epidemiology unit 3

Five approaches to handling confounding

DESIGN: 1. Restriction 2. Randomization (clinical trials) 3. Individual/group matching (case-control studies) ANALYSIS: 4. Stratificaton 5. Adjustment (Mantel-haenzsel adjustment, multivariate analysis)

example of effect modification

Does the association vary by a person's sex? Crude OR: 0.64 Women: 1.19 Men: 0.53 Interpretation: high phsycial activity appear to protect men from having an MI, whreas the effect for women is close to null. Therefore, a person's sex appear to modify the relationship between levels of physical activity and MI

Example 2

H0: there is no relationship between a person's sex and depression Ha: a person's sex is associated with depression RR= 2.77 -interpretation: females appear to be at 2.77 times higher risk of depression than males -however is this elevated risk similar among young persons and older persons? young: 1.22 Old: 2.85 this examples suggests risk ratios are heterogenous: this is effect modification by age

calculating AR %

I1 - I0 x 100 = .166-.045 / .166 x 100 73%

AR

I1 = 750/4500 = .166 I0 = 250/5500 = .045 AR = 750/4500 - 250/5500 = .1212 x 1000 = 121.2 /1000

Calculating PAR and PAR% with only I1, I0 and Pe I1 = 9/1000 10 = 1/1000 Pe = 45%

I1 x Pe = (9/1000) x .45 = .00405 I0 x Pne = 1/1000 x .55 = .00055 It = .00405 + .00055 = .0046 PAR: incidence of lunger cancer attributable to smoking in the total population PAR = .0046 - .001 = .0036 = 36/1000 PAR%: the proportion of the risk in the total population that is attributable to smoking PAR% = .0046 - .001 / .0046 x 100 = 78.3%

PAR

It - I0 It = 1000/10000 (1000 people have diabetes in total) I0 = 250/5500 = .45 PAR = 100/1000 - 45/1000 = 55/1000

PAR%

It-I0 / It x 100 (100/1000) - (45/1000) / (100/1000) x 100 55%

Odds ratio interpretation

OR = 5.95 People who ate chilli peppers were 5.95 times more likely to develop gastric cancer than people who did not eat chilli peppers

stratified analysis of catecholamine levels and CHD

OR m-h = 2.11 -conclusion: after adjusting for the effects of age, men with CHD were 2.11 times more likely to have high catecholamine levels than maen who did not have CHD

Other ways to calculate PAR%: levin's formula

PAR% = p(r-1) / (p-1)+1 x 100 - p = proportion of the population with the exposure r = relative risk or odds ratio

PAR% for case control study

Pe(OR-1)/ [Pe(OR-1)+1] x 100 Pe = proportion of exposed controls (assuming that the proportion of exposed controls is representative of proportion of exposed in the source population)

3. Dose response relationship or biologic gradient

Pro: logically, most harmful exposures could be expected to increase the risk of disease in a gradient fashion (e.g. if a little is bad, a lot should be worse) Note; some associations show a single jump (threshold) rather than a monotonic trend Note: some associations show a U or J shaped trend (e.g. alcohol consumptions and mortality; maternal age at time of birth and DS)

2. Strength of the association

Pro: the stronger the association, the less likely the relationship is due merely to an unsuspected or uncontrolled confounding variable or bias Con: strong but non casual associations are common (example: non casual relation between down's syndrome and birth rank; which is confounded by maternal age) Con: ratio measures (e.g. RR) may be comparatively small for common exposures and diseases (e.g. smoking and cardio vascular disease) but are casual

Calculating PAR% using levin's formula

RR = 9/1000 / 1/1000 = 9 PAR% = .45(9-1) / .45(9-1)+1 x 100 3.6/4.6 x 100 = 78.3%

interpret relative risk

RR=1: then risk in the exposed equals the risk in the unexposed (no association: null hypothesis) RR > 1: the risk in the exposed is greater than the risk in the unexposed (positive association; possibly casual) RR < 1: the risk in the exposed is less than the risk in the unexposed (negative association; possibly protective)

1. If U is sufficient and necessary

U -> DISEASE

3. If U is sufficient but is not necessary

U -> disease A -> disease C -> Disease

4. If U is not sufficient but is necessary

U/B -> disease U/A -> disease

interaction: multiplicative model

(look in notes)

Interpret as a percentage

(normally wouldn't be done for 6.0 because above 3.0) -RR = 6 > 1 -percent increase change = 6.00-1.00 x 100 = 500% -during the study, adults with hypertension were 500% times as likely to have a myocardial infarction than adults with no hypertension

