Module 7

Course Notes

Prognosis and Harms

(Observational Studies)

CEPI5100: Module 7 – Prognosis and Harms Page 1

 COMMONWEALTH OF AUSTRALIA

Copyright Regulations 1969

WARNING

This material has been reproduced and communicated to

you by or on behalf of the University of Sydney

pursuant to Part VB of the Copyright Act 1968 (the Act).

The material in this communication may be subject

to copyright under the Act. Any further reproduction or

communication of this material by you may be the

subject of copyright protection under the Act.

Do not remove this notice.

© Sydney School of Public Health, University of Sydney, 2017

CEPI5100: Module 7 – Prognosis and Harms Page 2

Learning outcomes
By the end of this module you will be able to:

 describe why we need observational studies

 locate the STROBE statement and explain why it is important
 differentiate between observational studies used to answer the different clinical
question types of prognosis, harm and aetiology
 critically appraise a cohort study used to answer prognosis, harm or aetiology
questions for risk of bias
 critically appraise a case control study used to answer harm or aetiology
questions for risk of bias

Learning activities
1. View video 1 Why do we need observational studies?
2. View video 2 Cohort studies – introduction
3. Complete activity 1 Cohort studies
4. View video 3 Cohort studies – quality 1
5. View video 4 Cohort studies – quality 2
6. View video 5 Case-control studies
7. Complete activity 2 Case-control studies
8. View video 6 Survival curves
9. Review the module 7 course notes.
10. Complete the module 7 quiz by the due date in the timetable
11. Complete the module 7 tutorial

Additional resources

Straus SE, Glasziou P, Richardson, WS, Haynes RB. Evidence-based Medicine: How
to practice and teach it. 4th ed. Sydney: Churchill Livingstone, 2011. Chapters 6 and 7.

CEPI5100: Module 7 – Prognosis and Harms Page 3

In this module you will learn about how to assess the risk of bias in studies used to
answer questions of prognosis, harm or aetiology/causality. These types of clinical
research questions are usually answered using observational studies, particularly
cohort and case-control studies, so the focus of this module will be how to assess the
risk of bias in cohort and case-control studies.

Cohort Studies – Introduction

In this first section, you will learn about the different types of clinical questions that can
be answered by cohort studies. The type of clinical question we are asking can have
an impact on the design of the cohort study and also to some extent how we assess
the risk of bias so it’s important to keep the question type in mind.

Prognosis refers to the course of a disease after its onset. The key feature that is
important to remember about a prognosis study is that because we are interested in
the course of a disease, the cohort or group of people in the study are a group of
people who all have that particular disease. We then follow this cohort of people with
disease over time to see who develops outcomes of interest. In a prognosis study, we
might be interested in the natural history of a disease (which is what happens without
medical intervention) or the clinical course of a disease (which is how the disease
progresses over time with medical care).

The study may be purely descriptive and produce results such as the case-fatality rate
– a measure of the rate of death in the sample; or the % survival at a specified time-
point like 1 or 5 years, or the median survival which is the time by which 50% of
patients have died.

In other prognosis studies we might be interested in the ability of particular prognostic

markers to predict survival or another important outcome, such as comparing survival
between younger and older patients. This would be an analytic study as we would be
comparing survival between groups of patients with and without a particular prognostic
factor.

In studies about harm or aetiology, in contrast to prognosis studies, we are interested

in the development of disease or health outcomes. So in these studies we start off with
a group of people without a particular disease and then follow them over time to see
who develops it. In a study about harm we are interested in whether exposure to a
potentially harmful agent increases the risk of developing a particular disease or
negative health outcome, for example whether exposure to smoking increases the risk
of lung cancer.

This kind of question can also be considered as an aetiology question as we are

looking at the risk factors or causes of particular diseases or health outcomes. So
harm and aetiology questions can sometimes overlap. However, harm questions may
also refer to the harmful effects of something that we might we using such as an
intervention or diagnostic test. For example, a study looking at the increased incidence
of bleeding in those taking warfarin compared to those not taking warfarin is also a
study about harms.

CEPI5100: Module 7 – Prognosis and Harms Page 4

Sometimes information about the harms of a treatment might come from an RCT or a
meta-analysis of RCTs. However, RCTs are usually not designed to focus on harms
and are too small or have periods of follow up that are too short to detect harms that
are rare or that take a long time to develop. In these instances, observational studies
are useful for these kinds of questions. For example, retrospective cohort studies that
use large national datasets of adverse drug reactions might be used to answer harm
questions.

