Sources of Systematic Error or Bias: Information Bias

ERIC at the UNC CH Department of Epidemiology Medical Center

Sources of Systematic Error or Bias:

Information Bias

E R I C N O T E B O O K S E R I E S

Information bias is one type of

systematic error that can occur in

epidemiologic studies. Bias is any

systematic error in an epidemiologic

study that results in an incorrect

estimate of the association between

exposure and the health outcome.

Bias occurs when an estimated

association (risk ratio, rate ratio,

odds ratio, difference in means, etc.)

deviates from the true measure of

association.

Bias is caused by systematic

variation, while chance is caused by

random variation. The consequence

of bias is systematic error in the risk

ratio, rate ratio, or odds ratio

estimate. Bias may be introduced at

the design or analysis phase of a

study. We should try to eliminate or

minimize bias through study design

and conduct.

Major types of systematic error

include the following:

 Selection bias

 Confounding bias

 Information bias

In this issue we present information

bias. Selection bias and confounding

are covered in separate ERIC

Notebooks.

Information bias is a distortion in the

measure of association caused by a

lack of accurate measurements of key

study variables. Information bias, also

called measurement bias, arises when

key study variables (exposure, health

outcome, or confounders) are

inaccurately measured or classified.

Bias in the risk ratio, rate ratio, or

odds ratio can be produced even if

measured errors are equal between

exposed and unexposed or between

study participants that have or do not

have the health outcome.

Non-differential misclassification

occurs if there is equal

misclassification of exposure between

subjects that have or do not have the

health outcome or if there is equal

misclassification of the health

outcome between exposed and

unexposed subjects. If exposure or

the health outcome is dichotomous,

then non-differential misclassification

causes a bias of the risk ratio, rate

ratio, or odds ratio towards the null.

Non-differential misclassification of

exposure status

Non-differential misclassification of

exposure status in a case-control

study occurs when exposure status is

Second Edition Authors:

Lorraine K. Alexander, DrPH

Brettania Lopes, MPH

Kristen Ricchetti-Masterson, MSPH

Karin B. Yeatts, PhD, MS

Second Edition

ERIC at the UNC CH Department of Epidemiology Medical Center

equally misclassified among cases and controls. Non-

differential misclassification in a cohort study occurs when

exposure status is equally misclassified among persons

who develop and persons who do not develop the health

outcome.

Non-differential misclassification of health outcome status

occurs in a case-control study when the health outcome

status is equally misclassified among exposed and

unexposed subjects. Non-differential misclassification of

the health outcome status occurs in a cohort study when a

study subject who develops the health outcome is equally

misclassified among exposed and unexposed cohorts.

Effect of non-differential misclassification of exposure

Non-differential misclassification biases the risk ratio, rate

ratio, or odds ratio towards the null if the exposure

classification is dichotomous, i.e., either exposed or

unexposed. If exposure is classified into 3 or more

categories, intermediate exposure groups may be biased

away from the null, but the overall exposure-response trend

will usually be biased towards the null.

Effect of non-differential misclassification of the health

outcome

In most cases, non-differential misclassification of the

health outcome will produce bias toward the null, i.e. the

risk ratio, rate ratio or odds ratio will be biased towards

1.0. If errors in detecting the presence of the health

outcome are equal between exposed and unexposed

subjects (i.e. sensitivity is less than 100%) but no errors

are made in the classification of health outcome status (i.e.

specificity is 100%), the risk ratio or rate ratio in a cohort

study will not be biased, but the risk difference will be

biased towards the null.

Effect of non-differential misclassification of health

outcome status

If no errors are made in detecting the presence of the

health outcome (i.e. 100% sensitivity), but equal errors are

made among exposed and unexposed in the classification

of health outcome status (i.e. specificity less than 100%),

the risk ratio, rate ratio, and risk difference (as applicable)

will be biased towards the null.

Combined errors in both sensitivity and specificity further

increase the bias towards the null, but specificity errors

produce larger biases overall.

