Epidemiology 8

Fundamentals of Statistical & Epidemiological Inference

The Null Hypothesis (H0)

• The null hypothesis is the NEGATION of the actual hypothesis of

interest. It is the formal basis for statistical significance testing which
assesses the compatibility of the data with the null hypothesis.

• Unique – there is only one and there are infinitely many alternatives
 Examples:

Hypothesis 1: Cigarette smoking increases risk for idiopathic pulmonary fibrosis.

• Null Hypothesis 1: Cigarette smoking does not increase risk for idiopathic
pulmonary fibrosis.

Hypothesis 2: Insulin resistance increases risk for coronary artery disease

• Null Hypothesis 2: Insulin resistance does not increase risk for coronary artery
disease
Real Truth vs Your Sample

• Hypothesis: Hemoglobin levels are higher among male compared with

female medical students
Your sample Real Truth HA is true Real Truth H0 is true
Male Hgb > female Hgb Male Hgb = female Hgb

Your sample finds Correct conclusion Wrong conclusion

Male Hgb > Female Hgb Power = 1 – β (80%) Type I error α (5%)
Type I error
Your sample finds Wrong conclusion Correct conclusion
Male Hgb = Female Hgb Type II error β (20%) 1 – α (95%)
Type II error
P values
• The significance level represents the chance that the null hypothesis is
rejected when it is actually true.

• The dividing line between an ‘acceptable’ and ‘unacceptable’ P value

is called the significance level (or probability of type I error);

• The significance level is usually given as a percentage and the 5% level

of significance is generally accepted to be a reasonable yardstick to
use.
 Strength of evidence for rejecting the
null hypothesis

Very strong evidence Strong Evidence Weak evidence Little or no

evidence evidence

P value 0 0.001 0.01 0.05 0.1 1

Significance testing
A hypothesis test seeks to discover whether an assertion about the
population appears (on the basis of the sample data) to be
unreasonable.

1. Set up a null hypothesis H0 of no effect

(e.g. no effect of a drug in a clinical trial of cholesterol lowering
treatments);

2. collect the data to test the hypothesis;

3. Calculate the probability, P, of observing the data we have, or more
extreme data, if H0 was true;

4. If P is large, the data are consistent with the null hypothesis – it

may be true or we just don’t have enough data to say otherwise;

5. If P is small, we reject the null hypothesis. We conclude that there

is a statistically significant result.
Misuse of p-values
• Claims of no difference

• A large p-value indicates that any observed difference MAY be due entirely to
chance, not that it IS due entirely to chance
• P-values DO NOT measure strength of association

• A p-value will be smaller when there is:

• A larger observed difference between 2 groups

• A larger sample size
• Less variation of exposure or outcome (continuous)
Confounding
• A confounding variable, or confounder, is an extraneous factor that
wholly or partially accounts for the observed effect of the risk
factor on disease status

1. Effect → apparent relationship

• confounder is causing the relationship to appear

2. Effect → apparent lack of relationship

• confounder is masking a true relationship.
• The relative risks for the various levels of the confounder are all the same, and this common value
is different from the relative risk when the confounding variable is ignored

• Perfect confounding is extremely unlikely in real-life epidemiological data

Some degree of confounding:

• More likely to find some, but not perfect, confounding in real life data

• May lead to over- or under-estimation of risk

• Not a serious problem when the risk estimates by level of confounder, and overall risk estimate,
are all very similar
Identification of a confounder
Denote disease by D, risk factor by F and the third variable by C

If C is to be a confounder it must:

• be related to the disease, but not be a consequence of the disease;

• be related to the risk factor, but not be a consequence of the risk

factor
Path diagrams
Woodward (2005)

A B A causes B

A is non-casually related
A B to B
Confounding type 1
• Confounder causes both potential risk factor and disease

Grey stroke
F D hair

C age

• Grey hair tends to be related to any disease, such as stroke, which is age-related. This does
not mean that grey hair is an independent risk factor for such diseases!
Confounder Type 2
• Confounder causes disease and is related to risk factor

