Occ Epi
Cumulative Incidence
# of events / average # of persons over that time period
Prevalence
# of existing events in a timeframe / average populatin in that time
Incidence Rate
# of new events in a timeframe / # of person-time units at risk
Nested Case-Control Study
- A hybrid study design in which a case-control study is nested in a cohort study. - More efficient to obtain dataset of exposure => lower cost - Controls are sometimes obtained from a different population
Cross-Sectional Study
- A study in which a representative cross-section of the population is tested or surveyed at one specific time. - Main outcome measure is prevalence - Relatively inexpensive and fast - Can study multiple outcomes - Prior disease status/state is often susceptible to recall bias - Exposure status is nearly always susceptible to recall bias - Nearly impossible to establish temporality - Can be a good pilot for a cohort study
Types of Selection Bias
- Ascertainment - Berkson's -Design - Detection of Patient Surveillance - Response - Sampling - Workup
Retrospective Cohort Study
- Begins at end and looks back at records taken in the past - Faster and less expensive than prospective cohort studies - Difficult to ensure a disease-free baseline - Less adequate follow-up of subjects
Proportional Mortality Ratio (PMR)
- Commonly used to study disease patterns by cause in settings where population denominators are not available. - Is the ratio of PM's in the two comparison populations. - PMR = PM Pop A / PM Pop B - Indicates the burden of disease within a population
Non-epidemiological / anecdotal study designs:
- Consecutive Case Series - Case Reports - A Single Case Report
Three Ways that Censoring Occurs
- Death - Loss to follow-up - End of study observation
Types of Information Bias
- Design - Interviewer - Lead-time - Length - Measurement - Observer - Recall - Reporting
Examples of Information Bias
- Digit Preference - Instrumental Error - Handling Outliers - Of an estimator
Case-Crossover Study
- Each subject serves as their own control - Subjects cross over between periods of risk and lack of risk
Bonus question, 4 points. When evaluating an exposure-outcome relationship, there are options for the impact of a 3rd variable. What are they? - Effect Modifier - Gender - Bias - Intermediary - Selection - Confounder - No relationship - Group - Standardized measure - Secondary exposure
- Effect Modifier - Intermediary - Confounder - No relationship
Retrospective Data Sources
- Employment personnel records - Job descriptions - Occupational hygiene monitoring data - Process descriptions and flow diagrams - Plant production records - Inspection and exposure release reports - Engineering control and protective equipment documentation - Biological monitoring reports
Systematic Error
- Error that shifts all measurements in a standardized way - Decreases accuracy. - Can result in bias - May invalidate either a positive or negative result
Random Error
- Is not a bias - Does not invalidate positive results - May invalidate negative results - Can be mitigated with larger sample size
Assumptions for Linear Regression
- Observations are statistically interdependent of each other - Random errors are normally distributed - Deviation of the observed values from the fitted values remains similar for all values of X (homoscedasticity)
Bias
- Result of a systematic error which results in false or misleading conclusions -Can either falsely raise or lower the estimates of risk
Hypothesis generating study designs:
- Retrospective Cohort - Case-Control - Cross-Sectional - Ecological - Proportionate Mortality Rate
Types of Bias
- Selection Bias - Information Bias - Confounding
Case-Control Study
- Study population defined by disease status (case vs control) - Strongest study design for rare outcomes - Temporality is difficult/impossible to establish - Shows associations or correlations rather than true Risk Factors - Faster than cohort studies - More efficient - Lower cost - Can only measure one disease at a time and cannot measure prevalence - Can be difficult to find suitable controls - Exposure is always assessed retrospectively - Matching between cases and controls removed confounding factors - Provides Odds Ratios
Prospective Cohort Study
- Study population defined by exposure status - Follow a population forward in time - Observe who develops the disease and when - Can be used for relative risk - Ca evaluate multiple outcomes simultaneously and shows temporality - Expensive and time consuming
Confounding
- Third variable related to the exposure and the disease. - Produces biased estimate of the effect of the exposure of interest
Examples of Selection Bias
