--- title: "Creating Survey Objects in surveycore" output: rmarkdown::html_vignette: toc: true toc_depth: 3 bibliography: references.bib link-citations: true vignette: > %\VignetteIndexEntry{Creating Survey Objects in surveycore} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include=FALSE} library(surveycore) ``` ## Introduction Every analysis function in surveycore — `get_means()`, `get_totals()`, `get_freqs()`, `get_ratios()`, `get_corr()` — takes a **survey design object** as its first argument. That object encodes how your data was collected: which units were clustered together, which strata were defined, what weights apply, and how variance should be estimated. Without it, point estimates may be biased and standard errors are almost certainly wrong [@lumley2010; @lohr2022]. This vignette answers one question: *given my data, which constructor do I call and how do I call it?* It is written for three audiences: - **Academic researchers** working with named public surveys (NHANES, ANES, ACS, GSS). Jump to the relevant worked example in each section. - **Practitioners** running surveys of schools, businesses, or organizations. The conceptual explanations in each section are for you. - **Non-probability panel users** — if you run message-testing or attitudinal research on Lucid, Dynata, or a similar platform and have vendor-provided raking weights, skip ahead to [Section 6](#sec-calibrated). This vignette covers object *creation* only. Estimation functions (`get_means()`, `get_totals()`, etc.) are covered in `vignette("getting-started")`. --- ## 1. Decision Guide {#sec-decision} Read the first row that matches your data. | My data... | Constructor | Why | |----------------------------------------------------------------------|--------------------------|--------------------------------------------------| | Has cluster IDs, strata, and/or design weights | `as_survey()` | Taylor series linearization — the general case | | Comes with pre-built replicate weight columns (repwt_1, repwt_2, …) | `as_survey_replicate()` | Uses the agency-supplied variance replicates | | Is a pure SRS — equal probability, no clustering, no strata | `as_survey()` | Omit `ids` and `strata`; creates an SRS design | | Is a non-probability panel or opt-in sample with calibration weights | `as_survey_nonprob()` | Calibrated design; SEs are approximate | | Was sampled in two stages with an expensive Phase 2 measurement | `as_survey_twophase()` | Two-phase variance accounting for both stages | ### Common surveys at a glance | Survey | Constructor | Design | |---------------------------|--------------------------|-----------------------------------------------| | NHANES | `as_survey()` | Stratified cluster, Taylor series | | ANES | `as_survey()` | Stratified cluster, Taylor series | | GSS | `as_survey()` | Stratified multi-stage cluster | | Pew NPORS | `as_survey()` | Stratified address-based sample (no PSU) | | ACS PUMS (1-year) | `as_survey_replicate()` | 80 successive-difference replicate weights | | Pew Jewish Americans 2020 | `as_survey_replicate()` | 100 JK1 jackknife replicate weights | | BRFSS | `as_survey_replicate()` | Bootstrap replicate weights | | NAEP / PISA | `as_survey_replicate()` | JK2 jackknife replicate weights | | Nationscape (Democracy Fund + UCLA) | `as_survey_nonprob()` | Non-probability quota panel; ACS-calibrated raking weights | | Opt-in online panels | `as_survey_nonprob()` | Non-probability; vendor-supplied raking weights | --- ## 2. `as_survey()` — Taylor Series Designs {#sec-taylor} `as_survey()` is the right constructor for probability surveys with cluster and/or stratum information but no pre-computed replicate weights. It uses **Taylor series linearization** (also called the linearization or delta-method estimator), the standard approach for complex probability surveys [@lumley2010, ch. 2; @lohr2022, ch. 9]. ### 2.1 Core arguments | Argument | Codebook term | What it does | |-----------|-----------------------------------------------------|------------------------------------------| | `ids` | "PSU", "primary sampling unit", "cluster ID" | Stage-1 cluster identifier | | `weights` | "sampling weight", "person weight", "design weight" | Inverse of selection probability | | `strata` | "stratum", "design stratum", "sampling stratum" | Stratification variable | | `fpc` | "FPC", "finite population correction", "N" | Population size or sampling fraction | | `nest` | (see below) | Whether PSU IDs are locally unique | All arguments accept bare column names — no `~formula` syntax required. ### 2.2 The `nest` argument Many government surveys assign PSU IDs locally within each stratum. NHANES, for