arf foq2 day router presentation
DESCRIPTION
TRANSCRIPT
FOQ2: Routing Insights and Recommended Practices
Nancy Brigham, Ph.D.
Global Head of Research-on-Research Ipsos Interactive Services
Chuck Miller
President Digital Marketing & Measurement (DM2)
Agenda & Team
• Overview: 15 minutes
• Insights: 30 minutes
The Team:
• Co-chairs: Michael Fallig (Cogenti) & Chuck Miller (DM2)
• Quantitative Design: Nancy Brigham (Ipsos) group leader, Scott Porter (Added Value), Paul Johnson (SSI), Mike Conklin (GfK), Robin Murphy (Ipsos), John Bremer (Toluna)
• Qualitative Design: Paul Johnson (SSI), Efrain Ribeiro (Lightspeed), Steve Gittelman (MKTG Inc.)
• Support Deliverables: Peter Milla (Peter Milla Consulting)
What is Routing?
• Routing is a method of assigning potential respondents to a set of surveys that need sample
Just as we do with panels, but this is done in real-time in an automated manner
• The technology used to do this is a router
Routers use software and business rules to assign a stream of potential respondents to available surveys
Traditional vs. Routed Approach
Screener Study 8
Outcome: NQ
Outcome: QUALIFY
Completes
Study 8
DONE; No
study
completed
…and so on for each individual study
NQ = Not qualified (e.g., ineligible, over quota, failed security screen, study closed, etc.)
Screener 42
Outcome: NQ
Screener 23
Outcome: NQ
Screener 71
Outcome:
QUALIFIED
Enters and
Completes
Study 71
Routing Non-routing (traditional)
How Routing Works: Big Picture
SOURCE 1
SOURCE 2
SOURCE 6
SOURCE 5
SOURCE 4
SOURCE 3
SOURCE 7
Multiple Sample Sources*
Survey D
Survey E
Qualified Respondents Sent to Surveys
Router – Determines Survey Qualification
E
B
A C
D
Survey C
Survey B
Survey A Screening for Survey…
* Panel and/or river sourcing – using a variety of management techniques.
Changing Market / Internet Environment
Continued decline of response rates
Changing Internet landscape and respondent behaviors
Continued growth of online studies
Key challenge facing the online industry today:
Sample Capacity
Consequences of doing things the “the way we’ve always done it”
Inefficient use of sample – capacity not optimized Respondent engagement suffers – lowering capacity Projects are increasingly harder to complete
Why is the Industry Adopting Routing?
Benefits Noted by Providers
Increased Sample Capacity & Population
Reach
Easier Survey Adaptation to New
Online Realities
• Allows studies to be tailored to people accessing them using different devices (e.g., mobile)
• Allows more sophisticated management of a broad base of studies, and the balancing of sample supply and demand (e.g., applying usage rules to sources, and exclusions to respondents)
• More efficient use of sample in general, and especially lower-response demographics
• Allows access to parts of the population that don’t traditionally join panels
• Greater control over operational aspects (e.g., automation)
• Automation produces greater sampling replicability
• Allows for measurement and control of selection bias (inherent in all methodologies, including panels)
• Increased respondent engagement and satisfaction
Greater Consistency & Quality Control
Findings from Router and Sample Provider Focus Group
Some Definitions Before We Start…
• Screening
A set of questions usually asked/utilized prior to main questionnaire content to determine or verify study eligibility or branching within a specific study.
• Reallocation / Reassignment The act of reassigning a respondent to further screening upon disqualification during an
attempt.
• Priority Assignment The act of assigning a screener for attempted qualification in a non-random manner,
based on a set of predefined business rules.
A full Glossary of router-related terms was created by the Router team and is available now!
Serial vs. Parallel Routers
Routers in use today can be broadly classified as either Serial or Parallel
Insight / Conclusion
Routers have variables or characteristics that need to be managed according to business rules
Insight / Conclusion
Router Team Findings
Randomization & Prioritization
Common screening
Studies - # and type
Frequency of survey-taking
Study qualification correlation
Parallel or Serial
Reallocation
Such as…
The Common Survey Environment
There is selection bias in all methodologies, but in a router we can more easily measure it
Screening process is shared among studies
Respondents may qualify for multiple available studies
Studies “compete” with each other in terms of access to respondents
Some studies will need priority over others in getting respondents (e.g., if they only need a few more completes)
Selection bias comes from other studies taking priority over your study for respondents
Characteristics
Routers use Randomization and Prioritization for survey assignment, usually in combination
Insight / Conclusion
Router Team Findings
Based on business rules for a router
Used in varying degrees by different routers
• 100% randomization best minimizes bias, but may not be practical
• 100% prioritization usually leads to greatest bias
• Hybrid approach is most commonly used
Quick Review of Methodology
Created a router to test various scenarios in a controlled fashion
Four “live studies” and 20 screeners from other studies = 24 studies in router
The 4 live studies were chosen to represent diversity across the market research business
• Concept Test (CPG), Customer Satisfaction (Telecommunications), Usage & Awareness (Health), FOQ2 questionnaire (shortened to 15 minutes)
The 20 screeners were chosen to represent diversity typically seen in a large routing system
Respondents were randomly assigned to either the Router (Test) or Non-Router (Control) conditions
Mirrored the 4 live router studies with the same 4 studies fielded in the traditional way (non-routed)
Router condition: Respondents were assigned to either a Serial or Parallel screening process
Both direct comparisons and simulations were used to derive insights from the data
What is the impact of moving to a common survey environment?
