Title: WE PROBABLY COULD HAVE MORE FUN TALKING ABOUT THESE TRAFFIC STOPPERS
1WE PROBABLY COULD HAVE MORE FUN TALKING ABOUT
THESE TRAFFIC STOPPERS
2WHO CLEARLY HAVE THE RIGHT OF WAY!
BUT
3DESIGNING SURVEYS OVER TIME (PANEL
SURVEYS)VARIANCE, POWER and RELATED TOPICS
- N. Scott Urquhart
- Senior Research Scientist
- Department of Statistics
- Colorado State University
- Fort Collins, CO 80527-1877
4BRIEF COMMENTS ON MONITORING
- Monitoring is a long-term endeavor for most
(living) natural resources - Think of it as a legacy for your grandchildren
- Many of you have tried to use solid data
someone else gathered, but documented poorly. - Much good data dies when the person who
gathered it dies or retires! - Monitoring data requires three things to retain
its value - Metadata
- Storage in a retrievable, maintained, data
system - Backed up safe from fires, floods, or
earthquakes
5SHORT OUTLINE OF THE REST OF THIS TALK
- Inference Perspectives monitoring
- Trend, Variance Structures, and Power to Detect
Trend - Panels and Panel Structures
- Putting it all together Power Curves and
standard errors of estimated current status - Status Estimates from Data Over Years
6INFERENCE PERSPECTIVES
- Design Based
- Inferences rest on the probability
structure incorporated in the sampling plan - Completely defensible very minimal assumptions
- Limiting relative to using auxiliary information
- Model Assisted
- Uses models to complement underlying sampling
structure - Has opportunities for use of auxiliary
information - Model Based (eg spatial statistics)
- Ignores sampling plan
- Defensibility lies in defense of model
7APPROACH OF THIS PRESENTATION
- Use tools from the arena of
- Model-assisted and
- Model-based analyses
- To study the performance of
- Design based
- Model-assisted analyses
- WHY?
- Without models,
- performance evaluations need simulation
(Steve Garmans topic) - Before substantial data have been gathered
- Minimal basis for values to enter into simulation
studies
8STATUS TRENDS OVER TIME IN ECOLOGICAL
RESOURCES OF A REGIONMAJOR POINTS
- Regional trend ¹ site trend
- Detection of trend requires substantial elapsed
time - Regional OR intensive site
- Almost all indicators have substantial patterns
in their variability - Design to capitalize on this dont fight it.
- Minimize effect of site variability with
planned revisits specific plans will be
illustrated - Design tradeoffs TREND vs STATUS
9REGIONAL TREND ¹ SITE TREND
- The predominant theme of ecology
- Ecological processes
- How does a specific kind of ecosystem function
- Energy flows
- Food webs
- Nutrient cycling
- Most studies of such functions must be
- Temporally intensive
- What material goes from where to where?
- Consequently spatially restrictive
- In this situation Temporal trend site trend
10REGIONAL TREND ¹ SITE TREND( - CONTINUED)
- The predominant theme of ecology versus
- A Substantial (any) Agency Focus
- All of an ecological resource
- In an area or region
- Across all of the variability present there
- For Example, National Park Service
- All riparian areas in Olympic National Park
- All riparian areas in National Parks in the
coastal Northwest
11TREND ACROSS TIME - What is it?
- Any response which changes across time in a
generally - Increasing or
- Decreasing
- Manner shows trend
- Monotonic change is not essential.
- If trend of this sort is present, it WILL BE
detectable as linear trend. - This does NOT mean trend must be linear
(examples follow) - Any specified form is detectable
- Time years, here
12TREND ACROSS TIME - What is it?(continued)
13(No Transcript)
14TREND DETECTION REQUIRES SUBSTANTIAL ELAPSED TIME
- IT IS NEARLY IMPOSSIBLE TO DETECT TREND IN LESS
THAN FIVE YEARS. WHY?
15VARIANCE HAS A LOT OF STUCTURE IMPORTANT
COMPONENTS OF VARIANCE
-
- POPULATION VARIANCE
- YEAR VARIANCE
- RESIDUAL VARIANCE
16HOW SHOULD YOU ESTIMATE VARIANCE?
- Alternatives
- Designed-based
- Horwitz-Thompson extremely variable dont
use - Local Neighborhood Variance Estimator (NBH)
Stevens - Gives estimate of variance something like
residual component of variance, only - Power nearly impossible to evaluate in this
context - Model-assisted
- Linear models Urquhart Courbois ( Williams,
now) - Gives estimates of site, year and residual
variances - Which should you use?
17HOW SHOULD YOU ESTIMATE VARIANCE?(Continued)
- We really arent sure, but this topic is
under active investigation - Don Stevens and I are good friends, so this
topic isnt a professional conflict. - We both want to know the answer!
