Random and Systematic errors

Observe an association

Is it real?
Real Spurious

Is it due to presence Bias/systematic

of third variable? error

No Is a third variable involved?

third variable

Could it be due to chance? Confounding or interaction

Unlikely Likely

CAUSAL Evidence for causation not as strong

 Observe an association

Is it real?
Real Spurious

Is it due to presence Bias/systematic

of third variable? error

No Is a third variable involved?

third variable

Could it be due to chance? Confounding or interaction

Unlikely Likely

(maybe) causal Evidence for causation not as strong

Random Error
Random error can occur in any studies
Even if the disease under study is not
associated with an exposure, there may be a
“chance” association in a particular study (e.g.
disease may be more common in the exposed
than in the non-exposed group)
Random Error
The precision (lack of random error) of an
effect estimate (e.g. an odds ratio) is reflected
in the 95% confidence interval
Random error reduces, and precision
increases, as the study size increases
 Observe an association

Is it real?
Real Spurious

Is it due to presence Bias/systematic

of third variable? error

No Is a third variable involved?

third variable

Could it be due to chance? Confounding or interaction

Unlikely Likely

CAUSAL Evidence for causation not as strong

 Observe an association

Is it real?
Real Spurious

Is it due to presence Bias/systematic

of third variable? error

No Is a third variable involved?

third variable

Could it be due to chance? Confounding or interaction

Unlikely Likely

CAUSAL Evidence for causation not as strong

Systematic error
Systematic error in the design, conduct or
analysis of a study that results in a mistaken
estimate of an exposure’s effect on disease
Distinguished from random error (lack of
precision) in that it would be present even
with an infinity large study
Rothman KJ. Epidemiology: An Introduction. Oxford, 2002.
Systematic error
Internal validity: the validity of the inferences
drawn as they pertain to the actual subjects in
the study.
External validity (Generalisability): the
validity of the inferences as they pertain to the
people outside the study population
Evaluating bias
Why did it occur?
What effect does it have on the observed
association?
What can be done to control for bias in this
study and to prevent it in future studies?
Types of (important) bias
Selection bias
Error in selection of study participants
Information bias
Errors in procedures for gathering relevant
information
Selection Bias
Involves biases arising from the procedures by which
the study participants are selected from the source
population
Can be introduced at any stage of a study (bad
definition of eligible populations, lack of accuracy of
sampling frame, uneven diagnostic procedures) and
implementations
In an incidence (or prevalence study) selection bias
will not occur if there is a 100% response rate
However, if for example, cases of disease are more
likely to participate than non-cases, and this is related
to exposure, then selection bias will occur
Selection Bias in Case-Control Studies

In a case-control study, the controls are a

sample of the source population
Selection bias can occur if the sample is non-
random, and the selection of controls is
related to exposure status
In other words, selection bias can occur if the
controls are not representative of the
exposure in the source population
Selection Bias: Solutions
Achieve a response rate of 100% (and in a
case-control study, ensure that controls are a
random sample of the source population)
Control for the determinants of selection bias
as confounders in the analysis (e.g. if
response rates vary by social class then
control for social class as a confounder)
Assess the likely size of the selection bias
Selection Bias in Case-Control Studies:
Solutions
Selection bias can occur if the selection of
controls is related to exposure status
In the analysis, we can control for the
determinants of control selection (e.g. social
class)
An exception is when we have chosen “other
disease” controls and the other diseases are
directly caused by the main exposure of
interest: this selection bias cannot be
removed
Example on selection bias. True population
distribution
Example on selection bias. Unbiased
sampling in case-control study
Example on selection bias. Biased
sampling in case-control study
Information Bias
Originated when sensitivity and /or specificity
of the procedures to detect exposure and
effect is not perfect
Occurs during data collections
Misclassification bias
If misclassification of exposure (or disease) is
unrelated to disease (or exposure) then the
misclassification is non-differential
If misclassification of exposure (or disease) is
related to disease (or exposure) then the
misclassification is differential
Non-Differential Misclassification
When the probability of misclassification of
exposure is the same for cases and non-cases
(or when the probability of misclassification of
disease is the same for exposed and non-
exposed persons).
Non-differential misclassification of exposure
usually (but not always) biases the relative risk
estimate towards the null value of 1.0.
Example non-differential misclassification: exposure
classified correctly
Example non-differential misclassification: sensitivity
and specificity are 80%
Example non-differential misclassification: sensitivity
and specificity are 80%
Non-Differential Misclassification:
Exceptions to “Bias Towards the Null”

