Chapter 3-8. Bias and Confounding We consider error to be the difference between the unknown correct effect measure value, such as an incidence rate ratio, and the study’s observed effect measure value (Rothman, 2002, p.94). If True IRR – Estimated IRR = “small” , we say that the study is accurate and has little error. True IRR – Estimated IRR = “large” , we say the study is inaccurate and has large error. Error in a research study can be classified either as random error or systematic error. Random errors are those that can be reduced to zero if the sample size becomes infinitely large, such as the inaccuracy of an incidence proportion estimate. This is a principal called statistical regularity, which is demonstrated in the K30 Intro Biostatistics course. Systematic errors are those that remain even when the sample size is infinitely increased. For example, taking measurements with an improperly calibrated heart monitor, which always measures the heart rate too low, would be a systematic error that would remain no matter how large the sample size became (called bias due to instrumental error, Last, 1995) Bias is another term for systemic error. Bias can result from the investigator’s attitude, the way in which subjects have been selected, or the way study variables were measured. A study can also be biased due to some confounding factor that is not completely controlled-what Rothman and a few other experts refer to as confounding bias. It is useful to think of confounding as something separate from bias, however, since different approaches are used to avoid these errors in a research study. Unlike Rothman, Vandenbrouchke et al (2007, p.W-172), in the STROBE statement for reporting observational studies, STROBE_Observational_Studies_AnnIternMed2007.pdf, insist on the distinction between bias and confounding:make this distinction: “Bias and confounding are not synonymous. Bias arises from flawed information or subject selection so that a wrong association is found. Confounding produces relations that are factually right, but that cannot be interpreted causally because some underlying, unaccounted-for factor is associated with both exposure and outcome.” Some other examples of bias are listed in the following box. _____________________ Source: Stoddard GJ. Biostatistics and Epidemiology Using Stata: A Course Manual [unpublished manuscript] University of Utah School of Medicine, 2010. Chapter 3-8 (revision 16 May 2010) p. 1 A Sample of Specific Biases (Last, 1995) Here are the definitions of a few specific biases selected from a much larger list that Last included in his epidemiology dictionary. Bias in assumption: (synonym: conceptual bias) Error arising from faulty logic or premises or mistaken beliefs on the part of the investigator. Design bias: The difference between a true value and that obtained as a result of faulty design of study. Detection bias: Bias due to systematic error(s) in methods of ascertainment, diagnosis, or verification of cases in an epidemiologic study. Bias of Interpretation: Error arising from inference and speculation. Recall bias: Systematic error due to differences in accuracy or completeness of recall to memory of past events or experiences. Response bias: Systematic error due to differences in characteristics between those who choose or volunteer to take part in a study and those who do not. Bias due to withdrawals: A difference between the true value and that actually observed in a study due to the characteristics of those subjects who choose to withdraw. Bias can be classified into three broad categories: selection bias information bias confounding Exercise Look at the article by Delgado-Rodriquez and Llorca (2004). 1) Notice the long list of sources of bias, beginning on the second page. Sackett (1979) also published a list of biases, which has been widely cited. 2) Notice that the biases are categorized into the three broad categories. Exercise Look at the article by Grimes and Schulz (2002). 