deff, resp_rate, df
n_ratio: r, cv_num, cv_den, component_cor, moe, rmoe, cv, alpha, N, deff, resp_rate, df
n_change: moe, rmoe, cv, alpha, N, deff, resp_rate, ratio, overlap, overlap_cor, and change on the mean scale only, since two proportions determine it. The occasion variances are held at the values the result was built from, so vary the design rather than the estimand: sweeping overlap and overlap_cor is what prices a rotation against a fresh sample each round.
n_pooled: moe, rmoe, cv, alpha, N, deff, resp_rate, mu on the mean scale, and occasions only where the overlap and correlation are flat across lags, a shaped profile being a statement about a horizon that changing the horizon would contradict. The lag profiles themselves cannot be varied, a grid row having nowhere to hold a vector.
n_cluster: cv, budget, unit_relvar, resp_rate, fixed_cost, stage deltas (icc or icc_psu, plus icc_ssu for 3-stage), stage ratios (var_ratio or var_ratio_psu, plus var_ratio_ssu for 3-stage), and stage costs (cost_psu, cost_ssu, cost_tsu). For 2-stage designs, cost_tsu aliases cost_ssu.
power_prop: p1, p2, n, power, alpha, N, deff, alternative, overlap, overlap_cor, resp_rate (excluding the solved-for parameter). Not supported for objects with vector n.
power_mean: effect, var, n, power, alpha, N, deff, alternative, overlap, overlap_cor, resp_rate (excluding the solved-for parameter). Not supported for objects with vector n.
prec_prop: p, n, alpha, N, deff, resp_rate, df. A result that solved for p varies the target it solved from, cv or rmoe, in place of p.
prec_mean: var, n, mu, alpha, N, deff, resp_rate, df. A result that solved for mu varies cv in place of mu.
prec_ratio: r, n, cv_num, cv_den, component_cor, alpha, N, deff, resp_rate, df

For svyplan_n objects, moe, rmoe, and cv are mutually exclusive in newdata. If one appears, that mode is used. If none appears, the original mode is preserved.

Similarly, for svyplan_cluster objects, cv and budget are mutually exclusive.

Multi-indicator (n_multi) and multi-indicator cluster results are not supported. Use the underlying single-indicator functions instead.

A joint constrained allocation (n_alloc() with measures and targets) varies two controls, and each grid row re-solves the stored problem.

A fixed-budget objective fit varies budget, tracing the cost-versus-objective frontier. A fixed-take multistage fit varies the takes, n_per_psu for a two-stage design and either n_per_psu or n_per_ssu for a three-stage one, tracing what the fieldwork decision costs. A scalar take applies to every stratum, so the per-stratum takes the fit was built from are replaced rather than scaled. Sweep them asymmetrically by editing frame and calling n_alloc() per point. Both controls may appear together on a budget-objective fit, one row per combination.

The result reports the continuous n and cost, the operational n_int and cost_int, and .feasible. A multistage fit adds the PSU counts n_psu and n_psu_int, and budget mode adds objective_value, its equivalent cv, and whether the budget binds. A row the solver cannot satisfy, a budget too small to fund the hard targets or a take that exceeds the stage populations, gives an all-NA row with .feasible = FALSE and a warning, so one infeasible point does not discard the rest of the grid. A take that is not a positive whole number is a malformed grid rather than an infeasible design point, and stops.

A certainty-aware fit (n_alloc() with psu) varies n_per_psu too, even though the solver saw a one-stage problem, and reports n_psu_certain and n_psu_draw in place of the PSU counts. This is how the take is priced. It also sets the certainty threshold, so a larger take leaves fewer certainty PSUs and costs more sample, and the resulting cost curve has a minimum worth finding rather than assuming.

Precision targets stay out of reach of the grid. A one-stage minimum-cost fit therefore has nothing to vary. Modify targets and call n_alloc() again.

If evaluation fails for a particular row (e.g. invalid parameter combinations), that row's result columns are NA and a warning is issued.

