a review of (total) survey error models
DESCRIPTION
A Review of (Total) Survey Error Models. William D. Kalsbeek Survey Research Unit University of North Carolina. Purpose. To review the following for existing total survey error (TSE) models:. Composition and Structure Presentation Utility. Presentations of TSE Models. - PowerPoint PPT PresentationTRANSCRIPT
1
A Review of (Total) Survey Error Models
William D. KalsbeekSurvey Research Unit
University of North Carolina
2
PurposeTo review the following for existing
total survey error (TSE) models:
• Composition and Structure• Presentation• Utility
3
Presentations of TSE Models• TSE Model (a Definition): *
– A postulation to understand or predict, by theory or simulation, the properties or behavior of the survey process
• Presentations of TSE:– Practical:
• Process origins; plus statistical nature, impact, measurement and/or control of error
– Theoretical:• A formulary (usually MSE-based)
* Based on Kotz, et al. (1981-89).
4
ThesisTSE Models
• Have organized our thinking on the statistical effects of error sources
But
• Translation of this understanding into practical improvement has been limited and largely marginalized to individual sources of error
5
Thesis
For the Future:
• Greater research emphasis on TSE components and application of TSE findings for a broader array of data systems?
• Model re-direction needed?
6
Sources of Error *• Sampling• Frame• Measurement• Nonresponse (Unit/Item)
* One might also view the underlying stochastic model responsible for the data array in model-based inference as a source of error
7
A Review of TSE Presentations • Tracking presentations for 2+ sources• Structural basis
– Various decompositions of MSE• Grouping by number of sources and:
– Type of presentation (practical/theoretical)– Source interrelationship (separate/integrated)
• Question: – Which parts of the survey process have TSE
models accommodated?
8
Sources of Error
1. Sampling2. Frame3. Measurement4. Nonresponse (Unit/Item)
9
?
10
Washington Nationals: Season starts: 4/4/05 (at Phillies) Home opener: 4/14/05 (Diamondbacks)
11
12
13
AROUND THE HORN
14
TSE
AROUND THE HORN
Total Survey Error
15
STSE
Sampling
AROUND THE HORN
16
S
M
TSE
Measurement
AROUND THE HORN
17
S
M
F TSE
Frame
AROUND THE HORN
18
S
NR
M
F
UI
TSE
Nonresponse
Item
Unit
AROUND THE HORN
19
S
NR
M
F
UI
TSE
AROUND THE HORN
Variances
20
UI
S
NR
M
F TSE
AROUND THE HORN
Interfaces
21
UI
F
M
S
NR
TSE
AROUND THE HORN
Biases(additive)
22
UI
M
S
NR
F TSE
A HOME RUN
23
• Nonresponse Bias – Hansen and Hurwitz (1946)– Several extension to more
complex sample designs • El-Badry (1956)• Rao (1968, 1973)• Rao and Hughes (1983)
Two-Source Theoretical (Integrated):
UI
F
M
S
NR
TSE
24
• Measurement Error Model – Hansen, et al. (1951a,
1951b, 1961, and 1964) – Subsequent work by others
at the Census Bureau– Forsman (1989) review
Two-Source Theoretical (Integrated):
UI
F
NR
M
STSE
25
• Multiplicity Estimators:– Birnbaum and Sirken (1965)– Several subsequent papers by Sirken, et al.
Two-Source Theoretical (Integrated):
S
NR
M
F
UI
TSE
26
• Model-Based Inference with Missing Data– Little (1995)– Little and Rubin (2002)
Two-Source Theoretical (Integrated):
UI
F
M
S
NR
TSE
27
• Platek, et al. (1977, 1983)• Lessler (1983)
Three-Source Theoretical (Integrated):
UI
F
M
S
NR
TSE
28
• Following Kish (1965)– Anderson, et al (1979)– Groves (1989)– Groves, et al. (2004)
• Federal Committee on Statistical Methodology– FCSM (2001)– Kasprzyk & Giesbrecht (2003)– Other error profiles by
Bailar and colleagues for Census statistics
All-Source Practical (Separate):
UI
F
NR
M
STSE
29
• Lessler and Kalsbeek (1992)• Sarndahl, Swennsson, and
Wretman (1992)
All-Source Theoretical (Separate):
UI
SF
M
NR
TSE
30
• A general model appended to Lessler and Kalsbeek (1992)
All-Source Theoretical (Integrated):
UI
M
S
NR
F TSE
31
Utility of Existing Models• Provides a theoretical basis in survey
practice to:– Structure our thinking– Motivate preventive strategies– Suggest process quality indicators– Suggest measurement approaches– Catalog empirical findings
32
Limitations of Existing Models *• Compartments and smokestacks
– Marginalized treatment of error sources• Plausibility and complexity
– Inverse relationship between proximity to reality and complexity
• Context and comparability– Breadth of model utility
• Lack of Attention– Priorities and cost
* Inspiration and insight from Platek and Sarndahl (2001)
33
Questions for the Future• More emphasis on studying and
minimizing TSE?– For the major and minor leagues
• Greater integration of TSE and practice?– Cataloging and lessons learned
• New directions in TSE model structure?– All sources jointly TSE– Action-directed models TQM? – More process indicators