1) Notice these three broad categories of bias used as section headings. Chapter 3-8 (revision 16 May 2010) p. 2 Selection Bias Selection bias is a systematic error in a study resulting from the procedures used to select subjects and from factors that influence study participation (Rothman, 2002, p.96). For example, if the association between exposure and disease, such as that measured with the odds ratio, is different between participants and non-participants in a study, then selection bias is introduced into the study. An example of a selection bias is what is often referred to as the healthy worker effect. When workers of a specific occupation are compared to the general population, the occupation tends to have a lower overall death rate. This is because people in the occupation are healthy, while people in the general population include many people who cannot work due to ill health. The correct way to design such a study is to make a comparison with workers in another occupation. Berkson’s Bias This is a selection bias, a classic bias found in nearly all epidemiology textbooks, that can occur in a case-control design that uses hospital controls. Definition: Berkson’s bias (Last, 1995, p.15) A form of selection bias that leads hospital cases and controls in a case control study to be systematically different from one another.1 This occurs when the combination of exposure and disease under study increases the risk of admission to hospital, leading to a systematically higher exposure rate among the hospital cases than the hospital controls; this, in turn, systematically distorts the odds ratio. _______________________ 1 Berkson J. Limitations of the application of fourfold table analysis to hospital data. Biometrics Bull 1946;2:47-53. Of course the bias can go in the other direction as well, leading to a systematically higher exposure rate among the hospital controls. This bias is also referred to “Berkson’s fallacy”, “Berksonian bias”, and “selection bias”. Here is how the bias could operate. Suppose a researcher came up with the idea to study whether or not water pills (diuretics) were a risk factor for bladder cancer. The researcher finds N=500 cases of bladder cancer in from a cancer registry. For controls, he decides to use N=500 patients from the hospital he works at, admitted for any reason but cancer. Chapter 3-8 (revision 16 May 2010) p. 3 Suppose that in the general population, the incidence of bladder cancer is equal among those who use diuretics and those who don’t (odds ratio = 1, or no associaton). Designing a case-control study, now, the researcher begins with the 500 bladder cancer cases. Starting with the cases (row totals shown) Past or Current Diuretic Use Bladder Yes No Cancer (exposed) (unexposed) Yes 75 (15%) 425 (85%) No Row Total 500 If there really is no association, a random sample of N=500 controls from the general population (not hospitalized controls) would have the same distribution of Diuretics. If Had Used General Population Controls (row totals shown) Past or Current Diuretic Use Bladder Yes No Row Cancer (exposed) (unexposed) Total Yes 75 (15%) 425 (85%) 500 No 75 (15%) 425 (85%) 500 Odds Ratio = (75 425)/(75 425) = 1.0 Hypertension (or high blood pressure) is an early stage of heart disease. Diuretics are usually given as the initial treatment for hypertension, and many patients are still using direutics when later hospitalized for heart disease events. Since heart disease makes up a large share of hospitializations, we should expect a random sample of hospital controls to be more frequent users of diuretics. Thus, our data might look like: Using Hospital Controls (row totals shown) Past or Current Diuretic Use Bladder Yes No Row Cancer (exposed) (unexposed) Total Yes 75 (15%) 425 (85%) 500 No 100 (20%) 400 (80%) 500 Odds Ratio = (75 400)/(425 100) = 0.71 Chi-square test , p = 0.038 Since no association exists in the population, this observed protective effect is attributable to Berkson’s bias. Chapter 3-8 (revision 16 May 2010) p. 4 Such a spurious association will not arise if either (Kraus, 1954; Lilienfeld and Stolley, 1994, p.233): 1) the exposure does not affect hospitalization, that is, no person is hospitalized simply because of the exposure; or 2) the rate of admission to the hospital for those persons with the disease is equal to those without the disease. Lilienfeld and Stolley (1994, p.233) clarify that condition 1) need only be an association, which makes it difficult to judge, and they provide this example: If the exposure is eye color, it could easily be assumed that this could not influence the probably of hospitalization. It is possible, however, that persons with a particular eye color more frequently belong to an ethnic group whose members are more frequently of a particular social class, which in turn may influence the probability of hospitalization. Lilienfeld and Stolley (1994, p.233) point out that condition 2 is rarely, if ever, met since different diseases usually have different probabilities of hospitalization. Exercise (good example of selecting controls to avoid