Upload
tadeo
View
20
Download
1
Embed Size (px)
DESCRIPTION
Coverage Rates and Coverage Bias When Interviewers Create Frames. Stephanie Eckman Joint Program in Survey Methodology University of Maryland June 15, 2009 @ NCHS. Interviewers as Source of Error. Interviewers contribute to nonresponse, measurement error - PowerPoint PPT Presentation
Citation preview
1
Coverage Rates and Coverage BiasWhen Interviewers Create Frames
Stephanie EckmanJoint Program in Survey MethodologyUniversity of Maryland
June 15, 2009 @ NCHS
2
Interviewers as Source of Error
Interviewers contribute to nonresponse, measurement error
Create frames in area-probability surveys Housing unit listing Missed housing unit procedure Household rostering Screening for eligibility
2
3
Literature Review
Previous work concentrates on: How many people are missed? What kinds of people are missed?
Errors by respondents Definitional Motivated
3
4
Research Questions
Compare listing methods Scratch: start with blank frame Update: start with address list
Mechanisms of lister error How are errors made? Incentives of interviewers
Bias & variance
4
5
Hypotheses on Mechanisms
What makes listing easier or more comfortable for the lister? Race or language match between
residents and lister High crime areas Driving
5
6
Hypotheses on Mechanisms
Confirmation bias in dependent listing Failure to add
HUs missing from list Failure to delete
Inappropriate units on list
6
7
Hypotheses on Mechanisms
Undercoverage and nonresponse Anecdotal evidence Hainer 1987 (CPS) Difficult to test
7
8
Pretest of Methods
Pretest with masters students 14 segments in SE Michigan 2 listings: traditional, dependent
Overall disagreement: 12% Evidence of confirmation bias
FTA: 13% less likely to add FTD: 11% less likely to delete
8
9
Census Dataset: Lister Agreement
2 identical listings 211 blocks Dependent on MAF, where available
Overall: 79% agreement
9
10
Census: Block Agreement Rates
10
11
Census Dataset: Drawbacks
Lister characteristics not available No manipulation to test for
confirmation bias
11
12
NSFG Data Collection
National Survey of Family Growth Three listings of 49 segments
1st listing by project 2nd listing: traditional 3rd listing: dependent
Manipulate input: add & delete lines
12
13
Paper 2: Mechanisms of Lister Error
Undercoverage and overcoverage rates Overall By housing unit and block
characteristics By listing method
Test hypotheses
13
14
Paper 3: Coverage Bias
Does lister error lead to bias in survey estimates?
2 sources of data on undercovered NSFG response data (27% selected) Experian data (60% match)
14
15
Paper 3: Estimating Bias
Direct estimates Re-estimate key variables without
cases undercovered by 1 or 2 listers Indirect estimates
Correlation between listing propensity & key variables
16
Questions and Concerns
Analysis of repeated listings Latent class analysis?
Interviewer debriefing What do I want to know?
NR and coverage Available datasets? How to test for this?
18
Pretest: FTA Conf Bias
D listers missed 11 suppressed lines T listers missed only 4
Difference-in-differences estimate Suppression leads to 13% decrease in inclusion
18
HUs D listed T listed
Suppressed HUs 58 47 54
19
Pretest: FTD Conf Bias
D listers confirmed 4 added lines All in multi-unit buildings
T listers included only 1 (??)
Difference-in-differences estimate Bad lines in D lead to 11% increase in inclusion
19
HUs D listed T listed
Added HUs 24 4 1
20
Difference-in-Differences Estimate
Lines D listed T listedUnmanipulated HUs 732
655 (89%)
645 (88%)
Added HUs 244
(17%)1
(4%)
(0.17-0.89) – (0.04-0.88) = -0.13
rounding
20