HC 1: Study design and controls
- Experimental > you’re in control. Experimental study.
o If who gets drug a and who gets b is completely random (roll of the
dice, completely by chance), you talk about a randomized controlled
trial.
- Observational > Not in control. Naturally occurring groups.
o If you have a control (e.g. pregnant vs non-pregnant), you do an
analytical study.
o If there is no control (e.g. only pregnant), you do a descriptive study.
Types of designs
Newer studies say it doesn’t matter what design you do, as long as it is well-
conducted. Prefer a well done case report design over a bad cohort study for
example.
Case series
Description of a (few) patients with a certain condition. Aim is that you want
to describe the condition of interest. No exposure, only describe what you
,see. Description can lead to a hypothesis. Descriptions in MMWR describes 5
patients, but tells you the entire pathogenesis of HIV (population, symptoms,
etc). Useful, but mainly generates hypothesis.
Cross-sectional
Typical surveys. E.g. ask people their opinion, measure high blood pressure,
etc. Mainly descriptive. Measure potential outcomes of interest in certain
population. Aim is to get the prevalence of this outcome. Exposure (e.g.
buying something at bol.com) and outcome (e.g. being satisfied) are
measured at the same time (might take e.g. a year, but directly measure
outcome after exposure).
Structure: define population > go to the individual > measure the outcome.
Ecological study
Measurement on group level. Use prevalence on a group level. All about an
estimate. Looks like a cross sectional study. Can say e.g. people who live at a
farm land in general have a higher risk for … Aim: exploring association.
Example > is a person that eats more meat at a higher risk for colon cancer?
See it is correlated at the national level, but can’t say anything about
individuals. This is what we call ecological fallacy. If you measure at group
level, you can never say something about the individual level. Outcomes are
reliable, but the interpretation is only reliable on the same level as you
measure. Can then do another study design to see if it is also seen at an
individual level to draw conclusions about the individuals.
Case-control
Put people in a group based on their outcome. Have 2 groups (1 with
outcome of interest, 1 without). Look for different exposures in these 2
groups (back in time). Can do a formal statistical comparison. Can calculate if
the exposure is significantly higher in the disease group compared to healthy
group. Easy to compare the groups statistically.
Very often used, strong design. 1 problem > selecting controls. Finding
controls to compare to the condition group requires a lot of attention.
Example > selected malaria vs non-malaria. Compared factors between the
groups.
Cohort
Looks at 2 different groups. Don’t start with the outcome, but with the
exposure. Measure the outcome over time. E.g. groups like covid vs non-
covid, then send questionnaires over time. Also can do statistical analysis.
Can also use history as a control, if controls are hard to find. Here there is
some flexibility (in case-control however it is really strict).
3 main study designs (case-control, cohort, cross-sectional). How do
they related to each other?
,- Case control > everything has happened.
- Cohort > you follow the development of the outcome.
- Cross-sectional > time doesn’t matter.
Time
- Retrospective > all data collection has been done (the outcomes). Cant
control how measurements are done, what exposure is, etc.
- Prospective > can control how the measurements are done, how often,
what the exposure is, etc.
- Lower line > exposures already happened, outcomes still need to be
measured. Very common situation.
Case-control is by definition a retrospective study!
Cohort not necessarily always prospective. Can also be retrospective!
Don’t rely solely on this terminology, because it is hard to describe the hybrid
situation (lower line). Just describe what you did.
Randomized controlled trial
Fancy cohort study. Very expensive. Allocation of placebo vs drug is
completely by chance (key characteristic) to reduce bias. Exposure controlled
by the investigator. Need to do this before the study starts. High level of
evidence of the effect of an intervention. Once the analysis is done, you just
compare the means or proportions.
Randomization > can be on individual level, but also e.g. schools or
households. By randomizing groups, they have an similar underlying risk of
the outcome. If you then find differences between the groups, it is most likely
that this is dependent on the exposure.
Meta analysis
, Compilation of available evidence with a similar research question. Pooled
analysis. Not interested in single group/individual, but level of inclusion is
individual studies. Tries to make sense of all different estimates. Aim:
summarize academic literature. Awful lot of work.
Comparing study designs
Randomized controlled trials usually have limited generalizability, because
they have strict inclusion criteria (so there are less confounding factors).
Makes it harder to say something about the general population.
Why cross-sectional not for rare diseases > prevalence is really low. Would
need to have a very large group to make the data make sense (just take a
group of people and ask what they have). Not enough events in a random
group of people. Would need to wait decades to get enough people with the
outcome.
Why can case-control handle rare diseases nicely > select people based on
outcome. Can e.g. just pick every Dutch person with HIV from the registry.
Would be harder to find controls.
Cohort hard to do for multiple exposures > select on the exposure (e.g.
smoking non-smoking) and follow them over time. If you also want to look at
e.g. alcohol use, can t use the same cohort. Groups wont overlap perfectly
(are people in e.g. the smoking group that do drink and people that don’t).
This is not a problem for the case-control studies. Crucial difference between
case-control and cohort.
Lecture 01b alternative designs GEEN studie material
Lecture 01c controls WEL studie materiaal, gaan we t over hebben tijdens wg.
Controls selection
Selection participants for a case-control study is based on 3 principles.
- Study base principle (prevents selection bias) > controls should be
collected from the study base that also gave rise to the cases. Have to ask
yourself “if this control also had the outcome, would they qualify as the
case? Controls should be selected independently of the exposure (e.g. if
you study the effect of red meat, selecting controls from a vegan health
club would not make sense).
- Deconfounding principle (prevents bias due to confounding) > exposures
can be related to a lot of other factors. If we do not takes this into account,
the association between exposure and outcome in our study will be
biased. Need to minimize this. Can be done by matching (select controls
- Experimental > you’re in control. Experimental study.
o If who gets drug a and who gets b is completely random (roll of the
dice, completely by chance), you talk about a randomized controlled
trial.
