drawing causal inferences in epidemiological studies of early life … · 2013. 5. 22. · maternal...
TRANSCRIPT
Drawing causal inferences in epidemiological studies of early
life influences
Andy NessAlex GriffithsLaura HoweSam Leary
Note: for non-commercial purposes only
Structure of this talk
• Analytic studies, chance, bias and confounding
• Examples of chance, bias and confounding
• Approaches to chance, bias and confounding
Analytic studies, chance bias and confounding
Cohort and case-control studies
PAST PRESENT FUTURE
PROSPECTIVE COHORT STUDY
HISTORICAL COHORT STUDY
RETROSPECTIVE CASE CONTROL
STUDY
Cases and controls ascertainedExposure recalled
Disease recorded
Disease recorded
Bias
A trend in the collection, analysis, interpretation, publication or review of data leading to conclusions systematically different from the truth.
Missing data in early life studies
� May be years/decades between the early life exposure and the outcome of interest
� Loss to follow-up (usually) increases with time
040
0080
0012
000
Num
ber of
que
stio
nnai
res
retu
rned
0 5 10 15
Age (years)
ALSPAC questionnaire
responses: birth to 15yrs
ConfoundingA confounder is a factor associated with the exposure which independently affects the risk of developing the disease.
Exposuree.g. Birthweight
Diseasee.g. Blood pressure
Counfoundere.g. Social class
Strength of the associationBiological credibilityConsistency with other studiesTime sequenceDose responseSpecificity
Causal Not causal
ChanceBiasConfounding
Is the association causal ?
Examples of chance, bias and confounding
Subgroup analyses Calcium intake during pregnancy and offspring blood
pressure at age 7, according to body mass index
* Intervention (2g calcium/day during pregnancy) vs placebo * Mean calcium intake during pregnancy approx 1g day
Findings from subgroup analysis may not be replicable
ALSPAC (observational), Leary et al. (N=6884)
-12
-10
-8
-6
-4
-2
0
2
4
6
≤14.8 14.8-15.8 15.8-17.0 >17.0
Body mass index age 7 years (kg/m2)
Ch
ang
e in
sys
toli
c b
loo
d p
ress
ure
(m
mH
g)
per
1g
in
crea
se i
n c
alci
um
*
Follow-up of calcium trial, Belizan et al. (N = 591)
-12
-10
-8
-6
-4
-2
0
2
4
6
≤14.4 14.4-15.7 15.7-17.5 >17.5
Body mass index age 7 years (kg/m2)
Dif
fere
nce
in
sys
toli
c b
loo
d p
ress
ure
(m
mH
g)
bet
wee
n g
rou
ps
*
Belizan et al. BMJ 1997;315:281-5 Leary et al. Archives of Disease in Childhood 2005;90:492-3
Birthweight and blood pressure
Rachel Huxley et al Lancet 2002
Advertisement in Boston Globe
0.1 0.5 0.75 1 1.25 1.5 1.75
Male health workers
Social insurance, men
Male chemical workers
Hyperlipidaemic men
Nursing home residents
Social insurance, women
Male physicians
Male smokers
(Ex)-smokers, asbestos workers
Trials
Cohorts
Skin cancer patients
USA
Finland
Switzerland
USA
USA
Finland
Cohorts combined
Trials combined
Finland
USA
USA
USA
Relative risk (95% CI)
Beta-carotene and Cardiovascular mortality
Egger et al BMJ;1998:316:61-66
Vitamin C and birth weight: Observational study and RCT
-150
-100
-50
0
50
100
150
Mathews et al1999
(Observational)
Poston et al2006
(RCT)
grams
“Well, so much for antioxidants.”
Approaches to chance, bias and confounding
Observational epidemiology?
Avoiding chance
• Design adequately powered studies• Do the correct analysis• Avoid statistical significance• Present p values and confidence intervals• Look at effect sizes and their clinical importance• Focus on pre-specified main effects• Report exploratory analyses as such• Replicate subgroup findings
The smaller the p-value, the stronger
the evidence against the null
hypothesis
P v
alu
e
.0001
.001
.01
.1
1
Weak evidence against the null hypothesis
Strong evidence against the null hypothesis
Increasing evidence against the null hypothesis withdecreasing P value
Interpretation of p values
Sterne JAC, Davey Smith G. British Medical Journal 2001;322:226-31Sifting the evidence: what’s wrong with significance tests?
Clinical significance?
Weight at birth and systolic blood pressure at age 3
1860 children (ALSPAC)
After adjustment for current size:Regression coefficient = -1.9 mmHg/kg
Confidence interval = -2.61, -1.21 mmHg/kgP < 0.0001
Interpretation:
Strong evidence against null hypothesis, blood pressure BUT differences small considering birthweight changes achievable
Avoiding bias (in cohorts)
• Reduce losses to follow up
• Report characteristics of those lost to follow up
• Compare complete case analysis versus analysis in subjects with missing data
• If missingness is related to values of observeddata, consider multiple imputation
• If missingness is related to values of unobserveddata, unbiased effect estimates are not possible
05
1015
2025
Wei
ght (
kg)
0 1 2 3 4 5Age (years)
Growth curves for weight, birth to 5 years
Unpublished data ALSPAC
Henderson et al, Thorax 2008;63:974-980
Avoiding confounding
• Follow up of trials • Temporal trends • Ecological explanatory power• Heterogeneity of confounding structure• Instrumental variable approaches• Specificity • Critical time periods• Sibling and twin studies in early origins
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
<1000
1000-1500
1501-2000
2001-2500
2501-3000
3001-3500
3501-4000
4001-4500
4501-5000>5000
NS
B ir th w e ig h t (g ram s)B ir th w e ig h t (g ram s)B ir th w e ig h t (g ram s)B ir th w e ig h t (g ram s)
1951
1967
2000
US Birthweight Distribution 1950-2000
Personal Communication Professor John Lynch
Owen et al. IJE (2005)
Mean Birthweight and systolic blood pressure (INTERSALT)
Breastfeeding and outcomes
ALSPAC Pelotas Belarus
SES gradient No SES gradient Trial
BP Inverse No association No effect
BMI Inverse No association No effect
IQ Higher IQ Higher IQ Higher IQ
Preliminary data Marie-Jo Brion et al
Instrumental variables
Need to find a variable which:
• is associated with the exposure of interest
• is not related to confounders
• has no direct effect on the outcome of interest (i.e. affects the outcome only indirectly through the exposure)
FTO acts via obesity
Maternal FTO and fat mass
6000
7000
8000
9000
Med
ian
tota
l fat
mas
s at
age
9 (
gram
s) TT AT AA
Offspring FTO genotype
TT AT TT AT AA AT AA
Maternal FTO genotype
Maternal smoking and offspring birth weight
• Maternal smoking associated with greater decrease in birth weight than paternal smoking
• Paternal smoking has little effect after adjusting for
maternal smoking
Davey Smith, BCPT 2008;102:245-256
Maternal smoking and offspring BP
Maternal and paternal smoking have similar
associations with offspring BP
Davey Smith, BCPT 2008;102:245-256
The New
Yorker,
November
2001
Any questions?