Slide 1: Introduction Issues of of Interpretation in in Epidemiologic Studies Developed through APTR Initiative to Enhance Prevention and Population Health Developed Education through in collaboration APTR Initiative with to Brody Enhance School Prevention of Medicine and Population at East Carolina Health University Education in with collaboration funding from with Centers Brody for School Disease of Medicine Control and at East Prevention Carolina University with funding from Centers for Disease Control and Prevention Hello and welcome to Module 2 Fundamentals of Epidemiology. This presentation will focus on issues of interpretation in epidemiologic studies. My name is Jeffrey Bel, Assistant Professor of Epidemiology at East Carolina University, Brody School of Medicine Slide 2: Acknowledgements APTR APTR wishes wishes to to acknowledge following following individual individual that that developed developed this this module: module: Jeffrey Jeffrey Bel, Bel, PhD PhD Department of Public Health Department of Public Health Brody School of Medicine at East Carolina University Brody School of Medicine at East Carolina University This education module is made possible through Centers for Disease Control and Prevention (CDC) and Association This education for Prevention module is Teaching made possible and Research through (APTR) Centers Cooperative for Disease Agreement, Control and No. Prevention 5U50CD300860. (CDC) and The module represents Association for opinions Prevention of Teaching author(s) and and Research does not (APTR) necessarily Cooperative represent Agreement, views of No. 5U50CD300860. Centers for Disease The module Control represents and Prevention opinions or of Association author(s) and for Prevention does not necessarily Teaching and represent Research. views of Centers for Disease Control and Prevention or Association for Prevention Teaching and Research. Slide 3: Presentation Objectives 1. 1. Describe key key features of of selection and and information bias bias 2. 2. Identify ways ways selection and and information bias bias can can be be minimized or or avoided 3. 3. Implement methods for for assessing and and controlling confounding 4. 4. Identify uses uses of of Surgeon General s Guidelines for for establishing causality for establishing causality. There are four objectives for this presentation. They are to describe key features of information and selection bias; identify ways selection and information bias can be minimized or avoided; Implement methods for assessing and controlling confounding; and identify uses of Surgeon General s Guidelines 1
Smith, AH. The Epidemiologic Research Sequence. 1984 Smith, AH. The Epidemiologic Research Sequence. 1984 Slide 4: Hierarchy of Study Designs The goal of public health as well as clinical medicine is to modify natural history of disease, to decrease morbidity and mortality. So how do researchers and public health practitioners do this? They carry out formal studies to identify risk factors for various diseases or or health outcomes. So here is an informative figure showing hierarchy of various epidemiologic study designs. The goal of all se studies is same and that is to identify exposures or characteristics that are associated with disease or or health related outcomes. Slide 5: Linking Exposure to Outcome Exposure or or Characteristic Disease or or Outcome Observed Observed Association Association Here is a figure depicting this goal of linking exposures to outcomes. The result of se formal epidemiologic studies is observed associations. That is associations between exposures or characteristics and diseases or outcomes. Exposure or or Characteristic Disease or or Outcome Observed Observed Association Association Is Is it: it: biased, biased, confounded, or or causal? causal? Slide 6: Linking Exposure to Outcome After study is carried out, and a disease is linked to an exposure or outcome, you need to ask if observed association is biased confounding or causal. We re going to go through each one of se items individually. And we re going to start with bias first. 2
Slide 7: Bias Any systematic error in in design, conduct, or or analysis of of a study that results in in a mistaken estimate of of an an exposure s effect on on risk risk of of disease So bias is any systematic error in design, conduct, or analysis of a study that results in a mistaken estimate of an exposure s effect on risk of disease. Slide 8: Bias Can Can create spurious association when when re re really really is is Now se errors or bias can make it appear as if not not one one (bias (bias away away from from null) null) an exposure is associated with an outcome when Can Can mask mask an an association when re re really really is is one one (bias (bias towards null) null) exposure and outcome are really not Primarily introduced by by investigators or or study study participants associated. That is also thought of as biasing Selection and and information bias bias away from null. The errors in bias can also mask an association where re really is an association. And that is called biasing towards null. There really is an association however a bias estimate makes it appear as though re is no association. As we talked about in definition of bias, bias is primarily introduced by investigators or participants, in design, conduct, or analysis of study. And we re going to talk about two forms of bias in a study; selection and information bias. Slide 9: Selection Bias Selection bias occurs when methods that Results from from procedures used used to to study study participants that that lead lead to to a result different from from what what would have have been been obtained from from entire population targeted investigators use to identify study participants for for study study results in a mistaken estimate. That is Systematic error error made in in selecting one one or or more more of of study study groups to to be be compared estimate that investigators generate is different than what would have been obtained from entire population targeted for study. Again this gets back to a systematic error made by investigators. Specifically error was made in selecting one or more of study groups that are being compared in epidemiologic study. 3