Berkson’s bias) Look at the case-control study published by Zhang et al (2005) in the American Journal of Epidemiology. These researchers studied the following assocation: Breast Cancer Use of Nonsteroidal Antiinflammatory Drugs (NSAID) [includes aspirin, Ibuprofen, etc Yes No Yes (cases) No (controls) Under the section heading Controls (p.166) they state: “Controls were selected from a pool of 3,906 women aged 30-79 years with no history of cancer who had been admitted to the hospital for nonmalignant diseases that we considered unrelated to NSAID use. Eligible diagnoses included appendicities, hernia of of the abdominal cavity, and traumatic injury.” Notice they are avoiding the exposure-hospitalization association (meeting condition 1 above). Chapter 3-8 (revision 16 May 2010) p. 5 In the Discussion section, 3rd paragraph, page 169, they state: “Several characteristics of our study are noteworthy. Controls were selected from women whose diagnoses were unrelated to NSAID use. The distributions of NSAID use among subgroups of the controls were similar, suggesting the absence of selection bias.” Information Bias Information bias results when a systematic error is made in information collected on study subjects. For a categorical scale measurement, this biased information is often referred to as misclassified, since the study subject will be placed in an incorrect category. For example, a heavy smoker who is categorized as a light smoker is misclassified. Misclassification for either exposure or disease can be differential or nondifferential, which refer to the mechanism for misclassification. For exposure misclassification, the misclassification is nondifferential if it is unrelated to the occurrence or presence of disease; if the misclassification of exposure is different for those with and without disease, it is differential. Similarly, misclassification of disease is nondifferential if it is unrelated to exposure; otherwise, it is differential. Chapter 3-8 (revision 16 May 2010) p. 6 A common type of information bias is recall bias, which occurs in case-control studies where a subject is interviewed to obtain exposure information after disease has occurred. An example is a case-control study where mothers of babies with birth defects are asked to recall exposures during pregnancy, such as taking nonprescription drugs. Given the stimulus of an adverse pregnancy outcome, these mothers recall vividly what they wondered might have caused the birth defect. Mothers of normal babies, however, have no stimulus for recall and thus forget such exposures. This particular version of recall bias is known as maternal recall bias. Exercise 1) Maternal recall bias is an example of: a) b) c) d) Nondifferential misclassification of disease Differential misclassification of disease Nondifferential misclassification of exposure Differential misclassification of exposure 2) Consider the following data table. Disease (Birth Defect) Exposure (Extra Strength Tylenol used in 1st Trimester) Yes No Yes No a c b d Would we expect the OR (OR = ad/bc) to be too large or too small? Chapter 3-8 (revision 16 May 2010) p. 7 Nondifferential Misclassification With nondifferential misclassification, either exposure or disease (or both) is misclassified, but the misclassification does not depend on a person’s status for the other variable. In contrast to maternal recall bias, all people to some extent have difficulty remembering when responding to survey questions. This tends to result in non-differential misclassification. Nondifferential misclassification of a dichotomous exposure will always bias an effect, if there is one, toward the null value. (Rothman, 2002, p.100) Nondifferential misclassification in a hypothetical case-control study Correct Classification Nondifferential Misclassification 20% of 20% of No Yes No Yes* 20% of Yes No High-Fat Diet High-Fat Diet High-Fat Diet Yes No Yes No Yes No Heart attack cases 250 450 340 360 290 410 Controls 100 900 280 720 260 740 Odds Ratio 5.0 2.4 2.0 *At first this column might appear as differential misclassification. It is nondifferential, however, since both cases and controls are misclassified equally by 20%. If the exposure is not dichotomous, there may be bias toward the null value; but there may also be bias away from the null value, depending on the categories to which individuals are misclassified. In general, nondifferential misclassification between two exposure categories will make the effect estimates for those two categories converge toward one another. In contrast, differential misclassification can either exaggerate (take the effect away from the null in either the protective or deleterious direction) or underestimate an effect (take the effect towards the null). (Rothman, 2002, p.99) Exercise. Look at the article by Millard (1999). 