Transcript Combining trends - De Vlinderstichting

Detecting trends in dragonfly data
- Difficulties and opportunities -
Arco van Strien
Statistics Netherlands (CBS)
Introduction
Statistics Netherlands’ involvement in monitoring
National:
Breeding birds
Waterbirds
Flora
Amphibians
Reptiles
Butterflies
Dragonflies
Bats and a few other mammals
Lichens
Fungi
•
•
•
International:
European Wild Bird Indicators
European Grassland Butterfly Indicator
European Bats (in development)
Close co-oparation with NGO’s for each species group
NGO’s responsible for field work (mainly by volunteers)
Statistics Netherlands responsible for statististical analysis
Introduction
Structure of talk
• Difficulty 1: Statistical power
• Difficulty 2: Bias
Three monitoring alternatives:
• No standardisation of field method
• Strong standardisation & analysis method TRIM
• Medium standardisation & analysis method Occupancy
modeling
• Combining trends (European perspective)
• Conclusions
Introduction
Difficulties in trend detection:
• Inability to detect existing trends (low statistical
power)
• Even more worse: Biased trend estimates
(increase or decline larger or smaller than in
reality)
Monitoring schemes need to have:
• Sufficient statistical power to detect
trends
• No or negligible bias in trend estimates
Introduction
Statistical power is low if:
• Yearly fluctuations in abundance are high
• Sites have different year-to-year changes
Dragonflies have considerable
fluctuations in abundance …
Power
Ischnura elegans
Pyronia tithonus
Turdus merula
190
170
150
130
110
90
70
50
1999 2000 2001 2002 2003 2004 2005
Power
Statistical power is low if:
• Yearly fluctuations in abundance are high
• Sites have different year-to-year changes
Remedies:
• Sufficient sampling effort (= no. of sites;
exact no. sites required depends on a.o.
field method)
• Use longer detection period (= wait longer)
Power
A longer detection period leads to more accurate trend estimates
2,5
trend (1=stable)
standard error of trend
2
1,5
1
0,5
0
3
4
5
6
7
8
9 Years
Length of time series of smooth snake Coronella austria
Power
Monitoring schemes need to have:
 Sufficient statistical power to detect trends
• No or negligible bias in trend estimates
Risk of bias is higher if:
• Sampling effort per site is not constant
across years
• Detectability of species is not constant
during the season and between years
• Inadequate sampling strategy applied e.g.
dragonfly-rich areas oversampled
Bias
Increasing sampling effort leads to artificial increase
High sampling effort
Index of
species
300
Low sampling effort
250
200
150
100
50
0
1990
1992
1994
1996
1998
2000
2002
Bias
Risk of bias is higher if:
•
•
•
Sampling effort per site is not constant across years
Detectability of species is not constant during the season & years
Inadequate sampling strategy applied e.g. dragon-rich areas
oversampled
Remedies:
• Standardize sampling effort (field method)
• Take into account variation in detectability
during the season (= multiple visits)
• Apply adequate sampling strategy (or
adjust a posteriori any bias due to unequal
sampling)
Bias
Structure of talk
 Difficulty 1: Statistical power
 Difficulty 2: Bias
Three monitoring alternatives:
• No standardisation of field method
• Strong standardisation & analysis method TRIM
• Medium standardisation & analysis method Occupancy
modeling
• Combining trends (European perspective)
• Conclusions
Alternatives
Monitoring alternatives
Field method
Data collection
Example
no
standardisation
presence data e.g.
per grid cell (e.g.
5x5 km2)
Comparison of
distribution between
two periods
strong
standardisation
count data per site
Dutch dragonlfly
scheme
medium
standardisation
presence and
absence data per
grid cell or site
?
Alternatives
No standardisation of field method
•
•
•
•
•
•
As in Atlas studies or studies to compile Red Lists
No fixed sites
Sampling efforts vary between years
No prescription of field method
No formal sampling strategy
Collecting presence data only
>>>>
• Statistical power low (only sensitive to pick up strong declines
& increases in distribution)
• Risk of bias considerable due to not constant sampling efforts
• Statistical analysis: simple comparison of distribution data (or a
statistical method)
Alternatives
Strong standardisation of field method
•
•
•
•
•
As in dragonfly scheme in the Netherlands
Fixed sites (500 m - 1 km long)
Yearly surveys
Multiple visits per year (during the season)
Detailed prescription of field method (fixed sampling effort per
site)
• Sampling strategy: preferably (stratified) random choice of sites
• Collecting count data
>>>>
• Statistical power high
• Risk of bias low
• Statistical analysis: TRIM
TRIM
TRIM version 3
Poisson regression (loglinear models,
GLM) for count data
Pannekoek, J. & A van Strien, 2001. TRIM 3. Statistics Netherlands, Voorburg
TRIM
TRIM: Trends and Indices for Monitoring data
• Specially developed by Statistics Netherlands for
wildlife monitoring based on count data
• Statistical heart of wildlife monitoring data analysis
• Internationally accepted and in use in many European
countries
• Easy to use
• Freeware
• Calculates yearly indices
TRIM
INDEX: the total (= sum of al sites) for a
year divided by the total of the base year
Site
Year 1 Year 2 Year 3 Year 4
Year 5
1
2
3
4
5
20
20
16
8
10
10
10
8
4
5
8
12
10
6
7
2
3
3
6
7
3
2
3
5
8
Sum
74
37
43
21
21
Index
100
50
58
28
28
TRIM
Statistical characteristics of TRIM
•
•
•
•
•
•
Produces yearly indices and overall trends per species
Produces confidence intervals
Include overdispersion & serial correlation in models
Goodness-of-fit tests for comparing models
Covariates to test trends between sets of sites
Weight factors may be included to improve
representativeness if sites are not randomly selected
• Imputation of missing values
TRIM
Imputation of missing counts required to
compute correct indices
Site
1
2
3
4
5
Year 1 Year 2 Year 3 Year 4
20
10
20
10
16 (7.5) ?