Slide 10: Selection Bias Example Control selection bias bias Self-selection bias bias Differential referral, surveillance, or or diagnosis bias bias Loss Loss to to follow-up And we re going to talk about four types of selection bias. Control selection bias, selfselection bias, differential referral, surveillance, or diagnosis bias, and loss to follow-up. Slide 11: Selection Bias in a Case-Control Study Question: Question: Do Do Pap Pap smears smears prevent prevent cervical cervical cancer? cancer? Cases Cases diagnosed diagnosed at at a a city city hospital. Now here is an example in which selection bias hospital. Controls Controls randomly randomly sampled sampled from from household household in in same same city city by by canvassing canvassing neighborhood neighborhood on on foot. foot. Here Here is is observed observed relationship: relationship: Cases Cases Controls Controls may occur in a case-control study. Specifically Pap Pap Smear Smear 100 100 100 100 control selection bias. So research question No No Pap Pap 150 150 150 Smear 150 Smear was, do Pap smears prevent cervical cancer? So Total Total 250 250 250 250 OR OR = = (100 (100 x x 150) 150) //(100 (100 x x 50) 50) = = 1.0 1.0 There There is is no no association association between between Pap Pap in this study cases were diagnosed at a city smears smears and and risk risk of of cervical cervical cancer cancer (40% (40% of of cases cases and and 40% 40% of of controls controls had had Pap Pap smears) smears) hospital. Controls were randomly sampled from households in same city by canvassing neighborhood on foot. So here was observed association. Here are data from that hypotical study. Women who were cases and controls, and women who received Pap smears and no Pap smears. So 250 cases, 250 controls, 100 of cases received a Pap smear and 100 of controls received a Pap smear. The resulting odds ratio was 1.0. That is re is no association between Pap smears and risk of cervical cancer. Slide 12: Selection Bias in a Case-Control Study However, cases came from hospital and Recall: Cases from from hospital and and controls come come from from neighborhood around hospital controls from neighborhood around Now Now for for bias: bias: Only Only controls who who were were at at home during recruitment for for study study were were actually included hospital. So here is where bias comes in. in in study. Women at at home were were less less likely likely to to work work and and less less likely likely to to have have regular checkups and and Pap Pap smears. Only controls who were at home during Therefore, being included in in study study as as a control is is not not independent of of exposure. recruitment for study were actually involved in study. Now se controls who were home while neighborhoods were being canvassed were less likely to work and less likely to have regular checkups and Pap smears. Therefore, inclusion in study as a control was not independent of exposure. That is that being included in study as a control was dependent on exposure, in that case Pap smear. 4
Slide 13: Selection Bias in a Case-Control Study Question: Question: Do Do Pap Pap smears smears prevent prevent cervical cervical cancer? cancer? Cases Cases diagnosed diagnosed at at a a city city hospital. So here are data for true relationship. hospital. Controls Controls randomly randomly sampled sampled from from household household in in same same city city by by canvassing canvassing neighborhood neighborhood on on foot. foot. Here Here is is true true relationship: relationship: Cases Cases Controls Controls Again, 250 cases, 250 controls. 100 of cases Pap Pap Smear Smear 100 100 150 150 received a Pap smear, 150 of controls No No Pap Pap 150 150 100 Smear 100 Smear received a Pap smear. The resulting odds ratio is Total Total 250 250 250 250 OR OR = = (100)(100) (100)(100)/ /(150)(150) = =.44.44 56% 56% reduced reduced risk risk of of cervical cervical cancer cancer 0.44. That is re is a 56% reduced risk of among among women women who who had had Pap Pap smears smears as as compared compared to to women women who who did did not not (40% (40% of of cases cases had had Pap Pap smears smears versus versus 60% 60% of of controls) controls) cervical cancer among women who received a Pap smear. This is true relationship. This is unbiased estimate. The odds ratio of 1.0 was biased estimate. Slide 14: Selection Bias in a Case-Control Study Anor form of selection bias is self-selection Refusal or or nonresponse by by participants that that is is related to to both both exposure and and disease bias. This occurs when refusal or non-response e.g. e.g. if if exposed exposed cases cases are are more/less more/less likely likely to to participate participate than than participants participants in in or or categories categories to being included in study is related to both Best Best way way to to avoid avoid is is to to obtain high high participation rates rates exposure and disease. For example, if exposed cases were more or less likely to participate than participants in or categories that can result in self-selection bias. The people, eir cases or controls, were more or less likely to participate. Now best way to avoid self-selection bias is to maintain high participation rates in all groups or categories of participants in study. Slide 15: Selection Bias in a Case-Control Study Now here is differential surveillance, diagnosis or Example related to to exposure CC CC study: venous thromboembolism (VT) (VT) and and oral oral referral bias. So here is an example related to contraceptive (OC) (OC) use use Cases: 20-44 yo, yo, hospitalized for for VT VT exposure. In a case-control study looking at Controls: 20-44 yo, yo, hospitalized for for acute acute illness or or elective venous thromboembolism (VT) and oral surgery at at same same hospitals Result: OR OR = 10.2 10.2 contraceptive (OC) use, cases were 20-44 year old women hospitalized for VT. Controls were same age group hospitalized for acute illness or elective surgery at same hospitals. In this instance hospital controls were used. The resulting odds ratio from this study was 10.2. That is OC use was highly associated with venous thromboembolism. 5
