Observer 1 versus 2

Visibility Bias and Population Estimation Technique

Protocol for Big Bend National Park Monitoring Field Crews

Visibility bias, the error associated with the failure to observe all animals during a specific survey, occurs in all studies which attempt to count the numbers of animals in the field.  How detectable or visible animals are depends on many factors, including "animal behavior and dispersion, observers, weather, habitat type, equipment, and methodology" (Cook and Jacobson 1979, p. 735).  A technique described by Cook and Jacobson (1973), what we call the "observer 1 versus 2" technique, allows the calculation of detection probabilities (the probability of not observing an animal, or p) and estimated population sizes (N).

The technique works as follows.  Observer 1 and 2 are designated at the start of the survey as primary or secondary observers, respectively.  The two observers must walk side by side at the same rate along the spring such that they can see the same section of spring simultaneously.  The primary observer behaves as the only observer present and points out all his/her tadpole and adult sightings to the second observer. The second observer confirms and records on the data sheet all sightings by the primary observer (observer 1 column), but additionally records in a separate column on the data sheet (observer 2 column) any sightings that the primary observer missed.  The secondary observer must not tell the primary observer what he/she missed or make any derogatory comments, noises, or motions whatsoever!  At the end of the first spring transect, the observer roles are switched, such that observer 1 becomes observer 2, and observer 2 becomes observer 1 for the second spring transect. Details on the calculation of p and N using this technique are presented in Cook and Jacobson (1973).

At Big Bend National Park, this technique only works at a subset (approximately 10%) of the springs and streams visited.  It works well at springs where the observers can walk side by side, or one right after the other, where there is little vegetation along the banks obscuring the view of the stream bank and water channel.  In order for this technique to work well, the water should also be clear.  In some springs, the technique works for both tadpoles and adults, but in others, the technique only works for adults.  It also does not work well in situations where there are hundreds or thousands of tadpoles.  In this case, the first observer will scare the tadpoles and the second observer will not have a chance to see them.

Counting Tadpoles

• Goal: Estimate number of observable tadpoles in springs.
• Test of “Index assumption:”
  • Relate count to actual number present: test whether same proportion of tadpoles are counted in each spring
  • Address one important component of observability -the proportion of the observable tadpoles that are actually seen during counting
• General Procedure: Use a multiple-observer approach to estimate the detection probability of tadpoles in springs

Design of Study
Multiple Observers (Cook and Jacobson 1979)

• Basic design:
  • Divide spring into 50 m transects
  • 2 counters
• On a transect, designate observers:
  • Primary observer
  • Secondary observer
  • Primary counter points out tadpoles, secondary observer notes only tadpoles not seen by first observer
• Observers switch roles between transects

Example Data Set
Cattail Spring, 25 May 1998 Survey

Observers: Robin Jung and Sam Droege

Estimate Detection Probability (p)

• 2 Goals:
  • Estimate detection probabilities (proportion of tadpoles seen by each observer)
  • Estimate population size of tadpoles
• Of relevance to the value of the index, want to know:
  • Do detection probabilities vary among observers?
    • Test within pairs for differences
  • Do detection probabilities vary among springs?
    • Test among springs for differences
• If the probabilities differ, have to incorporate estimation in operational surveys

Actually Estimating p

• For details see Cook and Jacobson (1979)
• Notion:
  • Each observer has an unknown detection probability p
  • We have number counted by observers- assume that these are binomial random variables (e.g., like flipping a coin that has a probability p of “heads.” Expected number seen is actual number (number of coin flips) times p (probability of “heads”).
  • Cook and Jacobson recognized that, by conditioning on total numbers counted, p can be estimated for each observer
  • Using p, can further estimate N (total population size) and SE(N) for each site

• We used Program SURVIV
  • Very flexible program
  • Allows estimation of p and its variance
  • Tests for differences in p among observers, springs
    • Use AIC (Akaike’s Information Criterion)-measure of “fit” of model.   Minimim AIC == best model
  • Estimate p for each observer in the pairs
  • Test for differences between observers
  • Test for differences among springs
• Take the “best” estimates of p, use it to adjust counts to provide population estimates

Reference

PRIMENet Amphibian Monitoring Program Homepage

Patuxent Homepage