example, assigns IDs 1 and 2 within *every* stratum — PSU 1 in stratum 31 is a completely different unit from PSU 1 in stratum 32. If you do not account for this, surveycore treats PSU 1 from stratum 31 and PSU 1 from stratum 32 as the same cluster, which produces incorrect variance estimates. Set `nest = TRUE` when PSU IDs are not globally unique across strata [@lumley2010, p. 28]. A quick diagnostic: ```{r nest-diagnostic} # NHANES: only two distinct PSU values, but 15 strata # Each stratum has its own PSU 1 and PSU 2 → nest = TRUE length(unique(nhanes_2017$sdmvpsu)) # 2 length(unique(nhanes_2017$sdmvstra)) # 15 ``` If the number of unique PSU values is much smaller than the number of strata, the IDs are almost certainly nested and you need `nest = TRUE`. ### 2.3 The `fpc` argument The finite population correction (FPC) reduces variance estimates when you have sampled a substantial fraction of the population [@cochran1977, §2.8; @lohr2022, §2.8]. Supply either: - An **integer column** with the total population size in each stratum - A **numeric column** (0–1) with the sampling fraction FPC has a meaningful effect when the sampling rate exceeds roughly 5% [@cochran1977]. For large national surveys like NHANES and ANES, the sampling fraction is tiny and FPC can be safely omitted (`fpc = NULL`). ### 2.4 Multi-level clustering For two-stage designs — counties then households, schools then students — pass both levels of IDs as a vector: ```r as_survey(data, ids = c(county_id, household_id), weights = wt, strata = region) ``` ### 2.5 Worked example: NHANES 2017–2018 NHANES uses a stratified, multistage probability cluster sample. The design variables are documented in the analytic notes on the NHANES website [@lumley2010, ch. 4]: | Variable | Role | Argument | |------------|-----------------------------------------------------------|-----------| | `sdmvpsu` | Masked variance PSU (cluster ID) | `ids` | | `sdmvstra` | Masked variance stratum | `strata` | | `wtmec2yr` | 2-year MEC examination weight (blood pressure, lab tests) | `weights` | | `wtint2yr` | 2-year interview weight (income, education, etc.) | `weights` | ```{r nhanes} # Subset to MEC exam participants (ridstatr == 2) before using wtmec2yr. # The 550 interview-only participants have wtmec2yr = 0 and are not part # of the exam sample. nhanes_exam <- nhanes_2017[nhanes_2017$ridstatr == 2, ] svy_nhanes <- as_survey( nhanes_exam, ids = sdmvpsu, strata = sdmvstra, weights = wtmec2yr, nest = TRUE # PSU IDs are locally unique within strata ) svy_nhanes ``` For interview-only variables (income, education), use the full dataset with `wtint2yr` — all 9,254 participants have a positive interview weight: ```{r nhanes-interview} svy_nhanes_int <- as_survey( nhanes_2017, ids = sdmvpsu, strata = sdmvstra, weights = wtint2yr, nest = TRUE ) ``` ### 2.6 Worked example: ANES 2024 The 2024 American National Election Studies uses a stratified cluster design with separate pre- and post-election weights. Use the correct weight for the variables you are analyzing: | Variable | Role | Argument | |------------|---------------------------------------------------------------|-----------| | `v240103c` | PSU (FTF+Web combined) — cluster ID | `ids` | | `v240103d` | Stratum (FTF+Web combined) | `strata` | | `v240103a` | Pre-election weight — use for pre-election variables | `weights` | | `v240103b` | Post-election weight — use for validated vote choice | `weights` | ```{r anes} # Pre-election analysis (party ID, ideology, candidate preference) svy_anes_pre <- as_survey( anes_2024, ids = v240103c, strata = v240103d, weights = v240103a ) # Post-election analysis (validated vote choice: v242066, v242067) svy_anes_post <- as_survey( anes_2024, ids = v240103c, strata = v240103d, weights = v240103b ) ``` **Missing values:** ANES uses negative integer codes throughout — `−9` = Refused, `−8` = Don't know, `−1` = Inapplicable. Recode these to `NA` before analysis. Check `attr(anes_2024$v241177, "labels")` for the full set of codes for any variable. ### 2.7 Worked example: GSS 2024 The General Social Survey uses a stratified multi-stage cluster design. Two weights are available depending on whether non-response bias is a concern: | Variable | Role | Argument | |-------------|--------------------------------------------------------------------|-----------| | `vpsu` | Variance primary sampling unit | `ids` | | `vstrat` | Variance stratum | `strata` | | `wtssps` | Person post-stratification weight — standard analysis weight | `weights` | | `wtssnrps` | Person post-stratification weight, non-response adjusted | `weights` | ```{r gss} # Standard analysis weight svy_gss <- as_survey( gss_2024, ids = vpsu, strata = vstrat, weights = wtssps ) # Non-response adjusted weight (preferred when non-response bias is a concern) svy_gss_nr <- as_survey( gss_2024, ids = vpsu, strata = vstrat, weights = wtssnrps ) ``` **Missing values:** GSS uses `−100` = Inapplicable, `−99` = No answer, `−98` = Don't know, `−90` = Refused. These are stored as value labels on every column — check `attr(gss_2024$happy, "labels")` and recode to `NA` before analysis. ### 2.8 Worked example: Pew NPORS 2025 The 2025 National Public Opinion Reference Survey is an **address-based sample (ABS)** — units are drawn directly from the USPS Computerized Delivery Sequence file with no intermediate cluster stage. Each address is its own sampling unit, so there is no PSU variable. Omit `ids`: | Variable | Role | Argument | |-----------|-------------------------------------------------------------------|-----------| | `stratum` | Sampling stratum (10 levels, defined by census block group) | `strata` | | `weight` | Final raked weight — base weight calibrated to Census targets | `weights` | ```{r npors} svy_npors <- as_survey( pew_npors_2025, strata = stratum, weights = weight ) ``` --- ## 3. `as_survey_replicate()` — Replicate Weight Designs {#sec-rep} Use `as_survey_replicate()` when your data provider has supplied pre-computed replicate weight columns — columns like `repwt_1`, `repwt_2`, ..., or `pwgtp1`–`pwgtp80`. Replicate-based variance estimation works by repeatedly re-estimating the target statistic under small perturbations of the sample, embedding variance information directly in the weights [@wolter2007, ch. 1]. **Use the agency-supplied replicate weights when they are available.** Survey agencies tune these weights for their specific design. Using them correctly replicates published point estimates and standard errors and is generally considered the preferred approach for variance estimation with major public surveys [@lohr2022, §9.4]. ### 3.1 The `type` argument The `type` argument specifies which replication variance formula applies. Getting this wrong produces systematically incorrect standard errors. Identify the correct type from your codebook's technical documentation. | Type | Full name | Identifying signs in codebook | Common surveys | |-------------------------|------------------------------|---------------------------------------------------|------------------------------------------| | `"JK1"` | Jackknife-1 | "JK1"; one PSU dropped per replicate | NHES, some Pew studies | | `"JK2"` | Jackknife-2 | "JK2"; paired PSUs; exactly 2 PSUs per stratum | NAEP, PISA, most NCES surveys | | `"JKn"` | Jackknife-n | One stratum dropped per replicate | Less common; some multi-PSU designs | | `"BRR"` | Balanced Repeated Replication| "BRR"; exactly 2 PSUs per stratum required | Some CPS variants | | `"Fay"` | Fay's Modified BRR | "Fay BRR" or "Fay's method"; BRR with epsilon | Some Census Bureau surveys [@fay1989; @judkins1990] | | `"bootstrap"` | Bootstrap | "bootstrap replication weights"; 100–500 replicates | BRFSS | | `"successive-difference"` | Successive Difference | "SDR" or "successive difference replication" | ACS 1-year PUMS [@census2022] | | `"ACS"` | ACS variant | Specific to ACS 5-year methodology | ACS 5-year PUMS | The Fay epsilon parameter (`fay_rho`) controls how much each replicate weight differs from the full-sample weight. Its value is specified in the survey's technical documentation [@fay1989; @judkins1990]. ### 3.2 Worked example: ACS PUMS 2022 — Wyoming The ACS 1-year PUMS provides 80 successive-difference replicate weights for variance estimation, documented in the ACS Design and Methodology report [@census2022]: | Variable | Role | Argument | |----------------------|---------------------------------------------------------|---------------| | `pwgtp` | Person weight | `weights` | | `pwgtp1`–`pwgtp80` | Successive-difference replicate weights (80 replicates) | `repweights` | ```{r acs} svy_acs <- as_survey_replicate( acs_pums_wy, weights = pwgtp, repweights = pwgtp1:pwgtp80, type = "successive-difference" ) svy_acs ``` ### 3.3 Worked example: Pew Jewish Americans 2020 This Pew study provides 100 jackknife-1 replicate weights alongside the full-sample weight: | Variable | Role | Argument | |--------------------------------|-----------------------------------------------|--------------| | `extweight` | Full-sample base weight | `weights` | | `extweight1`–`extweight100` | JK1 jackknife replicate weights (100 replicates) | `repweights` | ```{r pew-jewish} svy_jewish <- as_survey_replicate( pew_jewish_2020, weights = extweight, repweights = extweight1:extweight100, type = "JK1" ) svy_jewish ``` ### 3.4 The `scale` and `rscales` arguments Most users can omit `scale` and `rscales`. surveycore computes defaults based on `type` and the number of replicates. Override them only when your codebook's technical documentation specifies custom values [@wolter2007, ch. 3]. --- ## 4. `as_survey_twophase()` — Two-Phase Designs {#sec-twophase} > **If you are not sure whether your design is two-phase, it almost certainly > is not.