Slide 16: Selection Bias in a Case-Control Study However, authors acknowledged that this high Authors acknowledged high high OR OR might be be due due to to bias bias in in criteria for for hospital admission odds ratio might be due to bias in criteria for Previous studies linked VT VT to to OC OC hospital admission. That is who was admitted. Health care care provider more more likely likely to to hospitalize women with with VT VT symptoms who who were were taking OC OC than than symptomatic So previous studies had liked VT to OC use. women who who were were not not taking OC OC Therefore re was a tendency to hospitalize patients based on exposure status of oral contraceptive use and that led to a stronger association between VT and OC than what truly existed. Slide 17: Selection Bias in a Cohort Study Here s an example in form of selection bias in Compared HIV HIV incidence rates rates among IVDU IVDU in in NYC NYC from from 1992-97 through 10 10 incidence studies to to a cohort study that is loss to follow-up. One previous years years HIV HIV incidence rates rates (IR) (IR) range from from 0 to to 2.96 2.96 per per 100 100 study compared HIV incidence rates among IV person-years (py) (py) Well Well below IR IR in in NYC NYC from from late late 70s 70s and and early early 80s 80s (13 (13 drug users (IVDU) in New York City during a six per per 100 100 py) py) to to mid mid 80s 80s and and early early 90s 90s (4.4 (4.4 per per 100 100 py) py) year period using 10 incidence studies. The HIV Resulted in in funding cuts cuts to to drug drug treatment and and prevention programs rates ranged from 0 to 2.96 per 100 personyears. And se were well below incidence rates in NYC from late 70s and early 80s to mid-80s and early 90s. Now this study resulted in funding cuts to drug treatment and prevention programs because this study showed that HIV incidence rates were much lower than y were in 70s and 80s. Slide 18: Selection in a Cohort Study Now question is was this decline in HIV Was Was decline real? real? Follow-up rates rates in in 10 10 cohorts ranged from from 36% 36% to to incidence rates real? Closer inspection of se 95% 95% Only Only 22 reported reported >80% >80% 10 individual cohort studies found that follow-up Sample size size ranged from from 96 96 to to 1,671 rates ranged from 36% to 95% and only 2 of Solution: minimize loss loss to to follow-up 10 had follow-up rates greater than 80%. As we know, loss to follow-up can introduce bias into study. In this case only 2 of 10 had a high follow-up rate. Also, sample sizes ranged from 96 to 1,671 in se 10 studies. Now solution to loss to follow-up is to simply minimize loss to follow-up. That involves implementing methods and procedures in your study design to minimize loss to followup. You could maintain accurate contact information for participants, and contact information for people who will know where your participants are so as you do not lose m. 6
Slide 19: Selection Bias Can We Fix It? So selection bias, can we fix it? Can we do No No need need to to avoid avoid it it when when you you design design and and conduct conduct study study anything about it? The simple answer is no, you For For example simply need to put in measures into your design Use Use same same criteria criteria for for selecting cases cases and and controls controls Obtaining all all relevant participant records records and conduct study to avoid it in first Obtaining high high participation rates rates Taking Taking in in account diagnostic and and referral referral patterns of of place. For example you need to use same disease disease criteria for selecting cases and controls. Obtain all relevant participant records. That is, maintain complete data on everyone. Obtain high participation rates. That is, minimize loss to follow-up. And you need to take into account diagnostic and referral patterns of disease of interest, or diseases that you are using for your hospital controls. Slide 20: Information Bias The second major type of bias we re going to talk Arises from from a systematic difference in in way way that that exposure or or outcome is is measured between groups about is information bias. That arises when from Can Can bias bias towards or or away away from from null null a systematic difference in way that exposure Occurs in in prospective and and retrospective studies Includes recall recall bias bias and and interviewer bias bias or outcome is measured between various groups under investigation. Just like selection bias, information bias can bias your results towards or away from null. That is, mask an apparent association or create a spurious one. Information bias can occur in prospective as well as retrospective studies. And two types we re going to talk about include recall and interviewer bias. Slide 21: Recall Bias in a Case-Control Study Here s an example of recall bias in a case-control Case-control study study of of birth birth defects Controls: healthy infants study. In this example it is a case control study Cases: malformed infants examining birth defects. Controls in this study Exposure data data collected at at postpartum interviews with with infants mors were healthy infants. Cases were infants with Controls or orcases may may have have underreported exposure, depending birth defects. Exposure data was collected at on on nature of of exposure postpartum interviews with infants mors. Now in this situation re could be a differential level in accuracy of information provided by cases and controls which could result in an over or underestimate of measure of association. So data collected from cases could be more accurate than data collected from controls. That is mors of cases have been pouring over ir brains for history of various exposures during pregnancy which may have contributed to ir infants being malformed. Whereas controls, y have healthy infants, y haven t really spent a lot of time reviewing ir history during pregnancy. Or conversely, cases may underreport 7
