Study design Sample selection

1
Study design / Sample selection

• Summer Course
• Brian Healy

2
What have learned so far

• Types of data
• Descriptive statistics
• Tables and graphs
• Basics of R

3
What are we doing today?

• Association vs. causation
• Types of studies
• How do we choose a sample?
• Generalizability
• Sampling variability

4
Big picture

• The type of study and the sample selection
  determine the type of conclusions one can make
  from the study
• Sometimes the study design can be chosen by the
  experimenter, but often the study design is
  dictated by the question of interest. For
  example, if we are interested in determining the
  difference between men and women we cannot
  randomize people to gender

5
Association vs Causation

• Association
• Definition two factors are related, ex. height
  and weight
• There is no implied cause and effect relationship
• Causation
• Definition one factor causes a second factor,
  either directly or indirectly, ex. laying in the
  sun causes sun burn
• The causal factor is required for the outcome to
  occur

6
Types of studies

• Intervention
• Randomized trial
• Observational
• Cohort study
• Case-control study

7
Randomized clinical trial

• Definition study in which patients are randomly
  assigned their exposure
• One of the most common study types, especially
  for determining the effect of a treatment on an
  outcome because patients can be randomly assigned
  treatment
• Advantages
• Can draw causal conclusions
• Controls for known and unknown risk factors
  related to the outcome and exposure
• Disadvantages
• More difficult to design
• Only applies to certain types of exposures
• Adherence

8
Why randomize?

• We randomize subjects to treatment to ensure that
• Patients in each group are exchangeable, meaning
  that the two groups are exactly the same other
  than the treatment assignment
• This allows the investigator to say that the
  difference in the outcome over the two groups is
  caused by the treatment
• Therefore, we can say the CE Y(treatment group)
  Y(No treatment group)

9
How do we randomize?

• The choice of randomization technique is often
  based on the type of intervention
• For a drug treatment, we can randomize
  individuals
• For a group treatment, we may need to randomize
  at the hospital or school level
• The key to any randomization is to ensure that
  the groups are the same other than the exposure
  of interest
• Simplest technique is to use a random number
  generator and assign people based on this number

10
When can we not randomize?

• Exposures that cannot be changed
• Sex
• Age
• Exposures that are not ethical to assign
• Smoking
• Obesity
• Rare diseases
• If we randomize to exposure, we may never observe
  the disease

11
Observational studies

• Def Studies in which the group membership is
  simply observed, rather than assigned by the
  investigator
• Advantages
• Do not need to intervene on subjects
• Can complete studies after disease has already
  occurred
• Disadvantages
• More difficult to make causal conclusions
• Must collect many other factors to ensure you
  have controlled for confounders

12
Observational studies

• Case-control study Patients are broken into
  groups based on disease status and exposure
  status is compared
• Cohort study Patients are broken into groups
  based on exposure status and disease status is
  compared
• Prospective cohort study Patients are observed
  with a specific exposure and followed over time
  to determine if they develop the disease
• Retrospective cohort study Patients are observed
  after the disease may or may not have developed,
  but groups are still defined by exposure status

13
Case-control study

• Rare disease
• With a rare disease, you cannot match based on
  the exposure because you are rarely going to have
  anyone develop the disease
• This design allows the study participants to be
  found using all available cases and the
  appropriate controls. Finding appropriate
  controls is often the most important design
  aspect of the study.
• There is a link between the odds ratio of
  exposure given disease and the odds ratio of
  disease given exposure, which allows conclusions
  about the exposure

14
Choosing controls

• In case-control study, it is often difficult to
  determine what is the best control group
• Depending on exposure of interest can use
• Uninfected family members (Likely to participate)
• Other patients in the hospital (Convenient
  sample)
• Neighbors (Environmental factors)
• Random people
• The key to any set of controls is to ensure that
  selection bias is limited.
• Must also consider recall bias

15
Prospective cohort study

• Advantages
• Ensures a temporal relationship between the
  exposure and the outcome
• Easier to find controls
• Limited recall bias
• Disadvantages
• Can take significant time
• Require many factors to be collected
• Often many are lost to follow-up
• Only non-rare disease

16
Retrospective cohort study

• Advantages
• Can collect all of the data at one time, no
  problem with loss to follow-up
• Can use relative risk and risk difference, rather
  than odds ratio (case-control), which are more
  interesting
• Disadvantages
• Can have recall bias
• Temporal relationship may not be clear