1) Look at the possible sources of bias listed in the Data extraction section. 2) Look at the differences in RR for likely biased and likely not biased in the second half of the Main Results section. 3) Finally, look at the 2nd paragraph of Millard’ commentary, where he suggests the possibility of “diagnostic access bias”. Chapter 3-8 (revision 16 May 2010) p. 8 Exercise (design-related bias in studies of diagnostic tests) Look at the article by Lijmer et al (1999). Notice in the same pdf file that there is a correction to their DOR formula. Using the diagnostic odds ratio (DOR), which is another measure of the diagnostic accuracy of a test [see box], they show that studies of diagnostic tests are subject to design-related bias. Some Diagnostic Test Definitions With the data in the required form for Stata: Gold Standard “true value” disease present ( + ) disease absent ( - ) Test “probable value” disease present ( + ) disease absent ( - ) a (true positives) b (false negatives) c (false positives) d (true negatives) a+c b+d a+b c+d We define the following terminology, expressed as percents: sensitivity = (true positives)/(true positives plus false negatives) = (true positives)/(all those with the disease) = a / (a + b) 100 specificity = (true negatives)/(true negatives plus false positives) = (true negatives)/(all those without the disease) = d / (c + d) 100 likelihood ratio positive (LR+) = sensitivity / (1 – specificity) = odds that a positive test result would be found in a patient with, versus without, a disease likelihood ratio negative (LR-) = (1 – sensitivity) / specificity = odds that a negative test result would be found in a patient without, versus with, a disease diagnostic odds ratio (DOR) = LR+ / LR= (a/b)/(c/d) = ad/bc = odds of a positive test result in diseased persons relative to the odds of a positive test in nondiseased persons Chapter 3-8 (revision 16 May 2010) p. 9 Sensitivity Analysis Since the researcher cannot be sure of the extent that information bias has influenced the study results, if at all, a correction for the bias cannot be made in the data analysis. What clever researchers do in their papers, then, when the potential for information bias is a concern, is to present a sensitivity analysis in the Discussion section. Otherwise, the reader might just dismiss the paper altogether, being concerned that the results were too biased to draw conclusions. Greenland (1998, p.343) advises always including some level of sensitivity analysis in a scientific paper: “Potential biases due to unmeasured confounders, classification errors, and selection bias need to be addressed in any thorough discussion of study results.” Example: You are publishing a paper on a novel approach to convince patients, who visit your clinic for medical reasons, to quit smoking. You randomly present the approach to half of your patients, and then find that 15% of your study group quit smoking, compared to 0% of the control group (RR=0.85, p<0.001). You recognized that the reader may think this result is too good to be true, the reader perhaps wondering if some large part of the 15%, perhaps 14.5%, of the study group was lying to you. A clever thing to include in your paper is a sensitivity analysis, reporting the RR that would have been the result if 5% lied and if 10% lied. How to conduct a sensitivity analysis is presented in it’s own chapter of this course manual. Confounding A simple definition of confounding would be the confusion, or mixing, of effects This definition implies that the effect of the exposure is mixed together with the effect of another variable, leading to a bias. We call such variables confounding variables or confounders. Chapter 3-8 (revision 16 May 2010) p. 10 Example of Confounding Affected Babies per 1000 Live Births Rothman (2002, p.101) provides a classic example of confounding, which is the relation between birth order and the occurrence of Down syndrome. 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 1 2 3 Birth Order 4 5+ These data suggest that the prevalence of Down syndrome is associated with (perhaps “causally”) birth order. The effect of birth order, however, is a blend of whatever effect birth order has by itself and the effect of another variable that is closely correlated with birth order. Affected Babies per 1000 Live Births The other variable is the age of the mother. 