Let’s look at some data!
Research Questions
•Compared bias in the Serial and Parallel conditions through simulations, for 16 key metrics
•Compared survey data responses of Test (Serial and Parallel) conditions to Control condition for: 6 demographics, 12 behavioral variables, 10 attitude questions / batteries, 6 key survey metrics
•Compared length of time spent in the screening process
Does routing impact the data?
•Compared satisfaction with the survey experience for Test and Control respondents
Does routing impact the respondents?
•Examined sample capacity lifts due to routing
Does routing help sample capacity?
Is There Selection Bias in Routing?
We saw very little bias overall in both Serial and Parallel routing
Insight / Conclusion
Some variables are more biased than others
But did the bias we saw matter – did it affect survey responses ?
Distribution of Error Across All Surveys (Worst Questions in Each)
Router No Router
Does Routing Impact the Data?
Routing did not produce meaningfully different results vs. the traditional method
Of 252 comparisons, 26 were significant (10%)
Within what we would expect by chance given the large # of comparisons we tested.
30
138
72
12 2
14 8
2
Demographics Behaviors Attitudes Key metrics
# of comparisons
# significant
Insight / Conclusion
Does Routing Impact the Data?
About the same amount of comparisons for Serial and Parallel were different from the Control
Sometimes Control was higher, sometimes it was lower
# of significant comparisons 26
Control vs. Parallel 12
Control higher 8
Control lower 4
Control vs. Serial 14
Control higher 3
Control lower 11
There was no consistent pattern of differences
Two Major Types of Screening Processes
Parallel Serial Screener 42
Screener 23
Screener 71
Screener 8
QUALIFIED:
Screener 71
Screener 8
NQ:
Screener 42
Screener 23
Selection of Study 71
Screener 42
Outcome: NQ
Screener 23
Outcome: NQ
Screener 71
Outcome:
QUALIFIED
Enters and
completes
Study 71
Respondents attempt to qualify for one study at a time. If they don’t qualify for one study, they move to the screener for the next study.
Respondents attempt to qualify for several studies at one time. There can be several levels of qualification, but each level is for multiple studies
Enters and
completes
Study 71
Selection Bias for Serial Screening
For Serial routing, randomization is important. Adding 50% randomization decreases the bias by 20% on average.
Insight / Conclusion
-20%
-15%
-10%
-5%
0%
5%
10%
15%
20%
Error with 100% Priority Serial
-20%
-15%
-10%
-5%
0%
5%
10%
15%
20%
Error with 50% Priority Serial
Bias (distance from 0%)
Bias is very small compared to
sampling error
Selection Bias for Parallel Screening
For Parallel routing, randomization is important but it does not matter whether it is at the pre-screening or survey selection level
Insight / Conclusion
-20%
-15%
-10%
-5%
0%
5%
10%
15%
20%
Error with Random/Priority Parallel
-20%
-15%
-10%
-5%
0%
5%
10%
15%
20%
Error with Priority/Random Parallel
Pre-screening level is random Survey level is random
Bias
Does Type of Screening (Serial or Parallel) Impact Data?
Serial and Parallel screening may produce meaningfully different results for some variables
Of 126
comparisons, 15
were significant
(12%)
A bit more than what we
would expect by
chance, even given the
large # of comparisons
we tested.
Serial was higher/more
positive than Parallel for
all 15 differences
Nine of these differences
were on one variable –
Looking for health
information online
Insight / Conclusion
Time Spent in the Screening Process
Respondents spend longer in a router screening process than the traditional screening process
Insight / Conclusion
Numbers are specific to the router setup used for this study
3.87
1.20
Router Control
Average minutes spent in screening
process
Time Spent in the Screening Process
Respondents spend somewhat longer in a Serial screening process than a Parallel screening process
Insight / Conclusion
Numbers are specific to the router setup used for this study
4.16
3.58
Serial Parallel
Average minutes spent in screening process, by condition
BUT DOES IT REALLY MATTER?
