Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back...
Transcript of Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back...
![Page 1: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/1.jpg)
1/35
JJIIJI
Back
Close
Theory and Methods of NonparametricSurvey Regression Estimation
Jean Opsomer
Iowa State University
Jay Breidt
Colorado State University
June 21, 2004
![Page 2: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/2.jpg)
2/35
JJIIJI
Back
Close
Outline
1. Introduction
2. Generic estimation for surveys
3. Nonparametric model-assisted estimation
4. From theory to applications
(a) National Resources Inventory (NRI)
(b) Forest Inventory and Analysis (FIA)
5. Smoothing parameter selection
6. Conclusion
![Page 3: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/3.jpg)
3/35
JJIIJI
Back
Close
1. Introduction: Statistical Inference• Specific Inference:
– expensive, high quality, targeted
– using “custom-built” method (or
model) to achieve best possible
estimator for particular variable(s)
– willing to defend model
![Page 4: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/4.jpg)
3/35
JJIIJI
Back
Close
1. Introduction: Statistical Inference• Specific Inference:
– expensive, high quality, targeted
– using “custom-built” method (or
model) to achieve best possible
estimator for particular variable(s)
– willing to defend model
• Generic Inference:
– cheap, reasonable quality, good for
many purposes
– using method appropriate for large num-
ber of variables that need to be esti-
mated jointly
CornCornCornCorn FlakesFlakesFlakesFlakes
NET WT. 12 OZ.
![Page 5: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/5.jpg)
4/35
JJIIJI
Back
Close
Statistical Inference in Surveys
Number of Inference Modellingobservations
Large generic none
Moderate
{genericspecific
model-assistedmodel-based
Small specific small area estimation
• “Number of observations” depends on domain (subpopulation) size
• For moderate sample size, use of generic inference depends on model
goodness-of-fit
Nonparametric methods can improve generic infer-
ence
![Page 6: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/6.jpg)
5/35
JJIIJI
Back
Close
2. Generic Estimation for Survey Data
• Population U = {1, . . . , i, . . . , N} with unknown population “pa-
rameters”
yN =1
N
∑U
yi and zN , xN , . . .
• Sample s selected from U according to known sampling design p(s)
– stratification
– clustering
– multiple phases
![Page 7: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/7.jpg)
5/35
JJIIJI
Back
Close
2. Generic Estimation for Survey Data
• Population U = {1, . . . , i, . . . , N} with unknown population “pa-
rameters”
yN =1
N
∑U
yi and zN , xN , . . .
• Sample s selected from U according to known sampling design p(s)
– stratification
– clustering
– multiple phases
• Generic estimator for the population mean
ys =∑
s
wi yi
[zs =
∑s
wi zi, xs, . . .
]
![Page 8: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/8.jpg)
6/35
JJIIJI
Back
Close
Generic Estimation for Surveys: PropertiesThe “ideal” generic estimator would have the following properties
1. easy to compute
2. applicable to large numbers of variables
3. local/scale invariant
zi = a + byi ⇒ zs = a + bys
4. additive
U = {U1, U2} ⇒ Nys = N1ys1 + N2ys2
5. calibrated
xs = xN for known population quantities x
6. precise (low bias, low variance, consistent,...)
![Page 9: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/9.jpg)
7/35
JJIIJI
Back
Close
Simple Generic Estimation: Design-based
• Horvitz-Thompson estimator (1952)
yHT =1
N
∑s
1
πiyi
with inclusion probabilities πi = Pr(i ∈ s)
• Hajek estimator (1971)
yHA =
∑s
1
πiyi∑
s
1
πi
![Page 10: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/10.jpg)
8/35
JJIIJI
Back
Close
Better Generic Estimation: Model-assisted• Superpopulation model ξ: yi are iid with
– Eξ(yi) = β0 + β1xi = xTi β
– Varξ(yi) = σ2
• Least squares population fit for β
BU = (XTUXU)−1XUY U
![Page 11: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/11.jpg)
8/35
JJIIJI
Back
Close
Better Generic Estimation: Model-assisted• Superpopulation model ξ: yi are iid with
– Eξ(yi) = β0 + β1xi = xTi β
– Varξ(yi) = σ2
• Least squares population fit for β
BU = (XTUXU)−1XUY U
• BU is estimated by sample-based estimator
B =(XT
s Π−1s Xs
)−1XT
s Π−1s Y s
with Πs = diag{πi, i ∈ s}• GREG: Model-assisted estimator (Cassel et al., 1977)
yREG =1
N
∑U
xTi B +
1
N
∑s
yi − xTi B
πi
![Page 12: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/12.jpg)
9/35
JJIIJI
Back
Close
Properties of Regression Estimator• Generic estimator
yREG =∑
s
wi(s)yi
• Consistent, asymptotically design unbiased
Ep(yREG) ≈ yN
• Approximate design variance
Varp(yREG) ≈ 1
N 2
∑ ∑U
yi − xTi BU
πi
yj − xTj BU
πj(πij − πiπj)
![Page 13: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/13.jpg)
9/35
JJIIJI
Back
Close
Properties of Regression Estimator• Generic estimator
yREG =∑
s
wi(s)yi
• Consistent, asymptotically design unbiased
Ep(yREG) ≈ yN
• Approximate design variance
Varp(yREG) ≈ 1
N 2
∑ ∑U
yi − xTi BU
πi
yj − xTj BU
πj(πij − πiπj)
• Calibration
xREG =∑
s
wi(s)xi = xN
• Location/scale invariance, additivity,...