- Observational > Not in control. Naturally occurring groups.
o If you have a control (e.g. pregnant vs non-pregnant), you do an
analytical study.
o If there is no control (e.g. only pregnant), you do a descriptive study.
Types of designs
Newer studies say it doesn’t matter what design you do, as long as it is well-
conducted. Prefer a well done case report design over a bad cohort study for
example.
Case series
Description of a (few) patients with a certain condition. Aim is that you want
to describe the condition of interest. No exposure, only describe what you
,see. Description can lead to a hypothesis. Descriptions in MMWR describes 5
patients, but tells you the entire pathogenesis of HIV (population, symptoms,
etc). Useful, but mainly generates hypothesis.
Cross-sectional
Typical surveys. E.g. ask people their opinion, measure high blood pressure,
etc. Mainly descriptive. Measure potential outcomes of interest in certain
population. Aim is to get the prevalence of this outcome. Exposure (e.g.
buying something at bol.com) and outcome (e.g. being satisfied) are
measured at the same time (might take e.g. a year, but directly measure
outcome after exposure).
Structure: define population > go to the individual > measure the outcome.
Ecological study
Measurement on group level. Use prevalence on a group level. All about an
estimate. Looks like a cross sectional study. Can say e.g. people who live at a
farm land in general have a higher risk for … Aim: exploring association.
Example > is a person that eats more meat at a higher risk for colon cancer?
See it is correlated at the national level, but can’t say anything about
individuals. This is what we call ecological fallacy. If you measure at group
level, you can never say something about the individual level. Outcomes are
reliable, but the interpretation is only reliable on the same level as you
measure. Can then do another study design to see if it is also seen at an
individual level to draw conclusions about the individuals.
Case-control
Put people in a group based on their outcome. Have 2 groups (1 with
outcome of interest, 1 without). Look for different exposures in these 2
groups (back in time). Can do a formal statistical comparison. Can calculate if
the exposure is significantly higher in the disease group compared to healthy
group. Easy to compare the groups statistically.
Very often used, strong design. 1 problem > selecting controls. Finding
controls to compare to the condition group requires a lot of attention.
Example > selected malaria vs non-malaria. Compared factors between the
groups.
Cohort
Looks at 2 different groups. Don’t start with the outcome, but with the
exposure. Measure the outcome over time. E.g. groups like covid vs non-
covid, then send questionnaires over time. Also can do statistical analysis.
Can also use history as a control, if controls are hard to find. Here there is
some flexibility (in case-control however it is really strict).
3 main study designs (case-control, cohort, cross-sectional). How do
they related to each other?
,- Case control > everything has happened.
- Cohort > you follow the development of the outcome.
- Cross-sectional > time doesn’t matter.
Time
- Retrospective > all data collection has been done (the outcomes). Cant
control how measurements are done, what exposure is, etc.
- Prospective > can control how the measurements are done, how often,
what the exposure is, etc.
- Lower line > exposures already happened, outcomes still need to be
measured. Very common situation.
Case-control is by definition a retrospective study!
Cohort not necessarily always prospective. Can also be retrospective!
Don’t rely solely on this terminology, because it is hard to describe the hybrid
situation (lower line). Just describe what you did.
Randomized controlled trial
Fancy cohort study. Very expensive. Allocation of placebo vs drug is
completely by chance (key characteristic) to reduce bias. Exposure controlled
by the investigator. Need to do this before the study starts. High level of
evidence of the effect of an intervention. Once the analysis is done, you just
compare the means or proportions.
Randomization > can be on individual level, but also e.g. schools or
households. By randomizing groups, they have an similar underlying risk of
the outcome. If you then find differences between the groups, it is most likely
that this is dependent on the exposure.
Meta analysis
, Compilation of available evidence with a similar research question. Pooled
analysis. Not interested in single group/individual, but level of inclusion is
individual studies. Tries to make sense of all different estimates. Aim:
summarize academic literature. Awful lot of work.
Comparing study designs
Randomized controlled trials usually have limited generalizability, because
they have strict inclusion criteria (so there are less confounding factors).
Makes it harder to say something about the general population.
Why cross-sectional not for rare diseases > prevalence is really low. Would
need to have a very large group to make the data make sense (just take a
group of people and ask what they have). Not enough events in a random
group of people. Would need to wait decades to get enough people with the
outcome.
Why can case-control handle rare diseases nicely > select people based on
outcome. Can e.g. just pick every Dutch person with HIV from the registry.
Would be harder to find controls.
Cohort hard to do for multiple exposures > select on the exposure (e.g.
smoking non-smoking) and follow them over time. If you also want to look at
e.g. alcohol use, can t use the same cohort. Groups wont overlap perfectly
(are people in e.g. the smoking group that do drink and people that don’t).
This is not a problem for the case-control studies. Crucial difference between
case-control and cohort.
Lecture 01b alternative designs GEEN studie material
Lecture 01c controls WEL studie materiaal, gaan we t over hebben tijdens wg.
Controls selection
Selection participants for a case-control study is based on 3 principles.
- Study base principle (prevents selection bias) > controls should be
collected from the study base that also gave rise to the cases. Have to ask
yourself “if this control also had the outcome, would they qualify as the
case? Controls should be selected independently of the exposure (e.g. if
you study the effect of red meat, selecting controls from a vegan health
club would not make sense).
- Deconfounding principle (prevents bias due to confounding) > exposures
can be related to a lot of other factors. If we do not takes this into account,
the association between exposure and outcome in our study will be
biased. Need to minimize this. Can be done by matching (select controls