Does it impact survey data?
Does it impact respondent satisfaction?
Does Length of the Screening Process Impact Data?
Length of the Serial screening process did not impact survey data – respondents who spent longer did not answer differently
2.72
4.97
6.23
1.2
0
2
4
6
Brief Medium Long Control
Cell
Min
utes
Router
Control
Serial
Mean Minutes In ScreeningExamining the Serial condition (where respondents spent the most time)… Three screening length conditions were tested and compared to each other and to the control.
Insight / Conclusion
Length of the Screening Process
• 1314 pairwise t-tests run, across 37 tables of demos and KPIs.
2.5% were flagged as significant (less than what is expected by chance alone).
These differences were spread out across the variables (no concentration on any single variable).
0
500
1000
Actual Expected
StatusT
ests
Result
Not Significant
Significant
Number of Significant Stat Tests
Length of the Screening Process
There does appear to be an effect where males are more likely to drop out (i.e., not end up in the final sample) with longer screeners.
This may point to a need to control gender through quotas.
47.2
44
36.4
45.6
0
10
20
30
40
Brief Medium Long Control
Cell
Pct
.Mal
e Router
Control
Serial
Percent of Valid Respondents Who are Male
Does Routing Impact the Respondents?
Respondents were as satisfied with the routing survey experience as were control respondents with the traditional experience.
Time spent in the Serial screening process did not negatively impact enjoyment.
Screening % agree
Control 75%
Serial 74%
Parallel 73%
“I enjoyed filling out this survey”
Time spent in
screening (Serial)
% agree
Brief 74%
Mid 75%
Long 77%
Interestingly,
respondents who spent
the longest time in the
serial screening
process liked it
significantly more than
the Brief respondents!
Insight / Conclusion
How Does Routing Impact Sample Capacity?
The router’s overall incidence (the average % of studies for which a respondent qualifies) is what drives sample efficiency
Insight / Conclusion
33%
67%
Eligible
Non-Eligible
Within our study
router, on average, a
respondent was
eligible to complete
about 8-9 of the 24
studies
Routing Increases Sample Capacity
Routing makes more efficient use of respondents than fielding each survey separately
Insight / Conclusion
With 12 studies in our router, we only need
40% of the respondents
we would need if we were filling them all separately
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10 11 12
Ind
ex v
ers
us
s
ep
ara
te s
tud
ies
Number of studies in router
Respondents required on average versus fielding separate studies
Note: This is a conservative estimate! The actual impact would likely be higher for a larger router, and with new surveys being added as others complete.
Conclusions –Takeaways
Top Takeaways
Selection bias exists, but it is small, less than sampling error, and varies by question
Routing does not impact survey data in any meaningful way
There may be differences on some variables between Serial and Parallel routers
Quotas may be needed to control response differences by demographics
Longer time spent in the screening process does not impact survey data
Respondents were happy with the routing process, even though it took slightly longer
Routing increases sample capacity by making more efficient use of respondents
1
2 3 4
5 6
7
Recommended Practices
Reduce cognitive load and irritation for
respondents
• Standardized wording for screeners
• Store respondent profile information
• Give them periodic opportunities to leave the screening process,
• Length of screening process needs to be reasonable
Consistent and frequent monitoring & management
• Router business rules are important
• Rules are established by sampling experts
• Controls are in place to ensure rules are followed
What You Should Do Differently Tomorrow
• Ask your suppliers if they are using routers on your studies
• If so, use the checklist created by the router team to better understand the details
And…
• We encourage you to always ask questions about your sample - better sample leads to higher quality data, which leads to better business decisions!
APPENDIX
Serial Routing
Definition In its most basic form, serial routing is a process whereby a respondent is screened sequentially for the available studies on a router. Upon qualification for a study, the respondent is immediately sent into the full survey. Upon disqualification in any screener, another screener is served to the respondent – with the process repeating until qualification occurs.
Example Respondent enters a router, and is randomly served the screener for Study 42. He does not qualify and attempts Study 23, resulting in another disqualification. The router then randomly serves the screener for Study 71, the respondent meets the criteria, and is sent into Study 71 for a full and joyfully robust survey experience.
Parallel Routing
Definition In its most basic form, parallel routing is a process whereby a respondent is exposed to a set of screening questions from all or a subset of the surveys on the router, simultaneously, on the same page. After the respondent answers these questions, he or she is assigned to one of the studies for which he or she appears to be qualified.
Example Respondent enters a router, and a page of questions is presented that contains the screeners for Study 42, Study 23, Study 71, and Study 8. The questions were selected at random by the system with presentation order randomized. After completion of the page qualification is tested, with the respondent qualifying for only Study 71 and Study 8. Study 71 is randomly assigned for a full survey experience.