Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Keep in touch
@SEEC UCT
SEEC.UCT
ask to be added to mailing list([email protected])
www seec.uct.ac.za
Experimental and Survey Design
I Why?
I What?
I How?
Yoccoz, N. G., J. D. Nichols, and T. Boulinier. 2001. Monitoring of biological diversity in space and time. Trends
in Ecology & Evolution 16:446 - 453.
Experimental and Survey Designs
from Mackenzie et al 2006, Academic Press
6 / 55
Experiments: what they are and why we need them
I carefully controlled environment
I everything stays constant
I except the factor of interest (treatment factor)
I measure change in response
I infer causal relationships
7 / 55
Observational vs experimental studies
Hypothetical example: Vegetation in relation to temperature andaltitude
observational
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
temperature
altit
ude
low high
low
high
experimental
temperature
altit
ude
low high
low
high
* * *
* * *
* * *
8 / 55
Observational vs experimental studies
Hypothetical example: Vegetation in relation to temperature andaltitude
observational
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
temperature
altit
ude
low high
low
high
experimental
temperature
altit
ude
low high
low
high
* * *
* * *
* * *
9 / 55
Summary
Observational Experimental
↓ ↓
random sample randomize
↓ ↓
representativeequalize groups (average out
all other factors)
↓ ↓
make statements about thepopulation
causal inference (cannot sayanything about population in
general)
10 / 55
Three principles in experimental design
1. Randomization
2. Replication
3. Reduce (unexplained variation)
What is the experimental unit?
11 / 55
Three principles in experimental design
1. Randomization
2. Replication
3. Reduce (unexplained variation)
What is the experimental unit?
12 / 55
ED Principle 1: Randomization
Randomization
is the assignment of treatments to experimental units at random.
13 / 55
What happens if we don’t randomize?
14 / 55
What happens if we don’t randomize?
15 / 55
What happens if we don’t randomize order?
time
A A A B B B C C C
16 / 55
What happens if we don’t randomize order?
time
A A A B B B C C C
tem
pera
ture
17 / 55
Summary: Why is randomization important?
I average out systematic effects (anticipated or not), extraneousfactors not directly controlled (prevent confounding withunknown factors)
I ensures that treatment is independent of experimental unit’sproperties
18 / 55
ED Principle 2: Replication
each treatment is assigned to > 1 experimental unit, randomly andindependently
19 / 55
Uncertainty – Variability
If all experimental units were exactly identical, it would be easy toestablish cause and effect relationships.
A B
treatment
resp
onse
●
●
A B
treatment
resp
onse
●●●●●
●●●●●
A B
treatment
resp
onse
●
●
●
●
●
●●
●
●
●
20 / 55
Uncertainty – Variability
If all experimental units were exactly identical, it would be easy toestablish cause and effect relationships.
A B
treatment
resp
onse
●
●
A B
treatment
resp
onse
●●●●●
●●●●●
A B
treatment
resp
onse
●
●
●
●
●
●●
●
●
●
21 / 55
Is this proper replication?
22 / 55
Is this proper replication?
23 / 55
Is this proper replication?
24 / 55
Is this proper replication?
25 / 55
ED Principle 3: Reduce unexplained variation
1. keep all other factors constant
2. blocking
Blocking:
is the grouping of similar (homogeneous) experimental units intoblocks.
I The experimental units in one block aresimilar.
I Each treatment occurs once in each block.
I Common blocking factors: age, spatiallocation, time (day or year)
I each block is a replicate
26 / 55
ED Principle 3: Reduce unexplained variation
1. keep all other factors constant
2. blocking
Blocking:
is the grouping of similar (homogeneous) experimental units intoblocks.
A
C
B
A
B CI The experimental units in one block are
similar.
I Each treatment occurs once in each block.
