Selection bias

Example 3 (continued). Butler et al. also addressed the issue of selection bias in the case-control study on meat consumption. In this study, there was a 16% greater response rate among cases and they could not exclude the possibility of bias in their study. 

---

The answer is patently no —not because of any general sceptical-philosophical scruples about induction, but because of the particular fact that the impression of consistent predictive success for Mendeleev s scheme is a complete misrepresentation of history a classic example of an effect (Mendeleev s predictive success ) created by selection-bias. Mendeleev made any number of predictions on the basis of his scheme (or rather schemes—there are at least 65 versions of Mendeleev s table, published and unpublished). Many of these predictions (depending on how exactly they are individuated)—perhaps a majority—were unsuccessful. 

Differences exist in primary care between the patterns of prescribing fluoxetine, paroxetine or sertraline, which may influence cost outcomes. Sertraline-treated patients are more likely to have their dose increased (Sclar et al, 1995 Donoghue, 1998), and to drop out of treatment prematurely (Donoghue, 1998). The apparent need to titrate doses upwards with sertraline may require more involvement by the clinician and may delay response to treatment, with resultant increases in direct health costs (Sclar et al, 1995). However, these economic findings are retrospective, may suffer from selection bias, and being derived from HMO patients may not be generalizable to other populations confirmation in further studies is required. 

A subsequent large epidemiologic study found a greater risk for breast cancer with combined estrogen-progestogen use, as well as increased risk for estrogen-only therapy and tibolone, but selection bias was found in the study population. 

Bias Systemic error as opposed to a sampling error. For example, selection bias may occur when each member of the population does not have an equal chance of being selected for the sample... 

The segment of the population reporting adverse reactions cannot be considered representative of the entire population ingesting the product due to selection bias. 

Bias Occurs when there is a tendency to produce results that differ in a systematic manner from the true values. A study with small systematic bias is said to have high accuracy. Bias may lead to over- or underestimation of the strength of an association. The sources of bias in epidemiology are many and over 30 specific types of bias have been identified. The main biases are selection bias, information bias, and bias due to confounding. 

Selection bias Occurs when there is a systematic difference between the characteristics of the people selected for a study, or who agree to participate, and the characteristics of those who are not selected, or who do not agree to participate (e.g., in a study limited to volunteers). 

At one end of the spectrum, the event may be a simple dosage problem which could be an error on the part of the prescriber or an unanticipated hypersensitivity for that particular patient. At the other end of the spectrum, is an uncommon, serious adverse reaction not revealed in premarketing clinical trials. Somewhere between those two extremes are more or less serious adverse events which are not entirely unexpected but appear to be more common than is accepted for comparable products in the same therapeutic category. This maybe a real increase in frequency or may be due to patient selection bias. The later has arisen with new products which claim a lower incidence of certain adverse reactions which encourages doctors to precribe them preferentially for patients who have suffered such reactions with older products. 

As discussed by Hood and Miller (7), and in the lecture by Hood at the Maternal Toxicology Symposiums in 2009 (summarized in Beyer et al. (8)), it can be argued that it is Khera s interpretation, rather than the developmental toxicity study results themselves, that may be of concern. Khera s literature review indicated a possible association between maternal toxicity and embryo-fetal effects, but it did not establish a causal relationship between these two observations. Additional criticisms of Khera s hypothesis include the fact that his literature review was retrospective, there was a potential selection bias arising from the general tendency not to publish negative data, and the failure to adequately address maternal toxicity endpoints in the published literature of the time. In fact, Khera himself stated that in 40% of the studies he evaluated in support of his hypothesis the maternal toxicity data were insufficient or nonexistent (9). 

One consistent finding has been that patients with psychotic depression treated with antidepressant monotherapy or even the combination of an antidepressant plus antipsychotic have a lower response rate than those depressed patients without psychotic symptoms (45). Although evidence supports an improved response when these patients undergo treatment with ECT, this apparent superiority may be related to a selection bias ( 46). Thus, it may be that patients with psychotic depression are more likely to receive ECT earlier in their course of illness and therefore the extent of their true drug resistance is unknown ( 1). 

Selection bias (i.e., unobserved heterogeneity) exists when case and control groups differ in unobservable ways because participants were assembled after the occurrence of the causal process rather than before. For example, because hazardous occupations generally employ healthier workers, the health of those workers before exposure fo workplace hazards is not the same as the health of a control group of workers in nonhazardous occupations. ... 

A 1978 WHO report quoted an unpublished study by March in which pituitary adenomas were found in 26% of women with secondary amenorrhea following the use of oral contraceptives, yet in only 13% of cases who had not used these products (118). The difference was significant, but selection bias might have explained the results. 

Olanzapine had significantly positive diabetic effects, based on both duration of treatment and dosage. All patients had been exposed to antipsychotic drugs for more than 60 days because there was less awareness of the diabetic effects of atypical antipsychotic drugs during that period, the use of these data reduced the possibility of selection bias. 

Selection bias is the Achilles heel of such samples. Therefore when researchers use convenience samples for assessing population characteristics such as prevalence, incidence, or causal relationships, they must justify the validity of the sample. 

If the study is exploratory, it is not uncommon to have a multitude of creative ideas about the nature of the biomarker response and its relationship to the exposure. Ideally, some comparisons of particular interest are specified in advance. The analyses can then be divided into confirmatory and exploratory phases. Two key considerations in such exploratory studies are selection bias and confounding both were discussed in the section on the sam-... 

Reduces selection bias and observer bias (nonrandom error). 

Berger, V.W., 2005, Selection bias and covariate imbalances in randomized clinical trials, John Wiley Sons. 

The cohorts for comparison were chosen by a procedure established at UWG s Institutional Research Office to produce a set of non-LC students with similar entering characteristics (GPA, test scores, majors) to minimize program selection bias. 

Cohort Longitudinal design that begins with ascertainment of exposure(s) for cohort members with follow-up for occurrence of Less subject to selection bias Can establish a temporal order between exposure and outcome Loss to follow-up may bias results Depending upon length of follow-up needed, can be costly... 

---