** Skip to [Section 5](#sec-srs) or [Section 6](#sec-calibrated). ### 4.1 What two-phase sampling is Two-phase (or double-sampling) designs collect data in two stages [@lumley2010, ch. 9]: 1. **Phase 1:** A large, inexpensive sample that records basic variables (demographics, a screening question, administrative records). 2. **Phase 2:** A subsample drawn from Phase 1 that collects expensive or difficult measurements — lab tests, in-person interviews, expert coding. The variance estimator accounts for uncertainty from both sampling stages [@saegusa2013]. You must have retained the Phase 1 data and know which Phase 1 units were selected into Phase 2. Common contexts: case-cohort studies, medical validation studies, surveys with a screening phase [@breslow1988]. ### 4.2 Arguments | Argument | What it does | |-----------------------------------|--------------------------------------------------------------------------| | `phase1` | A `survey_taylor` object representing the Phase 1 design | | `subset` | Bare name of a logical column: `TRUE` = selected into Phase 2 | | `ids2`, `strata2`, `probs2`, `fpc2` | Phase 2 design variables (all optional) | | `method` | `"full"` (default), `"approx"`, or `"simple"` | The `method` argument: - `"full"`: Correct variance accounting for both phases. Requires Phase 1 cluster information. - `"approx"`: Faster approximation; adequate when the Phase 1 sampling fraction is small. - `"simple"`: Ignores the Phase 1 design. Use only if Phase 1 is a census. ### 4.3 Worked example: National Wilms Tumor Study The `nwtco` dataset from the `survival` package records outcomes for 4,028 children enrolled in the National Wilms Tumor Study — a multi-institution clinical trial. This is a case-cohort design: a random subcohort was selected from all enrolled children (Phase 1), and expensive central-laboratory histology was measured only for subcohort members plus all relapse cases [@breslow1988]. ```{r nwtco, eval=requireNamespace("survival", quietly=TRUE)} nwtco <- survival::nwtco # in.subcohort is stored as 0/1 — must be logical for as_survey_twophase() nwtco$in.subcohort <- as.logical(nwtco$in.subcohort) # Phase 1: all 4,028 enrolled patients (each patient is their own unit) phase1 <- as_survey(nwtco, ids = seqno) # Phase 2: subcohort, with Phase 2 sampling stratified by relapse status svy_twophase <- as_survey_twophase( phase1, strata2 = rel, # Phase 2 strata: cases (rel=1) vs. non-cases (rel=0) subset = in.subcohort, # Logical column: TRUE = selected into Phase 2 method = "full" ) svy_twophase ``` --- ## 5. Simple Random Sample with `as_survey()` {#sec-srs} Use `as_survey()` without `ids` or `strata` when every unit in your target population had an equal, known probability of selection — no clustering, no stratification [@cochran1977, ch. 2; @lohr2022, ch. 2]. This design is common in: - Surveys of a complete organizational roster (all employees at a company, all students at a school) where units are drawn directly from a list - Small-scale research with a well-defined, numbered sampling frame - Pilot studies and classroom experiments When neither `ids` nor `strata` is specified, `as_survey()` creates a `survey_taylor` object with no cluster or stratum structure — the SRS special case of the Taylor series estimator. ### 5.1 The `fpc` argument matters more here Without clustering or stratification, the FPC has a proportionally larger effect on variance estimates than in complex designs [@cochran1977, §2.8]. Supply it when you know the population size or sampling fraction. For the example below, the population is N = 400 schools. ### 5.2 Worked example: School district survey A district administrator draws a simple random sample of 80 schools from a complete roster of 400 schools. Every school has an equal probability of selection (80/400 = 0.20) — the textbook SRS case [@cochran1977, ch. 2; @lohr2022, ch. 2]: ```{r apisrs} set.seed(101) N <- 400 # total schools in district n <- 80 # schools sampled school_survey <- data.frame( school_id = sample(seq_len(N), n), avg_score = round(rnorm(n, mean = 72, sd = 11), 1), pct_frpl = round(runif(n, 0.10, 0.85), 2), # % free/reduced price lunch enrollment = round(runif(n, 180, 850)), sw = N / n, # equal sampling weight = 400/80 = 5.0 fpc = N # population size for FPC ) svy_srs <- as_survey( school_survey, weights = sw, # each sampled school represents 5 schools in the population fpc = fpc # reduces SEs: we sampled 20% of the population ) svy_srs ``` Two things worth making explicit so this example is not misread: **The unit of analysis is the school, not the student.