If there is misclassification of a positive

confounder then control for that confounder
will only partially remove the confounding;
hence there will be bias away from the null
(compared with what would have been
obtained with full control of confounding)
Differential Misclassification
When the probability of misclassification of exposure
is different in diseased and non-diseased persons, or
the probability of misclassification of disease is
different in exposed and non-exposed persons.
Bias can be in any direction, i.e. towards or away
from the null value
It is usually therefore desirable to ensure that
misclassification is non-differential, even if this means
not using some information on one group, i.e. it is
important to collect the information in a similar
manner in the groups being compared
 Observe an association

Is it real?
Real Spurious

Is it due to presence Bias/systematic

of third variable? error

No Is a third variable involved?

third variable

Could it be due to chance? Confounding or interaction

Unlikely Likely

(maybe) causal Evidence for causation not as strong

 Observe an association

Is it real?
Real Spurious

Is it due to presence Bias/systematic

of third variable? error

No Is a third variable involved?

third variable

Could it be due to chance? Confounding or interaction

Unlikely Likely

CAUSAL Evidence for causation not as strong

 Treatment of Renal Calculi

Successes Failures Total Successes

patients (%)
All Stones (n=700)
Open surgery 273 77 350 78
Percutaneous nephrolithotomy 289 61 350 83
Stones < 2 cm (n=357)
Open surgery 81 6 87 93
Percutaneous nephrolithotomy 234 36 270 87
Stone >2 cm (n=343)
Open surgery 192 71 263 73
Percutaneous nephrolithotomy 55 25 80 69

Charig CR et al., BMJ 1986; 292:879-82

STONE SIZE
Is a determinant of success rate (the larger the stone,
the lower the success)
Is associated with surgical procedure/ is unevenly
distributed among the two surgical procedures to be
compared
Note1: “Mixing of effects”: the effect of the exposures of
interest is distorted because it is mixed with the effect
of an alternative cause (or a correlate)
Note 2: confounding is the distortion that results from
failure to control for the effects of such alternative
cause (or its correlate)
 SURGICAL PROCEDURE SUCCESS
RATE
(Open surgery)

STONE SIZE (large)

• Open surgery gives better success

• An erroneous impression of a lower success rate is created by the

pooled data
• Open surgery was used much more often on patients with large
stones
• These patients had a lower success rate irrespective of the technique
used
 .3 small size
Open Surgery
.7 large size

.8 small size
Percut. Nephro
.2 large size
Confounding

Confounding occurs when an estimate of the

association between an exposure and an
outcome is mixed up with the real effect of
another exposure on the same outcome, the
two exposures being correlated.
Schematically...

EXPOSURE OUTCOME

“THIRD
VARIABLE”
Confounding factor:

1. It must be predictive of occurrence of

disease: the effect need not be casual
A correlate of a casual factor is identified as a
confounder:
- Age (artificial masters of biological changes)
- Social class
2. It must be associated with both the exposure
and the disease under study
Confounding factor (2):
3. The potential confounder must not be an intermediate
step in the casual pathway between exposure and
disease:
No mixing of effects, but only one effect
Controlling for an intermediate factor will make it
impossible to measure this effect
Moderate alcohol level of risk of
consumption HDL MI

But: depending from the strategy of analysis, HDL could

be treated as a confounder
Example of confounder -1

?
Occupation Lung cancer

Is the occupation a cause of lung cancer?

Example of confounder - 2

?
Occupation Lung cancer

1
Smoking

1. Smoking is associated with occupation

Example of confounder - 3

?
Occupation Lung cancer

1 2

Smoking

1. Smoking is associated with occupation

2. Smoking is a cause of lung cancer
Example of confounder - 4

3
Occupation X Lung cancer

1 2

Smoking
1. Smoking is associated with occupation
2. Smoking is a cause of lung cancer
3. Smoking is a confounder (occupation is not a cause
of lung cancer)
Example of non-confounder