- Withdrawals - In autopsy series - Healthy worker effect
JEM Process
1. Classify industrial processes broadly 2. Sub-divide to get common types of exposures in each group 3. Get job titles to sort the jobs into each of these distinct exposure groupings 4. Develop a Jem 5. Develop as many distinct JEMs as there are exposures to study 6. Different measures of exposure may be used but consistency is important
Assessment Hierarchy
1. Quantified personal measurement 2. Quantified area measurements in the vicinity of the activity 3. Quantified surrogates of exposure 4. Distance from the site and duration of exposure 5. Distance or duration of residence 6. Residence/employment in the geographical area in reasonable proximity to the assumed exposure 7. Residency/employment in the defined geographical area of the site
Hawthorne Effect
A change in a subject's behavior caused simply by the awareness of being studied
Bias Due to Digit Preference
An error especially present in rounding off numbers that produces more responses in one category than is random
Bias Due to Instrumental Error
Any lab based systematic error
When you standardize rates to a population, your rates are then comparable to: - That population which you are standardizing to - Other rates that are also standardized to the same population - All other standardized rates - Both 1 and 2 - None of the above
Both 1 and 2
You conduct a study to evaluate the rate of cancer death between firefighters and police officers in the state of Utah. You directly standardize to the 2000 United States Census population. Crude Death Rate per 100,000 person-years Age-adjusted death rate per 100,000 person-years Police Officers 85 238 Firefighters 284 204 US Census 84 84 What can you conclude from these results? - Firefighters have a higher death rate from cancer as compared to the US population - Police officers have a higher death rate from cancer as compared to the US population - Firefighters have a lower death rate from cancer as compared to the US population - Police officers have a lower death rate from cancer as compared to the US population - Both 1 and 2 - Both 3 and 4 - None of the above
Both 1 and 2
You conduct a study to evaluate the rate of cancer death between firefighters and police officers in the state of Utah. You directly standardize to the 2000 United States Census population. Crude Death Rate per 100,000 person-years Age-adjusted death rate per 100,000 person-years Police Officers 85 238 Firefighters 284 204 US Census 84 84 What would you report for the findings from this study? - Both Crude and Age-adjusted rates - Only Crude rates - Only Age-adjusted rates - Neither Crude nor Age-adjusted rates - Only calculated differences in death rates
Both Crude and Age-adjusted rates
Measurement Bias
Broad category including any misclassification based upon measurement/classification of exposure(s)
What is a p-value, and what does it mean? - Calculated probability of an effect as large or larger than the observed effect, assuming the null hypothesis is true - The probability of rejecting the null hypothesis. - Probability of replicating these study results in a study conducted on a similar population with a similar study design. - The potential for committing a type I or type II error based on the clinical effect and the sample size of the study. - The probability of accepting the null hypothesis.
Calculated probability of an effect as large or larger than the observed effect, assuming the null hypothesis is true
Berkson's Bias (Selection)
Case-Control Study Only - Hospitalized cases and controls may not be similar in exposures
If you have RANDOM error in your study your calculated risk estimate (Odds Ratio, Risk Ratio, etc.) will always be _________ compared to the true risk estimate (the risk estimate that would be present if there was not random error). - Farther from the Null (1.0) - You Answered Lower - Closer to the Null (1.0) - Higher - It depends on the direction
Closer to the Null (1.0)
The three main categories of bias are: - Random, systematic and selection - Confounding, selection and information - Reporting, selection and confounding - Systematic, confounding, and recall - Information, observer, and recall
Confounding, selection and information
When you standardize you: - Create a fair comparison between groups by adjusting for their distribution - Create an unfair comparison between groups by adjusting for their distribution - Create a measure that can be compared to all other standardized rates. - Create a measure that has no meaning when compared with other rates - Create a unfair comparison by artificially imposing a specific distribution on your population.