8
4
10
5
8
12
10
6 (2.3)
7
Sum
74
36
Index
100
49
Year 5
2
3
3
?
7
3
2
3
5
8
43
17
21
58
23
28
TRIM
Medium standardisation of field method
To be developed, but think of:
•
•
•
•
•
•
Preferably fixed sites
Survey per site once every 2-3 years
Multiple visits per year (during the season)
Limited prescription of sampling effort per site, e.g. 1 hour field work
Sampling strategy: preferably (stratified) random choice of sites
Collecting presence/absence data per site per visit (or abundance
categories)
>>>
• Power not high
• Risk of bias low
• Statistical analysis: Occupancy modeling to adjust for bias due to
limited standardisation
Occupancy
Occupancy modeling:
Recent developments in statistical methods make it
possible to estimate area of occupancy while taking into
account the detectability of species (which may differ
according to e.g. not constant sampling efforts)
Based on absence/presence data from repeated visits
(capture-recapture)
Statistical method is in development
Freeware (PRESENCE, MARK)
MacKenzie, D.I., J.D. Nichols, J.A. Royle, K.H. Pollock, L.L.
Bailey & J.E. Hines, 2006. Occupancy estimation and
modeling. Elsevier, Amsterdam.
Occupancy
Occupancy modeling uses capture histories per site to
separate occupancy and detectability
Simple example
Site
Visit 1 Visit 2
1
0
1
2
1
0
3
0
1
4
1
0
Site
Visit 1 Visit 2
1
1
1
2
1
1
3
0
0
4
0
0
area of occupancy 100%
detection probability per visit 50%
area of occupancy 50%
detection probability per visit 100%
Occupancy
Statistical characteristics of Occupancy modeling
• Produces estimate of area of occupancy per year (or
period), taking into account detectability of species
• Comparing area of occupancy per period >> trend
• Produces confidence intervals
• Allows missing values
• Covariates to allow for effect of e.g. temperature during
visit, incompleteness of survey etc.
• Weighting procedure (if sites are not randomly selected)
to be developed
Occupancy
Structure of talk
 Difficulty 1: Statistical power
 Difficulty 2: Bias
Three monitoring alternatives:
 No standardisation of field method
 Strong standardisation & analysis method TRIM
 Medium standardisation & analysis method Occupancy
modeling
• Combining trends (European perspective)
• Conclusions
Combining trends
Combining TRIM results per country, weighted by
population sizes, is well-developed
Yearly population size of
species A in country 1
Yearly population size of
species A in country 2
Yearly population size of
species A in country 3
European
population trend of
species A
Yearly population size of
species A in country 4
A. van Strien, J. Pannekoek & D. Gibbons, 2001. Bird Study 48:200-213
Combining trends
Example of combining TRIM results of countries:
Pan-Euromonitoring Common Bird Monitoring project producing
Farmland Wild Bird Indicator (EU biodiversity indicator)
140
Population Index (1980=100)
Other common birds (25)
120
100
Common forest birds (33)
80
Common farmland birds (19)
60
40
1980
1985
1990
1995
2000
2005
Year
Gregory R.D., van Strien, A., Vorisek P. et al., 2005. Phil. Trans. R. Soc. B. 360: 269-288
Gregory, R.D., Vorisek, P., van Strien, A. et al., 2007. Ibis, 49, s2, 78-97
Combining trends
Combining areas of occupancy per country,
weighted by areas, appears possible (but needs
to be developed)
Yearly occupancy area of
species A in country 1
Yearly occupancy area of
species A in country 2
Yearly occupancy area of
species A in country 3
European trend in
occupancy area of
species A
Yearly occupancy area of
species A in country 4
Combining trends
Scores of alternatives
Field method
Power to
detect
trends
Risk of
bias
Sampling
effort
needed
no standardisation
(presence data)
strong standardisation
(count data)
medium standardisation
(pres/absence data)
Conclusions
Conclusions
• Statistical enemies of monitoring: low power and bias
• Standardisation of sampling effort helps to increase
power and to reduce bias
• Monitoring based on strong standardisation: high power
& little bias. But it requires considerable sampling efforts
• If strong standardisation is not feasible, consider medium
standardisation: lower power, but again little bias (if
detection probabilities are taken into account)
• For both alternatives statistical methods are available
• Both alternatives enable to combine trends across
countries
Conclusions