Differential misclassification

Differential misclassification occurs when

misclassification of exposure is not equal between

subjects that have or do not have the health outcome, or

when misclassification of the health outcome is not equal

between exposed and unexposed subjects.

Differential misclassification causes a bias in the risk

ratio, rate ratio, or odds ratio either towards or away from

the null, depending on the proportions of subjects

misclassified.

Effect of differential misclassification of exposure or

health outcome

Differential misclassification of the exposure or health

outcome can bias the risk ratio, rate ratio, or odds ratio

either towards or away from the null. The direction of bias

is towards the null if fewer cases are considered to be

exposed or if fewer exposed are considered to have the

health outcome. The direction of bias is away from the

null if more cases are considered to be exposed or if more

exposed are considered to have the health outcome.

The effect of differential misclassification of the exposure

or health outcome can bias the risk ratio, rate ratio, or

odds ratio in either direction. The direction of bias is

towards null if fewer cases are considered to be exposed

or if fewer exposed subjects are considered to have the

health outcome. The direction of bias is away from the

null if more cases are considered to be exposed or if more

exposed subjects are considered to have the health

outcome.

Interviewer bias

Interviewer bias is a form of information bias due to:

1. lack of equal probing for exposure history

between cases and controls (exposure

suspicion bias); or

E R I C N O T E B O O K PA G E 2

ERIC at the UNC CH Department of Epidemiology Medical Center

may result in bias away from null, though this is less likely

than bias towards the null.

Non-differential misclassification of a health outcome

limited to a loss of sensitivity of detecting the health

outcome without any loss in specificity does not bias

toward null, whereas a loss of specificity always biases

toward the null.

Conclusions

Some inaccuracies of measurement of exposure and

health outcome occur in all studies.

If a positive exposure-health outcome association is found

and non-differential measurement errors are more likely

than differential ones, measurement error itself cannot

account for the positive finding since non-differential error

nearly always biases towards the null.

Strive to reduce errors in measurement:

1. develop well standardized protocols

2. train interviewers and technicians well

3. perform pilot studies to identify problems

with questionnaires and measuring

instruments

4. attempt to assess the direction of bias by

considering likelihood of non-differential or

differential misclassification

2. lack of equal measurement of health

outcome status between exposed and

unexposed (diagnostic suspicion bias)

Solutions:

1. blind data collectors regarding exposure or

health outcome status

2. develop well standardized data collection

protocols

3. train interviewers to obtain data in a

standardized manner

4. seek same information about exposure from

two different sources, e.g. index subject and

spouse in case-control study

Recall or reporting bias

Recall or reporting bias is another form of information bias

due to differences in accuracy of recall between cases and

non-cases or of differential reporting of a health outcome

between exposed and unexposed.

Cases may have greater incentive, due to their health

concerns, to recall past exposures. Exposed persons in a

cohort study may be concerned about their exposure and

may over-report or more accurately report the occurrence

of symptoms or the health outcome.

Solutions:

1. add a case group unlikely to be related to

exposure

2. add measures of symptoms or health

outcomes unlikely to be related to exposure

Complications in predicting direction of misclassification

bias

Misclassification of confounders results in unpredictable

direction of bias. Non-differential misclassification of a

polychotomous exposure variable (3 or more categories)

E R I C N O T E B O O K PA G E 3

ERIC at the UNC CH Department of Epidemiology Medical Center

Practice Questions

Answers are at the end of this notebook

1) Researchers conduct a case-control study. The

following table shows the true classification of exposure

and the health outcome (Note: these data are hypothetical

and typically researchers would not know the true

unbiased distribution of exposure and outcome).

a) Calculate the odds ratio

Now imagine that 50 people with the health outcome were

misclassified as being unexposed and 20 people with the

health outcome were misclassified as being exposed.

b) Create the corrected 2x2 table

Have health out-

come

Do not have health out-

come

Exposed 200 210

Unex-

posed

340 500

Have health out-

come

Do not have health

outcome

Exposed

Unex-

posed

E R I C N O T E B O O K PA G E 4

Terminology

Information bias: A distortion in the measure of

association caused by a lack of accurate measurements

of exposure or health outcome status which can result

from poor interviewing techniques or differing levels of

recall by participants.