9 8 7 6 5 4 3 2 1 0 <20 20-24 25-29 30-34 35-39 Maternal Age 40+ This figure gives the relation between mother’s age and the occurrence of Down syndrome from the same data. Chapter 3-8 (revision 16 May 2010) p. 11 1.8 Affected Babies per 1000 Live Births Affected Babies per 1000 Live Births It indicates a much stronger relationship (8.5 per 1000 for the highest category vs 1.7 per 1000 for the highest category from the previous graph—notice difference in scale of Y axis). 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 1 2 3 Birth Order 4 5+ 9 8 7 6 5 4 3 2 1 0 <20 20-24 25-29 30-34 35-39 Maternal Age 40+ Because birth order and the age of the mother are highly correlated, we can expect that mothers giving birth to their fifth baby are, as a group, considerably older than mothers giving birth to their first baby. Thus, the birth order effect is mixed with the mother’s age effect. We call this mixing of effects confounding. In this example, the birth order effect is confounded with maternal age. We can resolve this confounding by consider both effects simultaneously. Chapter 3-8 (revision 16 May 2010) p. 12 In this graph, we see a striking trend with maternal age (cases increase for each birth order), while there is no trend with birth order (cases essentially constant for each maternal age). This last graph is an example of a stratified display, in contrast to the previous graphs which are examples of a crude display. The stratified display reveals that the erroneous birth order effect was due to confounding with maternal age. Chapter 3-8 (revision 16 May 2010) p. 13 Properties of a confounding factor A confounding factor must have an effect on disease and it must be imbalanced between the exposure groups to be compared. That is, a confounding factor must have two associations: 1) A confounder must be associated with the disease 2) A confounder must be associated with exposure. Diagrammatically, the two necessary associations for confounding are: Confounder association association Exposure Disease confounded effect There is also a third requirement. A factor that is an effect of the exposure and an intermediate step in the causal pathway from exposure to disease will have the above associations, but causal intermediates are not confounders; they are part of the effect that we wish to study. Thus, the third property of a confounder is as follows: 3) A confounder must not be an effect of the exposure. Rothman (2002, p.164) points out that the degree of confounding is not dependent upon statistical significance, but rather upon the strength of the associations between the confounder and both exposure and disease. He advocates that a better way to evaluate confounding is to statistically control for the potential confounder, using stratification or regression analyses, and determine whether the unconfounded result (adjusted model) differs from the potentially confounded result (the unadjusted model). If they differ, then confounding is present in the unadjusted model. Chapter 3-8 (revision 16 May 2010) p. 14 Example (Stoddard’s hypothetical data) A researcher wants to study the association of a diet high in pizza and the disease outcome skin acne. The following data are collected using a cross-sectional study design: Eat Pizza (at least twice per month) Noticeable Skin Acne Yes No Yes 80 40 No 120 149 Odds Ratio = 2.48 , p < 0.001 Thus, an association between pizza consumption and skin acne is observed in this crude analysis. One might consider, however, if the association is confounded by age. The data are stratified and these are the results: Teenager (age 13-19) Eat Pizza at least twice per month Noticeable Skin Acne Yes No Yes 60 10 No 20 4 Odds Ratio = 1.20 , p = 0.78 Adult (age 20+) Eat Pizza at least twice per month Noticeable Skin Acne Yes No Yes 20 30 No 100 145 Odds Ratio = 0.97 , p = 0.91 Now we see that the association does not hold in either age group. Since the crude display gives a different result than the stratified display of the data, the result is confounded. That is, the pizza-acne association observed in the crude display is confounded by age. Chapter 3-8 (revision 16 May 2010) p. 15 Filling in the labels to our diagram, Confounder association association Exposure Disease confounded effect we have Age association association Pizza Acne confounded effect Chapter 3-8 (revision 16 May 2010) p. 16 There is an association between age and pizza, as 