Just like cohort studies used for prognosis questions, cohort studies used for harm or
aetiology questions can be descriptive such as those that report the incidence of a
particular disease. They can also be analytic where we are interested in the
relationship between particular risk factors and the development of disease, such as if
we were comparing the incidence of lung cancer between smokers and non-smokers.

You may remember in previous modules the importance of adequate reporting of trials
in publications so that it’s possible for the risk of bias in studies to be adequately
assessed. This appropriate reporting is supported by the Consort statement for RCTs
and Prisma for systematic reviews. For observational studies, the STROBE statement
stands for strengthening the reporting of observational studies in epidemiology
[Link] STROBE provides different check-lists of items to
report in publications of cohort studies, case-control studies and cross-sectional
studies, as each of these study types have different weaknesses in terms of risk of
bias.

Quality – Defining the study sample

In this section you will learn how to assess the risk of bias in a cohort study in terms of
how well the study sample is defined.

The first question to ask when assessing the risk of bias in a cohort study about
prognosis is whether there was a defined, representative sample of patients recruited
at a common point in the course of their disease. There are really 3 important points
covered by this question: defined, representative, and common time point; and so we’ll
go through each of these points separately.

A defined group of patients is important not just for the internal validity or risk of bias
in a cohort study but also for determining how generalisable the results of the study
will be to our own patients. Having a well-defined patient group also helps ensure that
the patients in the study are a fairly homogeneous group, i.e. that patients are not
vastly different from one another. Part of this comes from providing clear criteria for
what constitutes a disease case as well as the particular inclusion and exclusion
criteria used.

The study sample should be defined in terms of patient characteristics, the setting in
which patients are recruited from and the time period of recruitment (or in other words
the person, place and time of the study sample).

CEPI5100: Module 7 – Prognosis and Harms Page 5

The next point to consider is the representativeness of the patients with disease
included in the study. In a cohort study looking at prognosis, as we are dealing with
people that have a particular disease, ideally we would want this group of patients to
be representative of all those with a particular disease (e.g. women with breast
cancer) or perhaps a defined subset of those with a particular disease (e.g. women
with oestrogen receptor positive breast cancer or even women with metastatic breast
cancer).

Unfortunately, getting a representative sample of those with a particular disease can

sometimes be difficult. If we start off with a hypothetical sample of all new cases of a
disease in the study population, there may be those who never seek medical attention
and hence are never diagnosed, or maybe they seek medical attention but are
misdiagnosed, so they won’t be diagnosed as a case of the disease and therefore
won’t be recruited to enter the study. There might also be patients that are so unwell
that they die before being diagnosed or being recruited into the study or there might be
patients who have such mild disease that they get better before being diagnosed or
having a chance to enter the study. So depending on the disease in question, the
sample of new cases of the disease that we identify for the study could be different
from the total sample of new cases of disease in the study population.

An important consideration when judging the representativeness of the sample of

cases is to consider the setting where patients were recruited from and the referral
process involved. For example hospital patients in a tertiary referral centre are not
likely to be representative of all people with a particular disease. They are likely to be
patients that are not responding well to treatment or that have multiple comorbidities
that make their condition more difficult to manage.

As well as where participants are recruited from, we would also want to know how they
were recruited, such as did they take all patients in the clinic or hospital, a random
sample, consecutive patients starting from a particular date or a convenience sample.
The last approach, a convenience sample, which refers to recruiting those patients
that are most easily accessible, is the approach that is less likely to result in a
representative sample. We would also want to know how many eligible patients
refused to participate as this can also impact how representative the sample of
patients is.

Next we look at if patients included in the study were recruited at a common time
point in the course of the disease. Ideally we would want to start with a group of
patients that are identified early enough in the course of the disease so, that as we
follow them, we can observe all of the important events. If we recruited people that
have had the disease for some time already, it might be that many of the events that
we are interested in have happened already. Of even greater importance is that we
are starting the observations of the different patients in the study at a uniform point in
the course of the disease.