** Variables like `avg_score`, `pct_frpl`, and `enrollment` are school-level aggregates drawn from administrative records for each sampled school. This is a survey *of schools*. If you wanted individual student-level data from each selected school, you would need a two-stage cluster design — sample schools, then sample students within each school — and use `as_survey()` with `ids = school_id` to account for the clustering. **The weight is constant because this is SRS.** Each school was selected with probability 80/400 = 0.20, so each receives weight 1/0.20 = 5.0. The weight is the same for every school because no school was oversampled or undersampled relative to any other. Uniform weights are not a simplification — they are the defining signature of simple random sampling. --- ## 6. `as_survey_nonprob()` — Non-Probability and Calibrated Samples {#sec-calibrated} If you conduct research on opt-in panels — Lucid, Dynata, Qualtrics panels, Prolific, or similar — and your vendor has provided raking or post-stratification weights, this section is for you. The short answer: **you are probably doing it roughly right, and `as_survey_nonprob()` is the correct constructor to use.** Here is what you can and cannot claim from your estimates, and how to report them honestly. ### 6.1 The fundamental distinction A **probability sample** gives every unit in the target population a known, positive inclusion probability. Design-based variance estimators are valid because the randomness that justifies them comes from the sampling mechanism itself [@cochran1977, ch. 1; @lohr2022, ch. 1]. A **non-probability sample** — an opt-in online panel — has unknown inclusion probabilities. The decision to join a panel and to complete a particular survey is self-selected. No mechanical property of the data guarantees representativeness [@baker2013; @elliott2017]. ### 6.2 What your vendor's weights actually are Regardless of where they come from, `as_survey_nonprob()` is the right constructor whenever weights were derived *after* data collection to make the sample resemble a target population. Common forms include [@valliant2018, ch. 3]: - **Raking** (iterative proportional fitting): adjusts sample marginals to match population marginals on age, gender, education, race/ethnicity, etc. The standard approach used by most panel vendors. - **Post-stratification**: assigns a single weight to all respondents within a demographic cell defined by the cross-product of variables. - **Propensity score weighting (PSW)**: fits a model predicting the probability of being in the sample, then weights each respondent by the inverse of their predicted probability. Functionally equivalent to calibration — the weights make the sample resemble the population on the modeled covariates. - **Matching-based weights**: assigns weights based on similarity to a reference population sample (e.g., entropy balancing, MatchIt outputs). Another approach to demographic alignment. All four share the same fundamental property: the weights were computed from the data, not fixed by the sampling protocol. Use `as_survey_nonprob()` for all of them. What calibration weights accomplish [@mercer2018; @mcphee2023]: - They reduce bias from *measured* demographic confounders - Point estimates for outcomes correlated with calibration variables improve meaningfully compared to unweighted estimates - They do **not** correct for selection on unobserved variables - They do **not** make the design a probability sample ### 6.3 What you can and cannot claim | Claim | Valid? | Notes | |-------------------------------------------------------------------|----------------|---------------------------------------------------------------------------| | Point estimates representative of calibration margins | ✅ Yes | Calibrated to age, gender, education, etc. targets | | Estimates more accurate than unweighted | ✅ Usually | Especially for outcomes correlated with demographic variables | | Standard errors reflect true sampling uncertainty | ⚠️ Approximately | SEs computed under approximate variance model; likely underestimated | | Results equivalent to a probability-sample estimate | ❌ No | Selection mechanism is unknown and cannot be fully corrected | This is the standard practice across the industry — used routinely by academic researchers, major survey organizations, and commercial firms [@baker2013; @mcphee2023]. The key is transparency: **your methods section should state that you used a non-probability sample with vendor-supplied calibration weights, describe the calibration targets, and acknowledge that standard errors are approximate.