You are on page 1of 25

Model- vs.

design-based sampling and


variance estimation
on continuous domains

Cynthia Cooper
OSU Statistics
September 11, 2004

 R82­9096­01
Introduction
• Research on model- and design-based sampling
and estimation on continuous domains

Compare ...
• Basis of inference of each
• Sampling concepts
• Interpretation of variance
• Variance estimation

2
Duality in Environmental Monitoring
• Design-based Estimates
– Status and trend
– No model of underlying stochastic process
• Defensible
– Probability sample
• Avoid selection bias
• Control sample process variance
• Model-based predictions
– Stochastic behavior of response
– Forecasting/prediction conditional on the observed data

3
General Outline
• Introduction
• Summary comparison of approaches
• Summary characterization of variance estimators
• Proposed model-assisted variance estimator
• Simulation methods
• Design-based context results
• Model-based (kriging) results
• Conclusion

4
Comparison of approaches - Design-based

• Probability samples – unbiased estimates


• Basis for long-run frequency properties
– Design-induced randomness – sample process variance
• Basic linear estimator scales up sample responses
to extrapolate to population
– Inclusion probabilities
• Examples
– EPA EMAP
– ODFW Monitoring Plan Augmented Rotating Panel
– USFS Forest Inventory and Analysis
5
Comparison of approaches - Design-based
• Inclusion probability
– Element-wise – Sum of probabilities of all samples
which include the ith element
∀ πi
– Pair-wise -- Sum of … which include ith & jth elements
∀ πij

• For continuous domains


– Inclusion probability densities (IPD) (Cordy (1993))

6
Comparison of approaches - Model-based

• Response generated by a stochastic process


• Likelihood-based approaches to estimating
parameters of model
• BLUP
– Conditional on values observed in sample

• Examples
– Mining surveys
– Soil and hydrology surveys

7
Variance estimators - Design-based
• Quantifies variability induced by sampling process
• Variance of linear estimators
– Scale up square and cross-product terms with inverse
marginal and pair-wise inclusion probability densities
(IPDs)

• For continuous domains


– Congruent tessellation stratified samples w/ one
observation per stratum
• Require randomized grid origin to achieve non-zero
cross-product terms (πij-πiπj) (Stevens (1997))

8
Variance estimators - Design-based
• Horvitz-Thompson (HT)
Cordy
zi2 1 zi z j
(π ij − π iπ j )
(1993)
VˆHT [ w' z | S ] = V [τˆHT ] = ∑ 2 + ∑∑
i πi i j ≠ i π ij π i π j

• Can be negative
– Especially samples with a point pair in close proximity
• Requires randomly-located tessellation grid

9
Variance estimators - Design-based
• Yates-Grundy (YG)
Cordy 2
 zi z j 
1  −  (π iπ j − π ij )
(1993)
VˆYG [ w' z | S ] = V [τˆYG ] = ∑∑
i j <i π ij
π π 
 i j 

• Assumes fixed effective sample size


• Point pairs with close proximity can destabilize
(Stevens (2003))
• Requires randomly-located tessellation grid

10
Variance estimators - Model-based
• Estimating MSPE of BLUP
– Involves variances and covariances associated with
square and cross-product terms of error
• Assume form of covariance that describes rate of
decay of covariance
• Exponential
• Spherical
• Must result in positive-definite covariance matrix
• Incremental stationarity
– E[(z(si) -z(so))2] = g(||si-so||) = g(h)
– Typically, h ↓ ⇒ E[…] ↓

11
Variance estimators - Model-based
• Variance
– Quantifies stochastic variability of expected value of
response
– Vanishes as ||si-so|| → 0
• Mean-square prediction error (MSPE)
– a.k.a. MSE
– Variance + bias2
• Sample process variability of BLUP
– Weighted averages vary less
– Varies more as sample range increases relative to
resolution

