Sampling designs

A design describes how to sample, independently of what you are sampling. Build one with a design_*() constructor, then hand it to draw() with the data. The same design can be reused across data sets, stored, printed and passed around.

The designs

design_simple(): Rows uniformly at random.
design_stratified(): A share of each stratum.
design_systematic(): Every k-th row.
design_cluster(): Whole clusters.
design_multistage(): Clusters, then rows within them.
design_weighted(): Rows with probability governed by a weight.
design_certainty(): Everything above a threshold, plus a sample of the rest.
design_reservoir(): A fixed-size sample from a stream.
design_bootstrap(): Resampled replicates.
design_temporal(): A share of each time interval.
design_spatial(): Rows inside a region.
design_spread(): A sample spread evenly across a map or across auxiliary variables.

What every design guarantees

Arguments mean the same thing everywhere they appear:

n is always the total number of rows drawn, never a per-group figure. design_temporal() uses per_interval precisely because that one is per group, and design_cluster() has no n at all because the row count follows from which clusters were selected.
allocation always says how a total is split across groups, either "proportional" or "equal". Splitting is exact: the largest-remainder method is used, so a request for 60 rows returns 60 rather than drifting with per-group rounding.
na_rm always decides whether rows with a missing key are dropped or raise an error. It is never silently assumed either way.
replace always means sampling with replacement within whatever group the design works on, and always rules out draw(weights = TRUE): an inclusion probability describes distinct units, and a sample holding duplicates cannot be weighted by one.
draw() always restores the caller's random number stream before returning, and always gives back a data frame with the input's class and column order.
Rows come back in frame order for every design that selects a set of rows, including design_systematic() with order_by. The two exceptions are the ones where draw order is meaningful: design_simple() and design_weighted() return rows in the order they were drawn, and design_bootstrap() returns replicates in order with a leading .replicate column.

Estimating from a sample

Because a design is a value, it can report its own inclusion probabilities before any sampling happens. inclusion_prob() gives first-order probabilities and sampling_weight() their reciprocals — the number of population rows each sampled row stands for. joint_prob() gives second-order probabilities, and ht_total() and ht_mean() turn a sample into a population total, mean or proportion — overall or by domain — with a standard error from the variance estimator suited to the design and a confidence interval on the design's degrees of freedom. deff() reports what the design cost in precision against simple random sampling, and sample_summary() reports what was actually drawn against what was in the frame.

Going the other way, plan_size() solves for the sample size a given margin of error requires — the step before choosing a design. For analysis this package does not do, as_svydesign() hands a sample to the survey package.

Not every design has a closed form for these — see inclusion_prob() for which, why, and what to do instead. Those that do not say so rather than returning an approximation.

Choosing one

Every unit equally likely: design_simple(), or design_systematic() when the frame has a useful order, or design_reservoir() when it does not fit in memory.
Guaranteed coverage of subgroups: design_stratified(), with min_per_stratum if rare groups must appear.
Fieldwork cost matters more than efficiency: design_cluster() or design_multistage() — visiting five sites is cheaper than visiting fifty, at the cost of precision.
Large units matter more: design_weighted(), with method = "systematic" or "poisson" if you intend to estimate.
A few units dominate the total: design_certainty() — take those with certainty and sample the tail, which removes them from the variance entirely.
Coverage across time or space: design_temporal() for time; design_spread() for an even spread over a map, which is usually far more precise than a simple random sample when what you measure varies smoothly over space; design_spatial() to restrict sampling to a region.
Balance on known covariates without choosing strata: design_spread() across those covariates.
Uncertainty of a statistic, not a population total: design_bootstrap().

The designs

What every design guarantees

Estimating from a sample

Choosing one

See also