Create a fair comparison between groups by adjusting for their distribution
Additive vs Multiplicative EM
Depends upon measure of association - Ratio measure (multiplicative interaction) - Difference measure (additive interaction) - Absence of multiplicative interaction implies presence of additive interaction
Response Bias (Selection)
Differences in the subjects volunteering for a study
Biased Follow-up (Selection)
Different rates of follow-up OR different aggressiveness of follow-up in seeking disease status
Recall Bias
Differential remembrance of exposure
What is the best method of standardization? - Direct - Indirect - Unbiased - Unconfounded - Approximate
Direct
Case-Cohort Study
Each individual in one group is matched to an individual in another group who shares as many characteristics as possible
Bias in Handling Outliers
Either including (that which should have been excluded) or excluding (that which should have been included) outlier variables
Lead-Time Bias
Error, especially in survival estimates that occurs from the early detection of disease
What is type II error? - False rejection of the null hypothesis when the null hypothesis is true - False acceptance of the alternative hypothesis when the null hypothesis is true - Stating that there is a statistically significant relationship when there truely is no relationship - Having too large of a sample size which results in over fit of the model and artifically small confidence bounds. - False acceptance of the null hypothesis when the null hypothesis is false
False acceptance of the null hypothesis when the null hypothesis is false
Selection Bias
Individuals have different probabilities of being included in the study sample according to relevant study characteristics.
Interviewer Bias
Interviewer's sub/conscious errors in the collection of data
If you have SYSTEMATIC error in your study your calculated risk estimate (Odds Ratio, Risk Ratio, etc.) will always be _________ compared to the true risk estimate (the risk estimate that would be present if there was not random error). - It depends on the direction - Lower - Closer to the Null (1.0) - Farther from the Null (1.0) - Higher
It depends on the direction
What is an alpha level used for? Choose only the best answer. Multiple answers will not be accepted. - It is always 0.05 which is the acceptable standard set by Fischer. - Your level of statistical significance for a test - The probability of committing a type II error if you have a sufficient sample size - It is the level at which you are willing to accept type I error - It is essentially the same as your p-value. It tells you your probability for that statistical test.
It is the level at which you are willing to accept type I error: The alpha level is also the level of type I error you are willing to accept. You compare your p-value to your alpha level to see if you accept or reject the null hypothesis: You set your alpha level before you begin analyses to compare your p-value to after analyses. This will determine whether you reject or fail to reject the null hypothesis.
A study is being developed to determine if the early detection of heavy metal neurological disorders with a manditory screening using a chelating test (a blood test) will result in decreased incidence of acute lead toxicity in a brazing plant (joining metal pieces together using lead solder). This study is most likely subject to which of the following biases: - Reporting Bias - Migration bias - Recall bias - Length time bias - Lead time bias
Lead time bias
What are the 4 ways to control for confounding? - Match, Exclude, Stratify, Statistically Adjust - Minimize, Evaluate, Standardize, Spread - Reduce, Consider, Adjust, Fabricate - Control, Consider, Correlate, Conclude - Analyze, Design, Scrutinize, Combine
Match, Exclude, Stratify, Statistically Adjust
When you utilize multivariate logistic regression, that includes 3 variables, 1 of which is an exposure variable and two of which are confounders. What do you get from the model regarding the relationship between the exposure and the outcome (hint: it's what you get when you exponentiate the negative beta estimate)? - Odds Ratio - Goodness of Fit - Integrated Case Prediction - Sum of Squares - Maximum Likelihood
Maximum Likelihood
Choosing to use multiplicative or additive measures.
Multiplicative - Favored measure when looking for causal association Additive - Readily translated into impact of an exposure or intervention in terms of absolute number of outcomes prevented
Detecting Multiplicative Effect Modification
No Multiplicative effect modification - Multiplicative EM looks at risk estimate - Risk difference is not the same Multiplicative Effect - Multiplicative EM looks at risk estimate - OR's are meaningfully different
When you match or exclude to control confounding, can you analyze your results for that factor? For example, if you match on gender in a case-control study looking at artificial growth hormone and breast cancer, can you analyze a relationship between gender and breast cancer after adjusting for artificial growth hormone exposure? Why or why not? - No because you don't have a large enough sample size to detect a difference with an alpha of 0.05 and a beta of 0.10. - Yes because you collected that information and it may be a potential confounder. - Yes, you can still stratify or statistically adjust to control for residual confounding. - Yes because there is enough distribution to measure a potential relationship. - No because you matched (or excluded) that factor so there is no difference between your two groups.
