1

YouGov Sampling Methodology

Sampling and Sample Matching

the American Community Survey. In other contexts, this is known as the sampling frame,

though, unlike conventional sampling, the sample is not drawn from the frame. Traditional

sampling, then, selects individuals from the sampling frame at random for participation in the

study. This may not be feasible or economical as the contact information, especially email

addresses, is not available for all individuals in the frame and refusals to participate increase

the costs of sampling in this way.

Sample selection using the matching methodology is a two-stage process. First, a random

sample is drawn from the target population. We call this sample the target sample. Details on

how the target sample is drawn are provided below, but the essential idea is that this sample is

The purpose of matching is to find an available respondent who is as similar as possible to the

selected member of the target sample. The result is a sample of respondents who have the

same measured characteristics as the target sample. Under certain conditions, described

below, the matched sample will have similar properties to a true random sample. That is, the

matched sample mimics the characteristics of the target sample. It is, as far as we can tell,

“representative” of the target population (because it is similar to the target sample).

The Distance Function

When choosing the matched sample, it is necessary to find the closest matching respondent in

the panel of opt-ins to each member of the target sample. Various types of matching could be

estimated propensity scores can become unstable and lead to poor samples.

2

YouGov employs the proximity matching method. For each variable used for matching, we

define a distance function, d(x,y), which describes how “close” the values x and y are on a

particular attribute. The overall distance between a member of the target sample and a

member of the panel is a weighted sum of the individual distance functions on each attribute.

The weights can be adjusted for each study based upon which variables are thought to be

important for that study, though, for the most part, we have not found the matching procedure

to be sensitive to small adjustments of the weights. A large weight, on the other hand, forces

target sample as a simple random sample (SRS) from the target population. The SRS yields

unbiased estimates because the selection mechanism is unrelated to particular characteristics

of the population. The efficiency of the SRS can be improved by using stratified sampling in

place of simple random sampling. SRS is generally less efficient than stratified sampling because

the size of population subgroups varies in the target sample.

Stratified random sampling partitions the population into a set of categories that are believed

to be more homogeneous than the overall population, called strata. For example, we might

divide the population into race, age, and gender categories. The cross-classification of these

three attributes divides the overall population into a set of mutually exclusive and exhaustive

are similar on a large number of characteristics tend to be similar on other items for which we

lack data, then substituting one for the other should have little impact upon the sample. This

intuition can be made rigorous under certain assumptions.

Assumption 1: Ignorability. Panel participation is assumed to be ignorable with respect to the

variables measured by survey conditional upon the variables used for matching. What this

means is that if we examined panel participants and non-participants who have exactly the

same values of the matching variables, then on average there would be no difference between

how these sets of respondents answered the survey. This does not imply that panel participants

and non-participants are identical, but only that the differences are captured by the variables

3

Assumption 2: Smoothness. The expected value of the survey items given the variables used

for matching is a “smooth” function. Smoothness is a technical term meaning that the function

is continuously differentiable with bounded first derivative. In practice, this means that that the

expected value function doesn’t have any kinks or jumps.

Assumption 3: Common Support. The variables used for matching need to have a distribution

that covers the same range of values for panelists and non-panelists. More precisely, the

probability distribution of the matching variables must be bounded away from zero for

sample provides consistent estimates for survey measurements. The sampling variances will

depend upon how close the matches are if the number of variables used for matching is large.

In this study, over 150,000 respondents to YouGov’s Internet surveys were used for the pool

from which to construct the matches for the final sample.

Current Sampling Frame and Target Sample

YouGov has constructed a sampling frame of U.S. Citizens from the 2016 American Community

Survey, including data on age, race, gender, education, marital status, number of children under

ideology were matched from the 2014 Pew U.S. Religious Landscape Survey. Characteristics of

target samples vary based on the requirements of the projects. Typical general population

target samples are selected by stratification by age, race, gender, education, and voter

registration, and by simple random sampling within strata. At the matching stage, the final set

of completed interviews are matched to the target frame, using a weighted Euclidean distances

metric.

4

Weighting

The matched cases are weighted to the sampling frame using propensity scores. The matched

cases and the frame are combined and a logistic regression is estimated for inclusion in the

frame. The propensity score function may include a number of variables, including age, years of

education, gender, race/ethnicity, predicted voter registration, interest in politics, born again

status, ideological self-placement and inability to place oneself on an ideological scale, and

baseline party identification (i.e., the profiled party identification that was collected before the