Tip:
Highlight text to annotate it
X
Art Reingold: Morning. Good morning. So today I
need to finish up on what I should have finished up op Monday.
This is an important issue with regards to understanding case
control studies an what the odds ratio does or does not
estimate. So we've already talked a little bit about the fact
in case control studies controls are intended to represent the
source population from which the case came. That is the
purpose in choosing controls to be representative of a source
population in which the cases originated. I've also said that
all case control studies occur within a population or if you
will within a cohort even though we may not be able to always
define exactly who is in that cohort.
So we talked about different approaches to selecting
controls, friends, neighbors, random digit dialling. Random
samples of the population. Another important aspect about
thinking of control selection is when they are selected with
respect to this notion of everyone is basically in a cohort.
So if you think about this cohort or population within which
case control studies occur and that's a cohort of people
marching along together in time you can think of three options
for choosing controls in terms of the timing of when they are.
One of these is referred to as survivor sampling.
Survivor sampling is when you choose the controls from among
everyone who has come to the end of their risk period. They no
longer are at risk of getting the disease in question. And
they have survived. They haven't gotten the disease. They are
survivors.
This is true in many outbreak situations where once you
get past your incubation if you didn't get salmonella or
campylobacter you won't get it. You might get it later
unrelated but you aren't going to get it as part of this
outbreak. You are survived, made it to the end of the risk
period and not gotten sick. Many case control studies that we
do, lung cancer and smoking, at the time you choose controls
who don't have lung cancer, those people could get lung cancer
tomorrow. They could get lung cancer next week. They could
get lung cancer next year. Their risk period doesn't end for
lung cancer until they die of something else. Right?
So in many case control studies what we're actually
doing is sampling controls from among the individuals in the at
risk population who have not yet experienced the outcome by the
time a given case is diagnosed but who remain at risk of that
outcome.
Okay? And we refer to that as risk set or density based sampling. The
third thing you can think of is choosing the controls from among
everyone who is in the cohort at the baseline. Irrespective of what
happened to them later, choosing a random sample or a sample of people
who are in the cohort at the baseline and we refer to that as case based
or case cohort sampling. Okay. So to illustrate this if you envision a
cohort marching along in time, survivor sampling, people come to the end
of their risk period. They haven't gotten the disease and we choose
controls from among those survivors. Risk set sampling we are following
people over time. Some people get disease at the time a case develops.
We choose controls from among the people in that population or cohort
who have not yet gotten disease but who might get it later.
Okay. And in base sampling we have a cohort following
over in time. We choose the controls from among everybody who
is present at the baseline in the cohort.
It's important to get straight this difference and
understand it. Um, and it is when you do survivor sampling the
odds ratio in the case control study that you will calculate if
you choose an unbiased set of cases and an unbiased set of
controls the odds ratio will be an estimate of the odds ratio
in the population.
Okay. And the more common the outcome is the higher
the odds ratio will be the further from one it will be in
comparison to the risk ratio or the rate ratio.
So, and if it is not a rare condition then the odds
ratio may substantially overestimate the risk ratio or the rate
ratio. Okay. So in this situation with survivor sampling this
idea of the rare disease assumption you may have read about in
textbooks is correct. Okay? If it's not a rare disease the
odds ratio will overestimate the risk ratio or rate ratio. But
in these two situations when you're sampling controls either in
density based sampling or case cohort sampling you don't need
the rare disease assumption for the odds ratio to estimate the
relative risk. In fact in the density based sampling is odds
ratio is a good estimate of the incidence rate ratio. In case
cohort sampling the odds ratio is a good estimate of the risk
ratio. I'm not going to derive and show you mathematically.
If you are interested you'll have to take 250 B for that. The
point is in these circumstances you do not need the rare
disease assumption in order for the odds ratio to be a good
estimate of the rate ratio or risk ratio. Okay?
So, um, in studies using survivor sampling to choose
controls the odds ratio while still a valid measure of
association between exposure and outcome will always be further
from one, from unity than the risk ratio.
Okay. Now that's just a fact. And it's frequently
overlooked in many textbooks. Um, in studies using this risk
set or case based sampling which I'm going to illustrate in a
little bit, the individuals who subsequently develop the
outcome of interest may be included as controls. This bothers
people a lot. The idea that somebody who goes onto become a
case is included in the study as a control.
