Biostatistics & Research Method

Reno Rudiman

Concepts to take home

Recognize types of studies Strengths and weaknesses of study designs Risk calculations/confidence intervals P values and Power

Types of studies
Observational studies
Case report Case series Case control Cohort

Experimental studies

Case reports and series

Case report: describes an observation in a single patient.
I had a patient with a cold who drank lots of orange juice and got better. Therefore, orange juice may cure colds.

Case series: same thing as a case report, only with more people in it.
I had 10 patients with a cold who drank orange juice.

Value of case reports/series

May generate a hypothesis: maybe orange juice cures colds. Weakness: cannot test the hypothesis
no control group often too few people to make generalizations

Inferential studies
Most research is done not to generate hypothesis but to test them or make an inference from the results When setting up a question, usually start with NULL HYPOTHESIS Null hypothesis is that the treatment has NO effect (remember that null = no) If the study shows an effect then you REJECT the null hypothesis

Case control studies

Attempt to make inference from existing observations (retrospective) Compares patients with outcome/disease with those without and attempts to identify factors that influenced that outcome (or caused that disease) Important concept: start with the result (disease) and work backwards for the cause

Example of case-control study

In 1987, several children were diagnosed with a new neurological disease called Reyes Syndrome that appeared to be related to chicken pox. Hurwitz et al took 27 children who had gotten Reyes Syndrome and compared them to 140 who had not. He looked at what they had used to treat their fever.

Results
Cases (Reyes) # who use ASA 26 Controls 53

# who didnt
TOTAL

1
27

87
140

Strengths of case-control design

Best study when have rare disease or outcome Relatively quick and inexpensive

Weaknesses (potential biases)

Selection (confounding) bias: controls must be as similar to cases as possible Representativeness bias: cases should be typical Recall bias: cases may be able to remember events better because of its significance or may be prompted to remember by investigators Survival bias: dead people dont make it into many case-cohort studies; and if they do, they dont remember things very well

Ways to combat weakness

Matching: for each case, find a control that looks just like him/her in all other possible ways except for the disease (same age, race, economic class, etc.) Blinding: individual assessing exposures should be blinded to whether the person is a case or control

Cohort studies
Studies whether exposure to a risk factor is associated with a subsequent outcome Select two populations who seem the same except for the hypothesized risk factor Follow them ahead in time and see how many have the outcome or disease Important concept: Start with the risk, then look for the outcome

Example
In late 1840s/early 1850s London had several outbreaks of cholera Snow hypothesized that the source of water from a polluted area of the Thames was the cause of the cholera He identified three water companies that supplied the same neighborhoods in London: 2 got their water from one site on the Thames, the other from a different area

Results
Southwark & Vauxhall Cos
Lambeth Co # houses 40,046 # deaths Rate 1,263 3.15%

26,107

98

0.37%

Strengths of cohort study

Not only can you look at risk, you can calculate how many people actually get the disease (incidence rates) Since you enroll subjects before the outcome, you can measure multiple exposures without recall bias

Potential biases in cohort studies

Selection (confounding) bias: have to match similar groups Detection bias: measurement of outcomes needs to be objective and similar for both groups Length-time bias: study has to be long enough for outcome to happen Excursion bias: subjects may disappear or dropout (lost to follow-up)

How to deal with lost subjects?

Worst way: ignore them (sure to cost you at least one question on the final exam) Better way: assume they are like everyone else and that they would have the same rate of outcomes as those who did finish Best way: assume that they all did fine (in the exposed group) and all died (in the control group): AKA worst case analysis

An example of worse case

Study follows 100 people who eat chicken every day and 100 people who eat beef every day; plan is to measure their cholesterol after 1 year Study tracks down 90 people in each group with the following results:
Chicken group: 30/90 have high cholesterol Beef group: 60/90 have high cholesterol

What about the lost people?

Assume that all 10 in the chicken group had high cholesterol and all 10 in the beef group had normal cholesterol Now:
Chicken: 40/100 had high cholesterol Beef: 60/100 had high cholesterol

Conclusion remains the same despite the worst case

Some other problems

Cohort studies may take a long time Cohort studies may require a large number of people especially if the outcome is uncommon Both of these make cohort studies expensive

Retrospective or historical cohort

Prospective cohort studies start with the exposure, then follow patients over time Retrospective (or historical) cohort studies start with an exposure that happened some time ago, then look at the outcomes today Important point: Even though this is retrospective, it starts with the exposure or risk and then measures the outcome

An example of historical cohort

Polycythemia vera which was treated with radiation was found to be associated with leukemia Question arose of whether it was the disease (PV) or the treatment that might be causing the leukemia

Study design
Investigators went back to records from 1947-1955 and identified patients with PV They categorized patients into 4 groups: (1) No radiation treatment, (2) Treatment with x-rays, (3) Treatment with a radioactive isotope, (4) Treatment with both x-ray and isotope They then looked ahead 25 years (in 1980) to see how many in each group got leukemia