![Page 14: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/14.jpg)
10/35
JJIIJI
Back
Close
3. Nonparametric Regression Estimation?
• Superpopulation model ξ:
– Eξ(yi) = xTi β
– Varξ(yi) = σ2
![Page 15: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/15.jpg)
10/35
JJIIJI
Back
Close
3. Nonparametric Regression Estimation?
• Superpopulation model ξ:
– Eξ(yi) = xTi β
– Varξ(yi) = σ2
Replace by:
• Superpopulation model ξ:
– Eξ(yi) = m(xi)
– Varξ(yi) = v(xi)
![Page 16: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/16.jpg)
11/35
JJIIJI
Back
Close
Nonparametric Model-assisted Estimator• Superpopulation model ξ:
– Eξ(yi) = m(xi)
– Varξ(yi) = v(xi)
• Population fit for m(·) at xi, i ∈ U
mi = sUiY U
![Page 17: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/17.jpg)
11/35
JJIIJI
Back
Close
Nonparametric Model-assisted Estimator• Superpopulation model ξ:
– Eξ(yi) = m(xi)
– Varξ(yi) = v(xi)
• Population fit for m(·) at xi, i ∈ U
mi = sUiY U
• The mi, i ∈ U are estimated by design-weighted estimators
mi = ssiY s
• Model-assisted estimator
yNP =1
N
∑U
mi +1
N
∑s
yi − mi
πi
![Page 18: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/18.jpg)
12/35
JJIIJI
Back
Close
Nonparametric Model-assisted Estimator (2)
• Theoretical properties derived for
– kernel-based methods (Breidt and Opsomer, 2000)
– spline-based methods (Breidt, Claeskens and Opsomer, 2003)
• Nonparametric model-assisted estimator has same design properties
as GREG
– weighted (generic) form yNP =∑
s wi(s)yi
– design consistency, variance
– calibration, invariance
![Page 19: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/19.jpg)
12/35
JJIIJI
Back
Close
Nonparametric Model-assisted Estimator (2)
• Theoretical properties derived for
– kernel-based methods (Breidt and Opsomer, 2000)
– spline-based methods (Breidt, Claeskens and Opsomer, 2003)
• Nonparametric model-assisted estimator has same design properties
as GREG
– weighted (generic) form yNP =∑
s wi(s)yi
– design consistency, variance
– calibration, invariance
• Differences with GREG
– requires continuous auxiliary variable, available for all i ∈ U
– smoothing parameter selection
![Page 20: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/20.jpg)
13/35
JJIIJI
Back
Close
Efficiency Gains from Modelling
Varp(yHT ) =1
N 2
∑ ∑U
yi
πi
yj
πj(πij − πiπj)
Varp(yHA) ≈ 1
N 2
∑ ∑U
yi − yN
πi
yj − yN
πj(πij − πiπj)
![Page 21: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/21.jpg)
13/35
JJIIJI
Back
Close
Efficiency Gains from Modelling
Varp(yHT ) =1
N 2
∑ ∑U
yi
πi
yj
πj(πij − πiπj)
Varp(yHA) ≈ 1
N 2
∑ ∑U
yi − yN
πi
yj − yN
πj(πij − πiπj)
Varp(yREG) ≈ 1
N 2
∑ ∑U
yi − xTi BU
πi
yj − xTj BU
πj(πij − πiπj)
Varp(yNP ) ≈ 1
N 2
∑ ∑U
yi −mi
πi
yj −mi
πj(πij − πiπj)
![Page 22: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/22.jpg)
