Document Sample
Bias Powered By Docstoc
					                                                                             Epidemiologic Methods (Epi 203)
                                                          Problem Set: “Selection Bias and Measurement Bias”
                                                                                  Due 11/13/01 at 1 pm section

1. A recent study evaluated the association between antidepressant use and breast cancer risk
(Cotterchio M, Krieger N, Darlington G, Steingart A; Antidepressant medication use and breast
cancer risk. Am J Epidemiol 2000; 151: 951-7).

   Experimental and epidemiologic studies suggest that antidepressant medication use may be associated with
   breast cancer risk. This hypothesis was investigated using a population-based case-control study; cases
   diagnosed in 1995-1996 were identified using the Ontario Cancer Registry, and controls were randomly
   sampled from an Ontario Ministry of Finance database. Data were collected using a self-administered
   questionnaire, and multivariate logistic regression was used to estimate odds ratios and 95% confidence
   intervals. Adjusted odds ratio estimates were not statistically significant for “ever” use of antidepressants,
   tricyclics, and selective serotonin reuptake inhibitors. Compared with no anti-depressant use, use of trycyclic
   antidepressants for greater than 2 years’ duration was associated with an elevated risk of breast cancer (OR=2.1,
   955 CI, 0.9-5.0). Of the six most commonly reported antidepressant medications, only paroxetine use was
   associated with an increase in breast cancer risk (OR=7.2, 95% CI, 0.9-58.3). Results from this study do not
   support the hypothesis that “ever” use of any antidepressant medications is associated with breast cancer risk.
   Use of tricyclic medications for greater than 2 years, however, may be associated with a two-fold substantial
   increase in breast cancer risk.

   Cases: age-stratified random sample of women aged 25-74 years, diagnosed with primary breast cancer during
   1995 and 1996 (pathology report confirmed) and recorded in the population based Ontario Cancer Registry.

   Controls: population-based, aged 25 to 74, randomly sampled from the property assessment rolls of the Ontario
   Ministry of Finance; includes all home owners and tenants.

   Data were collected through mailed self-administered questionnaires. Antidepressant use included information
   on duration, dosage, timing and type of medication. Subjects were asked if they took them at any time in their
   life for at least 2 weeks. 11 examples of antidepressant agents were included.

   a) An apparent association was found between Paroxetine use (introduced in 1993) and
      breast cancer (OR=7.2, 95% CI, 0.9-58.3).
       (1) Explain a way that selection bias could have accounted for the apparent association?
           (2 pt)
       Breast cancer may initially be asymptomatic and detected only by volitional screening tests. For the study
       base, the authors chose the entire population of Ontario, which includes all women including those who
       never undergo routine screening for breast cancer. To the extent that the cases include any women who
       were asymptomatic and detected via a screening test, the cases are apt to be enriched for women who are
       under the active care of a physician and thus more likely than the controls to be taking anti-depressants
       (one needs to be under medical care to get anti-depressants). Looked at from the perspective of a cohort of
       women moving through time, women who are taking anti-depressants are definition under medical
       treatment and more likely to be undergoing routine screening examinations for a variety of things
       (increased medical surveillance) including breast cancer. Relative to a reference group of all women in the
       population, those women taking anti-depressants will be more likely to have breast cancer diagnosed
       simply because of increased medical surveillance.

       (2) How would you design a similar study to reduce the potential for selection bias?
           (2 pt)

       Controls could be selected from a study base that would include everyone for whom a diagnosis of breast
       cancer would have been made and detected by the registry if the disease had indeed developed. The key
       here is in the “a diagnosis . . .would have been made”. One approach would be to include in the control
                                                                                  Epidemiologic Methods (Epi 203)
                                                          Problem Set: “Selection Bias and Measurement Bias”
                                                                                        Due 11/13/01 at 1 pm section
         group only women who had seen a primary care provider in the past year. Another might be to include as
         controls only women with other types of cancer detected in the last year (via the cancer registry) that also
         have an asymptomatic phase that can be picked up with routine screening tests (e.g., colon cancer or
         cervical cancer).

         (3) Explain a way that measurement bias could have accounted for the apparent
             association? (2 pt)

         It is likely that many times courses of paroxetine are started but used for only a short period of time. Cases
         may better remember taking paroxetine compared to controls due to their increased attention to their
         medical condition. They may feel at greater ease revealing their use of antidepressant medication
         compared to controls for the same reason. This differential misclassification of exposure could result in the
         apparent association between anti-depressant use and breast cancer.

         (4) If you were designing the study, how would you perform the measurements to avoid
             this bias? (2 pt)

         Medical records could have been reviewed to ascertain exactly what antidepressant medications had been
         prescribed for the cases and controls rather than relying on self report.

    b) If the antidepressant medication use -- breast cancer association were well-known in the
       community, how would early signs of breast cancer influence the association of “current”
       anti-depressant use and breast cancer? (2 pts)

         Breast cancer diagnosis could have led to cases discontinuing the use of anti-depressants, decreasing the
         proportion of current exposed-diseased (cell a) and resulting in an artificially low association between the
         “current” anti-depressant use and breast cancer. This should have no effect on the observed association
         between "ever" use of antidepressant medication and breast cancer.