Non-differential misclassification: Equal misclassification

of exposure between subjects that have or do not have

the health outcome, or equal misclassification of the

health outcome between exposed and unexposed

subjects.

Differential misclassification: Unequal misclassification

of exposure between subjects that have or do not have

the health outcome, or unequal misclassification of the

health outcome between exposed and unexposed

subjects.

Acknowledgement

The authors of the Second Edition of the ERIC Notebook

woul d lik e t o a c k n ow le dge th e auth ors of th e E RI C

Notebook, F irs t Edition : Michel I brah im , M D , Ph D ,

Lorrain e Al exa n d er, D rP H, Carl Shy, MD, Dr PH ,

Gay le Shi moku r a , MSPH and Sherry Far r , G RA,

Depart men t o f E p ide mi olog y a t the Universit y o f

North Ca ro l in a a t Ch a pe l H ill. Th e First Edi ti on of

the ERI C No te bo o k was produ ced by the E du ca t i on a l

Arm of th e E p ide mi olog ic Research an d I n for ma ti on

Center a t Du r h a m , NC. The fun d i ng fo r th e ERIC

Notebook F i rs t Edi tio n wa s provided b y t h e

Depart men t o f Ve ter a n s Af f a irs (DVA), V et era n s

Health A dm in is tra ti o n (V HA ) , Cooperative Stu d ie s

Program ( CS P) to p romote the strateg ic g r ow th o f

c) Calculate the odds ratio for the corrected table

d) In which direction was the misclassification bias?

2) Researchers conduct a case-control study of the

association between the diet of young children and

diagnosis of childhood cancer, by age 5 years. The

researchers are worried about the potential for recall

bias since parents are being asked to recall what their

children generally ate, over a period of 5 years. Which of

the following potential control groups would be most

likely to reduce the likelihood of recall bias?

a) Parents of children with no known health problems

b) Parents of children with other known, diagnosed

serious health problems (aside from childhood cancer)

c) Parents of children with other known, diagnosed

minor health problems

References

Dr. Carl M. Shy, Epidemiology 160/600 Introduction to

Epidemiology for Public Health course lectures, 1994-

2001, The University of North Carolina at Chapel Hill,

Department of Epidemiology

Rothman KJ, Greenland S. Modern Epidemiology. Second

Edition. Philadelphia: Lippincott Williams and Wilkins,

1998.

Dr. Steve Marshall and Dr. Jim Thomas Epidemiology

710, Fundamentals of Epidemiology course lectures,

2009-2013, The University of North Carolina at Chapel

Hill, Department of Epidemiology

Dr. David Richardson Epidemiology 718, Epidemiologic

Analysis of Binary Data course lectures, 2009-2013, The

ERIC at the UNC CH Department of Epidemiology Medical Center

Answers to Practice Questions

a) Calculate the odds ratio

Odds ratio = (200*500) / (340*210) =1.4

b) Create the corrected 2x2 table

c) Calculate the odds ratio for the corrected table

Odds ratio = (230*500) / (310*210) =1.8

d) In which direction was the misclassification bias?

The bias was away from the null (the null value is

1.0).

2. Answer choice b is the best choice. Researchers

should aim to have similar recall bias between both

the case and control groups. Parents of children who

have childhood cancer, which is a serious health

problem, are likely to be quite concerned about what

may have contributed to the cancer. Thus, if asked by

researchers, these parents are likely to think very

hard about what their child ate or did not eat in their

first years of life. Parents of children with other

serious health problems (aside from cancer) are also

likely to be quite concerned about any exposure that

researchers ask about. Therefore, these parents can

be expected to recall exposures in a way that is more

comparable with parents of children who have

cancer. In contrast, parents of children who have no

health problems or parents of children with only

minor health problems are less likely to be

concerned with carefully recalling any exposures.

Have health out-

come

Do not have health

outcome

Exposed 230 210

Unex-

posed

310 500

E R I C N O T E B O O K PA G E 5