Let us start seeing what is going on when there are no
infected individuals in the population, i.e. when .
In our rough reasoning none
of the 10000 sampled individual will be infected.
But 12% of them will be tagged as positive, exactly the critical value of
1200 we have seen above.
In reality we have neglected the fact that 1200 is an expectation,
in the probabilistic meaning of expected value, but that
other values are also possible. In fact, given the assumed properties
of the test, the number of individuals which shall
result positive to the test is uncertain,
and precisely described by the well known binomial distribution
with `probability parameter' (see Ref. [19] for clarifications)
. The expectation has therefore an uncertainty, that we quantify
with the standard uncertainty [29], i.e. the
standard deviation of the related probability distribution.
Using the well known formula
resulting from the binomial distribution, which in our
case reads as
,
we get, using our numbers,
.
Since we are dealing with reasonably large numbers,
the Gaussian approximation holds and
we can easily calculate that there is about
16% probability to get a number of positives
equal or below 1167, and so on. In particular we get 0.1% probability
to observe a number equal or below 1100, which we could consider
a safe limit for practical purposes.
But, unfortunately, the story is a bit longer. In fact we don't have
to forget that comes itself from measurements
and is therefore uncertain.
Therefore, although 0.12 is its `nominal value', also values below
0.10 are easily possible, yielding e.g. an expected number
of positives, among the not infected individuals, of
for
and for
(hereafter, unless indicated otherwise,
we quote standard uncertainties).
Then there is the question that we apply the tests
on the sample, and not on the entire population.
Therefore, unless the proportion of infectees in the population
is exactly 0 or 1, the proportion of infectees
in the sample (`'), will differ from .
For example, sticking to a reference ,
in the 10000 individuals sampled from a population
ten times larger
we do not expect exactly 1000 infected,
but
as we shall see in detail in Sec.
(we only anticipate, in answer to somebody who might
have quickly checked the numbers, that the standard uncertainty differs from
30, calculated from a binomial distribution, because this
kind of sampling belongs, contrary to the binomial, to the model
`extraction without reintroduction').