- Results by Courbois (JABES, 2004) suggest this
- Unless response values and inclusion
probabilities (p) are highly correlated, they
(p) can be ignored. - If this stands up, as I expect it to,
practically, linear model estimates of components
of variance will be fine. - Answer expected by summer.
18IMPORTANT COMPONENTS OF VARIANCE ( - CONTINUED)
- POPULATION VARIANCE
- Variation among values of an indicator (response)
across all sites in a park or group of related
parks, that is, across a population or
subpopulation of sites
19IMPORTANT COMPONENTS OF VARIANCE ( - CONTINUED II)
- YEAR VARIANCE
- Concordant variation among values of an indicator
(response) across years for ALL sites in a
regional population or subpopulation - NOT variation in an indicator across years at a
single site - Detrended remainder, if trend is present
- Effectively the deviation away from the trend
line (or other curve)
20IMPORTANT COMPONENTS OF VARIANCE ( - CONTINUED -
III)
- Residual component of variance
- Has several contributors
- YearSite interaction
- This contains most of what ecologists would call
year to year variation, i.e. the site specific
part - Index variation
- Measurement error
- Crew-to-crew variation (minimize with well
documented protocols and training) - Local spatial protocol variation
- Short term temporal variation
21SOURCE OF DATA FOR ESTIMATES OF COMPONENTS OF
VARIANCE
- EMAP Surface Waters Northeast Lakes Pilot
1991 - 1994 - About 450 observations
- Over four years
- Including about 350 distinct lakes
- Design allowed estimation of several residual
components - Lakes illustrate what is generally referred to
in this presentation as sites. - Similar patterns appear in other data sets
- Both aquatic and terrestrial
22(No Transcript)
23ALL VARIABILITY IS OF INTEREST
- The site component of variance is one of the
major descriptors of the regional population - The year component of variance often is small,
too small to estimate. It is a major enemy
for detecting trend over time. - If it has even a moderate size, sample size
reverts to the number of years. - In this case, the number of visits and/or number
of sites has no practical effect.
24ALL VARIABILITY IS OF INTEREST( - CONTINUED)
- Residual variance characterizes the
inherent variation in the response or
indicator. - But some of its subcomponents may contain
useful management information - CREW EFFECTS gt training
- VISIT EFFECTS gt need to reexamine definition
of index (time) window or evaluation protocol - MEASUREMENT ERROR gt work on laboratory/measurem
ent problems
25DESIGN TRADE-OFFS TREND vs STATUS
- How do we detect trend in spite of all of
this variation? - Recall two old statistical friends.
- Variance of a mean, and
- Blocking
26DESIGN TRADE-OFFS TREND vs STATUS( - CONTINUED)
- VARIANCE OF A MEAN
- Where m members of the associated population
have been randomly selected and their response
values averaged. - Here the mean is a regional average slope, so
"s2" refers to the variance of an estimated
slope ---
27DESIGN TRADE-OFFS TREND vs STATUS( - CONTINUED
- II)
- Consequently
- Becomes
- Note that the regional averaging of slopes has
the same effect as continuing to monitor at one
site for a much longer time period.
28DESIGN TRADE-OFFS TREND vs STATUS( - CONTINUED
- III)
- Now, s2, in total, frequently is large.
- If we take one regional sample of sites at one
time, and another at a subsequent time, the
site component of variance is included in s2. - Enter the concept of blocking, familiar
from experimental design. - Regard a site like a block
- Periodically revisit a site
- The site component of variance vanishes from
the variance of a slope.
29PANEL DESIGNS
- Question What kind of temporal design
should you use for National Parks? - A Panel is a Set of Sites which have the same
Revisit Schedule - Each panel ordinarily should have as good a
spatial coverage as possible (GRTS) - You have many usable and defensible
temporal designs - Choose one which fits your needs and resources
- Evaluation tools are available, and demonstrated
here
30A SINGLE PANEL
NUM OBS YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR YEAR
NUM OBS 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
11 X X X X X X X X X X X X X X X X X X X X X
-
- Conventional for trend detection
- Not very good for status
- (Call this design 1)
31AN AUGMENTED PANEL PLAN (Design 2)
NUM OBS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS
NUM OBS 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
5 X X X X X X X X X X X X X X X X X X X
- Add sites to the above annual revisit plane, as
32AN AUGMENTED PANEL PLAN (Design 2)
NUM OBS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS
NUM OBS 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
5 X X X X X X X X X X X X X X X X X X X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
6 X
33A POSSIBLE NPS PANEL PLAN (Design 3)
NUM OBS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS
NUM OBS 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
5 X X
5 X
5 X
5 X
34A POSSIBLE NPS PANEL PLAN (Design 3)
NUM OBS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS
NUM OBS 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
5 X X
5 X
5 X
5 X
4 X X X X X
1 X X X X X
1 X X X X X
1 X X X X
35A POSSIBLE NPS PANEL PLAN (Design 3)
NUM OBS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS YEARS
NUM OBS 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
5 X X
5 X
5 X
5 X
4 X X X X X
1 X X X X X
1 X X X X X
1 X X X X
2 X X X X
1 X X X X
1 X X X X
1 X X
2 X X
1 X X
1 X X
1 X X
2 X X
36WHY LOOK AT POWER?
