PHARMACOEPIDEMIOLOGY under consideration demonstrate a statistically STUDY SAMPLE significant difference, then one is said to have an association. The process of assessing whether random variation could have led to a study's CONCLUSION ABOUT A POPULATION findings is referred to as statistical inference, and (ASSOCIATION) represents the major role for statistical testing in the scientific method If there is no statistically significant differ (CAUSUT SCIENTIFIC THEORY ONCLUSION ABOl ence between the study groups, then the process in Figure 2.1 stops. If there is an association then one is tempted to generalize the results of Figure 2.1. Overview of the scientific method the study even further, to state that methyldopa is an antihypertensive drug, in general. This is referred to as scientific or biological infere and the result is a conclusion about causation pressure of the 20 men treated with the active that the drug really does lower blood pressure in drug decrease more than the blood pressure of a population of treated patients. To draw this the 20 men treated with a placebo. In this ex- type of conclusion, however, requires one to gen ample, the 40 study subjects would represent the eralize to populations other than that included in study sample, theoretically a random sample of the study, including types of people who were middle-aged hypertensive men. In reality, the not represented in the study sample, such as study sample is almost never a true random women, children, and the elderly. Although it sample of the underlying target population, may be obvious in this example that this is in every individual who belongs in the target popu- case. Unlike statistical inference, there are no lation and then randomly choose from among precise quantitative rules for biologic inference them. However, the study sample is usually trea- Rather, one needs to examine the data at hand in ted as if it were a random sample of the target light of all other relevant data in the rest of the population scientific literature, and make a subjective judg At this point one would be tempted to make a ment. To assist in making that judgment, how generalization that methyldopa lowers blood ever, one can use the"Criteria for the Causal pressure in middle-aged hypertensive men. How. Nature of an Association, "described below ever,one must explore whether this observation First, however, we will place causal associatio could have occurred simply by chance, that is into a proper perspective, by describing the dir- due to random variation. If due to chance, then ferent types of errors that can be made in per- the observation might not have been made if one forming a study and the different types of had chosen a sample of 40 different study sub. associations that each results in jects. Perhaps more importantly, it might not exist if one were able to study the entire theore- tical population of all middle-aged hypertensive TYPES OF ERRORS ONE CAN MAKE IN men. In order to evaluate this possibility, one PERFORMING A STUDY can perform a statistical test, which allows an investigator to quantitate the possibility that the There are four basic types of associations which difference seen between the two study groups can be observed in a study (Table 2.1). The basic could have happened simply by chance. There purpose of research is to differentiate among are explicit rules and procedures for how one them should properly make this determination: the science of statistics. If the results of any study First, of course, one could have no association Second, one could have an artifactual associa- 第181页
第 181 页
STUDY DESIGNS AVAILABLE FOR PHARMACOEPIDEMIOLOGY STUDIES Table 2.1. Types of associations between factors under systematic difference in recall is referred to as a recall bias. 5 Note that biases, once present, cannot be cor- independent 2 Artifactual (spurious or false) rected. They represent errors in the study design a. Chance (unsystematic variation) which can result in incorrect results in the study. b.Bias (systematic variation) It is important to note that a statistically sig- 3. Indirect(confounded nificant result is no protection against a bias; one 4. Causal (direct or true) can have a very precise measurement of an incorrect answer! The only protection against biases is proper study design. tion, that is a spurious or false association. This Third, one can have an indirect, or con- can occur by either of two mechanisms: chance founded, association. A confounding variable, or or bias. Chance is unsystematic, or random, vari- confounder, is a variable other than the risk ation. The purpose of statistical testing in science factor and outcome under study which is related is to evaluate this, estimating the probability independently to both the risk factor and the that the result observed in a study could have outcome variable, which may create an apparent ppened purely by chance association or mask a real one. For example,a The other possible mechanism for creating an study of risk factors for lung cancer could find a artifactual association is bias. Epidemiologists' very strong association between having yellow use of the term bias is different from that of the fingertips and developing lung cancer. This is lay public. To an epidemiologist, bias is systema- obviously not a causal association, but an indir- tic variation, a consistent manner in which two ect association, confounded by cigarette smoking. study groups are treated or evaluated differently. Specifically, cigarette smoking causes both yellow This consistent difference can create an apparent fingertips and lung cancer. Although this ex- association where one actually does not exist of ample is transparent, most examples of con- course, it also can mask a true association. founding are not. In designing a study, one must There are many different types of potential consider every variable which can be associated study in which the research assistant is aware of variable under study, in order to plan to deal the investigator's hypothesis. Attempting to with it as a potential confounding variable. Pre- please the boss, the research assistant might ferably, one will be able to specifically control study group carefully during interviews with one for the confounding variable, using one of the than during interviews with the techniques listed in Table 2.2. difference in how carefully the inter Fourth, and finally, there are true, causal asso- viewer probes could create an apparent but false ciations. association, which is referred to as an interviewer Thus, there are three possible types of errors bias. Another example would be a study of drug- that can be produced in a study: random error, induced birth defects which compares children bias, and confounding. The probability of with birth defects to children without birth defects.A mother of an abnormal child inter- Table 2. 2. Approaches to controlling confounding viewed about any drugs she took during her pregnancy may be likely to remember drug inges- 1. Random allocation tion during pregnancy with greater accuracy than 2. Subject selection a mother of a normal child, because of the unfortunate experience she has undergone. The 3. Data analysis improved recall in the mothers of the abnormal a. Stratification children may result in false apparent associations b. Mathematical modeling between drug exposure and birth defects. This 第182页
第 182 页