Okay. Nevertheless that is the case. And in addition,
some people if they are initially included as controls and then
become cases could be included in the study both as a control
and then subsequently as a case. Okay.
Um, so, there are a lot of older textbooks and even
some recent ones that talk about three conditions that have to
be met for the odds ratio to be an accurate estimate of the
relative risk. The controls have been representative of
non-cases in the population. That's always true. The cases
have to be representative of cases in the entire population
with respect to the exposure of interest. That's also true.
We'll talk about selection bias in a couple of weeks. But the
notion of the frequency of disease in the population must be
small, this is the rare disease assumption is only true when
you are doing survivor sampling. Okay?
And so this need for the disease to be rare is only
correct when you are using survivor sampling. And that's
typically the case in an outbreak investigation. Let me
illustrate for you when this is in fact true and when the odds
ratio and the risk ratio will be very different. I made up
these data. You have 200 people at a banquet. The chicken is
contaminated with salmonella. There is an outbreak of
salmonella gastroenteritis. Who is more likely to get sick?
People who ate the chicken or didn't eat the chicken? People
who ate the chicken. That's obvious. Simpleminded. 120
individuals ate the chicken. 60 became ill. What's the attack
rate or cumulative incidence? 60 out of 120 or 50 percent.
Out of the 80 individuals who didn't eat the chicken four
became ill. What's the attack rate in the unexposed? Five
percent. What's the risk ratio or relative risk in this
population? Ten. People who ate the chicken are ten times
more likely to get sick than people who didn't eat the chicken.
So, that's just what I'm showing here. The attack rate
in the exposed is 50 percent. The attack rate in the unexposed
is five percent. The risk ratio is ten. In this example if
people get past the incubation period and haven't gotten
salmonella they have survived. You are going to choose
controls from among the survivors. Standard situation in a
case control study. So if you did the case control study
including everybody, in other words if you just calculated the
odds ratio for the entire population here's the odds ratio.
You have 60 chicken eater who got six. Four non-chicken eaters
who didn't get sick. 60 chicken eaters who didn't get sick.
76 who didn't eat chicken that didn't get sick. The odds ratio
is 19. Which is markedly different from the relative risk
which is ten. It's almost twice the relative risk. In this
outbreak the outcome of interest is not a rare outcome. It's
common. Half the people who got exposed got sick. This would
be an example of survivor sampling with a common outcome the
odds ratio is not a good estimate of the risk ratio. Okay?
So in that one circumstance it is correct that the odds
ratio will always overestimate the risk ratio or rate ratio.
It's still a valid measure of association. If you did a case
control study and only chose a sample of cases and a sample of
controls and they were a representative sample of cases and a
representative sample of controls the odds ratio from that case
control study would also be 19. Right?
The odds ratio wouldn't change just because you took a
sample of cases and a sample of controls. It would be the same
and it would be much higher than the risk ratio.
Okay. So, when survivor sampling is used the odds
ratio will deviate from the risk ratio. It will always be away
from the null and increase in magnitude as cumulative incidence
increases.
In any case control study whenever the sampling
approach in the absence of bias the odds ratio from the case
control study will be an accurate estimate of the odds ratio in
the population.
Okay. And the odds ratio is still a valid effect
measure. You can calculate it and discuss it. But in survivor
sampling it is not reasonable to assume that it is a valid
estimate of the risk ratio.
But in other types of case control studies you don't
need the rare disease assumption in order for the odds ratio to
estimate the risk ratio or the rate ratio.
This is just meant to illustrate that with increasing
risk in the unexposed group here you've got what the true
relative risk is. Here's what the odds ratio would be. You
can see in a very rare disease, this is survivor sampling
again. You can see that the overestimate of the odds ratio
compared to the risk ratio is very small. But as the outcome
becomes more and more common, the odds ratio deviates more and
more from the risk ratio. When survivor sampling is what's
being used. Okay.
Um, but we're going to talk in a minute now about risk
set or case based sampling when the odds ratio is an accurate
estimate of the rate or risk ratio even if the disease is not
rare.
Okay. Just to sum up on case control studies their
advantages are typically they are very inexpensive and fast
compared to cohort studies. So they are inexpensive, they are
efficient and they are fast. You can study rare diseases that
can't be studied using other designs and you can look at many,
many different exposures all at the same time in a single case
control study.
