Statistics Theory L6 = Design Strategies For The Real World Flashcards
How do we define the appropriate spatial or temporal scale? (3)
- Identify an appropriate scale in time/space.
- Consider breadth/length & resolution.
- Design our study to capture the variability across the appropriate scale (1 scale? >1 scale?).
Egs of a Temporal scale? (5)
- Dynamics in insect abundance over time.
- Seasonal?
- Interannual?
- Abundance alone?
- Periodicity or trend?
Egs of Spatial scale? (6)
- Patchy use of landscape by a herbivore.
- Which processes/behaviour?
- Forage selection?
- Movement?
- Home range?
- Migration?
NB about Spatial scale?
Affects space (& time) over which we measure stuff.
Design considerations for time? (2)
- Long-term studies.
- Alternatives to long-term studies.
Long-term studies attributes? (3)
- Important for slow processes, rare events, subtle effects & complex phenomena.
- Important for long-lived organisms (long generation time).
- Systems with slow dynamics (eg, climate change).
Eg of slow processes?
Tectonic shifts.
Eg of rare events?
Volcanic eruptions.
Eg of subtle effects?
Need long time to detect the smallest changes.
Eg of complex phenomena?
Nature.
Alternatives to long-term studies? (4)
- Space for time substitution.
- Retrospective studies.
- Fast dynamics for slow.
- Modeling or simulation.
Design considerations for space?
Spatial replication.
Spatial replication attributes? (3)
- Avoid spatial autocorrelation.
- Avoid pseudo-replication.
- Consider spatially-representative samples.
Spatial autocorrelation?
= the degree to which a spatial variable is correlated with itself across a geographic space.
Spatial autocorrelation attributes? (2)
- Measures how similar or dissimilar nearby locations are in terms of a specific attribute.
- Helps identify patterns in spatial data, such as clustering & dispersion.
Why avoid spatial autocorrelation?
It’s because 2 replicates (eg, plots) that are closer together in a landscape are more similar than 2 distant replicates as they are more likely to be affected by the same processes (climate, soil, disturbance, etc).
Pseudo-replication?
= treating 2 replicate plots from the same site as independent samples.
Why avoid pseudo-replication?
It’s because it leads to inflated degrees of freedom, misleading statistical results & an increased inability to apply the results to the target population.
How to avoid pseudo-replication?
Treat replicates as subsamples within sampling units defined by site.
Why consider Spatially-representative samples?
If we want to make inferences about a particular study area, our samples must cover the whole geographic area.
Thing to note about spatial replication?
Lower correlation = plots are independent.
Error definitions? (3)
= means mistake OR variability OR randomness.
Types of sampling biases? (5)
- Sampling error.
- Non-sampling error.
- Observer bias.
- Measurement bias.
- Selection bias.
- Minimising bias.
Sampling error?
= occurs as the consequence of selecting a subset (i.e., sampling units) from a population for study.
Sampling error attribute?
Is the sampling variability that we address with an appropriate design & sample size to achieve a desired precision.
Non-sampling error?
= is more like the “mistake” sense of error.
Non-sampling error attributes? (2)
- Often results in bias.
- Parameters are consistently over-estimated or under-estimated, if there is a systematic bias in the study design.
Observer bias?
= inter-observer differences in terms of technique or skill.
Observer bias attributes? (5)
- Can come from several sources.
- Human error in collecting, recording, transcribing, entering data.
- Differing amounts of bias between observers.
- Differences in what is missed/overlooked (eg, fewer total detections vs fewer species detected in a bird survey).
- Differences between survey occasions for the same observer.
Types of Measurement bias? (2)
- Measurement error.
- Correlated measurement errors.
Measurement error?
= variation that occurs as a consequence of the measurement method.
Measurement error attributes? (3)
- If bias or error is independent of the observer, biases should average to zero.
- Affects estimate precision (SE or 95% CI).
- If precision is too poor for the objectives of the study, changes to data collection methods could help to improve precision.
Eg of Measurement error?
Hanging scale vs electronic balance.
Correlated measurement error?
= occurs if methods consistently measure too high or too low.