14/35
JJIIJI
Back
Close
4. From Theory to Applications...
• Adapt estimator to more complex designs
– multi-stage
– multi-phase
⇒ possible in model-assisted context
![Page 23: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/23.jpg)
14/35
JJIIJI
Back
Close
4. From Theory to Applications...
• Adapt estimator to more complex designs
– multi-stage
– multi-phase
⇒ possible in model-assisted context
• Extend model to incorporate different data types and multiple aux-
iliary variables
– semiparametric models
– multivariate smoothing techniques
⇒ wide range of nonparametric methods available
![Page 24: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/24.jpg)
14/35
JJIIJI
Back
Close
4. From Theory to Applications...
• Adapt estimator to more complex designs
– multi-stage
– multi-phase
⇒ possible in model-assisted context
• Extend model to incorporate different data types and multiple aux-
iliary variables
– semiparametric models
– multivariate smoothing techniques
⇒ wide range of nonparametric methods available
• Smoothing parameter selection
• Variance estimation
![Page 25: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/25.jpg)
15/35
JJIIJI
Back
Close
Application 1: 1995 NRI Special Study
Two-stage survey of agricultural lands with 1992 National Resources
Inventory as sampling frame
![Page 26: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/26.jpg)
16/35
JJIIJI
Back
Close
NRI and 1995 Special Study
• National Resources Inventory is stratified longitudinal survey of non-
federal land conducted by Natural Resources Conservation Service
(USDA)
• sampling units are 160-acre plots of land, and points within plots
• 1992 NRI contains 300,000 plots
![Page 27: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/27.jpg)
16/35
JJIIJI
Back
Close
NRI and 1995 Special Study
• National Resources Inventory is stratified longitudinal survey of non-
federal land conducted by Natural Resources Conservation Service
(USDA)
• sampling units are 160-acre plots of land, and points within plots
• 1992 NRI contains 300,000 plots
• 1995 NRI Special Study is sample of 1900 plots obtained by stratified
two-stage sampling
– states are strata (14)
– PSUs are counties (1357 total, 213 selected)
– PSU selection probabilities are proportional to measure of erosion
potential in county
– variables of interest: water erosion (USLE), wind erosion (WEQ)
![Page 28: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/28.jpg)
17/35
JJIIJI
Back
Close
Estimator in two-stage sampling
• Usual case: auxiliary information x available for PSUs only
• Superpopulation model ξ for ti (cluster total)
– Eξ(ti) = m(xi)
– Varξ(ti) = v(xi)
![Page 29: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/29.jpg)
17/35
JJIIJI
Back
Close
Estimator in two-stage sampling
• Usual case: auxiliary information x available for PSUs only
• Superpopulation model ξ for ti (cluster total)
– Eξ(ti) = m(xi)
– Varξ(ti) = v(xi)
• x= square root of measure of erosion potential
• Model-assisted estimator
yNP =1
N
∑UI
mi +1
N
∑sI
ti − mi
πIi
with ti =∑
siyki/πk|i, and mi obtained from local linear regression
of ti on xi, i ∈ sI
![Page 30: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/30.jpg)
18/35
JJIIJI
Back
Close
Nonparametric fits
sqrt(size measure)
M T
ons/
Acr
e/Y
r
2 4 6 8 10 12
05
1015
20
REG4LPR1(h=3)
WEQ
sqrt(size measure)
sqrt(
M T
ons/
Acr
e/Y
r)
2 4 6 8 10 12
01
23
4
REG4LPR1(h=3)
Transformed WEQ
sqrt(size measure)
M T
ons/
Acr
e/Y
r
2 4 6 8 10 12
0.0
0.5
1.0
1.5
2.0
2.5
3.0
REG4LPR1(h=3)
USLE
sqrt(size measure)
sqrt(
M T
ons/
Acr
e/Y
r)
2 4 6 8 10 12
0.5
1.0
1.5 REG4
LPR1(h=3)
Transformed USLE
![Page 31: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/31.jpg)
19/35
JJIIJI
Back
Close
EstimatesWEQ USLE
HT 443.6 551.5(49.4) (31.8)
REG2 ν(x) ∝ x2 442.5 537.8(50.7) (26.5)
REG4 ν(x) ∝ x4 442.1 537.7(50.1) (26.5)
REG8 ν(x) ∝ x8 441.8 540.1(50.3) (27.6)
LPR1 h=1 434.1 529.0(47.5) (24.4)
LPR1 h=3 427.4 532.3(48.9) (25.3)
LPR1 h=5 430.5 541.2(48.7) (27.6)
