Bias, confounding and causality in p'coepidemiological research

3 EVER-PRESENT ISSUES IN
PHARMACOEPIDEMIOLOGICAL RESEARCH
• BIAS
• CONFOUNDING
• CAUSALITY

 Major objective of pharmacoepidemiological research is to estimate drugs’
effects (when prescribed) after marketing.
 Drug exposure:
 is not a stable phenomenon;
 may be associated with factors related to the outcome of interest (e.g.,
indication for prescribing, patient compliance, publicity, and natural
course of the disease).
 Challenge of pharmacoepidemiological research is to obtain an accurate
estimate (without error), of the relationship between drug exposure and
health status.
 2 types of errors:
• Random error is related to precision and reliability,
• Systematic error is related to validity and bias.
[Accuracy is the absence of both random and systematic error].

1981 – National Childhood Encephalopathy Study (NCES)
 Results presented by Alderslade and Miller;
 A nationwide case–control study conducted in the UK by the
Committee on Safety of Medicines and the Joint Committee on
Vaccination and Immunization.
 Research Question: Any possible association between DTP vaccine
and the subsequent development of neurologic disorders?
 Findings:
 Risk of a severe acute neurologic event was significantly increased
within the seven days following DTP vaccine.
[RR 2.3; 95%(CI) 1.4–3.2],
 One year later, of the 241 cases in whom the disorder had begun
within the seven days following a DTP vaccine, 7 (2.9%) cases either
died or had a developmental deficit .
 In controls: only 3 of 478 controls (0.6%) died or had developmental
deficit.

 These results were used in many court trials by parents of disabled
children who were seeking compensation.
BUT…..
 Credibility of the study was compromised by suspicions of bias.
 Numerous potential biases were identified and were responsible
(either fully or partially), for the results observed.
 Referral bias: physicians were aware of the study objectives and this
might have influenced their referral of cases and increased the
apparent relative risk.
 Information bias:
• interviewers were not blinded to study objectives, subjects’ clinical
status;
• date of onset of the neurological disorder was occasionally difficult
to establish (potentially increasing the apparent relative risk).

 Protopathic bias: possible presence of subclinical neurological
disease prior to vaccination, could have falsely increased the relative
risk.
 Lack of precise disease definitions and inclusion criteria thought not
related to DTP vaccine, caused misinterpretation of results (Reye’s
syndrome, hypsarrhythmia, or acute viral encephalopathies).
 Issues in study design can affect the validity of results in
pharmacoepidemiology research.
 Pharmacoepidemiology studies may be affected by particular biases
more often than other epidemiologic studies.

BIASES IN PHARMACOEPIDEMIOLOGY
 3 categories:
 Selection bias (related to the recruitment of study subjects or loss of
study subjects in follow-up)
 Information bias (related to the accuracy of information collected on
exposure, health status);
 Confounding (covariates or effect modifiers related to the
pathophysiology of disease development, whereby one factor (or
several factors acting together) can produce an observed effect that
may be incorrectly attributed to an exposure of interest).

SELECTION BIAS (Sample distortion bias)
 Due to the selection (inclusion) of groups of subjects into the study
who differ in characteristics from those in the target population,
causing distortion of the measurement of an effect (outcome).
 4 types of selection bias:
 Referral bias
 Self-selection
 Prevalence study bias
 Protopathic bias

REFERRAL BIAS
 Can occur if the reasons for referring a patient by the physician to the
study are related to the patient’s exposure to (use of) the drug.
 Problematic when an illness presents in such a manner that an
accurate diagnosis is not always obtained.
 E.g.1, Hospital ‘XYZ’ with 2 groups of patients:
• Group 1: 1000 patients on NSAIDs presenting w/ abdominal pain
may be more likely to be suspected as having a GU.
• Group 2: 10 patients with similar pain who are not using NSAIDs
• Group1 patients are more likely to be tested for GU than Group2.
• A study using these patients in Hospital XYZ will show a strong, but
biased, association between mild non-bleeding GUs and NSAIDs
use.
WHY??????

 E.g.2, Association b/n DVT and oral contraceptives
• The association b/n DVT and oral contraceptives is already well
known.
• The use of oral contraceptives is a vital factor in this study.
• Exposed women (women on oral contraceptives) may be more likely
to be tested for DVT than women not on oral contraceptives.
• Earlier studies reporting a positive association b/n drug (oral
contraceptives) and disease (DVT) can begin the referral bias
phenomenon.
*** Identifying the potential for referral bias in initial stages of any study
is important for that study, as well as for future similar studies.