No of
Breast
children
F D cancer

C age

• Older women have had more children and also tend to have more breast cancer.
Having had more children is not necessarily a cause of breast cancer, but is
confounded with age
Confounder Type 3
• Confounder causes disease and risk factor is related to disease and confounder

Poor
housing CHD
F D

C smoking

• Poor housing is related to CHD, maybe not causally, but through other causal
factors, besides smoking, that are also associated with social disadvantage
Selecting confounders

• Confounder causes disease or is related to disease

AND

• Confounder causes risk factor or is related to risk factor

BUT NOT

• Disease causes confounder or risk factor causes confounder

Potential confounders

• All known risk factors for the specific disease

• Age and sex, where appropriate

Estimation and confidence intervals
• Estimation involves the use of some measure derived from the
sample to act as a representative measure for the parent population;

• Confidence interval is a range of values which we are fairly confident

will contain the true value of the measure of interest in the overall
population (typically we use 95% confidence intervals).
• Estimation is much more useful than hypothesis testing because:
• estimate gives precise information on the magnitude and direction of the effect

• Confidence interval can often be used to make the same decision

as would an associated hypothesis test:
• Calculating a 95% CI is equivalent (or sometimes approximately so) to performing a
significance test at the 5% level for all the values in that interval
• Estimating the true population mean: Statistical inference using 95%
confidence intervals.
• 95% of the intervals will contain the true population mean, µ
• Hence, these are called 95% confidence intervals
Three cautions
• Confidence intervals should not be used to simply judge estimates as
‘statistically significant’ or not.

• Confidence intervals tell us about how large a role chance may play,
but reveal nothing about bias. Interpretation of point estimates
should consider possible bias.

• Confidence intervals are not the only way to estimate precision

(advanced course topic)
Bias
• Any systematic error in the design, conduct or analysis of a study that
results in a mistaken estimate of an exposure's effect on the risk of
disease.
• (Schlesselman and Stolley, 1982)

• Types of bias (classification of bias)

• Selection bias
• Information bias
• Confounding (alternatively may be considered as distinct from other forms of
bias)
Selection Bias
• Bias introduced by the selection of participants into a study that distorts
the exposure/outcome relationship from that in the source population.
• [Not the same as non-response!]

• Example:
• We compared HPV prevalence in:
• Cases --- all cervical cancer patients in Maryland
• Controls --- age-matched non-cancer patients from the HIV clinic
Information Bias
Method for collecting information on exposure, outcome or other relevant
factors (confounders and modifiers) introduces bias
Different interviewers used for cases vs. controls
Diagnostic test performance changes over time in cohort study
Diseased and non-diseased people remember their exposure history differently
Sources of Information Bias In
Epidemiological Research
• Participant responses
• differential recall (recall bias)
• e.g., mothers of children with birth defects more likely to remember their exposures during
pregnancy vs. mothers with normal deliveries because they want to find a reason for the
defect
• socially desirable answers
• e.g., participants in HIV-prevention trial know that they are not supposed to have risky sex so
may under-report their behavior

• Interviewers
• leading questions
• Interviewers with knowledge of the hypothesis may probe more on primary exposure history
when interviewing people with disease than the comparison group without disease.
• Measurement devices
• calibration
• blood pressure from patients with heart disease is collected at the hospital vs.
collected by Walgreen machine for comparison group without heart disease.
• operator differences

• Laboratory Assays
• sensitivity/specificity
• falsely classify exposure or disease
• sample collection and storage issues
• blood collected from cancer cases in hospital and immediately frozen vs. carried
around in a cooler for 10 hours after community collection from comparison group.
• operator differences
Types and Consequences of
Information Bias
Types

4Non-differential (“noise”)
4 e.g., using an insensitive and non-specific test to classify exposure or disease
4Differential
8 recall bias
8 interviewer probing cases more than comparison group

Consequences

8 Non-differential: generally dilute effects

8 Differential: increase or decrease effects
Detecting and Controlling
Information Bias
1. Validation studies
2. Ongoing Quality Assurance/ Quality Control
• HOW???
• Standardized protocol for data collection

• Sources and methods of data collection should be similar for all study groups

• Maintain interviewers and study personnel as unaware of exposure/ disease

status as possible

• Adapt a strategy to assess potential information bias

Thank you