You could imagine if we recruited some patients at the beginning of the disease
process and others at the end, then it would be difficult to interpret differences in the
timing of particular events. We use the term inception cohort to describe a cohort of

CEPI5100: Module 7 – Prognosis and Harms Page 6

patients that are recruited at a uniform point near the onset or inception of the disease.
In real life it is often not possible to determine the exact point when a disease process
starts but this needs to be uniform across the patients in the study and may include
options such as the onset of symptoms, time of diagnosis, or beginning of treatment.

For cohort studies of harm or aetiology questions, the considerations are similar in that
we still want the sample to be well defined in terms of person, place and time to help
us judge the generalisability of the findings and to help ensure that the sample is
reasonably homogeneous. However, this time we are starting off with a group of
people without the disease, so the considerations about the representativeness of
cases and starting with an inception cohort do not apply. So instead we would simply
ask the question “Was a defined sample of people without the disease assembled?”

Quality – Confounders, follow up and measurement.

In this section you will learn more about how to assess the risk of bias in a cohort
study in terms of adjustment for confounders, follow up and measurement of
outcomes.

Another important question we ask when assessing the risk of bias in a cohort study is
if there has been any adjustment for confounders when comparing different
subgroups in the study. In a prognosis study, we might be interested in comparing
people with different potential prognostic factors such as tumour stage or presence or
absence of particular comorbidities.

In a harm or aetiology study, we might be interested in comparing people who have

and haven’t been exposed to a harmful agent or who do or don’t have particular risk
factors such as obesity or hypertension. When comparing groups of people, ideally we
would want the groups to be exactly the same as each other in all aspects apart from
the prognostic factor or risk factor that we are interested in. The only way to really
achieve this is through randomisation and this is why a randomised controlled trial
provides a higher level of evidence than observational studies. However, a cohort
study is the primary study that provides the highest level of evidence for harm or
aetiology questions when randomisation is not possible.

In a cohort study, because we can’t randomise people to a harmful exposure or to

particular sociodemographic characteristics like age or to particular clinical features of
disease such as tumour size, it is quite possible that the comparison groups differ in
terms of confounders and so we need to use specific techniques to try to minimise or
adjust for potential bias from confounding. The most common way of dealing with
confounding is to adjust for confounding in the analysis stage using multivariate
analysis.

There will usually be a discussion in the methods section of a paper as to what other
risk factors or prognostic factors were measured and then a discussion of adjusted
analyses in the results section. Assessment of appropriate adjustment of confounding
requires some content knowledge about the likely risk factors or prognostic factors that
are important in the disease so that you can assess whether an important confounder

CEPI5100: Module 7 – Prognosis and Harms Page 7

has been left out. You should also think about the chances of any measurement error
in confounders that will lead to under adjustment. Lack of adjustment for confounders
is a common weakness of retrospective cohort studies as this type of cohort study
uses data that has already been collected and hence there may be no data on
confounders or only data of a small number of confounders.

Follow up.
The next point to assess in a cohort study is follow up where we are interested in two
things. These first is the length of follow up and the 2nd is the completeness of follow
up.

Follow up needs to be of sufficient length so that the outcomes that we are interested
in have time to occur. This is particularly important in prognosis studies as we are
often looking at the course of a chronic disease that takes a long time from the onset
of disease to eventual death, in aetiology studies as we are looking at a disease that
may take quite a long time to develop following a particular exposure, and in harm
studies where we are looking at particular side effects or adverse events that may also
take a long time to develop.

You have learnt in the previous module on RCTs about the importance of loss to follow
up and how loss to follow up can result in selection bias if it results in the comparison
groups becoming different from each other in terms of their risk of the outcome. Loss
to follow up can also bias the results of descriptive questions such as the incidence of
disease or clinical course of disease if the losses are so great that they result in the
sample no longer being representative. Remember that if the outcome you are
interested in is very uncommon, even quite small losses to follow up have the potential
to bias the results.

Similar to how we assess loss to follow up in an RCT, we are interested in the same
questions when assessing loss to follow up in a cohort study, which are:

- How many were lost?

- Were the numbers lost different or the same between the two groups (if
different patient sub-groups are being compared)

- Were those lost different from those who remained in the study? Particularly is
being lost to follow up related to the outcome of interest in the study?

- What were the reasons for loss to follow up?

CEPI5100: Module 7 – Prognosis and Harms Page 8

Measurement of outcomes.
The next point that we want to consider in terms of potential bias in a cohort study is
how the outcomes are measured. Just like in an RCT we want outcomes to be
measured the same way in different groups that we are comparing in the study. The
best way of ensuring this is to make sure that outcomes are measured blind to the risk
or prognostic factors that we are interested in as potential predictors of this outcome.