Okay. Those are the advantages of case control
studies. The disadvantages as we'll talk about when we talk
about bias in a couple weeks are that it can be very difficult
to avoid selection bias either of the cases or of the controls.
It can be very difficult to avoid information biases in terms
of particularly collecting information about the exposure of
interest. So that can be biased.
Um, you tend to have a loss of precision in the
estimate of the risk ratio or odds ratio compared to studying
the entire cohort because you are sampling the population and
have a smaller sample than the whole population you basically
lose precision. Okay.
And of course case control studies may not be feasible
if the exposure of interest is very rare. Okay. If you just
think about that for a minute if the exposure is very, very
rare and you do a case control study the likelihood is none of
your cases or none of your controls or very few of them will
have the exposure of interest and the study won't answer the
question of whether there's a relationship between the exposure
and the outcome. So case control studies are very useful in
some settings, but they have their problems.
So I want to move on now to this idea of nested case
control studies and case cohort studies. Are there any
questions about this so far?
Okay. We may not get to case crossover studies today
in which case we'll talk about them next Monday. Friday Craig
Steinmaus is going to talk about meta analysis. Today we'll
talk about nested case control studies and case cohort studies.
All case control studies occur within a cohort. You may or may
not be able to define the cohort in many instances but they all
occur within a cohort. But there are times when they occur
within a well defined cohort. And they can be a very, very
effective study design even when you have an entire cohort at
your disposal. And you might ask, well how could that be? If
we have a whole cohort why couldn't you analyze the cohort as a
cohort study. Why would you do a case control study within a
cohort? How could that possibly be advantageous?
Okay. Um, so, again nested case control studies occur
within a defined cohort like the nurses health study or the
British physicians study. In a nested case control study you
choose either you take all the cases or a representative sample
of the cases just like in any case control study. The outcome
of interest in the defined cohort. And you choose controls
from among the cohort members who are still alive and haven't
been diagnosed with the outcome of interest as the date of
onset of the respective cases. This is going back to that type
of sampling. This is risk set or density sampling.
So somebody in the cohort develops lung cancer. You
choose controls from among all of the cohort members who as of
that date have not yet developed lung cancer. Right? They are
still at risk of getting lung cancer. They can get lung cancer
next month or next year or tomorrow but they don't have it as
of the date when the case is diagnosed.
In this instance with risk set or density sampling the
odds ratio is an estimate of the rate ratio in the base
population. You don't need the rare disease assumption. And
furthermore, individuals select as controls at one point in
time and included in the study as controls if they develop the
disease a month or year later are included the study as cases.
They can be in the study both as controls and subsequently as
cases. I'm sorry if that bothers you, nevertheless it bothers
a lot of people, nevertheless it's correct.
So, the basic idea here is like any case control study
we've got a cohort of people. Some people develop the disease.
We look at their risk factors absent or present and we choose
controls from among the cohort members who are disease free at
the time the case develops.
Okay. So it's this type of sampling, risk set
sampling.
Okay. So why might this make sense? Well, this is
typically useful when it is either difficult or impossible to
measure the exposure status in the entire cohort. So suppose
you've got a cohort of 100000 nurses. And you took a blood
sample of all hundred thousand nurses and stored it away in the
freezer. But suppose the thing you wanted to measure the
exposure of interest in the blood sample costs a thousand
dollars per sample test. Many of the things people like to
test, genetic markers, chemicals, all kinds of things are
expensive. What would it cost to do the test on 100000 people
if it cost $1,000 a person. What would it cost? What's 100000
times a thousand.
>>>: A hundred million.
Art Reingold: A hundred million dollars. It
would cost a hundred million dollars to do that test on all the
cohort members. Most studies don't have a hundred million
dollars to spend on laboratory testing on the specimens. It's
impractical to do the test on all 100000 people. Right?
Typically this is a situation where a nested case control study
might make sense either where the exposure you want to measure
is incredibly expensive and you can't afford to do it on
everybody in the cohort and that can be true either if it's
some test biological test in the lab or if it requires
additional interviewing and collection of data that you didn't
get from everyone in the cohort at the beginning. That's also
frequently the case.
So here's an example. Here the research question is,
is there an association between body burden of organochlorines
and the risk of non-Hodgkin lymphoma. Anybody know an example
of organochlorine.