80/94, or 85%, of teenagers eat pizza regularly, while 120/295, or 41%, of adults eat pizza regularly. This “imbalance” represents an “association” (as age increases, pizza eating decreases). Age association 85% vs 41% association Pizza Acne confounded effect There is a well-known assocation between age and acne, acne being mostly a teenager disease. In these data, 70/94, or 74%, of children have acne, while only 50/295, or 17%, of adults have acne. This imbalance represents an association (as age increases, acne decreases). Age association 85% vs 41% association 74% vs 17% Pizza Acne confounded effect We therefore have the two assocations that must be present for age to be a confounder 1) A confounder must be associated with the disease (age-acne association) 2) A confounder must be associated with exposure (age-pizza association) We also satisfied the third property of a confounder: 3) A confounder must not be an effect of the exposure. (This holds, since eating pizza does not cause a person to be a teenager.) This example illustrates what is usually the case in a confounded relationship. What we consider to be the putative cause of the disease is simply a surrogate for something else, the confounder, that is the cause, or more directly related to the cause, of disease. That is, pizza is a surrogate for teenager, since teenagers are the big pizza consumers; and something that occurs in the teenage years produces acne, perhaps related to rising hormonal levels. Chapter 3-8 (revision 16 May 2010) p. 17 Control of Confounding Confounding is a systematic error that investigators aim either to prevent or to remove from a study. There are two common methods to prevent confounding. One of them, randomization, or the random assignment of subjects to experimental groups, can only be used in experiments. Randomization produces study groups with nearly the same distribution of characteristics (creates balance), and so removes one of the two required associations present in confounding (the exposure-confounder association). The other, restriction, involves selecting subjects for a study who have the same value, or nearly the same value, for a variable that might be a confounder, thus achieving balance. Rothman (2002, pp.20,110-111) provides a very convincing argument to support restriction. Basically, the arguments proceeds as follows. The idea of “representativeness” comes from survey sampling, where the sample is supposed to look like the general population. So, when researchers restrict their sample to a more limited group of subjects, there is an impression in many people’s minds that this is a bad thing, where they think the study results do not generalized to the population as a whole. However, the idea is not to achieve representativeness, as a survey sample does. The idea is to test a theory, which can only be done if confounding is removed. By restriction, confounders are removed because the subjects have the same value on the confounder (balance on the confounder). This takes the study closer to the conterfactual ideal, so that a causal inference is more tenable. Matching A third method to prevent confounding, matching, is deferred to the Regression Models course, where the analysis of a matched study is discussed in the conditional logistic regression chapter. Sometimes it works better than just using regression models to control for confounding, while other times it performs worse. For example, in a case-control study, matching can introduce confounding into the study when there was none to begin with. (Rothman and Greenland, p. 151) Chapter 3-8 (revision 16 May 2010) p. 18 Exercise In the Rauscher (2000) article: 1) Notice the use of restriction in the fourth line of the Abstract. 2) Look at the last two sentences of the Selection of Controls section. Notice the use of interview type matching to minimize recall bias. 3) Look at Table 2: a) Notice how Table 2 is a collection of two-way stratifications (something analogous to the two-way histogram we used in the Down syndrome example above). b) Notice one result this approach produced (reported in the BMI As a Categorical Variable section). 