Survey Weight Analyzer

Project roadmap

1. 01

   Get a weighted survey dataset

   ~1.5h

   Use a public survey that includes weights, for example the apiclus1 or nhanes-style data included with the survey package, or a public extract from a national survey. Save it in data/, and document in the README what each weight represents and where the data comes from. Install survey, srvyr, dplyr, ggplot2 and shiny. Done when: the data loads, and the README explains the weight variable.

2. 02

   Check the weights

   ~1.5h

   Write check_weights(df, weight_col) that reports missing, zero or negative weights, the smallest and largest weight, their ratio, and Kish's effective sample size (sum of weights squared divided by the sum of squared weights). Flag very extreme weights. Done when: the function returns a clear summary for the dataset, and catches a copy with broken weights.

3. 03

   Compute weighted and unweighted estimates

   ~2.5h

   Create a survey design with srvyr::as_survey_design (including clusters and strata if the data has them) and compute weighted means and proportions with survey_mean, with 95% confidence intervals. Compute the same estimates without weights, using ordinary standard errors. Put both side by side in one tidy table with the difference between them. Done when: the table shows both estimates, their intervals and the difference for several variables.

4. 04

   Break results down by group

   ~1.5h

   Allow grouping by a chosen variable (for example region or age group) with group_by before survey_mean. Report the weighted and unweighted estimate for each group, the design effect (vartype = "deff" is available in survey_mean), and flag groups with fewer than 30 respondents as unreliable. Done when: grouped results include design effects and warnings for small groups.

5. 05

   Build the Shiny app

   ~2.5h

   Create a Shiny app with inputs for the variable to estimate, the optional grouping variable and whether to show proportions or means. Show the comparison table, a chart with both estimates and their confidence intervals side by side (ggplot2 with geom_pointrange), and the weight diagnostics from step 2. Done when: changing inputs updates the table, chart and diagnostics.

6. 06

   Explain the results

   ~1h

   Add a short text in the app that updates with the results and explains them in plain language, for example: Weighting lowers the estimate by 4 points, because the sample had too many older respondents, and widens the interval (design effect 1.8). Include a help tab explaining weights, confidence intervals and the design effect. Done when: the explanation changes correctly with different selections.

7. 07

   Write tests

   ~1.5h

   Test check_weights and the estimate functions with testthat, using a tiny dataset where you can calculate the weighted mean by hand, and one where all weights are equal (weighted and unweighted results must then match). Add a testServer test for the main reactive table. Done when: the tests pass.

Resources

• Docssrvyr package documentation
• Docssurvey package on CRAN
• DocsExploring Complex Survey Data Analysis Using R (free online book)
• DocsMastering Shiny (free online book)
• ArticleDesign effect (Wikipedia)