ir history particularly if it is a socially sensitive exposure such as alcohol consumption during pregnancy. Slide 22: Methods to Minimize Recall Bias Now what are methods to minimize recall Select diseased control group Design structured questionnaire bias? In context of a case-control study you Use Use self-administered questionnaire could use a diseased control group. So as Use Use biological measurements participants in diseased control group have Mask Mask participants to to study study hyposes some disease unrelated to outcome of interest could also be pouring over ir history of exposures. You could use and design a structured questionnaire. You could use a self-administered questionnaire so as to make it easier for participants to recall sensitive exposures such as consuming alcohol or smoking during pregnancy. You could bypass participant in terms of self-report and use biological measurements. It is tough to get around socially sensitive exposures. And lastly you could mask participants to study hyposes. So as to illicit more accurate information from study participants. Slide 23: Interviewer Bias The second type of information bias we re going Systematic difference in in soliciting, recording, interpreting information to talk about is interviewer bias. That results Case-control study: exposure information is is sought when outcome is is known from a systematic difference in soliciting, Cohort study: outcome information is is sought when when exposure is is known recording, and interpreting of data collected Solutions: Mask Mask interviewers, use use standardized questionnaires or or standardized methods of of outcome during a study. In a case-control study this could (or (or exposure) ascertainment occur when exposure information is sought or collected differently between cases and controls. In a cohort study, outcome or disease information is collected differently between exposed and unexposed individuals. There is a simple solution to this and that is to simply mask interviewers or people collecting data as to disease status or exposure status of participants. You can also use standardized questionnaires or standardized methods of outcome or exposure ascertainment. 8
Slide 24: Association Exposure or or Characteristic Disease or or Outcome Observed Observed Association Association Is Is it: it: biased, biased, confounded, or or causal? causal? So that covers examination of wher our observed association is biased or not. The next question we re going to ask about our observed association is wher this association is confounded or not. Slide 25: Confounding X Here is a diagram of a possibly confounded relationship between A and B, confounded by A B variable X. Confounding is really a mixing of A mixing of of effects association between exposure and and disease is is distorted by by effect of of a third third variable that that is is associated with with disease Alternate explanation for for observed association between an an exposure and and disease effects. That is association between exposure and disease of interest is distorted by effect of a third variable that is also associated with disease as well as exposure. You can also think about confounding as an alternate explanation between A and B in this diagram. That is an alternate explanation of observed association between exposure of interest A and disease of interest B. Slide 26: Criteria for Confounding X There are three criteria for determining wher a variable is confounding relationship A B In between A and B. So in order for this factor or In order order for for a factor factor (X) (X) to to be be a confounder, all all of of following must must be be TRUE: TRUE: variable X to be a confounder, all of following Factor Factor X is is associated associated with with Disease Disease B (risk (risk factor factor or or preventive preventive factor) factor) must be true. This factor must be associated Factor Factor X is is associated associated with with Factor Factor A (exposure) (exposure) Factor Factor X is is not not a a result result of of Factor Factor A (not (not on on causal causal pathway) pathway) with Disease B. That is it s a risk factor or a preventive factor. This factor must also be associated with Factor A. That is exposure. Finally this factor must not be a result of Factor A. That is, it must not lie on causal pathway from A to B. Or it may not be a result of A. 9
Slide 27: Example of Confounding Here s an example of a confounding relationship Smoking Smoking in a previous study that was trying to link coffee consumption with pancreatic cancer. And y Coffee Coffee Pancreatic Pancreatic Consumption Cancer Cancer thought that perhaps smoking was confounding Smoking Smoking is is associated associated with with pancreatic pancreatic cancer cancer Smoking Smoking is is associated associated with with coffee coffee drinking drinking relationship between coffee consumption Smoking Smoking is is not not a a result result of of coffee coffee drinking drinking and pancreatic cancer. So y went through three criteria to see if smoking was a confounder. Criterion 1: is this confounding variable, smoking, associated with disease of interest or outcome, pancreatic cancer? And y found in ir data that yes, it was. Smoking was a risk factor for pancreatic cancer. Next, was smoking associated with exposure or coffee consumption? And y found in ir data that yes, people consuming coffee were more likely to also smoke. And looking outside ir data, y knew that smoking was not a result of coffee drinking. That smoking is not on causal pathway from coffee consumption to pancreatic cancer. All three criteria were satisfied refore y concluded that smoking was indeed confounding relationship between coffee consumption and pancreatic cancer. It was an alternate explanation between coffee consumption and pancreatic cancer. Slide 28: Impact of Confounding So what is impact of confounding? It is Pulls Pulls observed association away away from from true true association similar to bias. You can pull association away Positive confounding from true association in eir direction. So Exaggerates Exaggerates true true association association True True relative relative risk risk (RR) (RR) = = 1.0 1.0 and and confounded confounded RR RR = = 2.0 2.0 in positive confounding confounded estimate Negative confounding or confounded variable exaggerates true Hides Hides true true association association True True RR RR = = 2.0 2.0 and and confounded confounded RR RR = = 1.0 1.0 association. So in context of a cohort study, true relative risk is 1. That is exposure is not associated with outcome. However, confounded relative risk could be 2. It is exaggerating true association. There is also negative confounding when it hides true association. For example, in a cohort study true relative risk could be 2. That is exposure is related to outcome. However, due to confounding, relative risk is 1. There is no association between exposure of interest and disease. Again, it is similar to bias. 10
