10-11: Random error, bias and confounding

Question 1

Random error
definition
causes
counteractions

Answer 1

Random error
- defines precision
- divergence of a measurement from the truth just by chance alone
- can be handled

Causes
- biological variation
- measurement error
- other unknown factors with impact

Counteractions

• sample size
• control sources (standardized data collection, precise instruments)
• exclusion / inclusion criteria
• quantifiy uncertainty (SD,SE, CI, p-value)

Question 2

Bias
definition
leads to

Answer 2

definition
any systematic error in the design, conduct or analysis of a study that result in a mistaken estimate of an exposures effect on the risk of disease

leads to under or overestimation of OR

Danger of bias
- bias is present, but researcher isnt aware
- identification of causes is difficult
- in contrast to lab exp., control options are limited

Counteractions
yes: consideration of possible biases in study design, and interpretation of result, recognition of biases when reading study result

no: sample size, special analyses

Question 3

Selection bias
definition
main causes

Answer 3

Definition
results from selection process of the study participants from the population of interest

Main causes
Cross-sectional study:
* unsuccessful contacting,
* refusal of participation,
* unuseable forms, etc.

Case-control study:
* incomplete recruitment of cases,
* unrepresentative controls for the population of interest
Cohort study:
* Follow-up:
* change of residence, death, drop-out

Counteractions
- repeated contacting
- incentives
- Multiple ways of contacting

Measures
OR influenced by W (weight)
W > 1 -> overestimation
W < 1 -> underestimation
W = 1 -> no bias

Question 4

Types of selection bias
Counteraction

Answer 4

Response bias
- bias by systematic non partcipation (low participation of unexposed healthy, high participation of exposed diseased)

Admission rate bias
- case control study bias
- chance of exposed cases being admitted to the study is differenzt to exposed controls (e.g. exposure may be not representatively distributed in control, so you have too many exposed controld)

Migration bias
- due to systematic migration between comparison groups (cross sectional)
(e.g. people with res. disease move from areas with high air to low air pollution)
- Healthy worker effect:
vulnerable or diseased persons are not exposed to straining conditions or leave them early (e.g. straining tasks in mine done by young healthy workers)

Counteraction
Problem: W is unknown
Analyse:
- sample characteristics
- non responder analysis (associations with exposure / outcome)
- does not corect bias, but gives idea where bias might have occured and led to under or overestimation

Question 5

Sources of error in classification / measurement
and Counteractions

Answer 5

Participants
* Inability to understand or articulate
* Inability to recall
* Intentional misinformation/ unwillingness to disclose, or social desirability
* “Central tendency“ in
questionnaires (extreme answers of questions with predefined answer categories are rarely chosen)

Data collector
* Unclear or ambiguous questions
* Expecting certain results
* Lack of a neutral demeanor
* Inaccurate transcription
* Fraud

Data managers
misreading, coding, programming errors, follow up mechanism incomplete, inacuurate transcription

Data analyst
variable coding / programming errors, invalid statistical method

Data interpreter
inadequate appreciation of characteristics of the measures or the relations being studied

Counteractions
- pre-test and vailidity studies
- standardized data collection
- staff training, certifications
- multiple independent data collections
- precise operationalisation of variables
- audit of data centres
- data cleaning
- plausibility controls
- peer review
- reanalysis before publication
- critical appraisal

Question 6

Different types of misclassification

Answer 6

Nondifferential misclassification (MC)

• MC of exposure independent from disease status
• MC of diseases independent from exposure
• OR gets closer to 1

Differential misclassification (MC)
- Measurement of exposure depends on the disease status (retrospective)
- Measurement of disease depends on expsoure status (prospective)
- OR over or underestimated

Question 7

Types of Information bias

Answer 7

Recall bias
- Information such as exposure ma be better recalled by a case but forgotten by a control
- differential
- e.g. case control studies, mothers of children with eye cataract and exposure rubella duing pregnancy

Counteraction:
- objective observation methods
- strucutred standardized interviews

Interviewer bias
- different interviewing of cases and controls
- differential MC of exposure status
- e.g. case control studies, interviewer convinced of harmfull effects of exposure

Counteractions:
- blinding of interviewer
- good training of interviewer
- standardized interviews with predefined answers

Detection bias
- due to better diagnosis in exposed people
- differential MC of disease status
- e.g. cohort studies, people with exposure visit doctor more frequently

Diagnostic suspicion bias
- differential MC of the disease
- more thorough examination if exposed
- e.g. cohort studies where knowledge of exposure is used as diagnostic criteria

Question 8

Confounding

Answer 8

• Confounder causes spurious association between exposure and outcome
• age, gender, status, smoking
• has effect on disease
• has an effect on exposure
• is not only a result of exposure
• is not only a step between exposure and disease

Counteractions

By study design
- Randomization
controlling known and unknown factors

• Restriction
  inclusion criteria for similar distribution of covariates in both group
• Matching
  groups are similar with respect to covariates

by analysis
- stratified analysis
stratification by covariate categories, subgroups
positive: subgroup statements, simple, can detect interactions
negative: small sample size in subgroups if several confounders

• regression analysis
  “if all factors are equal, the risk factor has an effect of …”
  positive: several confounders
  negative: more complex

but: confounder must be measured accurately

Question 9

Main difference between bias and confounding

Answer 9

Bias creates an association that doesnt exist, data is biased incorrect

Confounding creates an association that really exists, but interpretation is incorrect

Question 10

Design of an epidemiological study, pitfals (error, bias, …) and counteractions
Inclusion/Exclusion
Randomization
Participation rate
standardized data collection
stratification
statistical method
matching
sample size

Answer 10