Results
Leukemia rate PV without treatment 1.0% X-ray treatment 9.2% Radioisotope treatment 11.8% X-ray & radioisotope 17.3%

Expressing results
Study results are usually expressed as a ratio or rate Cohort studies: relative risk Case control studies: odds ratio (which is the same as relative risk for large groups and uncommon outcomes)

Relative risk calculation

Rate in exposed group Relative risk = ----------------------------------Rate in controls

Example: sample study

Is hepatitis B associated with hepatocellular CA? Two groups followed 9 years:
152 cancer cases in 3,454 patients with Hep B (4.2%) 9 cases in 19,253 without Hep B (0.04%)

Calculate RR
Relative risk= (152/3,454) / (9/19,254) RR= 98.4 Interpretation: patients with hepatitis B are 98 times more likely to develop hepatocellular cancer

Drawbacks in relative risk

Relative risk tells you your risk RELATIVE to people who dont have the risk factor Your ABSOLUTE risk may still be low if very few people get the disease in the first place In other words, 10 times more often than when hell freezes over is still pretty rare

Example
7% patients exposed to smoking get cancer in their lifetime compared to 3.5% who dont smoke (RR=2) 1% of patients exposed to radiation get thyroid cancer compared to 0.25% of people not exposed (RR=4) Which exposure hurts more people?

Absolute risk reduction

Absolute risk reduction (ARR) is the DIFFERENCE between the risk with treatment than without it ARR is the absolute value of: (Rate in treated group) - (rate in group without treatment)

Two studies with same RR

10% of people with disease A get better without treatment, but 30% get better with treatment (RR=3) 0.001% with disease B get better without treatment, but 0.003% get better with treatment (RR=3) ARR for treating A = 0.2 (or 20%) ARR for treating B = 0.00002 (or 0.002%)

Number needed to treat

ARR can be taken one step further in the NNT NNT is the number of people you would need to treat to benefit 1 person NNT = 1 / ARR In last example: Disease A NNT = 1 / 0.2 = 5 Disease B NNT = 1 / 0.00002 = 50,000

Experimental design
Investigator controls exposure to the risk or treatment by assigning subjects to one group (experimental group) or another (control group) Assignment to experimental or control attempts to make sure both group are similar in all ways except the experimental manipulation

Experiments and cohorts

In experimental studies, individuals are ASSIGNED to a group and the exposure is FORCED upon them In a cohort study, the exposure is NATURAL and individuals are NOT ASSIGNED to a group; they are in a group either by their own choosing (smoking vs. not) or by chance (exposure to radiation leak or not)

Randomization of subjects
To help assure that groups are similar, subjects are randomly assigned to experimental or control groups Randomization is performed to increase the likelihood that groups are matched in other, non-experimental ways Randomization does not assure that the groups are the same: still need to assess whether they are

Blinding
Even after randomization, it is possible that experimental subjects may be treated differently than controls To combat this, blinding is often used Blinding means that the subject, investigator, or both do not know what group the subject is assigned to

Placebos
Placebos are another way of trying to make both groups similar A placebo is a biologically inactive substance given to the control group so that they think they are being treated Placebo effect is important: many patients in the placebo group report getting better simply because they are taking the placebo!

Are placebos ethical?

When there is no known successful therapy, a placebo is ethical (but withholding a treatment known to be effective is not!) Using a placebo instead of the experimental drug is ethical since the experimental drug is not known to be beneficial and could actually be harmful

P value and Power

When most studies are done, they are performed using a sample and not everyone in the universe Because the results are based on only a sample, there is a possibility that the results are not representative of everyone Consequently, all results based on samples are only estimates of the truth

Type I and Type II errors

The interpretation of results based on samples can have two types of errors Reality Different Same Different Correct Type I Result Same Type II Correct

P and power
P is the probability that any difference between groups happened just by chance (i.e. probability of making a Type I error) Power is the probability that the two groups are the same (i.e. probability of not making a Type II error)

Whats a good P
If you are to believe that two groups are really different, the P (probability that this happened just by chance) should be low 0.05 (of a 5% probability that the results differ just by chance) is accepted as a standard cut-off

What influences P?
The P value is related to the difference between the groups, the sample size, and the variability within the sample P ONLY tells you about the probability that the difference occurred by chance It DOES NOT tell you anything about the size of the difference between two groups

Example to drive home this point

Study on Drug A with 100,000 people shows that it lowers SBP by 2 points and DBP by 1 point (P=0.001) Study on Drug B with 250 people shows that it lowers SBP by 10 points and DBP by 6 points (P=0.04) Which is the better drug?

Whats a good Power

We like studies that have a POWER of 80% of higher Effected by:
Amount of difference between the two groups Size of the sample Internal variation within the groups

Look at study results

Is P <0.05?
YES 2 are different Is Power 80% YES 2 groups the same NO Lousy study NO