- Power provides a tool for comparing various
designs - Looking at it does not imply that we have to
conduct tests of hypotheses - The computations displayed here use
- A components of variance statistical model
- To evaluate the underlying variances of estimated
slopes - And the temporal design being considered
- These computations are much more complex,
and suitable to your problems, than any of the
web-available tools - Use of those tools here would commit sins of
pseudoreplication!
37POWER FOR DESIGNS 1 3, 2A WITH PARAMETER
CONDITIONS BILLY GAVElarge is good
38POWER FOR DESIGNS 1 3, 2A WITH PARAMETER
CONDITIONS BILLY GAVElarge is good
39POWER FOR DESIGNS 1 3, 2A WITH PARAMETER
CONDITIONS BILLY GAVElarge is good
40POWER FOR DESIGNS 1 3, 2A WITH PARAMETER
CONDITIONS BILLY GAVElarge is good
41STANDARD ERROR OF STATUS FOR DESIGNS 1 3,
2AWITH PARAMETER CONDITIONS BILLY GAVEsmall is
good
42STANDARD ERROR OF STATUS FOR DESIGNS 1 3,
2AWITH PARAMETER CONDITIONS BILLY GAVEsmall is
good
43STANDARD ERROR OF STATUS FOR DESIGNS 1 3,
2AWITH PARAMETER CONDITIONS BILLY GAVEsmall is
good
44STANDARD ERROR OF STATUS FOR DESIGNS 1 3,
2AWITH PARAMETER CONDITIONS BILLY GAVEsmall is
good
45DESIGN 3 A PROPOSED NPS TEMPORAL DESIGN??
- Defensible design, given restrictions of
- Field resources sites visited per year
- Need to split field resources across two type of
aquatic systems - Streams
- Wetlands
46STATUS ESTIMATES FROM DATA OVER YEARS
- Idea
- Estimate year effects, and
- Adjust all sites to the latest year
- Idea very similar to adjusted treatment means
in the analysis of covariance - For example, if a line approximates the trend,
- The array of Sitei (yearnow
yearinitial)slope - Describes current status, accounting for
documented trend - This approach extends to models other than
lines. - Linkage of site revisits (connectedness) is
necessary to allow estimation of all differences
in year effects
47FUNDING ACKNOWLEDGEMENT
The work reported here today was developed under
the STAR Research Assistance Agreement CR-829095
awarded by the U.S. Environmental Protection
Agency (EPA) to Colorado State University. This
presentation has not been formally reviewed by
EPA. The views expressed here are solely those
of presenter and STARMAP, the Program he
represents. EPA does not endorse any products or
commercial services mentioned in this
presentation.
48RELATED INFORMATION ON THE WEB
- Web-available information from a graduate
course in environmental Sampling (ST571) I
taught at OSU - http//oregonstate.edu/instruct/st571/urquhart/ind
ex.html - Environmental Sampling
- Anatomy
- Variable Probability Sampling
- Cost Effective Resource Allocation
- Sampling Macroinvertebrates
- Maps Grids
- Spatial Sampling
- Support Regions
- Statistical Power - Concepts
- Power to Detect Trend in Ecological Resources
- Representative Sampling
- Statistical Aspects of Taxonomic Richness
- Sample Size to Estimate Taxonomic Richness
- Evaluating a Protocol for "Measuring" Physical
Habitat - Also see http//www.stat.colostate.edu/starmap/
- Also see NPS IM site, from Port Angeles
meeting, 2003
49VISUALIZING LINES AND YEAR EFFECTS
- N. Scott Urquhart
- STARMAP
- Colorado State University
- Fort Collins, Co 80523-1877
50WHAT WE SEE NO LINES NO DECOMPOSITION
51A LINE
52A LINE WITH YEAR EFFECTS
53A LINE WITH YEAR EFFECTS ANNUAL RESIDUALS
54A LINE WITH ONLY ANNUAL RESIDUALS SHOWN
55A LINE WITH ANNUAL RESIDUALS END MARKERS
56A LINE WITH YEAR EFFECTS, JIGGERED RESIDUALS
END MARKERS
57A REALITY
- With a single site
- The year effect
- And residual effect
- Can not be separated
- But with several sites
- These effects can be separated
- Following figures show the patterns
58TWO LINES
59TWO LINES WITH YEAR EFFECTS
60TWO LINES WITH YEAR EFFECTS JIGGERED RESIDUALS
61TWO LINES WITH YEAR EFFECTS JIGGERED
RESIDUALS ONLY SOME YEARS
62WHAT WE SEE NO LINES NO DECOMPOSITION