17
How do we pick our sample?

• Choosing exposure group
• Choosing control group
• Sample size

18
Reasons for differences between groups

• Actual effect-when there is a difference between
  the two groups (ex. the treatment has an effect)
• Chance
• Bias
• Confounding

19
Chance

• When we run a study, we can only take a sample of
  the population. Our conclusions are based on the
  sample we have drawn. Just by chance, sometimes
  we can draw an extreme sample from the
  population. If we had taken a different sample,
  we may have drawn different conclusions. We call
  this sampling variability.
• The key for this is that the sample must be
  different purely by chance, not because of a flaw
  in the study design, which is a type of bias.

20
Example

• Lets say we were interested in determining if
  there is a difference in the average height in
  Boston and New York.
• Assume that the distributions are the same and we
  take two random samples from the population (B
  and N).
• Just by chance, sometimes these will be very
  different

B
N
21
Chance

• The most common application of biostatistics
  determines the probability of observing a certain
  difference. If the probability of the event is
  sufficiently small, we say that the difference is
  likely not due simple to chance and we have an
  actual effect. We call the probability, the
  p-value, and the cut-off, the alpha level.

22
Bias

• Selection bias Patients in one group are more
  likely to have the disease
• Ex. Patients in exposure group are systematically
  more ill than patients in control group
• Recall bias Patients in one group have better or
  worse memory of the exposure
• Ex. Studies comparing the diet / exercise of
  mothers of babies with and without birth defects
• Observation bias Patients or doctors in one
  group are more likely to detect disease
• Ex. Doctors of patients with a disease
  susceptibility may be more likely to spot the
  disease because they are looking for it
• Interviewer bias People conducting interviews
  gather information differently in two groups

23

• You can usually not correct for bias after the
  study has been conducted
• There is some research into using ideas of DAGs
  and causal inference to control for selection
  bias (EPI 289), but you need to collect data on
  the reason for the bias
• You must anticipate the possible types of bias in
  your study design

24
Confounding

• Def when a factor is related to the exposure and
  the outcome and therefore can mask the true
  effect
• To control for confounding, you must collect
  information about all of the possible confounders.

25

• As an example, think about alcohol consumption
  and lung cancer. It can be shown that there is an
  association between alcohol consumption and lung
  cancer, but this effect may be caused by the fact
  that smoking causes lung caner and smoking is
  associated with alcohol use

Cigarette smoking
Alcohol use
Lung cancer
26
Controlling for confounding

• In a randomized trial, there should be no
  confounding because no factor is associated with
  the exposure. Remember the randomized trial is
  the best form of study because we can draw causal
  conclusions.
• There is significant research (much here at HSPH)
  into ways to control for confounding, which is
  trying to make an observational study like a
  randomized trial to allow causal conclusions.
• The objective is to make the exposure group and
  the control group similar so that the observed
  differences are caused by the exposure
• Much more on this in EPI 289 and EPI 207

27
Conclusions

• When we observe a difference between two groups
  we want to determine if this difference is
  important or due to chance, bias or confounding
• The main job of the statistician after the study
  is to determine if the difference is simply due
  to chance, which will be the focus of most of the
  remainder of the summer
• Control of bias and confounding requires the
  statistician and the subject area expert to
  ensure that the biases are limited and
  confounding factors are collected

28
Generalizability

• Assume that we have found a difference between
  our exposure and control group and we have shown
  that this result is not likely due to chance,
  bias or confounding.
• What does this mean for the general population?
  Specifically, to which group of people can we
  apply our results?
• This is often based on how the sample was
  originally collected.
• Ex. If exposed group were school children living
  near power lines and the control group were
  school children living elsewhere, can we
  generalize the findings to all children? adults?

29
Practice

• I have given out a couple of papers. We are going
  to go through
• question of interest
• study design
• study population
• potential biases
• confounders

30

• Now, try to write functions to do the following
  things.
• Take a vector input and find the mean of all of
  the values except the minimum and maximum
• Take a vector input and output a graph with a
  histogram and boxplot
• Take a matrix input. Find the mean and median of
  each column. Output the mean, median and column
  number as a list for the column with the highest
  median