Description
In the beginning of the 17th century, John Graunt wanted to determine the effect of the
plague on the population of England; two hundred years later, Pierre-Simon Laplace wanted
to estimate the population of France. Both Graunt and Laplace implemented what is now
called the capture-recapture method. This technique is used to not only count human populations (such as the homeless) but also animals in the wild.
In its simplest form, n1 individuals are “captured,” “tagged”, and released. A while later,
n2 individuals are “captured” and the number of “tagged” individuals, m2, is counted. If N
is the true total population size, we can estimate it with Nˆ
LP as follows:
Nˆ
LP =
n1n2
m2
(1)
using the relation n1
N =
m2
n2
. This is called the Lincoln-Peterson estimator1
.
We make several strong assumptions when we use this method: (a) each individual is
independently captured, (b) each individual is equally likely to be captured, (c) there are no
births, deaths, immigration, or emigration of individuals (i.e., a closed population), and (d)
the tags do not wear off (if it is a physical mark) and no tag goes unnoticed by a researcher.
Goal: In this assignment, you will develop a Monte-Carlo simulation of the capture-recapture
method and investigate the statistical properties of the Lincoln-Peterson and Chapman es1
Interestingly, this estimator is also the maximum likelihood estimate. As you probably guessed, more
complex versions of this idea have been developed since the 1600s.
1
timators of population size, N. (Since you are simulating your own data, you know the true
value of the population size N allowing you to study how well these estimators work.)
Note: It is helpful to save your R workspace to an “.RData” file so that you don’t have to
keep running all of your code every time you work on this assignment. See Lecture 8 for
more details.
1. Simulate the capture-recapture method for a population of size N = 5, 000 when n1 =
100 and n2 = 100 using the sample() function (we assume that each individual is
equally likely to be “captured”). Determine m2 and calculate Nˆ
LP using Eq.1. (Hint:
think of everyone in your population as having an assigned number from 1 to 5,000, then
when you sample from this population, you say you selected person 5, person 8, etc., for
example.)
2. Write a function to simulate the capture-recapture procedure using the inputs: N, n1,
n2, and the number of simulation runs. The function should output in list form (a)
a data frame with two columns: the values of m2 and Nˆ
LP for each iteration and (b)
N. Run your simulation for 1,000 iterations for a population of size N =5,000 where
n1 = n2 = 100 and make a histogram of the resulting Nˆ
LP vector2
. Indicate N on your
plot.
3. What percent of the estimated population values in question 2 were infinite? Why can
this occur?
4. An alternative to the Lincoln-Peterson estimator is the Chapman estimator:
Nˆ
C =
(n1 + 1) (n2 + 1)
m2 + 1
− 1 (2)
Use the saved m2 values from question 2 to compute the corresponding Chapman estimates for each iteration of your simulation. Construct a histogram of the resulting Nˆ
C
estimates, indicating N on your plot.
5. An estimator is considered unbiased if, on average, the estimator equals the true population value. For example, the sample mean x =
Pn
i=1 xi/n is unbiased because on
average the sample mean x equals the population mean µ (i.e., the sampling distribution is centered around µ). This is a desirable property for an estimator to have because
it means our estimator is not systematically wrong. To show that an estimator ˆθ is
2Basically, you are empirically constructing the sampling distribution for NˆLP here. Remember the
Central Limit Theorem which tells us the sampling distribution of the sampling mean? Each statistic has
a sampling distribution and we are simulating it here (but using frequency instead of probability on the
y-axis).
Page 2 of 3
an unbiased estimate of the true value θ, we would need to mathematically prove that
E[
ˆθ] − θ = 0 where E[·] is the expectation (i.e., theoretical average)3
. Instead, we will
investigate this property empirically by replacing the theoretical average E[
ˆθ] with the
sample average of the ˆθ values from our simulation (i.e., Pnsim
i=1
ˆθ/nsim where nsim is the
number of simulation runs; θ is N in this case, and ˆθ is either Nˆ
LP or Nˆ
C as both are
ways to estimate N)
4
.
Estimate the bias of the Lincoln-Peterson and Chapman estimators, based on the results
of your simulation. Is either estimator unbiased when n1, n2 = 100?
6. Based on your findings, is the Lincoln-Peterson or Chapman estimator better? Explain
your answer.
7. Explain why the assumptions (a), (b), and (c) listed on the first page are unrealistic.
3Note that the sample size n does not appear in this equation. For an estimator to be unbiased, this
property cannot depend on sample size.
4Note: This procedure is not a replacement for a mathematical proof, but it’s a good way to explore
statistical properties.
Page 3 of 3