12
Proposed model-assisted variance (VMA)
Use covariance structure of response to model
variability due to sampling process
• Predict variance within a stratum
• Variance is reduced by mean covariance
(assuming positively correlated elements)
– Similar to error variance computations (Ripley (1981))
• Within-stratum estimated as
– Sill reduced by within-stratum average covariance
• Linear estimator variance estimated as sum of
squared coefficients times within-stratum variance
13
Precursors of and precedence for
modeling covariance
• Cochran (1946)
– Finite population
– Serial correlation w/ discrete lags
• Bellhouse (1977)
– Continued extension of Cochran’s work to finite
populations ordered on two dimensions
• Small-area estimation model-assisted approaches
– J.N.K Rao (2003)

14
Methods – part 1

Random field (background) generated in R


• M. Schlather's GaussRF() of R package
RandomFields
• Exponential covariance structure b*exp(-h/r)
– (e.g. 4*exp(-h/2))
• h is distance; b and r are "sill" and "range"
parameters

15
Methods – part 1a
Repeat 1000 times per realization

• Stratified sample
– n=100; one observation per stratum; stratum size 2x2
– Simple square-grid tessellation
• Randomized origin
• Constant origin

• REML estimate of covariance parameters (b,r)

16
Methods – part 2
Repeat 1000 times per realization (continued)
• For the design-based context
– Estimate total (zhat)
• HT estimator for continuous domain
– Compute VHT, VYG and VMA
– Compare estimated variances with empirical variance
(V[zhat])
• For the model-based context example (Kriging)
– Randomly selected zo at fixed location over 1000 trials
– Obtain zhat, VOK, VMA
17
Random field overlayed by stratified sample w/ constant origin
Exponential covariance with range= 2 and sill= 1

20
15
10
5
0

0 5 10 15 20
18
Results – Design-based application
V...( 50) − V [ z hat ]
Empirical median relative error EMRE =
V [ z hat ]
Compares estimated variances with empirical variance of
estimate of total (V[zhat])
(Stratified sample with randomized origin)
Sill 4 1
Range 0.5 1 2 4 0.5 1 2 4

VHT 0.068 0.063 0.189 0.161 *0.044 0.185 *0.070 0.372


*
VYG -0.045 -0.122 -0.117 -0.176 -0.052 *-0.032 -0.228 -0.133
VMA 0.055 *-0.024 *-0.001 *-0.035 0.049 0.084 -0.138 *0.004

19
Results – Design-based application
Exponential covariance with range= 2 and sill= 4
Med
Avg

Med
Med

Avg
Avg

200
200

Observed V[zhat]
Observed V[zhat]
Observed V[zhat]

0
0

1000 2000 3000 4000 5000 1000 1500 2000 2500 3000 3500 4000 4500 -6000 -4000 -2000 0 2000 4000 6000

Model-assisted Variance Yates-Grundy Variance Horvitz-Thompson Variance

20
Results – Design-based application
Ratios of empirical standard deviations sd (V... )
sd (V HT )

(Stratified sample with randomized origin)

Sill 4 1
Range 0.5 1 2 4 0.5 1 2 4

MA/HT 0.56 0.43 0.27 0.24 0.66 0.36 0.20 0.14


YG/HT 0.77 0.62 0.35 0.28 0.84 0.43 0.27 0.17

21
Results – Model-based application

Exponential covariance with range= 1 and sill= 1


(stratified sample with randomized origin)
Avg

Avg
100

Observed V[zhat]
Observed V[zhat]

0
0

0.0 0.2 0.4 0.6 0.8 1.0 1.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2

Model-assisted variance Kriging variance (MSPE)

22
Concluding - Model-assisted approach
• Small-area precedence
• Application to systematic and one-observation-
per-stratum samples
• Effective alternative to direct estimators of
continuous-domain randomized-origin tessellation
stratified samples
– Empirical results – less bias, better efficiency
• Doesn’t require randomly-located tessellation grid
on continuous domain for non-zero πij

23
Acknowledgements
Thanks to
Don Stevens
Committee members
OSU Statistics Faculty
UW QERM Faculty

24
 R82­9096­01

The research described in this presentation has been funded by the U.S. Environmental Protection Agency

through the STAR Cooperative Agreement CR82-9096-01 National Research Program on Design-

Based/Model-Assisted Survey Methodology for Aquatic Resources at Oregon State University. It has not

been subjected to the Agency's review and therefore does not necessarily reflect the views of the Agency,

and no official endorsement should be inferred

25