![Page 32: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/32.jpg)
20/35
JJIIJI
Back
Close
Application 2: Forest Health Monitoring
FHM: part of Forest Inventory and Analysis (FIA) of Forest Service
![Page 33: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/33.jpg)
21/35
JJIIJI
Back
Close
FIA and FHM dataVariables
Geographic X,Y coordinates E-W, N-SInformation elev elevation (m)System asp aspect (deg)(GIS) slope slope (deg)
hillshd hillshade (solar radiation)N = 67, 216 nlcd vegetation cover type (class)Forest fortyp forest type (class)Inventory trees number of treesand agemax max tree age (years)Analysis ageavg avg tree age (years)(FIA) bamax max tree basal area (sq in)
crcov tree crown cover (%)nI = 3, 107 . . .Forest lichen lichen species present (count)Health . . .Monitoring (FHM)n = 71
![Page 34: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/34.jpg)
22/35
JJIIJI
Back
Close
Multi-phase Sampling
Population U(N elements)
Phase 1 sample s(nI elements)
Phase 2 sample r(n elements)
GeographicInformationSystem(GIS)
N = 67, 216
ForestInventoryandAnalysis(FIA)
nI = 3, 107
ForestHealthMonitoring(FHM)
n = 71
![Page 35: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/35.jpg)
23/35
JJIIJI
Back
Close
Model-assisted Estimation for Multi-phaseSamples
Different information available at different phases
population xai, zai, i ∈ U GIS variablesphase 1 xbi, zbi, i ∈ s FIA measurementsphase 2 yi, i ∈ r FHM measurements (lichen count)
(xbi, zbi contains xai, zai)
Goal: estimate yN with generic but efficient estimator
![Page 36: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/36.jpg)
23/35
JJIIJI
Back
Close
Model-assisted Estimation for Multi-phaseSamples
Different information available at different phases
population xai, zai, i ∈ U GIS variablesphase 1 xbi, zbi, i ∈ s FIA measurementsphase 2 yi, i ∈ r FHM measurements (lichen count)
(xbi, zbi contains xai, zai)
Goal: estimate yN with generic but efficient estimator
Approach:
1. use penalized spline regression to fit semiparametric additive model
for each “level” of auxiliary info
2. construct multi-phase model-assisted estimator
![Page 37: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/37.jpg)
24/35
JJIIJI
Back
Close
Model-assisted Estimation for Multi-phaseSamples (2)• Models
– Model a: using predictors available for U
Eξ(yi) = ga(xai, zai) = ma(xai; βa) + zaiγa
– Model b: using predictors available for s
Eξ(yi) = gb(xbi, zbi) = mb(xbi; βb) + zbiγb
![Page 38: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/38.jpg)
24/35
JJIIJI
Back
Close
Model-assisted Estimation for Multi-phaseSamples (2)• Models
– Model a: using predictors available for U
Eξ(yi) = ga(xai, zai) = ma(xai; βa) + zaiγa
– Model b: using predictors available for s
Eξ(yi) = gb(xbi, zbi) = mb(xbi; βb) + zbiγb
• Fit both models on data from r
• Model-assisted estimator
yNP =1
N
∑U
gai +1
N
∑s
gbi − gai
πi(s)+
1
N
∑r
yi − gbi
πi(s) πi(r|s)
![Page 39: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/39.jpg)
25/35
JJIIJI
Back
Close
Semiparametric Model a
Eξ(LICHEN) = m(HILLSHD; β) + zNLCDγ
hillshd
p
s(hi
llshd
)
100 150 200 250
-6-4
-20
24
6
parti
al fo
r nlc
d2-4
-20
2
nlcd2
0 41 42 43 51 71
![Page 40: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/40.jpg)
26/35
JJIIJI
Back
Close
Semiparametric Model bEξ(LICHEN) = m1(CRCOV; β1) + m2(AGEMAX; β2)
+m3(BAMAX; β3) + zNLCDγ
crcov
ps
p
s(cr
cov)
0 20 40 60 80
-4-2
02
46
8
agemax
ps(
agem
ax)
0 100 200 300 400
-4-2
02
bamax
ps(b
p
s(ba
max
)
0 200 400 600 800 1000
-4-3
-2-1
01
parti
al fo
r nlc
d2-8
-6-4
-20
2
nlcd2
0 41 42 43 51 71
![Page 41: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/41.jpg)
27/35
JJIIJI
Back
Close
Forest Health Monitoring Estimates
• HT = Generic estimator, ignores all auxiliary information
• Linear = Generic estimator, all models are fitted by linear regression
• Semiparametric = Generic estimator, semiparametric models
HT Linear SemiparametricEstimate 3.62 2.92 2.67Est. St. Dev. 0.36 0.25 0.16