No because you matched (or excluded) that factor so there is no difference between your two groups.
You suspect that job satisfaction may be confounding the relationship between breathing in particulate matter and developing influenza like illness because the particles inhibit the ability of the lungs to remove virus particles from the lung. You calculate an unadjusted odds ratio between particulate matter and influenza like illness and have an odds ratio of 3.32 (95% CI 2.84, 4.05). You then stratify for job satisfaction and find that if individuals are satisfied, the Odds Ratio is 3.29 (95% CI 2.24, 5.56) and if they are unsatisfied, the odds ratio is 3.38 (95% CI 2.55, 4.47). Based on these data is job satisfaction a confounder in the relationship between particulate matter and influenza like illness? Why or why not? - Yes, because there is a large enough difference between the stratified odds ratios. - No, because including a measure of job satisfaction is considered a selection bias. - No, because the strat
No, because the stratified odds ratios are similar and the crude odds ratio is in the middle of the two.
Odds Ratio
Odds of event in Treatment group / Odds of event in Control group = (a/b) / (c/d) = ad/bc
Length Bias
Over selecting long-duration cases in a cross-sectional or case-control study
Proportional Mortality
PM = # of deaths due to cause / Total # of deaths
Hypothesis testing study designs:
Prospective Cohort
Types of Error
Random & Systematic
In a case-control study where participants are selected to be in the study based on their case status (diseased or not diseased) for occupational cancer and then asked to report prior exposures on an online questionnaire is most subject to what kind of bias? - Recall bias - Berksons bias - Interviewer bias - Confounding - Participation bias
Recall bias
If a study has high reliability then it is: - Close to the truth - Dependable - Reproducible - Free from bias - Drawn from a large population
Reproducible
Risk Ratio (Relative Risk)
Risk of event in the Treatment group / Risk of event in the Control group = a/(a+b) / c(c+d)
Reporting Bias
Selective reporting or suppression of the past medical history
If the distribution of the population among the strata are different and the stratum specific rates are similar then: - Standardization is not needed - Standardization is indicated - Standardization would over-emphasize the difference in the distribution of the two populations. - Standardization would introduce bias - Standardization would create hypothetical rates that have no meaning
Standardization is indicated
For both indirect and direct standardization you need several pieces of information, including total rate for the sample and distribution of your standardization variable (often age) of your standard population. What additional piece of information do you need for direct standardization that you don't need for indirect standardization? - Attributable Fraction in your standardization population - Relative Risk in your standardization population - Odds Ratio in your standardization population - Standardization population-specific rates of disease - Sufficiently large true population to standardize to, such as 2000 US Census population
Standardization population-specific rates of disease
When you have an estimate of risk from a multivariate regression, your risk estimate is - Statistically adjusted for other variables in the model - Controlled for all potential confounders - The closest to the truth - Further from the null
Statistically adjusted for other variables in the model
What are the 4 ways to control for confounding, and when in the study process do you implement them? - Exclude and Stratify after the data is collected; Statistically adjust and Match when designing the study. - Match, Exclude, and Statistically Adjust when designing the study; Stratify after the data is collected. - Stratify and Statistically Adjust after data is collected; Exclude and Match when designing the study. - Statistically adjust and Stratify when designing the study; Exclude and Match after the data is collected. - Stratify and Match when designing the study; Statistically adjust and Exclude after the data is collected.
Stratify and Statistically Adjust after data is collected; Exclude and Match when designing the study.
