00:00
Now we have to consider this idea of false
positives and false negatives. If a test says
something as positive and in reality it's
a negative, we call that a false positive.
00:12
And if a test says something as negative,
but in reality it's a positive, we call that
a false negative. It can be confusing, stop
to think about that for a second. And false
positives and false negatives have certain
kinds of burdens in the healthcare system.
00:26
We don't like either one, but some are probably
worse than others. Think about this for a
second, if you are truly diseased, let's say
you have a horrible disease like cancer, but
the test finds that you're negative, that's
bad, you go about your life not getting the
treatment that you need. On the other hand,
if you're not truly diseased and the test
finds that you are, yes you may worry for
a bit, but the follow-up test will confirm
that you're truly disease-free. We kind of
like the second scenario a bit better in most
cases, so we are bit more forgiving towards
false positives than false negatives.
01:01
So let's go over the effects of false positives
and false negatives. A false positive is a
burden on the health care system, because
for everyone who is test positive, there's a
follow-up, there is a biopsy or an ultrasound
or some other kind of expensive deeper investigation,
that costs money. Also if you test positive,
you're going to be anxious and worry until
the confirmation is made that you're actually
disease-free. And of course there's the psychological
aspects of being told that you are a disease
carrier or a diseased individual. On the other
hand false negatives, that's pretty serious.
That means people are missed being diagnosed
when in fact they actually are diseased and they
miss perhaps a window for timely early treatment.
01:45
There is also the shock and disbelief when
you finally do find out that you are diseased.
01:51
Either one of those are pretty serious considerations
and I would argue that's more serious than
the negative considerations of false positives.
As a result, we like to keep our sensitivity
quite high. So which are the false positives
in our contingency table? That's going to
be B. That's going to be the people who test
positive on the screening test, but in reality
are negative, in the sense that they're not
diseased individuals. Our false negatives
are going to those who test negative on the
screening test, but in reality
are diseased individuals, that's going to
be C. So given that we have so many false
positives, we need another kind of measurement
now, something around precision, i.e. if you
test positive on the screening test, what's
the likelihood you actually are diseased?
We call this the precision rate and there
are two kinds of precision rates. There is
a positive predictive value and a negative
predictive value. So the positive value or
PPV or PV+ or precision rate or positive
precision rate, there are lots of things to
call it, is essentially if a patient tests
positive, what's the probability that he or
she really has the disease. Similarly, the
NPV or negative predictive value or PV-,
again, so many different things we can call
this, is a probability that if a patient tests
negative on the screening test, that he or
she really is disease-free. You see the distinction
between PPV, NPV and sensitivity, specificity,
they're testing subtly different constructs
within the context of a screening test. Now
my PPV, my positive predictive value is going
to be the proportion of everyone who tests
positive that are actually diseased. Similarly,
the NPV is going be a proportion of anyone
who tests negative who are actually disease-free.
03:43
So again back to our contingency table, we
always go back here, the PPV can be re-expressed
as a function of sensitivity and specificity.
If you don't trust me you can do the math
yourself, but arithmetically the PPV can be
expressed as the product of sensitivity and
prevalence, divided by sensitivity times prevalence
plus one minus sensitivity times one minus prevalence,
that is a lot of words and numbers in there,
trust me when I say, that's how the math works
out. Similarly the NPV can also be expressed
as a function of sensitivity and specificity.
04:16
So if you compute the sensitivity and specificity
and you know the prevalence of a disease in
your sample, you back compute the PPV, NPV
as well or you can do it using first principles
from the data on the contingency table.
04:29
So let's summarize what we've learned so far.
We've learned the formulas for sensitivity,
for specificity, for positive predictive value
and for negative predictive value. We've also
learned what the prevalence of a disease is
on our sample by looking at all the positive
test cases, divided by the total sample and
we've learned that the PPV and the NPV can
also be expressed as functions of sensitivity
and specificity. Great, let's work through
an example. One of my favorite examples to
use when talking about screening tests is
a pregnancy test, because it's not an unhappy
sort of test to apply. Some people like to
use cancer tests; I find that quite depressing,
pregnancy is a little more fun. So let's say
that 4,810 women take a home pregnancy test,
which is essentially a kind of a screening
test if you consider pregnancy to be the disease
state here, all of them get a follow-up ultrasound
scan. Again, in real life they wouldn't, in
real life those who test positive for pregnancy
would likely go on have an ultrasound,
those who test negative likely would not.
05:31
But in this particular study example, everyone
goes on to get an ultrasound, so we know what
the true state of disease is. In this case
disease is not a disease, it's pregnancy.
05:41
So let's look at our data. From our data,
9 women tests positive on a pregnancy test
and actually turn out to be pregnant, whereas
1 woman test negative and is actually pregnant.
05:52
351 women on the other hand tests positive
on the test and are not pregnant but a whopping
4,449 women tests negative and are indeed
not pregnant. Keep in mind this is simulated
data, it is not real, it's hypothetical. So
I've chosen these numbers to work out a certain
way to tell a story, the actual numbers on
the back of a pregnant test box may tell a
different story. I compute the totals, now
let's do our math. So let's ask ourselves,
if a woman is actually pregnant, what is the
probability that this particular test will
also show that she is pregnant, it'll test
positive. That's a measurement of, do you
know? Sensitivity, that's right. Sensitivity
is given by that 9 divided by the total number
of people who are actually pregnant. So that's
90%, that's a lot, that's a big number.
06:43
In other words, if a woman is actually pregnant,
there is a 90% probability that the screening
test will be positive, that's great. Now if
a woman is actually not pregnant, but she
suspects that she is, so she takes the pregnancy
test, what's the probability that the test
will show correctly that she is not pregnant,
what's that a measurement of? Specificity.
07:05
Alright let's compute our specificity, that's
going to be given by 4,449 divided by everyone
who truly is not pregnant, that gives us 0,
in other words, 92,7% of women, there is a
probability that if she isn't actually pregnant
the test will show that she isn't pregnant.
07:27
Right. Now if a woman tests positive on a
pregnancy test, what is the probability that
she is actually pregnant? Do you see the nuanced
difference between this question and the first
one? We're asking it differently now, if you
test positive on the test, what's the probability
that she actually is pregnant. So imagine
a woman has tested positive on this test,
maybe she's excited, maybe she is not excited,
depending upon her lifestyle choices and she's
going to now follow up with her ultrasound
technician to see if in fact she is truly
pregnant. This is a measurement of PPV, positive
predictive value. We computer our numbers
and we get 2,5%, in other words, if a woman
screens positive for pregnancy, there is actually
only a 2,5% probability that she actually
is pregnant. Remember, this is made up numbers,
this is not the way the actual pregnancy test
would probably play out. And if a woman is tested
negative on a pregnancy test, what's the probability
that she really is not pregnant? What's this?
Negative predictive value, obviously. So I
plug in my numbers into my formula and I get
99,9%. So if a woman screens negative, there
is a 99,9% chance she's actually not pregnant.
08:40
Now think about this if it wasn't pregnancy,
let's say we're talking about something a
bit more dire, like cancer or something equally
as frightening, we would want our NPV to be
as high as possible. We want to be almost
100%, so that if you test negative, you can
be pretty certain that you haven't got the
disease, we don't want you walking away from
the clinic or wherever you got the test thinking
you're disease-free when you're not. So a
high NPV is fantastic, it tells me this is
probably a very good test.