Chapter 14

-

Chapter 15: bias

-

Chapter 15: bias, confounding, and interaction/effect modification

-

Chapter 15: confounding

-

Different philosophies of casual inference

-

If a relationship is casual, 4 types of casual relationships are possible (consider an exposure: U)

-

Obesity example: 10,000 people 4500 (BMI > 30) -> 750 develop diabetes 5500 (BMI < 30) -> 250 develop diabetes

-

differential misclassification may lead to

- a biased estimate that is an overestimate of the true estimate -a biased estimate that is an underestimate of the truth estimate -by chance, the same as the true measure of association -this applies equally to any study design

range of values for RD for CI

-1 to 1

range of values for risk difference (RD) for prevalence

-1 to 1

Interpretation of AR% case control

-Assuming a causal association between smoking and the development of bladder cancer, 67% of the bladder cancer cases among smokers can be attributed to their smoking -Assuming a causal association between smoking and the development of bladder cancer, 67% of bladder cancer cases could be prevented if they did not smoke

Effect modification continued

-H0: there is no association between alcohol consumption and larynx cancer -Ha: evidence suggests that alcohol consumption is associated with larynx cancer design: cohort study -measure of association: RR (31/201)/(41/314) RR = 3.46 interpretation: persons with high alcohol consumption appear to be at 3.46 times the risk of developing the larynx cancer than persons without high alcohol consumption. However, is this elevated risk similar among smokers and non smokers?

interpretation of PAR% case control

-In this study population, 43.2% of bladder cancer cases can be attributed to smoking - In this study population 43.2% of bladder cancer cases could be prevented if people did not smoke

1. Temporal relationship

-Pros: by definition, a cause of disease must precede onset of the disease. An absolute must Not really a con but a problem: the existence of an appropriate time sequence can be difficult to establish (e.g. lifestyle factors are likely to be altered after the first symptoms of a disease occur) (what if lung cancer causes smoking: having lung cancer makes you want to smoke versus vice versa)

RD interpretations

-RD is positive when the risk for those exposed is greater than those unexposed -RD is negative when risk for those exposed is less than the risk for those unexposed -RD = 0 when the risk factor is unrelated to the risk of the disease/health outcome

misclassification: outcome

-alcohol consumption and diabetes -by some token, diagnosis of diabetes is based on self reported symptoms; thirst, hunger, frequent urination -i.e. someone who urinates a lot might say they are not urinating more than normal (1o times a day) while for other's that is a lot -concept of hunger and thirst is very different for people -self reports can get you very different outcomes

Differential misclassification (non random)

-all of the subjects are not just responding in one way - just one group is responding in a way -misclassification of one variable dpeends on the status of the other (measurement error in exposure depends on outcome or vice versa)

Selection bias: incidence prevalence bias

-also called neyman bias or selective survival bias -estiamte the risk of disease using data collected at a given time point in a series of survivors (prevalent cases) rather than on data gather during a certain time period in a group of incidence cases (example 1) OR -when the sample offers a distorted frequency of the exposure (example 2) -this type of bias is of particular relevance in case control studies -important when there is a time gap between exposure and selection of study subjects -the odds ratio may increase (exposure may lead to selective survival) or decrease (expsoure may lead to selective mortality)

Breast cancer example ID exposed: 41/28010 person years (.0014) ID not exposed: 15/19017 (.00078) RR = (.00146/.00078) = 1.86

-among women with tuberculosis the rate of breast cancer is 1.86 times higher among women who were repeatedly exposed to multiple x ray flouroscopies than women who did not have x ray flouroscopies -among women with tuberculosis, the rate of breast cancer is 86% higher among those who were repeatedly exposed to multiple x ray fluoroscopies than women who did not have x ray fluoroscopies

9. Specificity of the association

-an association is specific when a certain exposure is associated with only one disease -weakest of all guidelines -Con: smoking cigarattes vs many other health outcomes -Pro: infectious diseases Guinea worm, smallpox and polio eradication efforts -priors and varaint creutzfeldt jacob disease

Misclassification

-applies to both cohort and case control studies -misclassifcation of exposure, outcome or both (the worst kind)

Interpretation of AR

-assuming a casual relationship between BMI and diabetes -among individuals with BMI > 30, the excess occurrence of diabetes is 121 cases/1000 subjects -among individuals with a BMI > 30, we could prevent 121 cases/1000 subjects of diabetes if we could lower the BMI to under 30

AR% interpretation

-assuming a causal relationship between BMI and diabetes -among individuals with a BMI > 30, 73% of cases of diabetes is due to their obesity -among individuals with a BMI > 30, 73% of cases could be prevented if people with obesity lowered their BMI to <30.

PAR% interpretation

-assuming a causal relationship between BMI and diabetes -in austin, 55% of the cases of diabetes is due to obesity -in austin 55% of the cases of diabetes could be prevented if people living in Austin lowered the number of people with BMI > 30

PAR interpretation

-assuming a causal relationship between BMI and diabetes -in austin, the excess incidence of diabetes due to having a BMI > 30 is 55/1000 population -in austin, 55/1000 incidence cases of diabetes could be prevented if the population of Austin reduced BMI levels to under 30

measures of public health impact

-attributable risk (AR) -attributable risk percent (AR%) -population attributable risk (PAR) -population attributable risk percent (PAR%) -they all require that a cause effect relationship exists between the exposure and the outcome -may be calculated using prevalence, incidence and rates -some may be calculated using odds ratios and risk ratios

Differential misclassification: case control study example

-can occur if the information on exposure status depended on whether or not the subject had the disease -a case is more likely to overestimate/underestimate the level of exposure than the control -a control is more likely to overestimate/underestimate the level of exposure than a case

Restriction disadvantages

-can substantially reduce N (sample size) -residual confounding if category is not sufficiently narrow (e.g age: 45-64 have restricted but people who are 45 may have different levels of activity from someone who is 64) -cannot study restricted factor as effect modifier (e.g., sex) -association is not generalizable to all people, only subgroups who have been studied

example 2: confounding

-case control study -examines the relationship between Total Cholesterol (high, low) and myocardial infarction (MI, no MI) -we know form other studies that obesity can contribute to an MI (obesity is casually associated with having a heart attack) -we know also that total cholesterol correlates with obesity (there is an association between cholesterol and obesity) -we investigate to see if obesity is confounding the relationship between total cholesterol and myocardial infarction

incidence prevalence bias: example 2a

-case control study: chronic use of aspirin and occurence of end stage renal disease -in this study the selection process of cases and controls as well as the assessment of exposure status were performed in an appropriate manner, thus we can assume the OR reflects the situation in the reference population -OR = 1.74 -aspirin using does affect end stage renal disease

Example of misclassification: case control study: dietary fat intake and myocardial infarction (differential misclassification)

-cases just had heart attack - don't want to have another heart attack so more likely to tell you the truth because they don't want to risk getting another heart attack -but controls might still not tell the truth and might still tell socially acceptable answer

Restriction advantages

-cheap -straightforward -convenient -offers complete control if confounder is narrow (e.g. sex, race, - less so for continuous measures such as age education)

Frequent sequence of studies in human populations

-clinical observations (case studies/case reports, case series) -available data (ecological: aggregated, descriptive: surveys, questionnaires) -case control studies -cohort studies -randomized trials

always report if you think bias may have affected your results (or not): measurement bias (misclassification)

-collect information about exposure (or outcome) in the same way from both groups (exposed vs non exposed; cases vs controls) -analyze data putting individuals into the category you suspect they belong and measuring the association

rules to discern whether there is confounding using stratified analysis

-compare crude OR/RR to stratified ORs/RRs -does the crude OR/RR lie WITHIN THE INTERVAL of the stratified -if the answer is NO: the third variable is suspected of confounding the relationship between the exposure variable and the outcome variable -calculate an adjusted OR (OR m-h)

rules to discern whether there is confounding using stratified analysis

-compare crude RR/OR and adjusted RR/OR m-h -if they are not the same (within 10% of each other): there is confounding... report the adjusted -if they are the same (or very similar) there is no confounding

Comparing extent of disease between groups

-compare groups with respect to extent of disease or their likelihood of developing disease

Relative Risk

-compares incidence of disease among the exposed with the incidence of disease among the non exposed by means of a ratio -RR = Risk in exposed / risk in non exposed

Relationship of potential confounding factors

-confounding is a nuisance effect resulting in a distortion of the true association between the exposure and the disease/outcome -in order to estimate a valid association, we must eliminate confounding

4. Stratified analysis

-definition: an analytic technique to control confounding, involving the evaluation of the association within homogenous categories, or strata, of the confounding variable

selection bias

-differential criteria are used to enroll (select) participants in investigation -participants come from different underlying populations -can be particularly problematic in case control studies but can also occur cohort studies (particularly multi sample) -need to apply the same criteria to all groups

Cohort study odds ratio

-diseased odds -among exposed: a/b -among non exposed: c/d

Berkson's bias

-dr joseph berkson 1940s -known as admission rate bias -hospital cases and controls are recruited from among hospital patients, thus characteristics of both of those groups will be influenced by hospital admission rates -cases and controls are admitted to the hospital at different rates -controls in hospitals may have a different higher levels of exposure than the population from which the cases are drawn

Practice example: CI of exposed = 180/10,000 (.0180) CI of non exposed = 30/10000 (.00301) RR = 6.00

-during the study period, the incidence of myocardial infarction is 6 times higher maong adults who have hypertension as compared to adults who do not have hypertension -during the study period, the risk of myocardial infarction is 6 times high among adults who have a history of hypertension as compared to adults who do not have hypertension

Reporting effect modification

-effect modification should be explicitly described and reported and not controlled as in the case of confounding -note from previous discussion -a varaible may be: a confounder an an effect modifier(present stratified OR/RR as effect modification) , a confounder alone )present the adjusted OR/RR), an effect modifier alone (present stratified ORs/RRs as effect modifiers), neither (present crude OR/RR)

always report if you think bias may have affected your results (or not): non response bias

-encourage participation (don't be afraid to beg) -evaluate whether those participating differ from those choosing to participate

case control study odds ratio

-exposure odds -among cases: a/c among controls: b/d

what can groups be defined by

-exposure to some harmful agent (e.g., exposed or not exposed) -by socio demographic characteristics (e.g., men or women; young or old) -by a particular risk factor (e.g., current smoker, non smoker, obese or slim)

Calculating the risk ratio

-exposure: Hypertension (yes or no) -outcome: myocardial infarction during the one year follow up cohort study

Risk difference (RD)

-general difference: RD = I1-I0 -prevalence: RD = P1-P0 -Cumulative incidence: RD = CI1-CI0 Incidence density: RD = ID1-ID0

Information bias (observaiton bias)

-has the information been gathered in the same way? -a flaw in measuring exposure or outcome data that results in different quality (accuracy) of information between comparison groups (cases vs controls) (exposed vs unexposed) -can occur when there is a random or systematic inaccuracy in measurement

Population attributable risk (PAR)

-how much of the disease incidence in the total population is due to the exposure (looking at a greater population) -risk difference

attributable risk

-how much of the disease that occurs can be attributed to a certain exposure -how much of the risk (incidence of disease can we hope to prevent if we are able to eliminate exposure to the agent in question -answered as an amount, proportion or percent

Attributable risk (AR)

-how much of the total risk in exposed persons is due to the exposure -risk difference

1. Restriction

-idea: if you restrict subjects to certain characteristics, you limit the variability on that factor and exclude the possibility of confounding. However, you also lose the ability to analyze that variable -method: in design stage, restrict eligibility (inclusion criteria) to persons who are with certain categories of the confounder -example: you know that sex and race are potential confounding factors in your study: so limit the study to white women, if you study only white women there are no men or people of other races -thus by restricting your sample, you have eliminated the possibility of males and other races accounting for your study results

7. Cessation of exposure

-if the exposure is casually associated with the disease, the risk of disease should be reduced by reducing the exposure (stop smoking and the risk of developing lung cancer) (L tryptophan preparations vs Eosinophilla myalgia syndrom 1989)

8. Consistency with other knowledge

-if the exposure is casually associated with the outcome, one would expect that the findings would be consistent with other knowledge or data -some oils are mutagenic to human cells in the lab -exposure to cooking oil is carcinogenic -cigerette smoking and lung cancer among men -cigerette smoking and lung cancer among women -other cancers among people who smoke

relative risk : beware

-if the incidence measure being used in cumulative incidence, the relative risk is called a risk ratio -if the incidence measure being used is incidence density: the relative risk is called a rate ratio

matched odds ratio

-in calculating the matched odds ratio, the 2x2 is laid out differently and only the discordant pairs provide useful information -MATCHED ODDS RATIO = B/C (case exposed and control not exposed/ case not exposed and control exposed) -McNemars test = x^2 = (|b-c|-1)^2 / B+C tests whether the OR is different from 1 (H0: OR =1): can also calcualte confidence interval for OR

the odds ratio (relative odds)

-in case control studies, we can compare the odds for exposure among the cases to the odds for exposure among the controls -in case control studies the odds ratio is also called the relative odds or exposure odds -note the odds ratio can be calculated for many studies: cohort, case control, cross section

range of values for RD for incidence density

-infinity to infinity events per person time

Random error can be problematic but

-influence can be reduced: increased sample size, change study design, improve instrumentation -probability of some type of influence can be quantified (e.g. confidence interval width)

matching disadvantages

-is difficult, tedious to match exactly -expensive, except when can be done by computer -if matching on 3 factors: age (5 categories) sex (2 categories), race (3 categories) have 30 possible combinations to consider in finding a control (5x2x3). Also since age has only 5 categories, will have residual confoudning that will have to adjusts for that -once a factor has been matched, cannot study it as a risk factor for disease

Example 3: confounding

-is drinking alcohol associated with CHD -case control study -examine the association between alcohol consumption and coronary heart disease (CHD) -we know from other studies: smoking is associated with alcohol consumption, smoking is a risk factor for CHD -perhaps the elevated OR for consuming alcohol-CHD may be due to lack of controlling (adjusting for smoking status)

a causal association:

-is the association real or not -if so, is the association causal (is there sufficient evidence to infer that a casual association exists between the exposure and the disease)

5. Multivariate analysis

-it is often necessary to look for more than one potential confounder in analysis - this may make a stratified analysis impractical or inefficient - because cell sizes get too small -variables like age, sex, race and education are often included with other potential confounders -multivariate analysis is a tool to adjust simultaneously for multiple confounders (can be used for one independent variable and one dependent variable) -able to conserve the power of the test

Systematic error "bias" is problematic because

-it may be present without investigator being aware -sources may be difficult to identify -influence may be difficult to assess

Once you determine you suspect confoundingL how do you adjust for it (aka how to calculate adjusted OR)

-mantel-haenszel formula (always add across before dividing) -don't need to use data from crude (example in notes)

matching continued

-matching is used to eliminate biased comparisons between cases and controls... but only can be accomplished if accompanied by matched analysis. If analysis is not matched, resulting OR can be biased (as a result of matching) -2 step process: matched design -> matched analysis -immediate goal of matching: balance numbers of cases and controls that occur at each level of the matching variables. E.g., match on age and race: balance groups for both variables - thus age and race are no longer related to outcome - eliminating one criterion of confounding: PCF-D associaton -the blaance is achieved by forcing the matching variables to be the same between cases and controls -has great intuitive appeal and has been widely used in epidemiological studies. should only be used when matching factor is known to be a strong confounder

multivariate analysis continued

-multivariate analysis involves modeling the relationship between exposure and disease by means of a mathematical equation -in the equation, the outcome is considered the dependent (outcome) variable -the exposure of interest, other strong risk factors for the disease, and one or more potential confounders are considered independent (predictor or explanatory variables)

can i fix bias now that it is in my study

-no but can do things to try to minimize bias that occur -selection bias: carefully examine who you're choosing for your study -attrition bias: prevent lost to follow up as much as possible get outcome data for those lost (usually not possible; if so-they would not be lost to follow up) evaluate whether those lost to follow up differ systematically from those who continue to participate analyze data with those loast as 1)having the outcome and then 2) as not having the outcome, and see if the observed association significantly differs

Information bias

-non comparable information is obtained from the different study groups -observer bias/interviewer bias: the interviewer elicits or interpret information differently between cases/controls or exposed/unexposed recall bias: participants recalls exposures erroneously. IN a case control study, cases may over. report, controls under report -cases/proxies of cases seek info about causes of disease -healthy controls/proxies have no similar reason for seeking such information

Types of misclassification

-non differential misclassification -differential misclassification

When is OR = RR

-odds ratios estimate the RR when the conditions are met (if so the OR from a case control study is a possible estimate of the RR even though it can not be calculated directly) -when the rare disease assumption is not met; the RR and OR will be different (if so the OR from a case control study is typically not a good estimate of the RR; therefore it should be interpreted as an odds ratio) -for the course we assume the 3 assumptions are met aka OR = RR

odds ratio

-odds= probability that something will happen/probability htat something will not happen

the multiple regression model

-outcome = continuous measure (blood pressure, cholesterol level) -the coefficients of each independent variable are directly interpretable in relation to the outcome -commonly used for cross sectional data, cohort data looking at longitudinal change -there are many variations of the multiple regression model

the logistic regression model

-outcome = natural log odds of disease Log (odds of disease) = alpha + betax1.... etc -the coefficients of each independent variable can be directly converted to the odds ratio, which describes the independent association of that parameter with the outcome -commonly used for case control data, can be used for cohort data if time of event is not important (adjust for follow up time) -if dependent = dichotomous -> logistic regression

Example 1: Confounding

-overall crude mortality shows that a lot more people die in US than costa rica but its because they did not account for differences in the age structure of the populations -age is a strong risk factor for mortality -age adjusted mortality rates

4. Replication of findings

-preponderance of evidence pro: due to the inexact nature of epidemiologic investigations, evidence of causality is a persuasive when several studies yield similar results when conducted by 1. different investigators 2. at different times 3. in different populations -con: some effects are produced by their causes only under unusual circumstances -con: studies of the same phenomenon can be expected to yield different results simply because they differ in their methods and from random errors

effect modification

-process of stratification is used to evaluate both confounding and effect modification. note that there are other, more sophisticated ways to evaluate effect modification -definition: an association between E-D varies by different levels of a third variable

Selection bias - publication bias

-publication bias is generated in the sleection process of the information that is eventually pubished -several factors influence publication, the most being study size and design, quality, funding and prestige -people with good research if they don't see results - fail the null hypothesis - might not get published

2. Randomization

-randomization potentially equalizes groups for all confounding characteristics -nice because confounders that may be unknown are controlled by means of randomization -used in randomized clinical trials/intervention studies -note: use as a tool to balance confounders (known and unknown), but does not guarantee control

The RR/OR is not always informative in and of itself: measure of strength of association

-relative risk = 2 means risk/rate of disease/outcome is twice as high in exposed vs unexposed RR = 2 if p1/p2 could be .02/.01 or could be .00000002/.0000002 RR and OR are used to measure strength of association and used in judgement of validity and casual nature of an association Ex1: the incidence rate has increased from 1/100 to 2/100 at risk Ex 2: the incicence rate has increased from 1/100,000 to 2/100,1000 at risk

healthy worker effect

-selection bias: people who work are healtheir than people who do not work -occurs in occupational cohort studies when workers represent the unexposed group. this is because workers tend to be healthier than the general population -cell C might be inflated

Epidemiologic study presents many opportunities for systematic error in relation to

-selection of study participants -classification and measurement procedures -comparison and interpretation

Remember

-selection subjects for a study is different from selection bias: almost all studies select a study target population - this affects the generalizability or external validity of the study, this may not affect comparisons made within the study or the study's internal validity -however when making a systematic error in selection one or more of the study groups that will be compared may result in selection bias: could happen when selecting cases or control group, could happen when selecting exposed or unexposed group, may affect internal validity of the study, may affect inference back to the target population

misclassification: exposure

-some people in cells a and b should be in cells c and d (and vice versa) -ex: alcohol consumption and diabetes -measurement of alcohol consumption by self report -college student who is pressured might classify themselves as moderate drinker when they really don't drink that much -or if you have a drinking problem and don't want people to know say that you don't drink that much

recall that we have already learned other techniques to control for confounding in epidemiological studies: rate adjustment

-statistical techniques to compare rates among populations with differing underlying structures (e.g., age, gender, education) -there are two ways to adjust (standardize) rates: direct adjustment, indirect adjustment

example of misclassification case control study (non-differential); dietary fat intake and myocardial infarction

-subjects do not recall the amount of fatty foods eaten, but the errors in recall did not depend on whether they had a myocardial infarction -notice 20% of both cases and controls who ate high fat diets underreported fat intake -regardless of disease status, respondents may under report intake of foods with high fat content because they think that low fat diets are more acceptable to the investigators

example 2 continued

-take all subjects that have MI and NO MI and then split into obese and non obese groups -among people who were obese -85% had high total cholesterol -found that obesity was a potential confounder

Another way to interpret relative risks as percentages

-the % increase change= (RR-1) x 100 for RR > 1 -the % decrease change = (1-RR) x 100 for RR < 1 (typically used for RR/OR for 3 or less)

confounding

-the confounding variable is casually associated with the outcome and -associated with the exposure but not the result of the exposure (it is not in the causal pathway between the exposure and the outcome)

Consensus (Thomas Kuhn - 1962)

-the consensus of the scientific community determines which is considered accepted and what is refuted

Absolute risk

-the incidence of disease in a population -indicates the magnitude of the risk in a specific group of people based on the degree to which they have been exposed -does not indicate whether a particular exposure is associated with increased or decreased risk

Effect modification

-the magnitude or direction of an association between an exposure and health outcome varies according to levels of a third factor -also called: effect modification, interaction -unlike confounding, effect modification should be described and reported rather than controlled

Proportional hazards model

-the proportional hazards model has as the dependent variable the natural logarithm of the incidence of disease -log [incidence rate (t)] = ..... -commonly used for incidence density data in cohort studies when the "time to an event" is important to quantify

Measures of association

-the relative risk (risk ratio, rate ratio) -the odds ratio (relative odds) (disease odds ratio, exposure odds ratio)

Non-differential misclassification example: case control study

-the subject's memory of exposure is unrelated to whether or not the subject has the disease of interest -the subject answers question about the exposure with a socially acceptable but sometimes inaccurate response whether or not they have the disease of interest

definition of bias

-the systematic error in design, conduct or analysis of the study that results in a mistaken estimate of the exposure's' effect on the risk of disease -any process at any stage or inference which tends to produce results or conclusions that differ systematically from the truth

Selection bias: loss to follow up bias

-this bias occurs in prospective cohort studies when individuals lost to follow up do not have the same probability of having the clinical outcome of interest in comparison with individuals who remain under observations -people with an exposure might decide to stay longer because they want a solution, or possible there is a cure somewhere else and they leave

epidemiology goal

-to determine the degree to which there is an association between -specified exposure and disease or health outcome -risk factor and disease or health outcome -some other measure and disease or health outcome

6. Consideration of alternative explanations

-to what extent are other explanations taken into account when describing the exposure disease relationship -is it possible that the association we see is not due to the exposure of interest, but to some other exposures (coffee vs pancreatic cancer: confounded by consumption of alcohol/smoking?) (people who drink coffee are more likely to drink alcohol and smoke)

important

-try to avoid common types of bias when designing, conduct or analyzing a study -when you cannot avoid bias, recognize the limitations of study and predict in which direction the OR/RR is biased

3. Matching

-used in case control studies -definition: pairing individuals in sampling phase (design) of investigation so that only those controls who "match" cases are selected for study ex: study of risk factors for multiple sclerosis: female 30 yo w/ female 30 yo -if you have enough cases, will usually match 1:1. However you may want to match up to 1:4 if case numbers are small (increases statistical power). may end up with varying ratios across samples. Statistical methods to deal with this

causality judgement requires a(n):

-valid association -evaluation of hill's casual criteria

example: increased stress levels and risk of CHD

-we suspect that age is a confounder because we know 1. age is a strong independent risk factor for CHD 2. age is associated with levels of catecholamines (stress) -so we divided the participants into the following groups: young (men 25-54) and old (men 55-85) -crude OR: 2.86 -young OR: 1.84 -old OR: 2.17 -crude OR lies outside: high suspect that age is a confounder

Relative Risk and Relative Odds

-what is the magnitude (strength) of the association (if any)

Population attributable risk percent (PAR%)

-what proportion of the disease incidence in the total population is due to the exposure -percentage -population etiologic fraction

Attributuable risk percentage (AR%)

-what proportion of the risk in exposed persons is due to the exposure) -percentage -etiologic fraction

interaction

-when the incidence rate of disease in the presence of two or more risk factors differs from the incidence rate expected to result from their individual effects -positive interaction or synergistic: effects is greater than expected -negative interaction or antagonistic: effect is less than expected

incidence prevalence example 1

-whenever possible researchers should estimate the occurences of a disease in terms of incidence (i.e. by countain all causes that occur in a sample in a given time interval rather than in a given time point) -don't just look at one given time point -there is a picture in the notes

Steps to evaluate confounding and effect modification

1. Look at crude exposure-disease association 2. stratify E-D association by levels of the third variable 3. Calculate the stratum specific RRs/ORs 4. Evaluate for confounding first (evaluate whether crude is outside stratum specific interval, calculate an adjusted RR/OR, if the crude and adjusted are different by 10% report the adjusted value) 5. evaluate for effect modification (evaluate the stratified RRs/ORs for similarity, be aware of differences vs statistical, if OR/RR are not uniform across stratified data, you may have effect modification, there is no need to calculate an adjusted RR/OR if you detect effect modification, report the stratum specific values and describe your findings, remember there is usually a biologic reason that the effect is limited to a subgroup of people) -Note: report confidence intervals (and P values) with every RR/OR

Guidelines for judging whether an observation association is causal (Hill's casual criteria)

1. Strength of the association 2. Biologic credibility/plausibility 3. consistency with other knowledge/research 4. temporality 5. dose-response relationship 6. replication of findings 7. consideration of alternate explanations 8. cessation of exposure 9. specificity of the association

Summary of evaluating causality

1. multiple philosophies exist for evaluating causality. none is definitive 2. the set of casual criteria offered by Hill are useful, but also saddled with reservations and exceptions 3. always keep open mind when evaluating evidence from epidemiological studies

Henly (1840s) Koch (1880s) postulates for casuation

1. the organism is always found with the disease 2. the organism is not found with any other disease 3. the organism when isolated from one who has the disease and cultured through several generations, produces the disease (in experimental animals) Koch: even when an infectious disease cannot be transmitted to animals, the regular and exclusive presence of an organism (postulates 1 and 2) proves a casual relationship

When is the odds ratio a good estimate of the relative risk?

1. when the cases studied are representative with regard to history of exposure, of all people with the disease in the population in which the cases were drawn 2. when the controls studied are representative with regard to history of exposure, of all people without the disease in the population from which the cases were drawn 3. when the disease being studied does not occur frequently

odds ratio

AD/BC

AR% for case control study

AR% = (OR-1)/OR x 100

2. If us is neither sufficient nor necessary

C/A -> disease U/A -> disease -not sufficient because the exposure must also be with A -not necessary because the exposure C and A and that disease pathway is independent

Does smoking modify the relationship between alcohol consumption and larynx cancer?

CRUDE RR: 3.46 Non smokers RR: 1.96 Smokers RR: 3.60 -unlike the assessment of confounding, the crude estimate is not used to evaluate the presence of effect modification -instead the stratum-specific estimates are compared directly to see if they are different (heterogenous) -this example suggests risk ratio heterogeneity

advantages of matching

allows for strict control of confounding -unlike restriction, allows one to look for effect modification -is intuitvely appealing -good for small samples since multivariate analysis may not be possible -good when a non matched design would provide little overlap for confounders (i.e. matching cases to sibling controls for many environmental and genetic factors that couldn't control for using an unmatched design)

confounding (treated differently in analysis)

can do more to fix this than the other two

differential misclassification (non random)

can underestimate/overestimate/correctly estimate (by chance) observed association

evaluating casual associations

causality: a philosophical concept merged with practical guidelines, we can never prove causality we infer it -the presence of a valid statistical association does not imply that is a casual one (e.g., gastric acid and peptic ulcers) -a judgement of causality must be made in the presence of all available data and reevaluate with each new finding -different criteria and philosophical views have been proposed to assess causality

risk ratio

cumulative incidence of exposed/cumulative incidence of non exposed

Inductive-oriented criteria (Hill 1965)

employ a common set of criteria to attempt to distinguish casual from non casual associations

**i don't think we need to know this but** temporal bias (reverse causality bias)

errors resulting from errors in collecting information on exposure and outcome (e.g. at the same time )

Calculating population attributable risk

ex: we want to measure how much of the burden of diabetes among people living in Austin is due to obesity -caution: in order to calculate PAR and PAR%, we have to be reasonably sure that the results of the study can be generalized to the population of Austin

**i don't think we need to know this but** exclusion bias

exclusion criteria applied differently for cases and controls or exposed and non exposed

remember

for a cohort study both the odds ratio and relative risk are valid measures of association -for a case control study only the odds ratio can be calculated directly -under certain conditions the odds ratio in a case control study can be a good estimate of the relative risk

incidence prevalence example: 2b

hypothetical situation: the investigator influenced by his knowledge of exposure status, tends to gather cases mainly among individuals known to be aspirin users (inflating the aspirin users box) OR = 3.15 -to avoid this problem, the selection process of cases and controls from the target population should be rigorously the same and should be independent from the exposure status (i.e. the investigator should be blinded to the exposure status) -has hypothesis that maybe people who use aspirin have a higher chance of having ESRD -investigaor over recuirs people who are aspirin users to test this - cell A is flooded -now we have a bias away from the null - positive bias away from the null

reality

in reality we describe what we see -additive effect -multiplicative effect -more than additive but less than multiplicative -more than multiplicative -less than additive -synergistic effect -antagonistic effect

Rate ratio

incidence density of exposed/incidence density of non exposed

bias towards the null

observed value is closer to 1.0 than is the true value

bias away from the null

observed value is farther from 1.0 than is the true value

Positive bias

observed value is higher than the true value

negative bias

observed value is lower than the true value

odds ratio among cases

odds = probabbility that cases have the exposure / probability that cases do not have the exposure

**i don't think we need to know this but** Non response bias

one of the comparison groups is more/less likely to participate in the study

5. Biologic plausibility of the hypothesis

pro: a known or postulated biologic mechanism by which the exposure might reasonably alter the risk of developing the disease is intuitvely appealing -con: plausibility is often based on prior beliefs rather than logic or actual data con: what is considered biologically plausible at any given time depends on the current state of knowledge (as we move forward in time and advance knowledge- things we used to believe change)

odds ratio among controls

probability that controls have the exposure/probability that controls do not have the exposure

but wait

recall that we cannot calculate incidence directly for exposed and non exposed populations when we conduct a case control study -odds ratio

Information bias (misclassification bias)

results because the subject's disease or exposure or both were misclassified

Selection bias

results from the way subjects were recruited into the study (design phase or inclusion or exclusion of subjects) that affects the measure of association

Falsification (Karl Popper - 1959)

scientific hypotheses can never be proven or established as true. Therefore, science advances by a process of elimination (falsification)

Non-differential misclassification (random)

the errors in misclassification of exposure or disease status are independent of the level of the other variable -measurement error in exposure does not depend on outcome -measurement error in outcome does not depend on exposure -can apply to everyone in the group

**i don't think we need to know this but** Hawthorn effect

the patient changes behavior because they know they are in a study

Non differential misclassification (random)

will always underestimate observed association (groups appear more similar) -bias toward the null

multivariate analysis equation

y = outcome a = intercept (B0 is sometimes used instead of alpha) B = coefficients X1 = independent variable outcome = intercept + exposure + confounder 1 + .... + confounder n