Smoking
2 Lung cancer
X
1 3
Alcohol drinking

1. Alcohol drinking is associated with smoking

2. Smoking is a cause of lung cancer
3. Alcohol drinking is not a cause of lung cancer
Example of intermediate variable

Alcohol drinking Liver cirrhosis Liver cancer

Liver cirrhosis is on the causal pathway between

alcohol drinking and liver cancer.
It is an intermediate variable, not a confounder
Characteristics of confounding
3
Intermediate
Exposure Disease
Variable
1 2

Confounder

1. Associated with exposure in the study base

2. Risk factor (cause) of disease
3. Not an intermediate step in the causal path
Result of confounding

Bias in unpredictable direction

Direction depends on
- direction of association between
exposure and confounder
- effect of confounder on disease
Direction of confounding bias

Association Effect on disease Direction of bias

confounder-
exposure
Positive Risk factor

Negative Risk factor

Positive Protective

Negative Protective
Positive confounding
OC
use MI
+ +
OC users Smokers (irrespective of
Smoke more OC use) have higher risk
Than non users S of MI than non smokers

Negative confounding

OC MI
-
use +
-
OC users Obesity (irrespective of
Less obese OC use) increase the risk
Than non users Obesity or MI
Evaluation of confounding
Based on previous knowledge
- epidemiological evidence
- clinical, biological data
Based on the results of the study
changes in risk estimate of main exposure after
controlling for potential confounder
stratum-specific results
Assessment of confounding

Qualitative and quantitative assessment of

confounding
Lack of control for a confounder: bias
(Confounding generally involves biases that
are inherent in the source population, and
therefore would occur even if everyone in the
source population took part in the study)
Control for a non-confounder: lack of precision
Control for an intermediate variable: bias
toward the null
Control of confounding

Study design
- Randomization
- Restriction
- Matching
Analysis
- Stratification
- Multivariate modeling
Randomization

It suppresses the association between

exposure and confounder
Need for large sample size
Applicable only in experimental studies
Restriction

Participation in the study to people belonging

to one stratum of the confounder
Ex: A study of non-smokers eliminates
possible confounding by smoking
It precludes the assessment of interaction and
limits the external validity
Matching

Selection of controls in such a way that the

distribution of potential confounders is
identical to that of cases
Individual vs. group (frequency) matching
Stratified analysis

The association between the exposure and

the outcome is measured in each category
(Stratum) of the confounder
A weighted average is calculated
The confounding effect is adjusted or
controlled for ( the summary estimate takes
into account differences in the distribution of
the confounder)
Example- occupational cohort

Unexposed Exposed

Cases 15 51

Person-years 6,000 6,000

Rate/1,000 2.5 8.5

Rate ratio Ref. 3.4

 Unexp Exposed
Cases 15 51
PY 6,000 6,000
Rate 2.5 8.5
RR Ref. 3.4

Non-smokers Smokers

Unexp Exp Unexp Exp.

Cases 5 1 Cases 10 50
PY 5,000 1,000 PY 1,000 5,000
Rate 1 1 Rate 10 10
RR Ref. 1.0 RR Ref. 1.0
Misclassification of confounding

Limitation of ability to control

Non-differential misclassification reduces
adjustment, therefore bias in unpredictable
direction
Confounders should be measured as reliably
and validly as exposure and outcome variables
Some practical issues

Confounding is often complex

- partial confounding
It is prudent to consider all potential
confounders
Ability to efficiently control is limited, also in
multivariate modeling
- interpretation of ‘adjusted’ results
Summary: evaluating confounding

Is a covariate a confounder?

QUESTION 1: Is it associated with exposure?

QUESTION 2: Is it causally associated with outcome?

YES

STEP 1: Calculate Crude Measure of Association

STEP 2: Calculate Measure of Association within strata of
potential confounder
 Assess measure of association within strata

Are stratum specific measures same?

Yes No

Crude measure=stratum specific? Statistical interaction; report

stratum specific measures

Yes No

No confounding or interaction Confounding

Report crude measure report measure
adjusted for
confounding
 Observe an association

Is it real?
Real Spurious

Is it due to presence Bias/systematic

of third variable? error

No Is a third variable involved?

third variable

Could it be due to chance? Confounding or interaction

Unlikely Likely

CAUSAL Evidence for causation not as strong

This document was created with Win2PDF available at http://www.win2pdf.com.
The unregistered version of Win2PDF is for evaluation or non-commercial use only.
This page will not be added after purchasing Win2PDF.