NUTR 319 INTERMEDIATE EPIDEMIOLOGY SPRING 2013 HOMEWORK 2 TOPICS: Sample Size and power; Measurement Error SECTION

A: Sample size and power 1. In addition to Type I and Type II error, what other factors determine optimal sample size? 2. Estimate the statistical power for a study with the following parameters: 1) Range of sample sizes (50 to 300), case:control ratio=1 2) Type I error of 0.05 3) Range of exposure prevalence in controls is fixed at 30% 4) Range of detectable odds ratios (1.5 to 4.0) Fill in the Table 1 with the estimates of statistical power for each cell: TABLE 1. Sample size estimates Sample Size Detectable Difference 0 50 100 200 300 1.0 .025 .025 .025 .025 1.5 .024 .16 .27 2.0 .023 .65 .92 .98 3.0 .021 .76 >.99 >.99 4.0 .019 .93 .99 >.99 >.99 3. Using the data from Table 1 and the commands provided below, plot five power curves that are associated with different odds ratios (ORs from 1.0 to 4.0) and different sample sizes (n from 50 to 300) with a 30% prevalence of exposure among controls, type I error 0.05, and an equal number of cases and controls.

/* draw power curves */ data power2; input or n power; cards; 1.0 00 .025 1.5 00 .024 2.0 00 .023 3.0 00 .021 4.0 00 .019 1.0 50 .025 1.5 50 .157 . . . . . . . . . . . . . . . . . . 4.0 300 1.000 ; run;

input data from Table 1 here

proc gplot data=power2; plot power*n=or; symbol line=1 width=5 interpol=join; title Power curves; run; quit;

4. Examine the plots and answer the following questions: a. Describe the relationship among statistical power, sample size and detectable differences. b. Which parameter (n or OR) is relatively more important in terms of increasing statistical power? Support your answer. 5. An investigator is applying for a NIH grant and wants to show to his/her reviewers that the proposed study has at least 80% statistical power in detecting an association of 2.0 (OR=2.0) when the exposure prevalence among controls is 30% and type I error=0.05. The cost of recruiting cases is much more expensive than controls. Facing a budget constraint, the investigator has to limit the recruitment of cases to 100. a. Based on the information in Table 2 below, plot a power curve with varying control:case ratios for this particular study and suggest ways of increasing the statistical power.

TABLE 2. Statistical power with varying case:control ratio scenarios OR c n Power 2 1 100 .654 2 2 100 .784 2 4 100 .856 2 8 100 .892 2 10 100 .898
OR: odds ratio c: case:control ratio n: number of cases power: statistical power at the significance level of 5%

b. Referring to the power curve, how many cases and controls would this investigator need in order to have at least 80% power? c. Beyond which case:control ratio (c) would you not recommend? Why? SECTION B: Validity and reliability studies You are concerned about precision and validity in a case-control study of esophageal cancer (epidat1_clean.sas7bdat). In particular, you think that there may be error in the measurement of alcohol and tobacco use, as well as in the classification of the outcome. You conduct two different studies: STUDY 1: Validity Study of Alcohol Intake Measures. You select 100 subjects from among the 978 participants in the parent study and ask these subjects to record alcohol intake over 7 days. Treating the 7-day diet record as the gold standard, you cross classify the results for the 100 subjects with their original responses on the food frequency questionnaire (FFQ) into high and low consumption of alcohol. The results are as follows: ALCOHOL High: 40+ gm/day Low: 0-39 gm/day FFQ 7 Day Diet Record High: 40+ gm/day Low: 0-39 gm/day 55 2 7 36

6. Compute the sensitivity and specificity for the alcohol intake measurement: High Alcohol Sensitivity = 55/62 = _______ Specificity = _______ = 0.95 7. Using the calculated sensitivity and specificity for alcohol, we can correct the odds ratio for measurement error. Within your sample of 100, you have checked that the sensitivity and specificity for alcohol does not differ by outcome group. Now lets consider the measurement of alcohol.

a. Using the data from the parent study (N=978), calculate the measured exposure [i.e., the proportion of cases exposed (heavy alcohol drinkers)] and the odds ratio. Measured exposure prevalence (drinking 40+ gm/day versus drinking 0-39 gm/day): Cases = Controls = Measured odds ratio = b. Using the information of sensitivity and specificity to correct for the measurement error, the true exposure prevalence can be calculated in the following formula. True Exposure prevalence = (measured exposure + specificity 1) (sensitivity + specificity 1) Calculate the true exposure prevalence for: Cases = Controls = True odds ratio = (.86 + .95 1) / (.89 + .95 1) = = =

c. How did measurement error affect the measured OR in comparison to the true OR? Briefly explain any observed difference. 8. What type of misclassification is this? Can the direction of this bias always be predicted? If so, in what direction is the bias? If not, under what scenarios can it be predicted and in what direction? STUDY 2: You now turn to the outcome measure and realize that there may be measurement error in the classification of the outcome. For this esophageal cancer study, error in measuring the outcome variable can take on one of two forms: 1) error in classifying cases and controls 2) error in classifying histologic type among cases You are less concerned about the former because past studies have shown very good agreement among pathologists in distinguishing invasive cancer from no cancer. However, you are very concerned about the latter because some studies have suggested that certain histologic types of esophageal cancer may be associated with alcohol intake while other types are not. Therefore,

you decide to do a reliability study to examine the agreement between pathologists in classifying histologic types of esophageal cancer. The following are your results: Histologic Type Pathologist Squamous Pathologist A Adenocarcinoma Pathologist A 9. Compute Kappa and interpret. 10. If alcohol is only associated with a particular type of esophageal cancer, what is the likely impact that these data have on studies of alcohol and esophageal cancer where different histologic types are considered together? Squamous Pathologist B 60 10 Adenocarcinoma Pathologist B 10 20