Dementia Dementia No No Dementia Dementia TOTAL TOTAL Slide 29: Hypotical Cohort Study of Obesity and Dementia Obese Obese 400 400 600 600 1,000 1,000 So now let s work through a hypotical cohort study examining association between obesity Not NotObese 100 100 900 900 1,000 1,000 and dementia. So a hypotical study was TOTAL TOTAL 500 500 1,500 1,500 2,000 2,000 conducted among 2,000 people looking at obesity and dementia. Here are results of Relative Relative Risk Risk = 400/1,000 = 400/1,000 = = 4.0 4.0 (crude (crude measure) 100/1,000 measure) 100/1,000 data. We calculated relative risk of 4.0. That is incidence of dementia among obese participants was 4 times incidence of dementia in non-obese participants. So obesity was associated with dementia. Slide 30: Association Between Obesity and Dementia So now we should ask ourselves, was age confounding association between obesity and dementia? Slide 31: Dementia Diabetes Cohort Study Dementia Dementia TOTAL TOTAL Age Age Yes Yes No No 80-99 80-99 Years Years 400 400 600 600 1,000 1,000 45-79 45-79 Years Years 100 100 900 900 1,000 1,000 TOTAL TOTAL 1,000 1,000 1,000 1,000 2,000 2,000 Relative Relative Risk Risk = = 400/1,000 400/1,000 = = 4.0 100/1,000 4.0 100/1,000 So we can look at our data and look at Criterion 1 first: Is age associated with our outcome, dementia? We can summarize our data in a 2x2 table and calculate relative risk and we get 4.0. That is age, and in this case older age, was associated with dementia. 11
Slide 32: Dementia and Diabetes Cohort Study Obese Obese TOTAL TOTAL Age Age Yes Yes No No 80-99 80-99 Years Years 900 900 100 100 1,000 1,000 45-79 45-79 Years Years 100 100 900 900 1,000 1,000 TOTAL TOTAL 1,000 1,000 1,000 1,000 2,000 2,000 900/1000 Relative 900/1000 Relative Risk Odds ratio Risk = = 900 x Odds ratio = 900= 9.0 100/1000 9.0 100/1000 = 81.0 81.0 100 x 100 Criterion 2: Was age, potential confounder, associated with our outcome obesity? Here are resulting data. In examination of this association we found out that yes, age is highly associated with obesity. Age Slide 33: Were Criteria for Confounding Satisfied? Obesity Dementia So were criteria for confounding satisfied? Criterion 1: Age was associated with dementia. Age Age is is associated associated with with dementia dementia (RR=4.0) (RR=4.0) Criterion 2: Age was associated with obesity. Age Age is is associated associated with with obesity obesity (OR=81.0) (OR=81.0) Age Age is is not not a a result result of of obesity obesity (not (not from from data) data) Age is not a result of obesity. Again that is not from data in study. All 3 criteria were satisfied. Therefore, yes age was confounding observed association between obesity and dementia. It just so happens that obese participants in study were older and that was what accounted for observed association with dementia Slide 34: Controlling Confounding So really, confounding variables are nuisance Design phase Group or or individual matching on on suspected variables that get in way of relationship confounding factor e.g. e.g. matching on on age age in in case-control study study that we really want to study. So how do we Analysis phase control for it? We control for confounding in Stratification Standardization design and analysis phases. In design phase we Adjustment (multivariate analysis) can group or individually match on suspected confounding factor. This is specifically regarding a case-control study where we can match cases and controls on a variable such as age to make m similar so y do not differ. In analysis we can stratify our analyses based on this confounding variable. Record our data stratified by age for example, young and old. We can standardize our data. Or what s most commonly done is conduct multivariate analysis to adjust for confounding variable. 12
Slide 35: Thoughts on Confounding Some final thoughts on confounding. Again Not an an error in in study confounding variables are se nuisance Valid finding of of relationships between factors and disease variables that get in way of examining relationship that we really want to study. Unlike Failure to to take into account confounding IS IS an an error and can can bias results! bias, it is not an error in study. So se confounding relationships are really valid findings of relationships between factors and a disease. However, failure to take into account confounding or to control for confounding is an error and will bias results of your study. Exposure or or Characteristic Disease or or Outcome Observed Observed Association Association Is Is it: it: biased, biased, confounded, or or causal? causal? Slide 36: Association Between Exposure and Outcome So we ve assessed wher observed association is biased or confounded. Let s finally look at wher our observed association between our exposure and our outcome is causal. Slide 37: Epidemiologic Reasoning First let s talk a little about epidemiologic Determine wher a statistical association exists exists between characteristics or or exposures and and disease reasoning. The primary goal is to determine Study Study of of group group characteristics (ecologic (ecologic studies) studies) Study Study of of individual individual characteristics (case-control (case-control and and cohort cohort wher re is a statistical association between studies) studies) our exposure or characteristic and our disease. Derive inferences regarding possible causal relationship using using pre-determined criteria or or guidelines We assess that through various studies. We can study group characteristics in an ecological study, or we can study individual characteristics in a case-control, cross-sectional or cohort study. The key here is deriving inferences regarding possible causal relationships using pre-determined criteria or guidelines. Again we have this observed association, now we need to infer wher this association is causal or not and we can use pre-determined guidelines to do this. 13
