Anda di halaman 1dari 21

Observational Studies

Why the caution?


Dr Rao found that patients who had had a transfusion because of a low red blood-cell count had an 8% chance of dying within 30 days. Without a transfusion, only 3% died. Those numbers need to be treated with caution.
The Economist Magazine, 2007

Chance, or is something else going on?


Dr. Rao found a difference of 8% - 3% = 5% more deaths within 30 days with blood transfusion. Is this a real difference, or just chance? If not chance, what else might be going on? Dr. Rao found an association with blood transfusion, but he cautioned that the effect of severity of illness might have gotten mixed up with the effect of the exposure.

Confounding variables
Dr. Raos study of patients survival after blood transfusion is observational, and the severity of illness is a confounding factor Definition Listen, for the sound of two feet dropping One: severity of illness correlates with the exposure, blood transfusion Two: severity of illness correlates with the outcome, death

True or false, and explain


Some studies find an association between liver cancer and smoking, but alcohol consumption is a confounding variable.

Two strategies
I. Matching
Three levels of severity
Low Exposed Control

Medium

High

II. Randomized clinical trial


Treatment

Eligible Popn or the popn at risk

1/2

1/2

Control

The clofibrate trial for heart disease prevention


a randomized, controlled, double-blind experiment

Clofibrate
Number

Placebo
Number 1,813 882 2,789 Deaths 15% 28% 21%

Deaths 15%

Adherers Non-adherers

708 357 1,103

25% 20%

Total

There are two conclusions here.

Experiments and Observational Studies


Method of comparison
 T versus C or Exposed versus Unexposed

Experiments
 Key variables are held constant  Control over group assignment, where impartial chance, called randomization, works best  Treatments are interventions, where the goal is a description of the effect.

Observational studies
 No control over group assignment, often self selection or exposure by association  Observations, not interventions  Key variables may vary, so confounding is always a risk  Matching, strategic but not always effective

Why not always experiment?

Ethics and human experimentation


 smoking and lung cancer

Prohibitive and costly


 social policy (e.g. welfare) and economics (e.g. monetary policy)  natural experiments; $ and happiness in Switzerland

Modes of scientific discovery


 notebooks, laboratories, and theories

Smoking and health


The Public Health Service studied the effects of smoking on health in a large sample of representative households. For men and for women in each age group, those who had never smoked were on average somewhat healthier than current smokers, but the current smokers were on average much healthier than those who had recently stopped smoking. (a) (b) Why study men and women and the different age groups separately? The lesson seems to be that you shouldnt start smoking, but once youve started, dont stop. Comment.

Cervical cancer and circumcision


Cervical cancer was once quite common. Epidemiologists trying to identify the cause had noticed that in several different countries it was rare among Jews and Moslems. In the 1950s, some investigators concluded that circumcision of males was protective. Was this conclusion justified? Discussion *What is the response or outcome? the exposure or treatment? * Do these communities differ from members of other communities in ways besides circumcision?

Cervical cancer1970s
According to a study done at Kaiser Permanente in Walnut Creek, California, users of oral contraceptives have a higher rate of cervical cancer than non-users, even after adjusting for age, education, and marital status. Investigators concluded that the pill causes cervical cancer.* Experiment? Or, observational study? Why the adjustments for? Users were likely to differ from non-users on another factor affecting the risk.what other factor? Were the conclusions justified? Yes or no, explain.
*American Journal of Epidemiology, Vol. 106, 1977, pp. 462-69.adjustments were made for religion, smoking, number of PAP smears before entry, and selected infections.

What explains the observations?


What is the human papilloma virus? And how is it spread? Associations establish links, but they are not causal. They can even be spurious.

Vitaminsassociations and effects


People who get lots of vitamins by eating five or more servings of fresh fruits and vegetables (especially cruciferous vegetables like broccoli) have much lower death rates from colon cancer and lung cancer, according to many observational studies. These studies were so encouraging that two randomized controlled experiments were done. Findings were 1. No difference in death rate for colon cancer between T and C; 2. Beta carotene (as a dietary supplement) increased death rates due to lung cancer. True or false, and explain 1. Experiments confirm the results of the observational studies; 2. People who eat lots of fruits and vegetables have lifestyles that are different in many other ways tooso due to confounding the a. observational studies could easily have reached the wrong conclusion; or b. experiments could easily have reached the wrong conclusion. In what way might the observational studies have gotten it right?

Associations establish links


Here are two examples Dr. Rao had established an association or link between blood transfusion and survival. It did point to something causal: if transfusion was a cause of death, then youd expect.but this alone does not prove causation. Nurses Health Study found a link between hormones and cardiovascular disease. It did point to something causal: if hormones are protective, then youd expectbut the Womens Health Initiative did not confirm the protective effects in an randomize controlled experiment. WHI did, however, corroborate the risks

Experiments measure effects


Here are two examples Salk Vaccine Field Trial held certain key variables constant protocol, blinding, etc.while introducing the vaccine to the treatment group but not the controls. Randomization made the controls like the treatment group in important respects, except for the intervention, so that the difference in response is likely to be due to the effect of the vaccine itself. HIP Screening Trials held certain key variables constant, while introducing screening to treatment and not the controls. Randomization made.so the difference in death rates is likely to be due to the lives saved by screening and early detection.

International Rice Research Institute, Philippines


I. Subjects and the treatment Subjects: 20 experimental plots planted with IR 8 Treatment: 5 amounts of Nitrogen fertilizer  0 oz, 4 oz, 8 oz, 12 oz, 16 oz  light, moisture, grade, and key nutrients held constant II. Randomization and the response  Amount Nitrogen assigned randomly to plots  Ounces of rice, or yield, is the response III. Response schedule states the effect predicted yield = (20 oz rice per oz nitrogen ) nitrogen + 240 oz

Summary
1. In observational studies, the investigators do not assign subjects to treatment or controlas in the case of smoker versus non-smokers. 2. Observational studies can establish associations, which may point to causation: if exposure causes disease, then the exposed ought to be sicker than the unexposed. Why the caution? Confounding factors 3. With observational studies, and nonrandomized controlled experiments, try to find out how the subjects came to be in treatment or in control. Are the groups comparable? different? Were there confounding factors? adjustments? etc... 4. Study design is a central issue in applied statistics, as shown here by contrasting controlled experiments with observational studies. The great weakness of observational studies is confounding; randomized controlled experiments minimize this weakness; when things go as planned, experiments measures the effects of interventions.

Supplements
The following slide provides supplements to our study of experiments and observational studies.

Design is central in applied statistics


Which studies include an intervention? self selection? Which are observational? experimental?
 Salk Vaccine Field Trial  Nurses Health Study of the effects of long term oral contraceptive use  Women Health Initiative of HRT  IRRI, Philippines, study of IR8 The next two go ahead  Predicting a students 1st-year GPA from her Math SAT, again using linear regression  Predicting a mans height from his weight, using linear regression
If a man puts on 20 pounds, will he grow taller by 20 0.047 inches per lb 0.9 inches? How does this case differ from a study of Hookes Law? (next slide)

Robert Hooke (England, 1653-1703)


Hookes Law
Hang a weight on a length of piano wire and it will stretch. In one experiment at Berkeley, it turned out that length (0.05 cm per kg) load + 439.1 cm Increase the load by 2 kilograms, and the wire lengthens by about 2 0.05 = 0.10 centimeter.

Anda mungkin juga menyukai