I Common blocking factors: age, spatiallocation, time (day or year)
I each block is a replicate
27 / 55
Controls
I to measure the effect of a treatment, we need to compareresponse with to without treatment
I eliminate alternative explanations for results
28 / 55
Part 2: Constrained Design Studies
One or several principles of experimental design cannot be adheredto:
I impact studies
I observational studies
29 / 55
Sampling Designs for Impact Assessment
from Green 1979, John Wiley & Sons
30 / 55
Example: Impact of a new wind energy farm
Problem formulation:I not an experiment, but impact assessmentI cannot randomly choose locationI only one farm (no replication of ‘treatment’)I impact: effect of wind farm on bird behaviour around turbinesI hypothesis: change in the local density of raptors with
presence of wind farm compared to abscence of wind farmI sampling method: vantage point surveys
31 / 55
Control and Replicate Time and Location
before
temporal control
o o o: temporal replication
spatial control
o o o: temporal replication
after
o o o
spatial control
o o o
32 / 55
Control and Replicate Time and Location
before
temporal control
o o o: temporal replication
spatial control
o o o: temporal replication
after
o o o
spatial control
o o o
33 / 55
Control and Replicate Time and Location
before
temporal control
o o o: temporal replication
spatial control
o o o: temporal replication
after
o o o
spatial control
o o o
34 / 55
Control and Replicate Time and Location
before
temporal control
o o o: temporal replication
spatial control
o o o: temporal replication
after
o o o
spatial control
o o o
35 / 55
Random Sampling
Purpose: sample must be representative (of the population)
I where: where to place the vantage points
I when: times of day, season, weather conditions
36 / 55
Observational studies
No manipulation
I Cause and effect unclear
I Influence of unmeasured variables can never be ruled out(confounding)
Two things to worry about:
1. Observation process (Solution: distance sampling, repeatedvisits, multi-observer designs)
2. Representativeness
Observational studies
No manipulation
I Cause and effect unclear
I Influence of unmeasured variables can never be ruled out(confounding)
Two things to worry about:
1. Observation process (Solution: distance sampling, repeatedvisits, multi-observer designs)
2. Representativeness
Observational studies
1. Identify target population
2. Random sampling: no assumptions about the populationneeded
I every unit has a known probability of being in the sampleI sample is drawn according to these probabilitiesI selection probabilities are used when making estimates
3. Model-based sampling: unbiased if model describespopulation well
Observational studies
1. Identify target population
2. Random sampling: no assumptions about the populationneeded
I every unit has a known probability of being in the sampleI sample is drawn according to these probabilitiesI selection probabilities are used when making estimates
3. Model-based sampling: unbiased if model describespopulation well
Observational studies
1. Identify target population
2. Random sampling: no assumptions about the populationneeded
I every unit has a known probability of being in the sampleI sample is drawn according to these probabilitiesI selection probabilities are used when making estimates
3. Model-based sampling: unbiased if model describespopulation well
Observational studies
1. Identify target population
2. Random sampling: no assumptions about the populationneeded
I every unit has a known probability of being in the sampleI sample is drawn according to these probabilitiesI selection probabilities are used when making estimates
3. Model-based sampling: unbiased if model describespopulation well
Random sampling
I Simple random sampling
I Stratified random sampling
I Cluster sampling
I Multistage sampling
I Adaptive sampling
Density of hadeda ibises
What is the average density of hadeda ibises in this landscape?
Simple random sampling
x x
xx
xx
x x
x x
xx x x
x
xx x x x
µ = 1.98x̄ = 1.3595%CI : 0.21 − 2.49
Unbiased regardless of how densities are distributed.
Convenience sampling
x
x
xx x
x
x
x
x
x
x
xx
x
x
xx
x
x
x
µ = 1.98x̄ = 3.0595%CI : 1.91 − 4.18
Preferentially sampling high-density areas leads to estimates thatare biased high.
Stratified random sampling
x
x
xx x
x
x
x
x
x
x
xx
x
x
xx
x
x
x
µ = 1.98
x̄ = 1N
∑Lh=1Nhx̄h = 1.75
s2 =∑L
h=1(NhN )2(Nh−nh
Nh)s2hnh
95%CI : 1.05 − 2.44
Strata should be as homogeneous as possible; blocking is a relatedidea in experimental design.
Model-based sampling
x
x
xx x
x
x
x
x
x
x
xx
x
x
xx
x
x
x
x̄ = 1.71
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
habitat gradient
dens
ity
1 5 10 15 20
0
2
4
6
8
10
Can be unbiased if the model is good; no random sampling needed.
Cluster sampling
x
xx
x
x
x
xx
x
x
x
xx
x
x
x
xx
x
x
x
xx
x
x
x
xx
x
x
x
xx
x
x
x
xx
x
x
I Clusters selected at randomI Most efficient if variability within cluster is high and variability
among clusters is low
Multistage sampling
xx
xx
xxx
x
xx
x
x
x
x
x xx x
xx
1. Select large units at random
2. Select small units at randomwithin selected large units
Adaptive sampling
x
x
x
x
x
x
x
x
x
x
x
xx
xx
x
x x
x
x
●
●●
● ● ●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●
x
x
x
x
x
x
x
x
x
x
x
xx
xx
x
x x
x
x
●
●●
● ● ●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●
x x xx x xx x x xx x x xx x x x
1. Select units at random
2. If species detected, sample surrounding units
Random sampling
I Probability of being in the sample is known for each unit
I This probability is used when making inference
I Lots of designs to choose from
I Some designs are more effective than others for your situationI Things to think about:
I Scale of variability and spatial correlation (do a pilot)I Do you have informative covariates? (stratification;
model-based sampling)I Costs of travelling to sites
I Things get complicated quickly
Summary: Nine principles of good study design
1. formulate problem (design, predictors, hypothesis ...)
2. replicate time, location
3. random sample
4. control (space and time)
5. conduct a pilot study
6. worry about observation process
7. subsampling, stratified sampling where necessary
8. appropriate sample unit size
9. use an appropriate statistical modelAdapted from: Green 1979, John Wiley & Sons
Key references
Experimental design:
I Underwood, A. J. 1997. Experiments in ecology: their logical design andinterpretation using analysis of variance. Cambridge University Press.
I Montgomery, D. C. 2012. Design and Analysis of Experiments, 8th edition.Wiley.
Constrained designs / impact studies:
I Green, R. H. 1979. Sampling design and statistical methods for environmentalbiologists. John Wiley & Sons.
Survey design:
I Thompson, S. K. 2012. Sampling. 3rd edition. John Wiley & Sons.