(69%) (44%)
![Page 42: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/42.jpg)
28/35
JJIIJI
Back
Close
Estimators for Domains• “Domain”: subpopulation for which separate estimator is needed
• Models can improve precision of domain estimators, if they have
good local properties
• Nonparametric model better able to adapt to local features of data
••
•
•
•
•
•
•
•
••
•
•
•
•
•••
•
•
•
•
•
••
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
••
•
•
•
•
•
•
•
•
•
•
•
•
crcov
s(cr
cov)
0 20 40 60 80
-20
24
6
![Page 43: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/43.jpg)
29/35
JJIIJI
Back
Close
Model-Assisted Estimators for Domains
(Sarndal, 1984)
• Obtain sample-weighted model fit for complete sample s
• Estimator for domain Ud ⊂ U with realized sample sd ⊂ s is
yNP =1
Nd
∑Ud
gi +1
Nd
∑sd
yi − gi
πi
• Variance follows from model-assisted estimation theory
• Approach maintains additivity of domain estimates
![Page 44: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/44.jpg)
30/35
JJIIJI
Back
Close
Example: Estimation for Domain withNLCD > 50
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Phase 1
Latitude Index
Long
itude
Inde
x
Only n = 27 observations in domain
![Page 45: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/45.jpg)
31/35
JJIIJI
Back
Close
Example (2)
All Data (n = 71)
HT Linear SemiparametricEstimate 3.62 2.92 2.67Est. St. Dev. 0.36 0.25 0.16
Domain (n = 27)
HT Linear SemiparametricEstimate 2.00 1.78 1.72Est. St. Dev. 0.57 0.39 0.17
(1.58) (1.56) (1.06)
Nonparametric regression makes it possible to
maintain generic approach at smaller “scales”
![Page 46: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/46.jpg)
32/35
JJIIJI
Back
Close
5. Smoothing Parameter Selection• Smoothing parameter selection less important in generic estimation
⇒ optimal value depends on variable being estimated
⇒ but: single set of survey weights, many variables!
![Page 47: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/47.jpg)
32/35
JJIIJI
Back
Close
5. Smoothing Parameter Selection• Smoothing parameter selection less important in generic estimation
⇒ optimal value depends on variable being estimated
⇒ but: single set of survey weights, many variables!
• Minimizing estimate of asymptotic design variance is poor choice
V (yNP ; h) =1
N 2
∑ ∑s
yi − mi(h)
πi
yj − mj(h)
πj
πij − πiπj
πij
• Proposed approach based on “design-based cross-validation”
CV(yNP ; h) =1
N 2
∑ ∑s
yi − m−ii (h)
πi
yj − m−jj (h)
πj
πij − πiπj
πij
• “Leave-one-out” estimator m−ii is easy to compute for most non-
parametric regression techniques
![Page 48: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/48.jpg)
33/35
JJIIJI
Back
Close
Smoothing parameter selection (2)
0 1 2 3 4 5 6 7 8 90
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
band
wid
th
![Page 49: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/49.jpg)
34/35
JJIIJI
Back
Close
Smoothing parameter selection (3)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
10
20
30
40
50
60
MS
E
(a)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
10
20
30
40
50
60
MS
E
(b)
bandwidth
![Page 50: Theory and Methods of Nonparametric Survey …nsu/starmap/pps/Presentations/...1/35 JJ II J I Back Close Theory and Methods of Nonparametric Survey Regression Estimation Jean Opsomer](https://reader031.fdocuments.us/reader031/viewer/2022022503/5ab080617f8b9a59478ea905/html5/thumbnails/50.jpg)
35/35
JJIIJI
Back
Close
6. Conclusions
• Generic estimation can be improved with nonparametric methods
– more efficient when relationship exists but parametric model not
appropriate
– almost as efficient when parametric model is correct
• Nonparametric model-assisted estimation
– fits in current survey estimation paradigm
– shares properties of parametric methods
– complementary with parametric approaches
– easy to implement with currently available software
• Requires unit-level “frame” information
Contact: – [email protected]
– http//www.public.iastate.edu/˜jopsomer/home.html