An example of a study problem

An example of a study problem

Bias or truth?

How it all started 1968/69

Invited:

1622 women aged 38, 46, 50, 54 and 60 years

Examined:

1462 women (90.1%)

Psychiatric follow-up

1992/93:

Those born 1908, 1914, 1918 and 1922, i.e. the four oldest age cohorts aged 70, 74, 78 and 84,

Interview

526 of the survivors agreed (89.6%)

CT scan:

277 women agreed

Samples

• 120 ml blood

serum stored at –20°C in 2.5 ml covered polystyrene cups enclosed together in small batches, in firmly tied plastic bags for 28 years.

• Thawed once for other analyses at 25 y. and restored for two years at ‑80°C.

CT scan and Lacunar Infarcts (LI)

• Endpoint:– Lacunar infarcts– White matter lesions

Logistic regression

0 1 1 2 2

0 1 1 2 21 2 1 2( | , , , ) ( , , , ) ,1

p p

p p

x x x

p p x x x

eE Y x x x p x x x

e

Predictors – independent variables

• Homocysteine was analysed in tertiles using cut points previously calculated for the whole sample.

Predictors – independent variables

• Covariates included age, basic cardiovascular risk factors and influential factors for tHcy.

• The basic CVD risk factors were systolic blood pressure, diastolic blood pressure, serum cholesterol, serum triglycerides, BMI and smoking.

Predictors – independent variables

• Factors considered to be influential for tHcy were serum B-12, serum creatinine, coffee consumption and dietary folate.

Odds ratio for LI according to tHcy

    1st tHcy tertile 2nd tHcy tertile 3rd tHcy tertile

Range 3.3-9.7 9.8-12.6 12.7-41.5

Median 8.31 11.2 14.8

n n=87 n=110 n=80

Covariate studied

LI (n)      

Age alone 1 1.54(0.54-4.38) 3.65(1.34-9.90)

Systolic blood pressure* 1 1.50(0.52-4.24) 3.60(1.32-9.85)

Diastolic blood pressure* 1 1.44(0.50-4.13) 3.50(1.28-9.55)

Study design

• Type of study?

• Possible biases?

True or false

• The data must be representative

• if we want to

• generalize the results

• A homogenous sample

• makes generalization easier

Estimate of OR

valueTrue

0 ∞

OR = 1

Estimation, precision

Estimate

( | )

Estimate with confidence interval

Sample

95% confidence interval: 95% of repeated intervals will contain the true value

valueTrue

0 ∞

Precision and validity

• Measures of populations– precision - random error - statistics– validity - systematic error - epidemiology

Truevalue

Estimate

Precision

Bias

True or false

• It takes 2 to tango

• It takes 3 chords to play the blues

• It takes 4 numbers to be an epidemiologist

2 x 2 table

Exposure/test Disease No disease

positive a b

negative c d

Odds ratio for the study population;hypothesis – no effect

SA D

ORB C

No effect, OR=1

Odds ratio for the sample- observed effect

Da d

ORb c

OR> 1

Odds ratio for the sample

Da d

ORb c

OR> 1

2 x 2 table

Exposure/test Disease No disease

positive a b

negative c d

Odds ratio for the sample

Da d

ORb c

OR> 1

D too large? Those without disease and low tHcy more likely to be included?

Odds ratio for the sample

Da d

ORb c

OR> 1

a too large? Those with disease and high tHcy more likely to be included?

Conclusion 1

• If those ill and exposed are more likely to be examined, the effect estimate will be overestimated

• Likewise, if those unaffected and unexposed are more likely to be examined.

Odds ratio for the sample

Da d

ORb c

OR> 1

a too small? Those without disease and high tHcy less likely to be included?

Conclusion 2

• If those ill and exposed are less likely to be examined, the effect estimate will be attenuated

Generalization

• Do the results apply outside the sample?

• Statistical generalization– Representative sample

• Biological generalization– Information from outside the study– Homogenous sample