SELF-SELECTION BIAS
 May occur when study participants themselves decide to participate
in, or to leave a study (based on drug exposure effects, change in
health status of participants, personal reasons).
 So, the association observed in the study sample may not be
representative of the real association in the source population.
 This bias is very important in case–control studies or cohort studies.
 E.g., Association b/n drugs used during pregnancy and birth defects
• 2 groups:
• Group 1: mothers of ‘affected’ children, who used medications during
pregnancy.
• Group 2: mothers of ‘normal’ children, who used medications during
pregnancy.
• Group 1 will be more willing to participate in the study than group 2.
• Solution: systematically identify and recruit all eligible cases (for both
groups).

 Losses to follow-up (study participants dropping out) in cohort studies
can also induce bias, if those who drop out belong to a special
disease–exposure category (those who fulfill the Inclusion criteria).
Solution: Use population-based registries

PREVALENCE BIAS
 A type of selection bias that may occur in case–control studies when
prevalent cases (rather than new cases) are selected for a study.
 Prevalence is proportional to both incidence and duration of the
disease (But, it is related more to the duration of the disease rather
than to the incidence).
 In a group of incidence cases, significant association with prevalent
cases might not be confirmed.
 Recruiting only incident cases with recent documented data is
relevant only to disease incidence, not to prevalence.

PROTOPATHIC BIAS
 Feinstein (1985) – may occur “if a particular treatment or exposure
was started, stopped, or otherwise changed because of the baseline
manifestation caused by another disease or other factor.”
 If some other disease or risk factor produces the same symptoms or
signs that the researcher is analysing.
 E.g., Studying the association between blood in stool as an indicator
for colorectal cancer.
 BUT… excessive use of aspirin can also cause blood in stool
 Haemorrhoids cause blood in stool.

INFORMATION AND MISCLASSIFICATION BIAS
 Errors can occur if cases in a study are classified with regard to their
exposure and disease status…..
• unexposed people may be considered exposed and vice versa.
• health status may also be incorrectly classified.
 This type of error may lead to a misclassification bias.
 Equally affects case–control and cohort studies.
 Non-differential misclassification:
• When the misclassification error occurs randomly (i.e., independent
of the exposure–outcome relationship).
• Mostly occurs if study instrument is not very reliable.
• It may lead to a decrease in the strength of the association between
drug and outcome (bias toward the null hypothesis)

Exposure timing
• Inaccuracy in properly defining the exposure time can result in
information bias which may lead to a non-significant association
overall, even when there is a very strong association between the
drug and the outcome, within a specific time window.
• E.g., Anaphylactic reactions occur rapidly after drug exposure, very
high risk during this short time period, and null after this initial period.
• The risk mostly decreases with time.
• E.g., Sometimes, chronic long-term users of NSAIDs are likely to be
at a lower risk of gastrointestinal bleeding than new users, because
of a ‘survivor effect’.
• Sometimes, the risk steadily increases with time, due to the
cumulative effect of drug exposure
• E.g., risk of myocardial toxicity after the use of doxorubicin.

Differential misclassification:
 When this error is influenced by knowledge of the exposure (drug /
disease) and the outcome status.
 E.g.1, during data collection in cohort studies, knowledge of the
exposure influences the quality of the information collected
 E.g.2, in case–control studies when knowledge of the disease status
influences the quality of the information collected about exposure, it is
also called information bias.
 2 situations: Differential recall bias and Differential detection bias.
Differential Recall bias:
• mostly seen in retrospective studies,
• in case–control studies, cases and controls may have a selective
memory of their past exposures.
• E.g., In studies of birth defects, mothers with an impaired child may
give a more valid and complete report of their exposure to drugs
during pregnancy as a result of devoting more time to contemplating
the cause of the birth defect.

• This type of bias may be minimized by selecting controls who are
likely to have the same cognitive processes affecting memory of past
drug exposures.
Differential detection bias:
• can affect either cohort or case–control studies.
• In case–control studies: occurs when the procedures for exposure
assessment is more thorough among cases than controls.
• In cohort studies: occurs mostly due to difference in the follow-up for
detecting adverse events.
• E.g., women taking postmenopausal hormonal supplements are likely
to see their doctors more often than other women. They are more
likely to be examined for breast or endometrial cancer, or risk of CV
disease.
• This may lead to an excess number of diagnosed diseases in the
‘treated’ group (women who took postmenopausal hormonal
supplements) and a falsely elevated risk

CONFOUNDING
 Occurs when the association between drug exposure and health
status is distorted by the effect of one or several extraneous
variables that are also risk factors for the outcome of interest.
 E.g., Study of relationship b/n use of NSAIDs and GU
• Potential Confounders: Personal Hx of GU in patient, Chronic
alcoholism.
 For a variable to be a confounder, it must be associated with both the
drug exposure and the outcome of interest.

