Page 31 - Consumers and Mobile Financial Services 2014

23

Appendix 1: Technical Appendix on

Survey Methodology

In order to create a nationally representative demographic distributions from the most recent
probability-based sample, GfK’s KnowledgePanel® (November 2011 for re-interview cases and Octo-
has selected respondents based on both random digit ber 2012 for the fresh sample) data from the Current
dialing and address-based sampling (ABS). Since Population Survey (CPS). The variables used include
2009, new respondents have been recruited using gender, age, race/ethnicity, education, census region,
ABS. To recruit respondents, GfK (formerly Knowl- residence in a metropolitan area, and access to the
edge Networks) sends out mailings to a random Internet. The Panel Demographic Post-Stratification
selection of residential postal addresses. Out of 100 weight is applied prior to a probability proportional
mailings, approximately 14 households contact GfK to size (PPS) selection of a study sample from
and express an interest in joining the panel. Of those KnowledgePanel. This weight is designed for sample
who contact GfK, three-quarters complete the pro- selection purposes.
8
cess and become members of the panel. If the per-
son contacted is interested in participating but does Once the sample has been selected and fielded, and
not have a computer or Internet access, GfK provides all the study data are collected and made final, a
him or her with a laptop and Internet. Panel respon- post-stratification process is used to adjust for any
dents are continuously lost to attrition and added to survey non-response as well as any non-coverage or
replenish the panel, so the recruitment rate and under- and over-sampling resulting from the study-
enrollment rate may vary over time. specific sample design. Demographic and geographic
distributions for the non-institutionalized, civilian
For this survey, the number of KnowledgePanel® population ages 18 and over from the most recent
members who were invited to complete the survey, CPS are used as benchmarks in this adjustment.
and the invitation response rates, are presented in
table 1 (see main text). A total of 4,030 e-mail solici- Comparable distributions are calculated by using all
tations to participate in the survey were sent out to a completed cases from the field data. Using the base
random selection of KnowledgePanel respondents, weight as the starting weight, this procedure adjusts
and data collection was terminated when the quota the sample data back to the selected benchmark pro-
of 2,600 individuals completed the survey fully (a portions. Through an iterative convergence process,
“completion rate” yield of 65 percent). To enhance the weighted sample data are optimally fitted to the
the completion rate, GfK sent e-mail reminders to marginal distributions.
non-responders on days three and six of the field
period. After this final post-stratification adjustment, the dis-
tribution of the calculated weights are examined to
As with any survey method, probability-based Inter- identify and, if necessary, trim outliers at the extreme
net panel surveys are subject to potential survey upper and lower tails of the weight distribution. The
error, such as non-coverage and non-response due to post-stratified and trimmed weights are then scaled
the panel recruitment methods and to panel attrition. to the sum of the total sample size of all eligible
In order to address these potential sources of error, a respondents.
post-stratification adjustment is applied based on
There are several reasons that a probability-based
8 For further details on the KnowledgePanel sampling methodol- Internet panel was selected as the method for this
ogy and comparisons between KnowledgePanel and telephone
surveys, see www.knowledgenetworks.com/accuracy/spring2010/ survey rather than an alternative survey method. The
disogra-spring10.html. first reason is that these types of Internet surveys