My talk at International Congress for Conservation Biology 2015, in Montpellier.
Data collected through citizen science programs allow addressing many important questions in conservation biology related, e.g., to the shift in species range, the ecology of infectious disease or the effects of habitat loss and fragmentation on biodiversity. However, citizen science data are subject to serious statistical challenges when it comes to their analysis and the reliable extraction of the information they contain, mainly due to sampling biases generated by variation in the observation process. Numerous methods have been proposed to address this issue that can be split into two main strategies: either a new approach is developed to deal with a specific problem or an existing approach is used pending some pre-treatment of the data or post-processing of the results. I review these various methods, trying to make the links between them and emphasizing their advantages and drawbacks with respect to the question. I illustrate my talk with case studies drawn for the research conducted in our group, mainly on large carnivores. Based on this review, I end up this contribution by recommendations on the use of existing methods and by suggesting perspectives on future developments.
3. Mo3va3on
• Recent interest in large terrestrial and
marine mammals
• Hardly amenable to standard field protocols
• Growing curiosity in citizen science data
(CSD), but where to start?
4. What
are
the
biases
in
CSD?
• Observer bias
• Spatial bias
• Detection bias
You
see
me
You
don’t
see
me
5. Review
of
the
literature
• List all papers with ‘Citizen Science’ in them
• Scan and check those actually analysing CSD
• Add papers found randomly (ignoring
observer bias…)
• Can we build a taxonomy of methods?
• It’s going to be clumsy and
non-exhaustive
And
boring…
6. 1
-‐
the
‘compara3ve’
approach
• Comparison of results from (classic)
analyses of CSD vs. standardized protocols
- Deemed to be study/species specific
- Results are often convergent
• My review stops here then…
7. 2
-‐
‘filtering’
and
‘correc3on’
approaches
• Methods to filter, select data
• Correction methods: List Length Analysis,
Ball’s approach, Telfer’s approach,
Frescalo’s method, …
Sample
Completed
Least
Bi#ern
Survey
Data
Sheet
8. 2
-‐
‘filtering’
and
‘correc3on’
approaches
• These methods are not robust to bias in
CSD, except the Frescalo method
Check
out
our
paper,
it’s
awesome!
9. 3
-‐
the
‘simula3on’
approach
(Virtual
Ecologist)
• Simulate the bias, and check how your
favorite method behaves
• Case study with wolverine in Scandinavia
• Counts on den sites to infer abundance
• Accumulation of knowledge about the
sites falsely increases observed counts
V.
Gervasi
10. 3
-‐
the
‘simula3on’
approach
(Virtual
Ecologist)
Year
Log(N)
• Tool to design protocols adequately and
explore potential bias
• Convincing way to prove that raw indices
are biased
11. 4
-‐
the
‘regression’
approach
• Use relevant variables to account for biases
Ian
Renner
&
David
Warton
12. 4
-‐
the
‘regression’
approach
• Use relevant variables to account for biases
• Ecological variables
- Affect species’ presence
- Used for building models and predicting
• Observer bias variables
- Affect species detection
- Used only for building models
- Prediction with common level of bias
13. 4
-‐
the
‘regression’
approach
Maps of estimated intensity of Eucalyptus apiculata in Australia
(# detections / km2)
Ecological
variables
only
Ecological
+
observer
bias
variables,
condiFoning
on
a
common
level
of
bias
Sydney
Wollemi
Nat
Park
14. 5
-‐
the
‘combina3on’
approach
• Combine CSD with data collected via
standard protocols (detection/non-detection)
- DND data allow correcting for bias in
opportunistic data
- If no DND for one species, share information
with other species assuming similar bias
OpportunisFc
data
DetecFon/non-‐
detecFon
data
Actual
presence-‐
absence
of
the
species
Will
Fithian
15. 5
-‐
the
‘combina3on’
approach
• Combine CSD with data collected via
standard protocols (detection/non-detection)
- DND data allow correcting for bias in
opportunistic data
- If no DND for one species, share information
with other species assuming similar bias
• Several clever people are on it: Pagel,
Giraud, Dorazio, Fithian, O’Hara, …
16. 6
-‐
the
‘occupancy’
approach
• Correct for false-negatives, and
time/spatial variation in detection
- Account for false-positives
- Extension to multiple species
• How to get the non-detections?
- Relatively easy for checklist data
- But otherwise? You need to know something
about the observer effort…
17. • Typical example of human-wildlife conflict
• Network of observers all over the country
• Map its range, and assess its dynamics
Wolf
range
dynamics
in
France
21. Conclusions
• CSD are great!
• But, we need to deal with bias if we want
to extract meaningful ecological signal
22. Recommenda3ons
(at
your
own
risk)
• A myriad of approaches; no decision tree
• Use simulations to explore effect of bias
• If possible, incorporate detectability via
occupancy / capture-recapture models
• If not, the regression approach, with
covariates to correct for observer bias, is an
avenue to explore
23. Perspec3ves
• The combination approach holds great promise
• The (inhomogeneous) Poisson point process
modeling framework seems to be a unifying
framework
OpportunisFc
data
DetecFon/non-‐
detecFon
data
Actual
presence-‐
absence
of
the
species
24. Perspec3ves
• We should focus more on the citizens
- Fieldwork sheet for recording data on observers too?
- A protocol to collect/store data on both species and citizens
• Technology will help
• As well as social sciences
25. Thank
you!
… and Barney Stinson from How I met your mother, Tom from the Minions, a
random cute cat, Boromir from Lord of the Rings, James Montgomery Flagg
(Uncle Sam), Karine and Wesley, Anne-Sophie and Julie from our boulet
team, and the meme generators