>>>: DDT.
Art Reingold: DDT. Everybody knows the
pesticide. It's an organochlorine. You might be interested
does having that in your body increase the risk of getting this
kind of cancer. A lot of people are worried about the
carcinogenic effects of chemicals. Here the idea is if you
want to answer this question you need to measure these
chemicals in people before anybody gets cancer. Right?
If you wait until people have cancer and then take
samples from them you may be taking samples 20 or 30 years
after the exposure period of interest. And having cancer might
even influence the body burden of something you are interested
in measuring. You don't want to measure the samples at the
time people have cancer. You want to measure samples before
anybody gets cancer. So you have to find a cohort in which
somebody stored away all these samples. Okay. And now within
that cohort you are going to do a case control study. You need
to measure these levels in individuals before they get the
cancer. In addition the incidence of non-Hodgkin lymphoma is
low. Very few are going to get this cancer. Testing all
hundred thousand of them for DDT levels in their blood would be
expensive. At the end of the day you might have 50 cases and
900950 controls. And it wouldn't be very efficient to study,
to test all hundred thousand samples.
Okay. So in order to do this study you need a cohort
big enough to produce an adequate number of cases in the
outcome of interest. Somebody needs to have stored away
baseline samples from cohort members of adequate volume that
permits testing for organochlorines. You need to find the
cases of non-Hodgkin's lymphoma and choose your controls from
individuals in that cohort. You take this cohort. You find
all the cases of non-Hodgkin's lymphoma that develop over 20 or
30 years. You take the date of diagnosis of the case and from
among all the cohort members alive as of that date who don't
have that type of cancer you take a random sample. There are
your controls. What do you do in order to test the hypothesis?
You found all the cases. You've sampled the non-cases. And
identified them and now what do you do?
>>>: Check the blood for organochlorines.
Art Reingold: Go to the freezer and thaw out the
samples an measure the organochlorines. Not on a hundred
thousand people but on a couple of people.
What happened to the data from that study? I'll have
to dig those up. Somebody happened to the slides. Here it is.
It's out of order. Here you can see all together how many
cases do they have in this study. You can add them up.
They've got 74 cases of non-Hodgkin's lymphoma. 74 cases. How
many controls? 147 controls. Right?
So how many people have they studied all together in
this case control study? A couple hundred people. They don't
have to test organochlorines on a hundred thousand nurses.
They test them on a couple hundred people. And testing them on
an additional 99000 controls would cost an astronomical amount
of money and would actually add relatively little to the power
of the study.
Why? Because they are all in one cell. They are all
in the non-diseased, the non -- they don't have lymphoma
category. Right? They don't have much power but they cost an
astronomical amount of money. Here you study about 200 people.
Here you can see the quartiles of DDT from lowest to highest.
Cases, controls. Calculate an odds ratio. It's a matched odds
ratio in this case, conditional logistic regression and you can
see there's no relationship between how much DDT was in your
stored blood sample taken 20 years ago and your risk of getting
non-Hodgkin's lymphoma. As opposed to PCB, another form of
organochlorine where in fact there's increasing evidence of
odds of disease with increasing body burden. Right?
But in this nested case control study you don't have to
measure organochlorines on 100000 nurses you can study it on
200. Does that make sense why you would do this? Okay. I
don't know how that got out of order.
This is another example of a nested case control study.
Um, it's the same basic -- it got out of order. I apologize.
Here you can see the total number of serum samples tested.
76-case, 147 controls. 223. Sorry, this other cohort of 25000
cohort members. Somehow they've gotten confused. I apologize.
This is the same issue. Organochlorines and the non-Hodgkin's
lymphoma. You can answer the question by testing samples of
200 of them rather than 25000 of them. This got out of order.
That's the idea behind a nested control study. Within an
existing cohort you find the cases you use incidence density
sampling to choose controls and then you only measure the
exposures of interest in those cases and in those controls
rather than in the entire cohort.
Okay. Now a very similar study design is a case cohort
study. Okay. And in fact case cohort studies and case control
studies are identical except for one thing. That's, who is
eligible to be a control? In nested case control studies we
choose controls from among individuals who are disease free as
of the date of diagnosis of the case.
In case cohort studies we choose controls from everyone
in the cohort at the baseline. Okay?