Slide 38: Causation However, association between an exposure and Association is is not not equal to to causation an outcome is not equal to causation. Consider Consider following statement: If If rooster crows at at break of of dawn, n n rooster caused following statement: If rooster crows at sun sun to to rise rise break of dawn every morning, n Causation implies re re is is a true true mechanism from from exposure to to disease rooster caused sun to rise. Again se are associated but y re not causal. Causation implies re is a true mechanism from exposure to disease. More than just an observed association. Slide 39: Koch-Henle Postulates (1880s) So a little history lesson on causation. In this 1. 1. The The organism is is always found with with disease (regular) instance in context of infectious disease. 2. 2. The The organism is is not notfound with with any any or disease Jacob Henle and his student Robert Koch in (exclusive) 1880s built a set of postulates on disease 3. 3. The The organism, isolated from from one one who who has has disease, and and cultured through several generations, produces causation based on germ ory. There were disease (in (in experimental animals) 3 postulates. 1: organism is always found with disease. This is also known as regular postulate. Second: organism is not found with any or disease. That is, it is exclusive. Third, organism, isolated from one who has disease, and cultured through several generations, produces disease. That is, it is reproducible in experimental animals. Slide 40: Koch-Henle Postulates (1880s) Koch added that even when an infectious disease Koch Koch added that that Even when an an infectious disease cannot be be transmitted to to animals, regular and and exclusive cannot be transmitted to animals, ( third presence of of organism [postulates 1 and and 2] 2] proves a causal relationship postulate) regular and exclusive presence Unknown at at time time of of Koch-Henle (1840-1880) of organism ( first two postulates) can Carrier Carrier state state Asymptomatic infection infection Multifactorial causation causation prove a causal relationship. However, unknown Biologic Biologic spectrum spectrum of of disease disease at time of se postulates in mid 1800s, was idea of a carrier state, idea of asymptomatic infection, of multifactorial causation, or even biologic spectrum of disease. Those were not considered at this time. So causation needed to mature from this idea of se three postulates. 14
Slide 41: Understanding Causality Let s say say you you have determined: There is is a real real association You You believe it it to to be be causal (ruled out out confounding) NOW have you you proven CAUSALITY? Let s say you ve determined re is a real association between our exposure and our outcome. We believe it to be causal. That is we ve ruled out bias and confounding. Have we really proven causality? 1. 1. Temporal relationship 2. Strength of association Slide 42: Surgeon General s Guidelines for Establishing Causality 2. Strength of association 3. So have you proven causality if you ve ruled out 3. Dose-response relationship 4. 4. Replication of of findings 5. bias and confounding? Now Koch-Henle 5. Biologic plausibility 6. 6. Consideration of of alternate explanations 7. postulates are not applicable to chronic disease 7. Cessation of of exposure 8. 8. Consistency with with or or knowledge 9. 9. Specificity of of association epidemiology. But Doll and Hill in 1950s were trying to link smoking to lung cancer through a series of case-control and cohort studies. And se led to development of Surgeon General s Guidelines for establishing causality. These have since been through one round of revisions since 1950s. Here are nine guidelines. Slide 43: Temporal Relationship The first guideline is wher re is a temporal Exposure to to factor must must have have occurred before disease developed relationship between exposure and Easiest to to establish in in a prospective cohort study study disease of interest. That is exposure must Length of of interval between exposure and and disease very very important have occurred before disease developed. e.g. e.g. asbestos asbestos and and lung lung cancer cancer Lung Lung cancer cancer followed followed exposure exposure by by 33 or or 20 20 years? years? This is easiest to establish in a prospective cohort study in which all participants begin study disease free. There are couple considerations regarding this guideline. The length of interval between exposure and disease is very important. For example, asbestos exposure and lung cancer. Did lung cancer follow exposure by 3 years or 20 years? That is going to have an effect on wher this guideline is satisfied or not. 15
Slide 44: Strength of Association The next guideline is strength of The The stronger association, more more likely likely exposure is is causing disease association. The stronger observed Example: RR RR of of lung lung cancer in in smokers vs. vs. nonsmokers = 9; 9; RR RR of of lung lung cancer in in heavy vs. vs. non- non- association, more likely exposure is nonsmokers = 20 20 causing disease. That is, strong association is more likely to be causal because y are unlikely to be entirely due to bias and or confounding. So a weak association may be causal but it is harder to rule out bias and confounding. So here s an example of strong association. Relative risk of lung cancer in smokers vs. non-smokers is 9. The relative risk in heavy vs. non-smokers is 20. These are strong associations. So it is unlikely that y are entirely due to bias or confounding. Slide 45: Strength of Association So here are three odds ratios (OR) and Which odds odds ratio ratio (OR) (OR) would you you be be more more likely likely to to infer infer causation from? accompanying 95% confidence intervals (CI). OR#1: OR OR = 1.4 1.4 OR#2:OR = 9.8 9.8 OR#3:OR = 6.6 6.6 95% 95% CI CI = (1.2 (1.2--1.7) 1.7) 95% 95% CI CI = (1.8 (1.8--12.3) 12.3) 95% 95% CI CI = (5.9 (5.9--8.1) Which odds ratio would be more likely to infer causation? OR#1 is 1.4 with a very narrow CI. OR#2 is 9.8 with a very wide CI. OR#3 is 6.6 with a fairly narrow CI. The third OR would give us strongest evidence towards an association, OR of 6.6 with a fairly narrow CI. As compared with an OR of 9.8 which is stronger but has a very wide CI. Slide 46: Dose-Response Relationship The next guideline is wher a dose-response Persons who who have have increasingly higher exposure levels have have increasingly higher risks risks of of disease relationship exists. That is, persons who have Example: Lung Lung cancer death rates rates rise rise with with number of of cigarettes smoked higher exposure should have higher risks of disease. For example, lung cancer death rates rise with number of cigarettes smoked. This is not considered necessary for a causal relationship but it does provide additional evidence that a causal relationship exists. The higher exposure, higher risk of disease. 16
Slide 47: Age-Adjusted Mortality Rates Here is a figure depicting a dose response relationship between smoking and lung cancer. This graph shows age-adjusted mortality rates of bronchogenic carcinoma by current amount of smoking. We have mortality rate by 100,000 person-years on y-axis and amount of smoking on x-axis. As you can see, mortality rate per 100,000 among never smokers was 3.4. And it steadily increases with number of packs smoked a day up to 2+ packs/day where mortality rate is 217.3. Clearly a dose response. Slide 48: Replication of Findings The next guideline is replication of findings. Your The The association is is observed repeatedly in in different persons, places, times, and and circumstances observed association linking exposure to disease Replicating association in in different samples, with with different study study designs, and and different investigators should be observed repeatedly in different gives gives evidence of of causation Example: Smoking has has been been associated with with lung lung persons and populations, places, times, and cancer in in dozens of of retrospective and and prospective studies circumstances, using different study designs, and different investigators. For example, smoking has been associated with lung cancer in dozens of retrospective and prospective studies all around world using different study designs and among different populations. Slide 49: Biologic Plausibility The next guideline is biologic plausibility. There Biological or or social model exists exists to to explain association must be a biological or social model to explain Does Does not not conflict conflict with with current current knowledge knowledge of of natural natural history history and and biology biology of of disease disease your observed association. That is, your Example: Cigarettes contain many many carcinogenic substances observed association should not conflict with Many epidemiologic studies have have identified causal relationships before biological mechanisms were were current knowledge or natural history and biology identified of disease. For example cigarettes contain many carcinogenic substances. However, many epidemiologic studies have identified this causal relationship before biological mechanisms were identified. 17
Did Did investigators consider bias bias and and confounding? Slide 50: Considerations of Alternate Explanations Next, consideration of alternate Investigators must must consider or or possible explanations explanations. So did investigators rule out bias Example: Did Did investigators consider associations between smoking, coffee consumption and confounding? And se are alternate and and pancreatic cancer? explanations. So did investigators consider associations between smoking and coffee consumption and pancreatic cancer? This gets back to our earlier discussion of bias and confounding. These must be ruled out in order to determine wher a causal relationship exists. Slide 51: Cessation of Exposure Next, cessation of exposure. The risk of Risk Risk of of disease should decline when when exposure to to factor is is reduced or or eliminated developing disease should decline when In In certain cases, damage may may be be irreversible exposure has been reduced or eliminated. Example: Emphysema is is not not reversed with with cessation of of smoking, but but its its progression is is reduced However, in certain cases damage may be irreversible. For example, emphysema is not reversed with cessation of smoking, but its progression can be reduced. In famous example of John Snow investigating a risk factor of cholera being contaminated water, he removed pump handle at Broad St. pump and found that risk of disease decreased after handle was removed. Slide 52: Consistency With Or Knowledge If If a relationship is is causal, findings should be be consistent with or data If If lung lung cancer incidence increased as as cigarette use use was was on on decline, need to to explain how how this this was was consistent with a causal relationship case. Next, is your observed association consistent with or knowledge? So if relationship is causal, your relationship should be consistent with or data. So if lung cancer incidence increased as cigarette use was on decline, you need to explain how this was consistent with a causal relationship because this shouldn t be 18