This is particularly important for outcomes that are self-reported by patients or that
require some degree of clinical judgement. Blinding to risk or prognostic factors is also
important to ensure that extra tests or extra follow up is not given to those with
particular risk or prognostic factors, as if this is done they may be a greater chance of
diagnosing the outcome in this group. If blinding is not possible, having objective
outcomes such as death or having clear criteria for how the outcome is defined can
help reduce bias.

It is important to remember that lack of blinding can result in a different approach to

measurement in each of the groups being compared in the study. For example in a
cohort study of women exposed or not exposed to HRT, if the investigators are
unblinded as to who is taking HRT, they might put more effort into identifying cases of
breast cancers in those exposed to HRT. This difference in the way the outcome is
measured across groups, leads to differential misclassification, that is, the
measurement error is different between the two groups. This can bias the estimate of
effect in any direction so can lead to either an overestimation or underestimation of the
true effect.

There also might be some non-differential measurement error that occurs equally
between study groups when the same error prone approach to measurement is used
in each study group. For example, getting participants to self-report their physical
performance rather than measuring this objectively. Retrospective cohort studies or
even prospective studies that use pre-collected forms of data such as medical records
may have more measurement error. This is because these pre-collected forms of data
are not as rigorous as the data collection that takes place as part of an established
study so they are more likely to contain errors, missing information or there may be
different pre-determined approaches to measurement in groups being compared. Non-
differential misclassification biases the result towards no effect or underestimates the
true effect.

CEPI5100: Module 7 – Prognosis and Harms Page 9

Case-Control Studies
In this section you will learn how to assess the methodological quality or risk of bias of
another type of observational study – the case-control study.

Case-control studies are most commonly used to answer harm or aetiology questions
for diseases that are uncommon and where it is not ethical or feasible to conduct
higher quality studies such as RCTs or cohort studies.

Case-control studies differ from cohort studies in that they start off with a group of
people with the disease (the cases) and a group of people without the disease (the
controls). They then look back in time to determine past exposure to risk factors and
compare this exposure between the cases and controls. Because they are starting off
with the outcome, case-control studies are more prone to bias and the reasons for this
will become clearer as you progress through this module.

The first consideration when appraising the risk of bias in a case-control study is to
think about whether or not the cases and controls are from the same defined
population. This is important in terms of selection bias, as if cases and controls don’t
come from the same population, they may have a different chance of being exposed to
the risk factor of interest and so the comparison between them will be biased.

As an example to help clarify this, imagine that you wanted to do a case-control study
investigating the relationship between taking benzodiazepines (the exposure) and
urinary incontinence (the outcome) in women. For the cases (those with urinary
incontinence) women with urinary incontinence were selected from GP practices within
a defined geographical area. For the controls (those without urinary incontinence) a
random sample of the female population was taken in the same defined geographical
area but excluded those with a history of falls or anxiety. Are these 2 groups, the
cases and controls, from the same population?

No they are not, and this is important because they do not have the same risk of
exposure to benzodiazepines. Benzodiazepines are a risk factor for falls and may be
taken by those with anxiety, so by excluding women with a history of these 2
conditions from the control group, we have reduced the chances of exposure to
benzodiazepines in the control group. Therefore, any comparison between the cases
and controls in terms of benzodiazepine exposure is now going to be biased. Even if
there is no link between benzodiazepine use and incontinence, it might look like there
is one because the risk of exposure to this drug in the cases is higher than that in the
controls because of the way we have selected participants into these two groups.

The best way to avoid selection bias and to ensure the groups are comparable to each
other is to clearly define the population in terms of person, place and time. For
example, when doing a study on smoking and lung cancer we might define the
population as those aged 40 years and over, living in New South Wales between the
years 2000 to 2010.

CEPI5100: Module 7 – Prognosis and Harms Page 10

Once the population has been defined, the next thing to consider is if the cases and
controls were selected from this population in such a way as to be representative of
cases and controls in this population.

Selection of Cases

There needs to be a clear definition of what a case is and this is usually provided in
the methods section of a research publication. Ideally we want the cases to be either
all those in the defined study area or a representative sample of all cases. However,
as discussed earlier in this module for prognosis studies, getting all cases or a
representative sample of cases of a disease can be difficult as there will be those who
don’t seek medical care, those who are misdiagnosed, and those who die or get better
before being asked to enter the study.