** ### 6.4 Worked example: Democracy Fund + UCLA Nationscape The Nationscape is a large-scale non-probability survey conducted by Democracy Fund + UCLA, fielded weekly from July 2019 through January 2021. Each wave recruited approximately 6,250 respondents from the Lucid respondent exchange using a quota design, with raking weights calibrated to American Community Survey (ACS) marginals for age, gender, education, race/ethnicity, and region, plus 2016 presidential vote choice. This is the textbook use case for `as_survey_nonprob()`. | Variable | Role | Argument | |----------|-----------------------------------------------------------------------------|-----------| | `weight` | Raking weight calibrated to ACS demographic targets and 2016 presidential vote | `weights` | ```{r nationscape} svy_ns <- as_survey_nonprob(ns_wave1, weights = weight) svy_ns # Presidential approval rating (July 2019) get_freqs(svy_ns, pres_approval) ``` This produces a `survey_nonprob` object. Use it with `get_means()`, `get_freqs()`, and other estimation functions exactly as you would any other survey object. Standard errors are computed under an approximate variance model and should be interpreted with appropriate caution and disclosed in your methods section. The `weight` column is a raking weight, not a design weight — it was computed after data collection to match population marginals, not fixed by the sampling protocol. Using `as_survey_nonprob()` makes this explicit to both R and future readers of your code. ### 6.5 What not to do Do not use `as_survey()` for a non-probability sample and present standard errors as if the design were a probability sample: ```r # Creates a survey_taylor object, which misrepresents the design svy_wrong <- as_survey(ns_wave1, weights = weight) ``` Using `as_survey_nonprob()` makes the non-probability nature of the design explicit — both to R and to future readers of your code. This distinction matters for transparency in reporting and for correctly interpreting what your uncertainty estimates actually mean [@elliott2017; @baker2013]. ### 6.6 Worked example: University voluntary response survey A university sends an email to all 8,000 enrolled students inviting them to complete a campus climate survey. 2,400 respond (30%). The response is self-selected — students with strong opinions are more likely to complete the survey than those who are neutral. **If calibration weights are available:** If the university has computed post-stratification or raking weights using registrar demographics (year, major, housing status), use `as_survey_nonprob()`. This is the appropriate constructor whenever the weights were derived to make the respondents resemble the full student body: ```r svy_campus <- as_survey_nonprob(campus_survey, weights = ps_weight) ``` **If no calibration weights are available and you still want to use surveycore functions:** Add a column of 1s and use `as_survey()` without `ids` or `strata`: ```r campus_survey$wt <- 1 svy_campus <- as_survey(campus_survey, weights = wt) ``` This treats all respondents as equally weighted. The SEs it produces reflect variability *among the 2,400 respondents* — they do not measure how representative those respondents are of the full student body. This framing is valid when your target population is "students who chose to respond," not "all students at the university." **Disclosure:** Whether you use calibration weights or equal weights, your methods section should state the response rate, describe the weighting approach, and acknowledge the limitation: voluntary response bias cannot be fully corrected by any weighting strategy [@baker2013]. --- ## 7. Probability, SRS, and calibration weights: understanding the distinction {#sec-weight-types} The two constructor families most users encounter — `as_survey()` / `as_survey_replicate()` and `as_survey_nonprob()` — differ in one fundamental way: *where the weights come from*. | | `as_survey()` / `as_survey_replicate()` | `as_survey_nonprob()` | |---|---|---| | Weight source | Sampling protocol (1/π_i) | Post-hoc adjustment | | Selection probabilities | Known and controlled | Unknown or overridden by calibration | | Weight values | Vary across respondents (or uniform for SRS) | Vary (reflect adjustment, not design) | | Variance estimator | Design-based (exact) | Approximate | In `as_survey()`, every