Upload
rick-kelly
View
209
Download
0
Embed Size (px)
Citation preview
Differential Sampling Based on
Historical Individual-Level Data in
Online Panels
By Richard Kelly
Overview
The growth in the market has been rapid, and while methodological improvements have come fairly quickly, online survey research is not yet mature.
Online sampling can be improved by supplanting traditional marginal population quotas for a balanced population entering the survey by employing differential sampling
Online Research
Disadvantages
Advantages
Growth in online survey research is not
likely to ebb anytime soon
Still, there remains room for improvement
The survey “research industry is wedded
to quota controls, sometimes to the point
of obsession. It places them on Age and
Gender without a second thought as to
why or indeed whether they are doing any
good at all.” (Cape, 2011)
Researchers who do not know the
demographic composition of the
population they are attempting to study
should not enforce general population
quotas in an attempt to yield
“representative” data
However, we should not abandon controls
on demographics
Population quotas
force something
artificial onto a
sample group.
Differential Sampling
The sample entering the survey mimics the
population under study…each respondent has
an equal opportunity to enter the research effort
and is able to screen out in their natural
incidence.
optimizes sampling efficiency
Advantages of online access panels
Individual historical propensity scores
Determining invitations by demographic group
Test
Ran two surveys identical in content
Survey A = population quotas
Survey B = differentially sampled
Respondents screened on 12 month
purchase intent for auto insurance
Expectations
Demographics in final data would be
substantially different from Survey A to
Survey B
Survey health (incidence) of Survey A
would decline relative to Survey B
Significant difference in data between
Survey A and Survey B
ResultsSurvey A Survey B
Final Quota Target Final Actual Return Return Target
Male 49 49 62 50% 49%
Female 51 51 38 50% 51%
18-24 13 13 12 15% 13%
25-34 18 18 17 17% 18%
35-49 28 28 25 28% 28%
50+ 41 41 46 40% 41%
Northeast 18 18 11 17% 18%
Midwest 22 22 20 20% 22%
South 37 37 44 39% 37%
West 23 23 25 25% 23%
$29,999 or less 31 31 20 31% 31%
$30,000-$49,999 19 19 22 19% 19%
$50,000-$74,999 18 18 24 20% 18%
$75,000 or more 32 32 34 30% 32%
Results (con’t)
Survey Health
Survey A final incidence = 21.8%
Survey B final incidence = 28%
Results (con’t)
Satisfaction with current auto insurance provider
Survey A mean score: 4.06
Survey B mean score: 3.99(not statistically significant)
Average annual spend (ordinal scale)
Survey A mean score: 2.5
Survey B mean score: 2.84p<.05
Conclusions
Marginal population quotas in online
research are a persistent holdover from
other sampling frames that don’t
necessarily apply to a different frame.
Issues of representivity can be mitigated
by using differential sampling instead of
population quotas.
Much more future research is needed