The setting or approach used to recruit cases can depend on the disease. For
example, for some diseases such as cancer there may be registries available that can
be used to recruit cases, for other diseases primary care settings or hospitals may be
used. The appropriateness of these settings needs to take into account the likely
referral patterns for the particular disease. For example, for some diseases taking only
cases that go to hospital might lead to a sample of only those with severe disease but
for other diseases where it is the norm to go to hospital, recruiting cases from this
setting would be appropriate. The other important issue when selecting cases is that it
is better to select incident cases rather than prevalent cases in order to avoid a special
type of bias known as incidence-prevalence bias.

Back in module 3 the relationship between incidence and prevalence was

demonstrated using a bathtub analogy. The prevalence of a disease is related to the
incidence of disease but also to the survival rate or cure rate. This means that
prevalent cases (those that stay in the bathtub) are less likely to include those that
have severe disease that progresses rapidly to death. So if we use prevalent cases in
the study, we might be finding risk factors for slowly progressive disease rather than
risk factors for developing the disease in general.

Lastly, we want to look at the participation rates of cases. For example, even if we
managed to identify all new cases of the disease in the defined population in a
specified time period, if only 50% of these cases agreed to participate in the study, the
sample of cases is not going to be representative of all the cases in the population.

Selection of Controls

The best way of getting a representative sample of controls from the defined
population is to take a random sample. In Australia, as voting is compulsory, it may be
possible to take a random sample of the population using the electoral roll. Sometimes
a case control study may select cases from a hospital and then select controls as
those patients without the disease of interest in the same hospital. This is not ideal as
hospital patients are not representative of the general population. They are in hospital
because they are unwell and are likely to have a different chance of exposure to the
risk factor of interest than those who are healthy and living in the community, and this
would make the comparison between cases and controls biased.

CEPI5100: Module 7 – Prognosis and Harms Page 11

We also want to examine the participation rates of controls that are invited (just as was
done for cases). It might be that there is a high participation rate in cases because
they have the disease and are interested in participating in research about it, but there
is a low participation rate in healthy controls. Different participation rates between
cases and controls might mean that the comparison between these groups is biased
and would be a weakness of the study.

Adjustment for risk factors and confounders

The next thing we would need to consider in our assessment of the quality of a case-
control study is whether or not adjustments were made for any differences in risk
factors or confounders between the cases and controls. Because there was no
randomisation of people to be a case or a control, there are likely to be many
differences between those who develop a particular disease and those who don’t,
apart from the particular risk factor of interest.

Sometimes a case-control study may use methods such as matching in the study
design stage to avoid confounding. However, this technique can only control for
confounding by one or two important characteristics and therefore there is still likely to
be some confounding remaining in the study. An example of matching in a case-
control study would be each time a case of a particular age and sex is recruited, a
control of the same age and sex is selected. Note that the cases and controls would
still need to be coming from the same clearly defined population to avoid selection
bias, but within this population we would match our cases and controls according to a
couple of important potential confounders during the selection process.

The other approach to managing confounding is to measure confounders as part of

the study and then adjust for them in the analysis stage, usually by multivariate
analysis. When assessing how likely it is that confounding remains when using this
approach, you should consider whether all important confounders were measured and
adjusted for, and also the likelihood of having any measurement error in the
confounders.

Measurement bias

The next important cause of bias in a case-control study is measurement bias. In a

cohort study, because we started off with the exposure groups and followed over time
to measure an outcome, we were most interested in how the outcome was measured.
In a case-control study, since we are starting off with the outcome already and looking
back at past exposure, in terms of potential measurement bias we are interested in
how the exposure is measured. We want to make sure that the exposure has been
measured the same way in cases and controls and the best way of ensuring this is to
measure the exposure blind to outcome status.

Unfortunately this is often difficult and this is an important reason why case-control
studies are more prone to bias. Patients will naturally be aware whether or not they are
a case or control and this knowledge may influence how they remember their past
exposure. This gives rise to a special type of bias known as recall bias.

