Page 1 Measuring Survey Quality through Representativity Indicators using Sample and Population...
-
Upload
brooke-poole -
Category
Documents
-
view
213 -
download
0
Transcript of Page 1 Measuring Survey Quality through Representativity Indicators using Sample and Population...
Page 1
Measuring Survey Quality through Representativity Indicators using
Sample and Population based Information
Chris Skinner, Natalie Shlomo,
Barry Schouten, Jelke Bethlehem
Li-Chun Zhang,
Page 2
Representativity Indicators for Survey Qualitycollaboration between:
national statistical institutes of Netherlands, Norway and Slovenia
& universities of Leuven & Southampton
Page 3
Representativity Indicators
• quality indicators for survey non-response
• to supplement response rate
• to measure how well respondents represent population
• tools for use at different stages of survey process (data collection+)
Page 4
Aim of paper
• to consider estimation of representativity indicators using either:
• sample-based information (microdata for both respondents and nonrespondents), or
• population-based information (microdata for respondents and aggregate data for population)
Page 5
How to define representativity indicator?
two approaches
• both based on idea of response propensity
Page 6
Response propensity
Idea: R-indicator measures homogeneity of response propensities:
= probability of response for unit i
given values of auxiliary variables
i
Page 7
R-indicator
where
if constant if
Schouten, Cobben & Bethlehem (2009, Survey Methodology)
( ) 1R ρ
21( ) ( )
1 i UUS
N
ρ
( ) 1 2 ( )R S ρ ρ
( ) 0R ρi( ) 0.5S ρ
Page 8
Alternative Indicator
where
proposed by Särndal and Lundström (2008, JOS) in the context of selecting auxiliary variables for weighting adjustment
2 1 2( ) [ ] [ ( ) ]i i iU UQ ρ
1i i
Page 9
Estimated R-indicator
Sample-based – auxiliary variables recorded for whole sample
(1) estimate response propensities using e.g. logistic regression
(2) 21ˆ ˆˆ( ) 1 2 ( )1 i i Us
R dN
ρ
Page 10
Estimated R-indicator
Population-based – auxiliary variables only measured on respondents and in aggregated form for population
(1) Estimate response propensities using ordinary least squares
if population covariance matrix known. Estimate from respondents if only population
mean vector known
xxS
1'[ ( ')]OLSi i xx i ir
x N S xx d x
xxSx
Page 11
Estimated R-indicator
Population based (continued)
(2) 1 21ˆ ˆˆ ˆ( ) 1 2 ( )1r i i i rr
R dN
ρ
Page 12
Simulation Study
• Samples from Israel census data on 753000 individuals
• ‘realistic’ sampling with fractions 1:50, 1:100, 1:200
• ‘realistic’ non-response based on type of locality, household size, children in household – overall response rate 82%
Page 13
Simulation Means of for Sample (S), Population with Known Covariance matrix (PC) and Population with Unknown Covariance Matrix (PUC) of Auxiliary Variables
ˆ(ρ)R
0.8
0.81
0.82
0.83
0.84
0.85
0.86
0.87
0.88
0.89
0.5% 1.0% 2.0%
True Model S based True Model PC based
True Model PUC based Less Complex S based
Less Complex PC based Less Complex PUC based
Page 14
. 5% S . 5% PC . 5% PUC 1% S 1% PC 1% PUC 2% S 2% PC 2% PUC
0. 70
0. 75
0. 80
0. 85
0. 90
0. 95
R_indicator
gr oup
for Sample (S), Population with Known Covariance Matrix (PC) and Population with Unknown Covariance Matrix (PUC) of Auxiliary Variables – ‘True’ Model
)ρ(R̂
. 5% S . 5% PC . 5% PUC 1% S 1% PC 1% PUC 2% S 2% PC 2% PUC
0. 70
0. 75
0. 80
0. 85
0. 90
0. 95
R_indicator
gr oup
Page 15
for Sample (S), Population with Known Covariance Matrix (PC) and Population with Unknown Covariance Matrix (PUC) of Auxiliary Variables – Less Complex Model
)ρ(R̂
. 5% S . 5% PC . 5% PUC 1% S 1% PC 1% PUC 2% S 2% PC 2% PUC
0. 70
0. 75
0. 80
0. 85
0. 90
0. 95
1. 00
R_indicator
gr oup
Page 16
Simulation Means of for Sample (S) and Population (P) Based Auxiliary Variables
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0.5% 1.0% 2.0%
True Model S Based True Model P Based
Less Complex S Based Less Complex P Based
2q
Page 17
for Sample (S) and Population (P) Based Auxiliary Variables
0. 5% S 0. 5% P 1% S 1% P 2% S 2% P
0
0. 01
0. 02
0. 03
0. 04
q_squared
gr oup
2q
Page 18
Conclusions
• representativity indicators defined in terms of response propensities
• can estimate accurately given either sample-based or population-based information
• have also applied to surveys from RISQ countries
• and have examined bias-corrected estimators and estimators of standard errors of estimators