In an evaluation of the possibility for occupational lead exposure to be a risk factor for the development of neuropathy, two cohort studies were performed. Both studies reported the relative risk with 95% confidence intervals for the development of neuropathy in association with occupational lead exposure. In the first study (Study A), the relative risk and 95% confidence interval was 3.24 (0.38 - 7.41, p=0.683). In the second study (Study B), the relative risk and 95% confidence interval was 2.06 (1.34 - 2.97, p=0.015). The alpha level set a priori (before the study began) was 0.05 and the beta level was 0.10. The potential for type I statistical error is possible in which study(ies).? - Not enough information provided regarding the potential for type I error - Study A - Study B - Neither Study A nor Study B - Both Study A and Study B
Study B
Information Bias
Systematic Tendency for individuals selected for inclusion in the study to be erroneously placed in different exposure/outcome categories, leading to misclassification
As study size increases what generally happens to systematic and random error? - Systematic error stays the same and random error goes down - Both random and systematic error go up - Random error stays the same and systematic error goes down - Systematic error goes up and Random error goes down - Both random and systematic error go down - Random error goes up and systematic error goes down - Systematic error and random error stay about the same.
Systematic error stays the same and random error goes down
Observer Bias
Systematic errors in observation that occur because of an observer's expectations
Effect of Bias
Systematic mis-classification of exposure status - If differential mis-classification, then the estimate of risk may be either falsely elevated/protective (to a greater degree) - If the mis-classification of exposure status is non-differential, then the bias is towards the null
Placebo Effect
The phenomenon in which the expectations of the participants in a study can influence their behavior
Multiple Linear Regression
The regression coefficients are now 'partial' in that it signifies the amount of linear relationship of one independent variable to the dependent variable with all others in the equation.
In an evaluation of the possibility for occupational lead exposure to be a risk factor for the development of neuropathy, two cohort studies were performed. Both studies reported the relative risk with 95% confidence intervals for the development of neuropathy in association with occupational lead exposure. In the first study (Study A), the relative risk and 95% confidence interval was 3.24 (0.38 - 7.41, p=0.683). In the second study (Study B), the relative risk and 95% confidence interval was 2.06 (1.34 - 2.97, p=0.015). The alpha level set a priori (before the study began) was 0.05 and the beta level was 0.10. Which of the following statements is correct? - The results in both study A and study B would lead the investigators to reject the null hypothesis for both studies. - There is insufficient evidence for any conclusive decisions regarding accepting or rejecting the null hypothesis for either study. - The res
The results in study B would lead the investigators to reject the null hypothesis.
Chi-Square Test of Homogeneity
The test is applied to a single categorical variable from two or more different populations. It is used to determine whether frequency counts are distributed identically across different populations.
Health Worker Effect (Selection)
Those employed are generally healthier than the general population
Detection Bias (Selection)
Unequal case status determination due to - Ascertainment - Diagnosis - Verification problems
Ascertainment Bias (Selection)
Unequal inclusion of cases and controls
Least Squares Regression
Using minimum mean squared error as the criterion
This is a bonus point question, worth 2 bonus points. When you are designing a case-control study the best way to reduce recall bias is to - Abandon the study as recall bias is a fatal flaw - Utilize multiple sources of data - Rely on questionnaire data for all participants, not just controls. - Rely on questionnaire data only for controls, because they are disease-free. - There is no way to reduce recall bias, the best you can do is acknowledge it as a weakness
Utilize multiple sources of data
Bias of an Estimator
When an estimate is used as a surrogate, this is a systematic difference between that estimator and the true value
You conduct a cross-sectional study and calculate to assess the relationship, if any, between dust containing lead and memory loss. Your "exposed" group are demolition contractors who are potentially exposed to dust containing lead while demolishing houses built before 1978. Your unexposed group are concrete cutters who are exposed to dust, but that dust is primarily silica and does not contain lead. The crude odds ratio is 2.57 (95% CI 1.54, 3.11). You stratify based on smoking status (the theory is that smokers have reduced ability to remove particulate from the lungs so their "dose" is higher, cigarettes do not contain lead). The stratified odds ratios for smokers and non smokers are 5.80 (95% CI 3.88, 7.59) and 3.75 (95% CI 2.17, 5.22). Is there confounding present? - Yes, there is negative confounding present. - Yes, there is positive confounding present. - Yes, there is qualitative confouding present. - No,
Yes, there is negative confounding present.
A true confounder has to be related to - neither the exposure nor the outcome - the detection or measurement of the exposure - the detection or measurement of the disease - both the exposure and the outcome - selection bias - only the outcome - only the exposure
both the exposure and the outcome