So everyone who is in the cohort at the baseline is
eligible to be selected as a control. Every member of a cohort
has the same chance of being included as a control. The odds
ratio is an estimate of the risk ratio in the base population.
As in a nested case control study individuals selected as
controls can become cases.
Okay. And they can include in a study both as controls
and as cases.
So, again, here this is this cohort marching forward,
all of the controls are selected at random from among everybody
in a cohort at the baseline.
So if you are asked what's the difference? Why would
you choose one of these instead of the other? The primary
reason for choosing a case cohort instead of a case control
design is convenience.
Okay. The data that allow risk set sampling from a
specific risk set for each case may not be available. And also
if you want to study more than one health outcome you can use
the same control group to study multiple outcomes in a case
cohort design but not in a nested case control study. For a
nested case control study you would need to choose different
controls for each outcome you want in the study. Heart attack,
lung cancer, stroke. You would have to choose -- why would you
have to choose different controls for each of those outcomes in
a nested case control study?
Suppose you are interested in organochlorines in blood
or whatever it is. Is that associated with your risk of having
a stroke? That's one question. Is that associated with your
risk of having a heart attack? That's the second question. Is
it associated with risk of lung cancer? That's a third
question. Why would you need to choose different controls for
each of those research questions?
>>>: Because you are selecting for non-diseased.
Art Reingold: Because you are selecting for the
people who don't have the particular disease at the time of
onset of the cases. That's going to be completely different
for lung cancer than it is for heart disease and stroke. In
each one you would have to choose different controls. If you
do a case cohort study if you take a random sample of everyone
at baseline you can use the exact same control group for all
these different outcomes. Because it's a random sample of
everybody in the cohort at the baseline.
Okay. Here's an example of a case cohort study.
Here's a question of question is chronic infection with this
bacterium, chlamydia pneumoniae increase your risk of coronary
heart disease. Why might this be a plausible hypothesis?
Chronic infection with this bacterium increasing your risk of
coronary heart disease? Does that seem like a plausible
hypothesis? Anybody familiar with this hypothesis?
So one hypothesis is that chronic inflammation in your
blood vessels is an important aspect of your risk of having
coronary artery disease. Okay?
And so chronic infection might be one cause of chronic
inflammation. Your body is continuing to respond with some
infection that is there for years and years and years. This
was considered such an important hypothesis people actually did
randomized control trials to test this hypothesis. How would
you do a randomized control trial to test this hypothesis?
Chronic infection increases your risk of a heart attack. How
could you do a randomized control trial to test that
hypothesis? What could you randomize people to, to test that
hypothesis that makes any sense? You can't randomize them to
chronic infection, yes or no. That's typically not possible.
What could you randomize them to?
>>>: Could you randomize them into treatment
groups?
Art Reingold: What type of treatment?
Antibiotics effective against this bacterium. Who would you
potentially put into that study? If you just took everybody in
this room and randomly assigned antibiotics or no antibiotics
to see who had a heart attack and who didn't, what would be the
problem with that study. Pardon?
>>>: (Inaudible).
Art Reingold: So most of you are not going to
have your first heart attack for quite some time I hope. The
incidence rate of heart attacks in healthy young people is very
small. The first thing is you have to wait a really long time.
We don't want to take healthy young people like you. What's
the problem with taking healthy old people like me? I'm at
much more risk of having a heart attack.
>>>: You are not at risk for chronic infection
with chlamydia.
Art Reingold: We have to measure that and see.
We don't know what the prevalence is. The main problem is in
healthy old people like me the risk of heart attack is very
small. It would take an enormous study and a long time. The
studies were done in people who already had a heart attack.
They are at great risk of having a second heart attack. Those
studies were a dismal failure. Antibiotics against this
bacterium do not prevent a second heart attack. A randomized
control trial has been done to test that hypothesis. A good
study would be to randomize people like you to antibiotics and
no antibiotics and follow you for 50 years. You can tell that
would be a rather expensive study. These are the types of
studies that preceded that randomized control trial. Does
chronic infection with chlamydia influence the risk of coronary
heart disease. They take a cohort assembled for completely
different reasons. If somebody has an existing cohort and you
are going to graft onto that study something else. The ARIC
study, Atherosclerosis Risk in Communities Study, 14000 black
and white cohort members without evidence of prevalent
cardiovascular disease are included at these various sites.