Slide 53: Specificity of Association A single exposure should cause a single disease Next, specificity of association. A single Example: Smoking is is associated with with lung lung cancer as as well well as as many or diseases exposure should cause a single disease. This is a Lung Lung cancer cancer results results from from smoking smoking as as well well as as or or exposures exposures very narrow guideline. For example, smoking is When present, provides additional support for for causal inference associated with lung cancer as well as many When absent, does does not not preclude a causal relationship or diseases. And lung cancer results from smoking as well as many or exposures. So this guideline is very difficult to satisfy and it has its roots back in Koch-Henle postulates. However, when this is present, like or guidelines, it provides additional support for a causal relationship. However, if this guideline cannot be satisfied, it does not preclude a causal relationship. Remembering distinctions between association and and Slide 54: Uses of Surgeon General s Guidelines for Establishing Causality causation in in epidemiologic research How can we use Surgeon General s Critically reading epidemiologic studies Guidelines for Establishing Causality? We can use Designing epidemiologic studies se guidelines to remember distinctions Interpreting results of of your your own own study study between association and causation in epidemiologic research. It needs to be in back of every investigator s mind that if association is established it does not confirm causation. Next, se guidelines are useful when critically evaluating or reading epidemiologic studies. These guidelines can also be useful for designing epidemiologic studies. Different study design options or different methods within a study design can be used so that se guidelines can be satisfied. For example a dose-response or a temporal relationship between your exposure and your outcome. Next se guidelines can be useful when interpreting results of your own study. And finally, overall se guidelines are not meant to be rigid. If one guideline is not satisfied that does not preclude a causal relationship. These guidelines are used to essentially build a case for exposure causing a disease. One study is not going to prove causality. It s body of work or body of knowledge about an exposure and an outcome. 19
Slide 55: Does HIV Cause AIDS? Majority of of scientists believe HIV HIV causes AIDS AIDS Small Small group believes AIDS AIDS is is a behavioral rar than than an an infectious disease AIDS AIDS is is caused by by use use of of recreational drugs and and antiretroviral drugs in in U.S. U.S. and and Europe, and and by by malnutrition in in Africa and by malnutrition in Africa. Let s look at an example: Does HIV cause AIDS? The vast majority of scientists believe that HIV causes AIDS however; re is a small group that believes AIDS is a behavioral rar than an infectious disease. They believe that AIDS is caused by use of recreational drugs and treatment drugs offered in U.S. and Europe, Slide 56: Does HIV Cause AIDS? Let s look back at Koch-Henle Postulates. Regular and and exclusive Gallo Gallo et et al. al. routinely found HIV HIV in in people with with AIDS AIDS Regular and exclusive? Yes. Gallo et al. routinely symptoms and and failed to to find find HIV HIV among people who who eir lacked AIDS AIDS symptoms or or AIDS- found HIV in people with AIDS and failed to find associated risk risk factors Experimental model HIV among people who eir lacked AIDS 33 lab lab workers workers accidentally accidentally infected infected in in early early 1990s 1990s with with purely purely molecularly molecularly cloned cloned strain strain of of HIV HIV symptoms or AIDS-associated risk factors. One One developed developed pneumonia pneumonia (AIDS-defining disease) disease) before before started started antiretroviral rapy rapy Regarding experimental model with 3 rd postulate, 3 lab workers were accidentally infected in early 1990s with purely molecularly cloned HIV. One developed pneumonia (an AIDS-defining disease) before starting antiretroviral rapy. That would negate argument that rapy is causing disease. Slide 57: Does HIV Cause AIDS? Let s look at guidelines. There have been Epidemiologic studies have have established: Temporal Temporal relationship relationship numerous epidemiological studies which have Strong Strong association association Dose Dose response response established a temporal relationship. That is Replication Replication of of findings findings Biologic Biologic plausibility plausibility infection with virus precedes Cessation Cessation of of exposure exposure (decrease (decrease in in deaths deaths after after antiretroviral rapy) rapy) development of AIDS-related symptoms. There Specificity Specificity Consistency Consistency with with or or knowledge knowledge is a strong association between virus and symptoms. There is a dose-response and replication of findings. This is biologically plausible. Cessation of exposure: re is a decrease in deaths associated with AIDS after advent of antiretroviral rapy. This is a very specific association. And it s consistent with or knowledge. 20
Slide 58: Causation Associations are are observed A few final words on causation. Remember Causation is is inferred associations are observed, and causation is It It is is important to to remember that that se criteria provide evidence for for causal relationships inferred. So it s important to remember that All All of of evidence must must be be considered and and se guidelines provide evidence for causal criteria weighed against each each or or to to infer infer causal relationship relationships. They don t really prove a causal relationship. Next, all of evidence must be considered and criteria weighed against each or to infer that re is a causal relationship. Slide 59: Summary In summary, we discussed three aspects of an Bias Bias is is a systematic error error that that results in in an an incorrect estimate of of association association. Wher association is biased: Selection Selection and and information information bias bias was re a systematic error that resulted in a Confounding is is an an alternate explanation which can can be be controlled flawed or incorrect estimate of association? We Design Design and and analysis analysis phases phases Causation must must be be inferred talked about selection and information bias. Next we talked about wher this observed association was confounded, that re was an alternate explanation, which resulted in this observed association. We talked about controlling confounding in design and analysis phases. Lastly, if bias and confounding can be ruled out, is observed association causal? It s important to remember causation must be inferred, whereas associations are observed. Thank you and that concludes Module 2: Issues of Interpretation in Epidemiologic Studies. Center for Public Health Continuing Education University at Albany School of Public Health Department of Community & Family Medicine Duke University School of Medicine 21