Study: Death associated with use of Drug A;
Comparison group: Patients treated with Drug B

Study: Risk of allergy associated with use of a drug;
Comparison group: Patients not exposed (treated with) the drug.
EFFECT
MODIFICATION

Confounding by Indication for Prescription
 Synonyms: Indication bias, Contraindication bias, Channeling,
Confounding by severity.
 Indication for a prescription is the most important confounder in
pharmacoepidemiological research. WHY????
 Because there is always a ‘reason’ for a prescription and the reason
is often associated with the outcome of interest.
 Can induce selection bias in drug efficacy studies.
 Difficult to control.
 Often impossible to obtain a sufficiently accurate estimate of the
confounder’s effect.
 Miettinen (1983) – preventive use of warfarin was associated with a
27-fold increase in the risk of thrombotic events (conditions that
should actually be prevented by warfarin). This paradoxical result was
because only highly susceptible patients, or those already
experiencing the first symptoms of thrombosis were included in the
study.

Confounding by Comedication and Other Cofactors
 Patients often take more than one drug at a time and it is sometimes
difficult to isolate the effect of a specific drug, in research studies.
 Coronary Drug Project (1980) –the risks of death in placebo group
after 5 years were 15% (compliant cases) and 28.2% (non-compliant
cases).
Potential Confounders: Selection bias; patients compliant with one
drug were very likely to be compliant with other interventions (e.g.,
other very effective drugs, diet, physical exercise, etc.).

SOLUTIONS FOR SELECTION BIAS
 Must be prevented at the design stage, because it cannot be
corrected at the analysis stage.
 Selection bias can result in over- or under-representation of the
people who have a drug exposure–outcome relationship.
 Strategies:
• Random sampling of the cases and controls from the source
population.
• Systematically recruiting a series of consecutive subjects (to
prevent self-selection).
• Minimizing the number of subjects lost to follow-up in cohort
studies.
• Tracking procedure for drop-out cases (to identify the reasons).
• Selecting only incident cases of the condition.
• Assigning random allocation of drug exposure – Follow the
procedure (to prevent self-selection and referral bias).

SOLUTIONS FOR INFORMATION BIAS
 Must be resolved at the design stage.
 Strategies:
• Blinding (or masking) of relevant study personnel;
• Standardization of the measurement process for both cases and
controls
• use of standard structured questionnaires,
• specific training of interviewers,
• different observers for different measurements
• Standardized criteria for defining drug exposure and disease
outcomes.

SOLUTIONS FOR CONFOUNDING
 It is possible to control the effect of confounding at both the design
and the analysis levels.
 Strategies:
@ Design level:
• Randomization,
• Matching – ensure similarities in both case and control groups; be
cautious of ‘over matching’,
• Restrict confounding by studying only one level – e.g., studying
the drug effect among only one category of age will prevent against
the occurrence of confounding by age.
@ Analysis levels:
• Standardization
• Stratification

• Cause: a stimulus that produces an effect or outcome.
• Change in host-agent-environmental balance produces an
effect or outcome.
• Cause (statistical definition): a factor which varies (either
proportional or inversely proportional) to the health condition
of interest (health condition studied).
Statistical relationship
• E.g., a disease with risk factors X, Y and Z
• Is there a ‘statistical relationship’ b/n factors X and Y?
• Meaning: any association b/n X and Y occurs by chance or not
by chance (greater frequency than that by chance?)
• Next step: tests for independence or association by Chi-square
test or Correlation coefficient test
• If the test results are statistically significant, then X and Y are
not independent, but have an association that is not entirely
due to chance.

• Study: rates of developing complications after mastectomy for
women with and without anxious personalities.
• Test Result: Chi-square test is statistically significant at p<0.05
• Meaning: 95 times out of 100, this difference in complication
rates between the two samples is not due to chance.
• Inference: Personality and complication rates are not
independent.
• They have a significant statistical association.
• BUT….. This does not mean that only those with anxious
personality type suffer from complications post-mastectomy.
SAMPLE COMPLICATION RATES
(after mastectomy)
Women with anxious
personality
100 nos.
Women without anxious
personality
40 nos.

• It means that a woman with anxious personality is more
likely to suffer from complications after mastectomy than a
woman without anxious personality.
• Determining the ‘statistical significant association’ is the first
step in determining whether the relationship is causal.
Causal relationship
• Statistically significant factors (non-independent) can have non-
causal relationship or causal relationship.
Non-causal relationship
• Hypothetical variable varies with actual causal variable
• E.g., Relation b/n paternal age and infant birth weight
• paternal age (Hypothetical variable)
• Maternal weight (actual causal variable )and infant birth weight
• Even if there is a statistical significant relation b/n paternal age
and infant birth weight, there isn’t any logical explanation.