What do you need to do to answer this question within this
cohort? If this is the research question of interest and you
have this cohort what do you need to do? Whether you do a
nested case control study or case cohort study what's the first
thing you need to do?
You have to figure out what the outcome is and you need
to measure the outcome. Figure out who the cases are. Right?
Then you need to have a strategy for choosing the controls and
then you need to take their samples and test them for whatever
you think is the right thing to test them for.
Okay. So, they've got 246 individuals who have
incident coronary heart disease, a heart attack basically. As
of this particular date after median of about three years. So
they go to the lab and see who has a blood sample sitting in
the lab from when they were enrolled into the cohort. Most of
them do, but not everybody. They've got 246 people in this
cohort who have had a heart attack for whom there is a blood
sample sitting in the lab. They are going to go back, take out
those blood samples and measure antibodies to this bacterium.
Right? They could have done a nested case control study. They
choose to do a case cohort study. What that means is they took
a random sample of everybody in the cohort at baseline for whom
there was a blood sample in the laboratory.
Okay. So and they chose 550 people in the cohort at
baseline. Ten of whom actually were in the heart attack group.
Okay. So ten of the individuals chosen as controls are also
cases in this study. They are included both as cases of people
who got a heart attack but also as controls as a sample of the
baseline membership of the cohort. And now again instead of
testing 20 -- 14000 cohort members with this expensive test
they only test 246 plus 550 or about 800 people.
Right? And here are the results. This is the
distribution of antibody levels. The dark bars are the cases
the light bars are the controls. They basically didn't find
much of a difference. In one sense it's a disappointing study.
The antibody levels don't really predict who is going to have a
heart attack. Okay?
But it's an example of this idea of a case cohort study
where you choose the controls from among everybody in the
cohort at the baseline. Now they then play a lot of games to
see if they can save something out of this study. So they
adjust for a lot of demographic factors and risk factors and
after adjusting for all these various things maybe there's a
suggestion of increased risk of heart attack. Marginally
statistically significant. Maybe there's an effect of chronic
infection, but not a very powerful piece of evidence.
As I've said people have done randomized trials to
further test this hypothesis.
Okay. And then they also try to break it down by
smoking. They say that in never smokers the relative hazard is
substantial. Maybe it's a risk factor. Never nonsmokers, but
not among smokers. In an effort to salvage something out of
this study.
Significantly increased coronary heart disease hazard
associated with antibody levels greater than 1 to 64. This
says among non-smokers. Sorry, among non-smokers. Even after
adjusting for all the risk factors overall the results don't
provide strong support for this hypothesis. In this instance
they took a sample of everybody in the cohort at baseline so
it's a case cohort study.
Okay. Let's begin this. See if we can finish any
questions about this idea of nested case control studies or
case cohort studies. By definition you have to have a cohort
at your disposal. It has to be a large enough cohort to
produce a reasonable number of cases of the outcome you are
interested in. There need to be stored samples that can be
tested for whatever it is you want to test for.
Or the other circumstance when they are useful if you
didn't collect the risk factor information from everybody and
it's impractical to go back and interview all 100000 people in
the cohort it's also more practical to do a nested case control
study. As opposed to going back and interview one hundred
thousand people, go back and interview 300 people for more
detailed information about some exposure or whatever it is.
If you have all of the risk factor information on
everybody in the cohort there is no reason to do a nested case
control study or a case cohort study. You should analyze the
entire cohort. Right?
If you have all the information on everybody in the
cohort, why would you take a sample? You are just throwing
away data that you already have. There has to be some type of
additional data you need to collect in this cohort that's too
expensive to collect on everybody. That's what a nested case
control study or a case cohort study makes sense. The other
general type of study design you should be familiar with.
There are actually a number of these different types of
studies.
The most well known is called the case crossover study.
These are basically study designs in which everybody in the
study is a case. Everybody.
Okay? But we can get control time from the individuals
and we can actually ascertain relative risk or odds ratios by
simply comparing what happens in individuals with the outcome
depending on changes in their risk status. Okay. So case
crossover study is a study in which all the subjects have the
outcome of interest. And the relationship between the outcome
and the exposure is assessed by comparing the frequency of
exposure during a specified hazard interval immediately
preceding the outcome with the frequency of exposure during
some comparison time period.