4) Look at the 2nd paragraph of the 2nd column of the Discussion section. Notice they discuss a sensitivity analysis to account for bias (although their actual sensitivity analysis was poorly presented). Chapter 3-8 (revision 16 May 2010) p. 19 Stata Exercise Evans County Dataset (evans.xls) Data are from a cohort study in which n=609 white males were followed for 7 years, with coronary heart disease as the outcome of interest. Codebook n = 609 outcome chd coronary heart disease (1=presence, 0=absence) predictors cat catecholamine level (1=high, 0=normal) age age in years (continuous) chl cholesterol (continuous) smk smoker (1=ever smoked, 0=never smoked) ecg electrocardiogram abnormality (1=presence, 0=absence) dbp diastolic blood pressure (continuous) sbp systolic blood pressure (continuous) hpt high blood pressure (1=presence, 0=absence) defined as: DBP 160 or SBP 95 We will use the evans.dta dataset to illustrate confounding. File Open Find the directory where you copied the course CD Change to the subdirectory datasets & do-files Single click on evans.dta Open use "C:\Documents and Settings\u0032770.SRVR\Desktop\ Biostats & Epi With Stata\datasets & do-files\evans.dta", clear * which must be all on one line, or use: cd "C:\Documents and Settings\u0032770.SRVR\Desktop\” cd “Biostats & Epi With Stata\datasets & do-files" use evans.dta, clear Recall, these data were produced using a cohort study design, so the risk ratio is an appropriate measure of effect. We will illustrate with the odds ratio, however, so we can compare the results to a logistic regression approach. The exercise is to see if the smoking-CHD association is confounded by age in this dataset. Chapter 3-8 (revision 16 May 2010) p. 20 First fitting a univariable (one predictor variable) logistic regression, Statistics Binary outcomes Logistic regression (reporting odds ratios) Model tab: Dependent variable: chd Independent variables: smk OK logistic chd smk Logistic regression Log likelihood = -216.40647 Number of obs LR chi2(1) Prob > chi2 Pseudo R2 = = = = 609 5.75 0.0165 0.0131 -----------------------------------------------------------------------------chd | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------smk | 1.955485 .5708642 2.30 0.022 1.103477 3.465337 ------------------------------------------------------------------------------ we see that there is a significant smoking-CHD association (OR=1.96, p=0.022). Next, fitting a multivariable (two or more predictor variables) logistic regression, adjusting for the continuous variable age, Statistics Binary outcomes Logistic regression (reporting odds ratios) Model tab: Dependent variable: chd Independent variables: smk age OK logistic chd smk age Logistic regression Log likelihood = -209.31516 Number of obs LR chi2(2) Prob > chi2 Pseudo R2 = = = = 609 19.93 0.0000 0.0454 -----------------------------------------------------------------------------chd | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------smk | 2.30354 .6892501 2.79 0.005 1.281459 4.140823 age | 1.052105 .0141545 3.78 0.000 1.024725 1.080216 ------------------------------------------------------------------------------ Chapter 3-8 (revision 16 May 2010) p. 21 Checking to see if the adjusted effect changed the unadjusted effect by more than 10%, display 1.96*1.10 2.156 We see that the adjusted odds ratio (adjusted OR=2.30) differs by more than 10% from the unadjusted odds ratio (unadjusted OR= 1.96 × 1.1 = 2.16), so by definition, the unadjusted smoking-CHD assocation was confounded by age. Here we are using the “10% change in estimate” variable selection rule that has been proposed for determining if the putative confounder needs to be adjusted for in a regression model. (see box) “10% change in estimate” variable selection rule A variable selection rule consistent with this definition of confounding is the change-in-estimate method of variable selection. In this method, a potential confounder is included in the model if it changes the coefficient, or effect estimate, of the primary exposure variable by 10%. This method has been shown to produce more reliable models than variable selection methods based on statistical significance [Greenland, 1989]. In practice, we could stop here. For illustration, however, lets assess confounding by the “two associations” definition. We could compute odds ratios between 1) smoking and age, and between 2) age and CHD, to test for the associations. More quickly, we can just use Pearson correlation coefficients, another way to test for associatons. Statistics Summaries, tables, & tests Summary and descriptive statistics Pairwise correlations Main tab: Variables: chd age smk Options: Print significance level for each entry OK pwcorr chd age smk, sig | chd age smk -------------+--------------------------chd | 1.0000 | | age | 0.1393 1.0000 | 0.0006 | smk | 0.0944 -0.1391 1.0000 | 0.0198 0.0006 Chapter 3-8 (revision 16 May 2010) p. 