Causal relationships
• 2 types: a) direct and b) indirect
a) Direct causal relationship
• A factor directly causes a disease with no other intervening
factor.
• Causal factor Outcome
• E.g., Tubercle bacilli TB
• Sometimes, what is considered a direct causal association may
later on be identified as indirect causal association.
• E.g.1, Cholera outbreak in England (1883) – Dr. John Snow –
identified certain water sources as the causative factor;
Drinking water from those sources was banned. Later on, it was
identified that V.cholerae was the causative factor. Water was
only the vector.

E.g.2, Toxic Shock Syndrome (TSS)
• When TSS first emerged, clinicians identified tampons as the
causative factor.
• Later on, Staphylococcal spp. was identified as the causative
factor.
• Tampons were the vector (indirect contributory cause).
• Education programmes aimed at eliminating the use of
tampons; changing the way tampons were used;
• Women were advised to avoid super-absorbent tampons,
change tampons frequently, practice good hygiene.

b) Indirect causal relationships
• Extra variable/s (intervening variables) occupies/y an
intermediate stage b/n cause and effect.
• A B C D
• ‘A’ is causally associated with ‘D’, only after the interposition of
variables ‘B’ and ‘C’.
• E.g., Relationship b/n cigarette smoke and chronic bronchitis
Breathing the air polluted by cigarette smoke (A)
Damage to the respiratory epithelium (B)
Increased susceptibility of respiratory epithelium to infection (C)
Chronic bronchitis (D)

Multiple causative factors
• Diseases have many risk factors, all of which are involved in
development of the disease.
• Exposure to multiple causative factors can have additive or
multiplicative effect.
• E.g.1, Even though smoking is a major cause of lung cancer, it
is not the only cause. Non-smokers (either active or passive)
can also get lung cancer.
• Risk of cancer is higher among non-smokers exposed to
asbestos than smokers not exposed to asbestos.
• E.g.2, Automobile accidents can occur due to speeding, faulty
equipments, heavy traffic, poor visibility, driver’s inexperience,
drinking etc.
• Web of causation: association b/n all the causative factors
which have impact on the risk of developing a disease.

Establishing causality
• Through epidemiological studies and clinical trials.
• A factor is considered causal, when reducing it’s amount or
frequency, reduces the frequency of the effect.
• E.g., If treating hypertensive patients (keeping their BP low) can
reduce the frequency of stroke when compared to an
equivalent , untreated group of hypertensive patients, then HTN
is considered a risk factor for stroke (HTN is a causal factor for
stroke).

Naranjo Algorithm (ADR) Probability Scale
• A method by which to assess whether there is a causal relationship
between a DRUG and ADR.
• Use a simple questionnaire (10 qns.) to assign probability scores.
• Answers are either Yes, No, or “Do not know”.
• Different point values (-1, 0, +1 or +2) are assigned to each answer.
• Total scores range from -4 to +13;
– Definite reaction: 9 to 13; Probable reaction: 5 to 8;
– Possible reaction: 1 to 4; Doubtful reaction: 0 or less.
• The response ‘Do not know’ should be used
– sparingly;
– only when the quality of the data does not permit a ‘Yes’ or ‘No’
answer.
– if the information is not available
– if the question is inapplicable to the case.

• When more than one drug is involved or suspected,
• the ADR Probability Scale is applied separately to each of the
possible etiologic agents (drugs);
• the drug with the highest score should be considered the
causative agent;
• the potential of interaction should be evaluated;

Question Yes No
Do Not
Know
Score
1. Are there previous conclusive reports on
this reaction?
+1 0 0
2. Did the adverse event appear after the
suspected drug was administered?
+2 -1 0
3. Did the adverse event improve when the
drug was discontinued or a specific
antagonist was administered?
+1 0 0
4. Did the adverse event reappear when the
drug was readministered?
+2 -1 0
5. Are there alternative causes that could on
their own have caused the reaction?
-1 +2 0

Question Yes No
Do Not
Know
Score
6. Did the reaction reappear when a placebo was
given?
-1 +1 0
7. Was the drug detected in blood or other fluids
in concentrations known to be toxic?
+1 0 0
8. Was the reaction more severe when the dose
was increased or less severe when the dose was
decreased?
+1 0 0
9. Did the patient have a similar reaction to the
same or similar drugs in any previous exposure?
+1 0 0
10. Was the adverse event confirmed by any
objective evidence?
+1 0 0
Total Score:

THANK YOU!!!
Courtesy: Textbook of Pharmacoepidemiology by
Strom and Kimmel

Bias, confounding and causality in p'coepidemiological research

More Related Content

What's hot (20)

Viewers also liked (10)

Similar to Bias, confounding and causality in p'coepidemiological research (20)

More from samthamby79 (20)

Recently uploaded (20)

Bias, confounding and causality in p'coepidemiological research