CEPI5100: Module 7 – Prognosis and Harms Page 12

Recall bias will particularly be an issue for any exposure measurements that are self-
reported by patients such as diet or smoking history. It can be avoided by using
information that has been pre-recorded such as that available in medical records or
results of previous laboratory tests. Sometimes even though it will not be possible to
blind study participants to the outcome that they have, it may be possible to blind them
to the study hypothesis. This might be done by asking them about several different
exposures, not just the one of interest so that they remain unaware of the hypothetical
link between a particular exposure and the disease they have developed.

Because recall bias may lead to differences in how the exposure is measured between
the cases and the controls, it can result in differential misclassification which can bias
the result in any direction.

The last point to consider in terms of measurement bias is whether exposure

information is available on all participants and if this differs between the cases and
controls.

Survival Curves
In module 3 some of the common measures of effect that are used in clinical
epidemiological studies have been covered. In some cohort studies, particularly those
interested in comparing the time to particular events such as death, a special type of
analysis known as survival analysis is used. You don’t need to understand the details
of this analysis method as part of this unit, but it is important to know what the results
can look like so that you know how to interpret them when you come across them.

A common way to present results in prognosis studies are what are known as survival
curves. On the following page is an example of a survival curve that is comparing
survival in heart failure patients between different weight category groups.

CEPI5100: Module 7 – Prognosis and Harms Page 13

From: Rev Esp Cardiol 2010; 63(10): 1209-16

On the x axis is the follow up time and on the y axis the percent of the study sample
surviving. At the top left corner where the lines start, 100% (in this graph shown as
one) of the study sample is alive. Over time, as people in the study die you can see
that the % alive in the study decreases. If we followed the people in the study until
everyone was dead then the lines would all end up touching the x axis as the % alive
reached zero.

In this particular example you can see that survival is different in heart failure patients
depending on their weight with those that are underweight dying much faster than
those who are in the other weight categories. And that being overweight appears to be
protective against mortality in this patient group.

There is also lots of other information we can get from this curve. For example if we go
down to 0.5 or 50% on the y axis then draw a line across till we reach a survival curve,
and then draw the line down to the x axis, this would tell us the median survival time
which is the time at which only 50% of the sample are still alive, which for the
underweight group appears to be just over 2 years.

We can also go across to 5 years on the x axis and draw a line up to a survival curve
and then across to the y axis to tell us what percentage of patients are alive at 5 years,
which for the underweight group is just under 40%. These kinds of survival curves are
frequently found in prognosis studies and are a really useful way of understanding
visually the different survival patterns in different subgroups of patients.

CEPI5100: Module 7 – Prognosis and Harms Page 14

Another result that you may see in a prognosis study is something called a hazard
ratio (HR) that is produced by a type of analysis known as cox regression. This
analysis enables the comparison of time to death or other important events between
different prognosis subgroups with multivariate adjustment for differences in
confounders between these groups. Even though it’s estimated in a different way we
normally interpret a hazard ratio in the same way as a relative risk.

For example, the same study as that shown in the survival curves, found a hazard
ratio for death of 0.81 in obese compared to normal weight heart failure patients. This
shows that the risk of death was lower in obese patients (as the HR is less than 1). We
could say that obese patients had about 0.8 times the risk of death over time as
normal weight heart failure patients or that the risk of death was 20% lower in the
obese compared to the normal weight patient group.

CEPI5100: Module 7 – Prognosis and Harms Page 15

 FAQs
Observational studies use relative risks (RR), odds ratios (OR), and hazard
ratios (HR). What is the difference between these measures?
These are all measures of effect expressed in relative terms, or ratios between
exposed and unexposed groups, as opposed to measures of absolute effect such as
the absolute risk difference (ARD/RD) and number needed to treat (NNT).

RR is a ratio of the cumulative incidence during a certain period of time. RR can be

produced in cohort studies where measuring incidence is possible.

OR is a ratio of odds of exposure between those with and those without the disease.
ORs are estimated in case-control studies where it is not possible to estimate
incidence and therefore not possible to estimate a RR. Cohort studies may report an
OR because this is what is produced by many commonly used statistical packages
that perform multivariate analysis to adjust for confounders. This is okay if the disease
is rare, as in this circumstance the OR will approximate the RR. But if the disease or
outcome is common (usually >10%) then the OR will overestimate the RR.

HR is a ratio of hazards, and this can be calculated only when time to outcome is
known. A hazard is the probability of having an outcome at a specific time point, given
the outcome has not occurred up to then.

CEPI5100: Module 7 – Prognosis and Harms Page 16