Undercoverage plaques many frames - housing units are missed by listers or do not appear on the postal service list; persons with tenuous connections to households are not captured in rosters; persons hide their eligibility during screener interviews. The literature on undercoverage suggests several methods for improving the coverage of such frames, via a missed housing unit procedure, or detailed probes about household members, or disguising the target population in survey questions. However, each of these solutions introduces additional costs into the survey process. In this way, survey designers face a coverage-cost trade-off. In addition, there is increasing evidence that the cases found via these coverage-improvement measures are disproportionately nonresponders to the survey request. Thus there appears to be a coverage-nonresponse trade-off as well. Together these points raise the question of how much effort we should put into increasing coverage, when such efforts increase costs and nonresponse? This presentation will review empirical evidence for these trade-offs and search for clues to the mechanisms underlying the connection between nonresponse and undercoverage.
2. Motivation
Reducing undercoverage means:
‐ Adding cell numbers to RDD frame
‐ Including homeless, institutionalized in general population
surveys
‐ Providing tablet & internet access
Efforts are costly
Are the people included with these efforts
disproportionate nonresponders?
2
3. Examples
RDD + mobile phone surveys
‐ Lower response rates among mobile only HHs
(AAPOR Cell Phone Task Force Report)
LISS online panel
‐ Lower recruitment rates among cases without internet
(Leenheer & Scherpenzeel 2013)
Random walk
‐ Interviewers may skip HHs that look like NRs (Alt et al 1991,
Manheimer & Hyman 1949 )
Half open interval procedure
‐ Interviewers may fail to cover units that look like NRs (Eckman
& O’Muircheartaigh 2011)
3
4. Screening Study
2 versions of screener questions
‐ Direct “Is anyone in this household 35-55?”
‐ Full HH roster Age of all adults in HH
Condition
Screener
Completion
Rate
Eligibility
Rate
Interview
Completion
Rate
Response
Rate Yield
Direct 59.3 31.8 86.3 32.3 285
Roster 53.5 45.1 71.5 23.9 361
Trade-off: High response rate or high coverage?
Tourangeau, Kreuter & Eckman 2013 4
5. Mechanisms Behind Trade-Off
Respondent side
‐ Burden
‐ Learning to use internet/computer difficult
‐ Survey on cell phone annoying
‐ Hidden refusals: Respondents screen out rather than refuse
Interviewer side
‐ Judged by response rate, not coverage
‐ Coverage & response are different skill sets
5
6. Choice Faced by Survey Designers
6
Choose A or B?
High RR in A hides low
coverage rate
Cost considerations
Coverage
Rate
High Low
Response Rate
High A
*
Low B
7. Screening Study
Coming back to this example:
‐ Direct: high RR, low coverage
‐ Roster: low RR, high coverage
Condition
Screener
Completion
Rate
Eligibility
Rate
Interview
Completion
Rate
Response
Rate Yield
Direct 59.3 31.8 86.3 32.3 285
Roster 53.5 45.1 71.5 23.9 361
Tourangeau, Kreuter & Eckman (2013) 7
8. Propensity Models
Logit models, run separately for 2 screener conditions
IVs: female, age (squared), # calls (squared), party identification,
postal code, interviewer payment, refusal flag, mobile indicator
Coverage Model
‐ Case base: expected to be eligible (n=2,904)
‐ DV: screened & eligible
‐ Pseudo-R2: 15% direct; 10% roster
Response Model
‐ Case base: found to be eligible (n=735)
‐ DV: complete main interview
‐ Pseudo-R2: 20% direct; 33% roster
8
9. Propensity Models
Logit models, run separately for 2 screener conditions
IVs: female, age (squared), # calls (squared), party identification,
postal code, interviewer payment, refusal flag, mobile indicator
Coverage Model
‐ Case base: expected to be eligible (n=2,904)
‐ DV: screened & eligible
‐ Pseudo-R2: 15% direct; 10% roster
Response Model
‐ Case base: found to be eligible (n=735)
‐ DV: complete main interview
‐ Pseudo-R2: 20% direct; 33% roster
9
}}
CPdirect
CProster
RPdirect
RProster
15. Research Agenda
Mechanisms of connection between them
‐ Interviewers? Respondents?
Bias due to undercoverage & NR
How much should we spend to increase coverage,
if it only increases nonresponse?
15
17. References
17
Alt, C., Bien, W. & Krebs, D. (1991). “Wie zuverlässig
ist die Verwirklichung von Stichprobenverfahren?
Random route versus
Einwohnermeldeamtsstichprobe”. ZUMA
Nachrichten, 28, 65-72.
American Association for Public Opinion Research
(2010). “Cell Phone Task Force Report: New
Considerations for Survey Researchers When
Planning and Conducting RDD Telephone Surveys in
the U.S. With Respondents Reached via Cell Phone
Numbers”.
Eckman, S. & O’Muircheartaigh, C. (2011).
“Performance of the Half-Open Interval Missed
Housing Unit Procedure”. Survey Research Methods,
5(3), 125-131.
Hainer, P. (1987). “A Brief and Qualitative
Anthropological Study Exploring the Reasons for
Census Coverage Error Among Low Income Black
Households”. In Report prepared under contract
with the Census Bureau.
Leenheer, J. & Scherpenzeel, A. (2013). “Does it
Pay Off to Include Non-Internet Households in an
Internet Panel?” International Journal of Internet
Science 8 (1), 17–29.
Manheimer, D. & Hyman, H. (1949). “Interviewer
Performance in Area Sampling”. The Public
Opinion Quarterly, 13(1), 83-92.
Tourangeau, R., Kreuter, F. & Eckman, S. (2012).
"Motivated Underreporting in Screening
Interviews“. Public Opinion Quarterly 76(3), 453-
469.
Work in progress, happy to take suggestions
Forthcoming in TSE15 monograph
Bit about conference
As I said there are many ways surveys can try to reduce undercov
Best documented example of this trade-off
Roster is intrusive, high burden
Nice empirical evidence of tradeoff
What is going on here, what might cause this tradeoff?
Obviously we’d like to be in high/high cell
But previous slides have suggested we often have to choose between A & B
Usually only response rates reported
These two choices are just like Design A & B on earlier slide
Coverage propensities rather low
Not all cases on 45
– some have 60 or 70% in direct method, but 40% in roster method
Much higher in direct method
– if you say yes to eligible direct question, more likely to do full intw
IP is product of the 2
Also not on 45
Suggests the 2 approaches are including different people
Bias different with the 2 designs
IP is product of the 2
Also not on 45
Suggests the 2 approaches are including different people
Bias different with the 2 designs
For each of 74 Ys, have cov(direct) and cov(roster)
Plot these 2
If all on 45, then two versions are collecting the same sort of people (in terms of these 74 variables)
Patterns – intrigued by patterns I see in nature (part of big 5
All these graphs and correlations dependent on models
Has expected shape – there is a trade=off
All these graphs and correlations very dependent on models