In Verasight Self-Serve, Nationally Representative and Validated Voters surveys of 250 or more respondents include survey weights. Targeted Demographics and Convenience surveys, and surveys under 250 respondents, do not.
#What a survey weight does
Sampling and response variability means that not all groups are perfectly represented in the final data. The column labeled weight in your data file lets your analyses generalize to the population your sample represents. Respondents from groups that are underrepresented in the sample get larger weights, and respondents from overrepresented groups get smaller ones.
#Weighting method
We weight by raking, also called iterative proportional fitting. Raking adjusts each respondent's weight step by step until the weighted sample matches the population on each weighting variable. Pew Research Center describes the method.
Nationally Representative surveys are weighted to US adults, with demographic benchmarks from the US Census Bureau's Current Population Survey. Party identification uses a running three-year average of partisanship from the Pew Research Center's National Public Opinion Reference Survey. 2024 vote uses population benchmarks of the 2024 presidential vote.
Validated Voters surveys are weighted to registered-voter benchmarks.
#Weighting variables
Nationally Representative weights use eight variables, each measured by one or two profile variables in your data file:
| Variable | Profile variable(s) in your file | Benchmark source |
|---|---|---|
| Age | age_group4 | Current Population Survey |
| Education | education | Current Population Survey |
| Race and ethnicity | race, hispanic | Current Population Survey |
| Income | income | Current Population Survey |
| Census region | region | Current Population Survey |
| Metropolitan status | metro | Current Population Survey |
| Party identification | pid_base | Pew Research Center National Public Opinion Reference Survey, three-year average |
| 2024 vote | vote_2024, votechoice_2024 | Population benchmarks of the 2024 presidential vote |
Party identification is weighted with pid_base only. pid_lean, the follow-up question that asks which party a respondent's opinions are closer to, is not a weighting variable. Gender is in your data file but is not a weighting variable.
#When weights arrive
Your weights and profile variables are typically ready a couple of days after your survey finishes fielding. You then get an email titled Your Verasight Data is Ready. Then download your data file.
After your survey's end date passes and until your file is uploaded, you see these labels:
- On your dashboard, the survey's status is Pending Delivery.
- On the survey's details page, the weights card button reads Weights pending if your survey includes weights. Otherwise it reads Not yet available. When your file is uploaded, it reads Download Document.
Surveys without weights still get the profile variables.
#Use the weights in your analysis
First, join your survey export to your data file on vsid and remove duplicate and incomplete responses. Then tell your software to use the weight column, as in the code below.
In R, with the survey package:
library(survey)
# d is your merged data frame: your survey export joined to your data file on vsid
design <- svydesign(ids = ~1, weights = ~weight, data = d)
# Weighted mean of a numeric survey item
svymean(~my_item, design, na.rm = TRUE)In Stata, with svyset:
* Your merged dataset: your survey export joined to your data file on vsid
svyset [pweight = weight]
* Weighted proportions for a survey item
svy: proportion my_itemmy_item stands for any question from your own survey. ids = ~1 tells R that the sample has no clusters.
For SPSS, your ZIP includes a .sav file. The email titled Your Verasight Data is Ready links to guides on using survey weights in R, Stata and SPSS.
#What weighting cannot do
Keep these limits in mind:
- Weighting increases the variance of estimates. In the delivered file we checked, the design effect of the weights was 1.09 (1 plus the squared coefficient of variation of the weights).
- Design effects for individual estimates can be larger. If you report precision, compute it with survey software that uses the weights.