weight traces back to a specific moment in the sampling protocol — the moment each unit's selection probability was fixed. A PSU drawn with probability 1-in-10 gets weight 10. A school drawn from a roster of 400 with probability 1-in-5 gets weight 5. SRS designs are the special case where all weights are equal because every unit had the same selection probability. The randomness that makes design-based inference valid is mechanical and recorded. In `as_survey_nonprob()`, weights were computed *after* data collection to make the sample resemble a target population. The underlying selection mechanism is either unknown (opt-in panel, voluntary response) or was overridden by the calibration adjustment. Standard errors are approximate because the calibration step itself introduces additional uncertainty that standard variance formulas do not fully capture. The practical test: **if you can point to the sampling protocol that fixed each unit's probability of selection, use `as_survey()`.** If the weights were derived from the data after collection, use `as_survey_nonprob()`. --- ## 8. When no constructor applies: convenience and purposive samples {#sec-no-constructor} Not every data collection fits the survey design framework. ### 8.1 Example: program evaluation classrooms A researcher surveys students in five classrooms that volunteered to participate in a new educational program and wants to assess whether the program changed their attitudes. The classrooms were not randomly selected from any defined population. There is no sampling mechanism to justify a design-based variance estimator, and no calibration weights that would correct for the non-random selection. The inferential question — whether the program *caused* attitude change — is a **causal inference** problem requiring a control group and appropriate methods (difference-in-differences, matching, regression discontinuity), not a survey design object. If the goal is purely **descriptive** — summarizing the attitudes of students in these specific classrooms without generalizing — you can treat the participants as a census. Add a column of 1s and use `as_survey()` without `ids` or `strata`: ```r classroom_data$wt <- 1 svy_participants <- as_survey(classroom_data, weights = wt) ``` Equal weights treat all participants as equally represented. The SEs reflect variation *among participants*. Do not interpret results as representative of all students at the school. ### 8.2 General decision rule | Design | Appropriate tool | Notes | |---|---|---| | Probability sample with design weights | `as_survey()`, `as_survey_replicate()` | Exact variance | | Pure SRS — equal probability, no clustering/strata | `as_survey()` (no `ids` or `strata`) | Exact variance; SRS special case of Taylor | | Any sample with calibration/raking/PSW/matching weights | `as_survey_nonprob()` | Approximate variance | | Voluntary response or convenience sample, no weights | `as_survey()` with `weights = 1` (no `ids`/`strata`) | Conditional inference only; disclose | | Causal inference (treatment effect estimation) | Not surveycore | Use MatchIt, WeightIt, lme4, etc. | When you use `as_survey()` with equal weights and no `ids` or `strata` for a non-probability sample, surveycore produces estimates and SEs without error. The SEs are valid as a measure of variability *among the observed participants*. They should not be interpreted as uncertainty about a broader population unless the sample can be independently defended as representative. --- ## 9. Reference: Common Codebook Variables {#sec-reference} A lookup table for common codebook terms and how they map to constructor arguments: | Codebook term | Maps to | Notes | |------------------------------------------------------------------|--------------------------------------|-------------------------------------------| | "sampling weight", "survey weight", "person weight" | `weights =` | | | "PSU", "primary sampling unit", "cluster ID" | `ids =` | | | "stratum", "design stratum", "sampling stratum" | `strata =` | | | "FPC", "finite population correction", "population size" | `fpc =` | | | "replicate weights", "bootstrap weights", "BRR weights" | `repweights =` | Use `as_survey_replicate()` | | "base weight", "design weight" (with separate replicates) | `weights =` in `as_survey_replicate()` | | | "Fay coefficient", "Fay factor", "epsilon" | `fay_rho =` | With `type = "Fay"` | | "raking weights", "post-stratification weights", "cal weights" | `weights =` in `as_survey_nonprob()` | Non-probability design | | "two-phase", "double sampling", "case-cohort" | Phase 1 → `as_survey()`, then `as_survey_twophase()` | | --- ## References