Correlated measurement error attributes? (2)
- Systematic transcription or data-entry errors can also occur.
- Therefore, precision estimates are biased low.
Egs of data-entry errors? (3)
- European 1s, 7s and 9s (how they’re written).
- Comma vs full stop.
- North American dates.
Eg of Correlated measurement error?
Poor calibration of tools.
Selection bias?
= choosing correct/relevant variables to measure, not just what is convenient or easy to measure (or what we have budget for).
Selection bias attribute?
Should be based on a thorough review of the species & past work on it.
Best way to minimise bias?
Using a programme like Quality Assurance/Quality Control (QA/QC) that assists in ensuring that data are collected, analysed & reported to “best practices”.
QA?
= focuses on study planning, design, implementation, analysis & reporting (academic side).
QC?
= focuses on protocols & procedures (technical side), routine application of procedures, calibration of tools & instruments.
Aims of such an approach? (4)
- Reduces random or systematic errors.
- Allows the data generated to be analysed, interpreted & communicated to a set of standards.
- Allows the development of qualified personnel & their capacity to use appropriate methods & procedures.
- Specifies a requirement to consult a statistician (at all stages of study), to enter, manage & store data in a particular format (Access vs Excel vs plain text), data archiving standards.
How do we manage Observer bias? (5)
- Use repeatable methods, with little room for error, little judgement or subjectivity.
- Use skilled, qualified, motivated observers.
- Train observers & then retrain them (as they will forget details as data collection proceeds).
- Follow QC protocols (might require finding experts or delving into the literature).
- Standardise data entry (use electronic app like Cybertracker).
Problem you may encounter?
Missing data.
Missing data attributes? (3)
- Avoid! Use something like Cybertracker.
- Record observations as they are made.
- Avoid transcription errors, by reducing the number of steps between measuring & stats analysis.
Ways to deal with missing data? (2)
- Nonresponse errors.
- Deviating from a planned protocol.
Nonresponse errors?
= when one fails to observe or record an individual or unit that is part of the selected sample.
Nonresponse errors attributes? (3)
- Leads to bias.
- Might be possible to correct for missing observations after collection (if done correctly).
- Methods to estimate correction must also be reported.
Eg of Nonresponse errors?
“Imputation”.
Deviating from a planned protocol attributes? (4)
- If at all possible, don’t.
- Is sometimes unavoidable.
- Realise that changes in methods mid-stream could introduce unknown bias to the data set.
- Be very careful.
Things to consider when selecting a sampling protocol? (4)
- Usually no single “correct” approach.
- Better methods & worse methods depending on the objectives of the data collection.
- We ideally want a defensible approach that will hold up to scrutiny (do a thorough review of the pros & cons of each method).
- There could easily be more than one valid approach.
Sampling intensity?
-deals with how many, how long, how often sample units should be measured.
Sampling intensity attributes? (3)
- Depends on the objectives of data collection.
- Often a trade-off among them, more sample units to sample might mean you can measure them less often.
- Pilot study might help determine the sampling intensity.
What could possibly go wrong? (10)
- Weather not cooperating.
- Equipment fails/gets stolen.
- Samples go missing/get destroyed.
- Field methods not followed by assistants.
- No game guard; or yes, game guard but no rifle.
- Funding situation changes.
- People/collaborators “go dark”.
- All zebras in your study area are captured & moved to some other area.
- Wildlife sweeps through your study area.
- Global pandemic.
Okay, something went wrong. Now what? (2)
- Don’t panic (most work can be salvaged).
- Pivot (mid-point change?).
- Rarely will you have to redo from start with a whole new project (or quit altogether).
Things to consider if problems arise? (4)
- Can I still address my objectives, or must they be changed?
- How much of my existing data can be used? How much do I need to collect? Are data sets still comparable?
- Can I use the same study design or must I change it?
- How do options for analysis change & what are the consequences?
Some considerations if options for analysis change? (5)
- Such changes could mean estimates with reduced precision & increased possibility of bias.
- Reduced certainty in conclusions.
- Could lead to publication in a lower-quality journal.
- Do the best you can.
- Try to publish anyway.