Examples

# Sensitivity of sample size to deff and response rate
x <- n_prop(p = 0.3, moe = 0.05, deff = 1.5)
predict(x, expand.grid(
  deff = seq(1, 3, 0.5),
  resp_rate = c(0.7, 0.8, 0.9)
))
#>    deff resp_rate         n         se  moe         cv      rmoe
#> 1   1.0       0.7  460.9751 0.02551067 0.05 0.08503558 0.1666667
#> 2   1.5       0.7  691.4626 0.02551067 0.05 0.08503558 0.1666667
#> 3   2.0       0.7  921.9501 0.02551067 0.05 0.08503558 0.1666667
#> 4   2.5       0.7 1152.4376 0.02551067 0.05 0.08503558 0.1666667
#> 5   3.0       0.7 1382.9252 0.02551067 0.05 0.08503558 0.1666667
#> 6   1.0       0.8  403.3532 0.02551067 0.05 0.08503558 0.1666667
#> 7   1.5       0.8  605.0298 0.02551067 0.05 0.08503558 0.1666667
#> 8   2.0       0.8  806.7064 0.02551067 0.05 0.08503558 0.1666667
#> 9   2.5       0.8 1008.3829 0.02551067 0.05 0.08503558 0.1666667
#> 10  3.0       0.8 1210.0595 0.02551067 0.05 0.08503558 0.1666667
#> 11  1.0       0.9  358.5362 0.02551067 0.05 0.08503558 0.1666667
#> 12  1.5       0.9  537.8042 0.02551067 0.05 0.08503558 0.1666667
#> 13  2.0       0.9  717.0723 0.02551067 0.05 0.08503558 0.1666667
#> 14  2.5       0.9  896.3404 0.02551067 0.05 0.08503558 0.1666667
#> 15  3.0       0.9 1075.6085 0.02551067 0.05 0.08503558 0.1666667

# Power curve: how does power vary with sample size?
pw <- power_prop(p1 = 0.30, p2 = 0.35, n = 500, power = NULL)
predict(pw, data.frame(n = seq(100, 1000, 100)))
#>       n     power effect
#> 1   100 0.1175929   0.05
#> 2   200 0.1877131   0.05
#> 3   300 0.2581732   0.05
#> 4   400 0.3272968   0.05
#> 5   500 0.3938436   0.05
#> 6   600 0.4569385   0.05
#> 7   700 0.5160053   0.05
#> 8   800 0.5707088   0.05
#> 9   900 0.6209032   0.05
#> 10 1000 0.6665884   0.05

# Cluster design: sensitivity to icc (homogeneity)
cl <- n_cluster(stage_cost = c(500, 50), icc = 0.05, budget = 100000)
predict(cl, data.frame(icc = c(0.01, 0.03, 0.05, 0.10, 0.15)))
#>    icc     n_psu n_per_psu   total_n         cv  cost
#> 1 0.01  48.23430 31.464265 1517.6570 0.02931966 1e+05
#> 2 0.03  71.47587 17.981472 1285.2413 0.03427016 1e+05
#> 3 0.05  84.08997 13.784049 1159.1003 0.03760588 1e+05
#> 4 0.10 102.63340  9.486833  973.6660 0.04357388 1e+05
#> 5 0.15 114.10493  7.527727  858.9507 0.04800166 1e+05

# Allocation: how does the CV change with sample size?
frame <- data.frame(
  N    = c(4000, 3000, 3000),
  sd   = c(10, 15, 8),
  mean = c(50, 60, 55)
)
alloc <- n_alloc(frame, n = 600)
predict(alloc, data.frame(n = seq(200, 1000, 200)))
#>      n        se       moe          cv cost
#> 1  200 0.7624828 1.4944388 0.013990510  200
#> 2  400 0.5332495 1.0451498 0.009784394  400
#> 3  600 0.4305191 0.8438019 0.007899432  600
#> 4  800 0.3685682 0.7223803 0.006762719  800
#> 5 1000 0.3257913 0.6385393 0.005977823 1000

# Two-stage joint allocation: what does the per-PSU take cost?
cluster_frame <- data.frame(
  stratum  = c("Urban", "Rural"),
  N        = c(60000, 40000),
  N_psu    = c(600, 500),
  n_per_psu = c(10, 10),
  cost_psu = c(400, 550),
  cost_ssu = c(20, 25)
)
cluster_measures <- data.frame(
  stratum = rep(cluster_frame$stratum, 2),
  name    = rep(c("literacy", "income"), each = 2),
  p       = c(0.62, 0.48, NA, NA),
  mean    = c(NA, NA, 520, 380),
  sd      = c(NA, NA, 210, 160),
  icc_psu = rep(c(0.04, 0.07), 2)
)
cluster_targets <- data.frame(
  name = c("literacy", "income"),
  cv   = c(0.03, 0.04)
)
joint <- n_alloc(cluster_frame, measures = cluster_measures,
                 targets = cluster_targets)
predict(joint, data.frame(n_per_psu = c(6, 10, 14, 20)))
#>   n_per_psu     n_psu        n      cost n_psu_int n_int cost_int .feasible
#> 1         6 175.97204 1055.832 103715.64       176  1056   103760      TRUE
#> 2        10 122.67192 1226.719  83273.41       123  1230    83400      TRUE
#> 3        14  99.75419 1396.559  76626.54       100  1400    76800      TRUE
#> 4        20  82.48407 1649.681  74400.37        83  1660    74900      TRUE

Usage

Arguments

Value

Details

See also

Examples