2. A study in Finland evaluated life satisfaction and 20-year mortality (Koivumaa-Honkanen et
   al., Self-reported life satisfaction and 20-year mortality in health finish adults. Am J
   Epidemiol 2000; 152:983-91)

Abstract: The authors investigated the role of self-reported life satisfaction in mortality with a prospective cohort
study (1976-1995). A nationwide sample of healthy adults (18-64 years, n=22,461) from the finish Twin Cohort
responded to a questionnaire about life satisfaction and known predictors of mortality in 1975. A summary score for
life satisfaction (LS), defined as interest in life, happiness, loneliness, and general ease of living (scale range 4-20),
was determined and used as a three-category variable: the satisfied (LS, 4-6), the intermediate group (LS, 7-11)
(65%), and the dissatisfied (LS, 12-20) (14%). Mortality data were analyzed with Cox regression. Dissatisfaction
was linearly associated with increased mortality. The age-adjusted hazard ratio of all-cause, disease or injury
mortality among dissatisfied versus satisfied men were 2.11 (95% CI, 1.68-2.64), 1.93 (95% CI, 1.40-2.39) and 3.01
(95% CI, 1.94-4.69), respectively. Adjusting for marital status, social class, smoking alcohol use, and physical
activity diminished these risks to 1.49 (95% CI, 1.16-1.92), 1.35 (95% CI, 1.01-1.82), and 1.93 (95% CI, 1.19-3.12),
respectively. Dissatisfaction was associated with increased disease mortality, particularly in men with heavy alcohol
use (hazard ratio=3.76, 95% CI, 1.61-8.80). Women did not show similar associates between life satisfaction and
mortality. Life dissatisfaction may predict mortality and serve as a general health risk indicator. This effect seems
to be partially mediated through adverse health behavior.

1. Baseline questionnaire sent by mail in 1975.
2. 31,133 subjects responded. Response rate was 84% for those age 18-64 in 1975. Eligibility for this study: had
    to have baseline life satisfaction score available from 1975, alive at the start of follow-up in 1976 and self-
    reporting as “healthy” at baseline.

                                                                                         Epidemiologic Methods (Epi 203)
                                                                  Problem Set: “Selection Bias and Measurement Bias”
                                                                                              Due 11/13/01 at 1 pm section
3.   A self-reported life satisfaction score was determined by means of a scale with four questions: Do you feel that
     your life at present is (1) very interesting, fairly interesting, fairly boring, or very boring?; (2) very happy, fairly
     happy, fairly sad or very sad; (3) very easy, fairly easy, fairly hard, or very hard; (4) Do you feel that at the
     present moment you are: very lonely, fairly lonely or not at all lonely?

a. As an inclusion criterion, the study relied on self-report of being healthy at baseline. If
   dissatisfaction with life was associated with certain specific diseases that may have already
   been present at baseline but not yet diagnosed, such as cardiovascular disease, explain the
   possible bias (including direction of the bias) that might have been introduced? (1 pt)

If the presence of not-yet-diagnosed conditions was resulting in persons expressing dissatisfaction with life, then
those entering the cohort classified as “dissatisfied” would spuriously be found to develop more morbid conditions
(and likely have greater mortality).

         Is this selection or information bias? (1 pt)

This is a form of selection bias.

b. If the entity of life satisfaction was measured with perfect accuracy, the following data
would result:

          Exposure                                              Outcome

     Life satisfaction                         Death                           No Death

                                                  67                              1673
      (Exposed) Dissatisfied
                                                 187                              9412
      (Unexposed) Satisfied

Calculate a measure of association between life dissatisfaction and mortality using these data. (1

RR= 1.97

c. In the study, measurement of life satisfaction was performed by self-report by using 4 short
questions. Assume that a team of behavior scientists assessed the validity of the 4 questions
(versus a gold standard qualitative interview) and found it to be 90% sensitive for detecting
dissatisfaction and 85% specific. How would this change the cells in the 2 x 2 table? Please fill
in the “observed” table below. (4 pt)

                                                                       Epidemiologic Methods (Epi 203)
                                                    Problem Set: “Selection Bias and Measurement Bias”
                                                                            Due 11/13/01 at 1 pm section
    Exposure                                        Outcome

Life satisfaction                     Death                      No Death

                                         88                        2918
 (Exposed) Dissatisfied          sensitivity =0.9            sensitivity =0.9
                                 67(0.9) + 187(1-          1673(0.9) + 9412(1-
                                    .0.85)= 88                 0.85)=2918
                                        166                        8167
 (Unexposed) Satisfied           specificity=0.85            specificity=0.85
                                   67(1-0.9) +                1673(1-0.9) +
                                  187(0.85)=166             9412(0.85)=8167

 Calculate the measure of association in the new table. (1 pt)

 Misclassified RR= 1.4

 Is this differential or non-differential bias? (1 pt)

 Non-differential misclassification

 What is the effect of this bias on the estimate of the effect measure? (1 pt)
 Bias toward the null


Shared By: