Epidemiology Lecture 6 - Misclassifications

Choosing a Test/Measure

(Almost) all measures are imperfect Choice of test will depend on: - degree of measurement error - study resources *Quantifying AND Qualifying measurement error helps to interpret study results*

Assessing Measurement Error: *Reliability*

*Degree of agreement between multiple measures of the same characteristic on the same study subject* Answers the question: "How reproducible or consistent is the measure?"

Assessing Measurement Error: *Validity*

*Degree to which a measure yields the correct value of underlying characteristic* Answers the questions: - "Does your test measure what it is supposed to measure?" - "How close is the measure to the truth?"

Cohort Study Validity: Example

*Hypothesis*: Smoking marijuana is associated with higher risk of losing wallet within 1 year.

Hypothetical abstract: "We found a 5-fold higher risk of heart disease among men who drink more than 3 alcoholic beverages per day..." "In conclusion, excessive drinking may cause heart disease"

*Methods*: ... number of drinks per day obtained from an employment-based survey ... heart disease ascertained by questionnaire. Heart disease was presumed if the subject responded 'yes' or 'possibly' to the question, "Have you ever had a heart attack or stroke" *Reality*: 5-fold greater risk of checking the box next to the question "have you ever had a heart attack or stroke?" among men who self-reported >3 drinks/day on a work-related survey, compared to men who self-reported no drinking

Example Scenario 3 DIFFERENTIAL Misclassification 1: *Hypothesis*: Exposure to mildew in home is associated with higher risk of asthma development in young adulthood 40% of truly unexposed subjects that are diseased report being exposed

*Truth*: 15/24 exposed got disease 5/25 unexposed got disiease RR = 3 *Misclassification*: 17/29 "exposed" got disease 3/21 "unexposed" got disease RR = 4 *Relative Risk Increases, pushed away from the null*

Practice Example 2: Investigators identify 40 pre-term newborns and 120 to-term newborns - Ask mothers whether they took multivitamins during pregnancy - "We found multivitamin use to be associated with a greater risk of a pre-term birth (OR 1.6, 95% CI 1.1, 4.0)" 1.) *Did misclassification occur here? Of what?* 2.) *What type of misclassification do you expect to have occurred?* 3.) *What effect will misclassification have on the observed results?*

1.) Case-Control study. YES misclassification of exposure - mothers knew the outcome when they answered the Q, which may have affected their propensity toward telling the truth 2.) Differential classification 3.) Unable to tell bc it is differential

Practice Example 1: Investigators identify 20,000 pregnant women and ask them, during their first trimester, whether they are currently taking multivitamins - "We found multivitamin use to be associated with a decreased risk of a major birth defect (RR 0.8, 95% CI 0.7, 0.9)" 1.) *Did misclassification occur here? Of what?* 2.) *What type of misclassification do you expect to have occurred?* 3.) *What effect will misclassification have on the observed results?*

1.) Cohort study. YES misclassification of exposure - self-reported (may be over or under reported) 2.) Non-differential misclassification bc at the time they ask about vitamin use they don't know yet if the infant will have a birth defect or not 3.) RR pushed towards Null

Misclassification Key Points

1.) Misclassification occurs when measurement tools do not accurately capture the study data 2.) *Non-differential misclassification* represents random error and results in finding relative risks that are closer to 1.0 3.) Can NOT dismiss a positive study on the basis of non-differential misclassification - a negative study result you may be able to dismiss but NOT a positive one since nondiff pushes RR Towards null rather than away from it 4.) *Differential misclassification* represents preferential error and results in unpredictable change in the relative risk

Tests and Measures Used in Clinical Research

1.) Responses on a self-administered questionnaire - "How often do you exercise in a usual week?" 2.) Answers to an interview question - "Has a doctor ever told you that you have diabetes?" 3.) Conditions recorded in medical records - "History of MI" 4.) Diagnosis codes in a registry database - ICD-9 code 410.x for MI 5.) Laboratory result - CRP level

Non-Differential Misclassification

Degree of misclassification of the exposure is the same among subjects with and without the outcome *Consequences*: 1.) Non-differential misclassification produces a relative risk that is closer to 1.0 than the true relative risk ("bias toward the null") 2.) More misclassification more movement towards the null 3.) Non-differential misclassification can (and often does) lead to falsely null results

Measurement Error (Definition)

Difference between the observed value of a quantity and its true value

Misclassification and Measurement Errors

Errors in measurement or classification inevitable in any study, but, errors tend to fall into small number of categories - Once you know what kind of error it is, it will *Lead to predictable changes in study results*

Information Bias (Definition)

Impact of misclassification or measurement error on study results and inference

Assessing Validity

In order to assess validity, an *error-free measurement (i.e., "gold standard") is needed as a reference point* - Need to apply the gold standard and the measure in question to a set of study subjects

Misclassification (Definition)

Incorrect classification of an individual into category other than that to which they should be assigned (e.g. classifying someone as no-fracture when they did have a fracture)

Critique of a Research Article

Is the study hypothesis relevant? Is the study population appropriate? *How were the data collected?* Are the analyses valid? Is confounding present? Are the results applicable? Statistical inference?

Assessing Validity: Example

Kidney pts taking their meds or not? - Digoxin has higher % agreement and would be a more valid inventory agreement than aspiring consumption

Impact of Misclassification in the Marijuana Example

Misclassification led to a relative risk closer to 1.0 - When groups are mixed together, they look more alike - Key assumption in this example: misclassification of exposure occurring equally across outcome groups *Non-differential misclassification of the exposure*

Differential Misclassification of Exposure

Misclassification of the exposure differs among subjects with and without the outcome Relative risk can change towards or away from 1.0 (the null) - Depends on the situation - Not predictable

Cohort Study Validity: Example2 Self-Reported Asbestos

More realistic situation with misclassification - Self-reported asbestos exposure measurement: 30% of asbestos exposed miners do not know they were exposed Observed data indicated incidence rate among truly exposed stays the same but increases in those who were reported unexposed but really WERE exposed

Cohort Study Validity: Example2 True Asbestos Exposure *Hypothesis*: Asbestos exposure is associated with higher risk of laryngeal cancer

Recruit 1,500,000 miners and measure asbestos exposure *Gold-standard asbestos exposure measurement* Air in miners' work area drawn through filter to capture asbestos fibers; Filter is then analyzed by phase contrast microscopy - True prevalence of asbestos exposure: 2/3 *Interpretation*: Among this study population of miners, asbestos exposure was associated with a 5-fold greater risk of developing laryngeal cancer.

Cohort Study Validity: Example Scenario 1 *Hypothesis*: Smoking marijuana is associated with higher risk of losing wallet within 1 year.

Recruit 25 TRUE MJ users and 25 true non-users (as measured by blood THC) -> follow for 1 year - 15/25 users have lost wallet = 60% incidence proportion - 5/25 non-users lost wallet = 20% incidence proportion Relative Risk of Loosing your wallet as a user? RR = 3 = MJ users are 3X more likely to use their wallet than non-users.

Cohort Study Validity: Example Scenario 2 *Hypothesis*: Smoking marijuana is associated with higher risk of losing wallet within 1 year.

Recruit 25 reported MJ users and 25 reported non-users (as measured by drug checklist) -> follow for 1 year *say 20% of truly exposed report being unexposed lie/forget* - 12/20 users have lost wallet = 60% incidence proportion - 8/30 non-users lost wallet = 27% incidence proportion Relative Risk of Loosing your wallet as a user? RR = 2.2 = MJ users are 2.2X more likely to use their wallet than non-users. *Relative Risk is Decreased due to misclassification*

Cohort Study Validity: Example Scenario 3 *Hypothesis*: Smoking marijuana is associated with higher risk of losing wallet within 1 year.

Recruit 25 reported MJ users and 25 reported non-users (as measured by drug checklist) -> follow for 1 year *say 40 % of truly unexposed report being exposed* - 14/30 users have lost wallet = 47% incidence proportion - 6/20 non-users lost wallet = 30% incidence proportion Relative Risk of Loosing your wallet as a user? RR = 1.6 = MJ users are 1.6X more likely to use their wallet than non-users. *Relative Risk is Decreased due to misclassification*

Information regarding misclassification and measurement error is found...

Under Data Collection: - Exposure - Outcome - Other Study Data

Example Scenario 3 DIFFERENTIAL Misclassification: *Hypothesis*: Exposure to mildew in home is associated with higher risk of asthma development in young adulthood

What if we have DIFFERENTIAL misclassification of the exposure? - 40% of truly unexposed subjects that are diseased report being exposed - 10% of truly unexposed subjects that are not diseased report being exposed The misclassification of the exposure is differential based on disease/outcome status

Non-differential misclassification of the exposure

When there is misclassification of the exposure that does not differ across outcome groups it leads to a relative risk that is muted and pushed towards the "null"