So, one of the examples I'm going to show you is for,
what are the risk factors for having a heart attack?
Okay. In previous experience the cases substitutes for
a control series to estimate the person time distribution of
the exposure in the source population. And each individual
serves as his or her own control. Okay? This may seem a
little odd.
So, in one sense case crossover studies resemble both
cohort studies and case control studies. You can calculate
odds ratios or risk ratios or rate ratios but you can't
calculate risk or rates or risk or rate differences. They are
really only useful for studying the relationship between acute
outcomes and exposures that produce a transient increase in
your risk.
When the latency period is short.
So some examples of these that I will show you are does
exertion increase your risk of having a heart attack? Does
using a cell phone while driving increase your risk of having a
traffic accident? Does drinking increase your likelihood of
having unsafe sex? You can get a sense for all of these your
exposure status will be different at different times.
Sometimes you are using a cell phone and sometimes you are not.
Sometimes you are exerting yourself and sometimes you are not.
Sometimes you are drinking alcohol and other times you are not.
Your exposure status can change. Right?
And in general the time period between the exposure and
the outcome is in the order of minutes to hours. Not decades.
Right?
Okay. So some of you may be familiar with this issue.
We're busy passing lots of laws about cell phones while
driving. What's the law in California? You are all cell phone
users as I recall. You all raised your hands a few weeks back
and said you all have cell phones. What's the law about using
a cell phone while driving in California?
>>>: You can't.
>>>: Hands free device.
Art Reingold: It has to be a hands free device.
My wife got a ticket. You have to have a hands free device.
Is that a smart policy based on good science? That's not the
same in other states. In Georgia you can use either type of
cell phone, including a hand held device. It's legal. The law
varies from state to state. The question is, is there a risk
associated with using a cell phone while driving? How would
you figure that out?
>>>: Accident rates while using.
Art Reingold: Pardon?
>>>: Just looking at accident rates.
Art Reingold: How would you then figure out a
relationship with cell phones?
were using a cell phone or not while at the time of the
accident.
Art Reingold: How would you analyze that?
>>>: Compare it to people who were in the traffic
accident but were not using the cell phone.
Art Reingold: Okay. So I suppose that's one way
of doing it. The classic approach for studying this is one of
these case crossover studies. Let me show you how it works.
So, observational studies at any given time. The
figures have gone up a lot. This is nine years ago. I've done
my own survey standing before the law was passed and it was
more than one out of 20 people on the phone while driving.
Presumably it's gone down since then. Many people are on the
phone while they are driving. Experimental studies show that
phone use impairs your performance doing simulated driving
tasks in the laboratory. It's a distraction. The research
question is does mobile phone use while driving increase the
risk of being in an accident severe enough to result in
hospital attendance?
That's the outcome here. This was a case crossover
study done in Australia. So who are the participants? Drivers
over the age of 17 who were involved in a traffic accident
coming to the emergency room who own a cell phone. Okay.
Those are the people eligible. Everybody in this study has
been in a car accident. Everybody. There are no controls who
haven't had a car accident. Right?
They collected demographic data, gender, time and
description of the crash. Usual patterns of driving and mobile
phone use. Now you need to define what's the hazard interval?
During what time frame is cell phone use relevant? They chose
a ten-minute interval before the crash. Now for each person
who has been in a crash, the first question is were you on a
cell phone ten minutes before the crash? Then you need a
control interval in the same person.
So, for that they chose the same ten-minute interval in
a 24-hour day. The previous day, three days before, and seven
days before the crash. During that same ten-minute interval
were they on the phone or not?
Each person serves as his or her own control. So here
you've got the collision day. Here's the hazard interval
before the collision. Comparison day the exact same time of
day but the day before, three days before, etc.
This is taken from a study of heart disease. It's
basically the same basic issue. You have to decide what the
relevant time period is. I'll just show you, we'll come back
to this because we're almost out of time. Here are the data of
risk of injury while using a cell phone while driving. Here
are the odds ratios using various control periods. It doesn't
matter. Whatever comparison periods, one day before, two days
before, seven days before all are associated with an odds ratio
of 4 and 5 between using a cell phone and being in a crash
serious enough to go to the emergency room. We'll come back to
this issue on Monday. These are really interesting and
important studies. Just think about this for a little bit and
we'll talk about it on Monday and Friday we'll talk about meta
analysis.