22 We see that both assocations are present, indicating confounding by the “two assocations” definition. Note: In epidemiologic studies, effect measures such as the odds ratio are preferred to ordinary correlation coefficients. Although we can see that the association is significant, from examining the p values, it is difficult to get a feel for the size of the effect from the correlation coefficient. The odds ratio, on the other hand, provides such a feel. To investigate the association further, let’s next try a “stratified” analysis, where we are testing the smoking-CHD association within homogenous age subgroups, or age strata. First, creating a categorical age variable, recode age 40/49=1 50/59=2 60/69=3 70/76=4, gen(agecat) Then, requesting a logistic regression for each of these age categories, Statistics Binary outcomes Logistic regression (reporting odds ratios) Model tab: Dependent variable: chd Independent variables: smk by/if/in tab: Repeat command by groups Variables that define groups: agecat OK by agecat, sort : logistic chd smk * <or> bysort agecat: logistic chd smk Chapter 3-8 (revision 16 May 2010) p. 23 -> agecat = 1 Logistic regression Log likelihood = -67.277892 Number of obs LR chi2(1) Prob > chi2 Pseudo R2 = = = = 247 4.33 0.0375 0.0311 -----------------------------------------------------------------------------chd | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------smk | 3.849057 2.922623 1.78 0.076 .8690193 17.04823 ------------------------------------------------------------------------------> agecat = 2 Logistic regression Log likelihood = -62.733206 Number of obs LR chi2(1) Prob > chi2 Pseudo R2 = = = = 203 5.19 0.0227 0.0398 -----------------------------------------------------------------------------chd | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------smk | 3.675676 2.368044 2.02 0.043 1.039807 12.99336 ------------------------------------------------------------------------------> agecat = 3 Logistic regression Log likelihood = -58.821111 Number of obs LR chi2(1) Prob > chi2 Pseudo R2 = = = = 115 0.17 0.6809 0.0014 -----------------------------------------------------------------------------chd | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------smk | 1.208081 .5559709 0.41 0.681 .4901901 2.977333 ------------------------------------------------------------------------------> agecat = 4 Logistic regression Log likelihood = -17.466348 Number of obs LR chi2(1) Prob > chi2 Pseudo R2 = = = = 44 3.63 0.0569 0.0940 -----------------------------------------------------------------------------chd | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------smk | 6.333333 7.15092 1.63 0.102 .6927028 57.90522 ------------------------------------------------------------------------------ In Chapter 10, we will see how to do a more sophisticated stratified analysis, where the agespecific ORs are combined into a single summary measure. Chapter 3-8 (revision 16 May 2010) p. 24 References Delgado-Rodriquez M, Llorca J. (2004). Bias. J Epidemiol Community Health 58:635-641. Greenland S. (1989). Modeling and variable selection in epidemiologic analysis. Am J Public Health 79(3):340-349. Greenland S. (1998). Chapter 19. Basic methods for sensitivity analysis and external adjustment. In, Rothman KJ, Greenland S. Modern Epidemiology, 2nd ed. Philadelphia PA, Lippincott-Raven Publishers, pp. 343-357. Grimes DA, Schulz KF. (2002). Bias and causal associations in observational research. Lancet 359:248-52. Kraus AS. (1954). The use of hospital data in studying the association between a characteristic and a disease. Pub Health Rep 69:1211-1214. Last JM. (1995). A Dictionary of Epidemiology. 3rd ed. New York, Oxford University Press. Lilienfeld DE, Stolley PD. (1994). Foundations of Epidemiology, 3rd ed, New York, Oxford University Press. Millard PS. (1999). Review: bias may contribute to association of vasectomy with prostate cancer. West J Med 171:91. Rauscher GH, Mayne ST, Janerich DT. (2000). Relation between body mass index and lung cancer risk in men and women never and former smokers. Am J Epidemiol 152:506-13. Rothman KJ. (2002). Epidemiology: An Introduction. New York, Oxford University Press. Rothman KJ, Greenland S. (1998). Modern Epidemiology, 2nd ed. Philadelphia PA, Lippincott-Raven Publishers. Vandenbrouchke JP, von Elm E, Altman DG, et al. (2007). Strengthening and reporting of observational studies in epidemiology (STROBE): explanation and elaboration. Ann Intern Med 147(8):W-163 to W-194. Chapter 3-8 (revision 16 May 2010) p. 25