Friedo Dekker

Dept. Clinical Epidemiology

Leiden University Medical Center
F.W.Dekker@lumc.nl
 Is grey hair a risk factor for death?

Death Alive
Grey + 45 55

Grey – 30 70
200

• Risk to die in Grey + = 45%

• Risk to die in Grey – = 30%
 Relative Risk (RR) = 1.5

 Is this true?
 What’s the problem?
 What do we do about it?
 Is grey hair a risk factor for death?

OLD Death Alive YOUNG Death Alive

Grey + 40 40 Grey + 5 15

Grey – 10 10 Grey – 20 60
100 100

• In old people more grey hair (80% versus 20%)

• In old people more deaths (50% versus 25%)

Age

?
Grey
†
 Is grey hair a risk factor for death?

OLD Death Alive YOUNG Death Alive

Grey + 40 40 Grey + 5 15

Grey – 10 10 Grey – 20 60
100 100

• In old people more grey hair (80% versus 20%)

• In old people more deaths (50% versus 25%)
• Within old people, grey hair is no risk factor (50%/50%  RR=1)
• Within young people, grey hair is no risk factor (25%/25%  RR=1)

 Overall (adjusted) Relative Risk (RR) = 1.0 (crude RR was 1.5)

 Effect of grey hair on mortality fully explained by age
 Age is a confounder
 Confounding
to confuse and very much surprise someone,
so that they are unable to explain or deal with a situation
http://dictionary.cambridge.org

“Confound thy enemies…”

C

?
E D

• One effect is obscured by another effect

• Interference of another effect in interpretation
• Something you want to get rid of
 Confounder
1. Additional risk factor for outcome
and

2. Unequally distributed on levels of risk

factor under study
and

3. No causal effect of risk factor itself

 Confounding (1)

C
Confounder
?
E D

C No confounder
(other risk factor)
?
E D
 Confounding (2)

C No Confounder
(other disease)
?
E D

No confounder
E C D (intermediate /
in causal pathway)
 Confounding? (1)

Association between grey hair and death.

Is age a confounder?  YES

Association between age and death.

Is grey hair a confounder?  NO

C
?
E D
 Confounding? (2)
Association between HD/PD and survival.
Is GFR at start of dialysis a confounder?  YES

Association between HD/PD and survival.

Is GFR at 6 months a confounder?  NO

Association between HD/PD and survival.

Is ACE-polymorfism a confounder?  YES

Association between ACE-polymorfism and survival.

Is HD/PD a confounder?
C
 NO
?
E D
 Confounding? (3)
• Study effect of ace-inhibitor on survival. Is subsequent
blood pressure (on ace-inhibitors) a confounder? NO

• Study effect of blood pressure on mortality. Is use of

subsequent medication a confounder?  NO

• Study effect of blood pressure on mortality. Is color of

your eyes a confounder?  Yes
• And color of your socks?
 Yes
 Confounding? (4)
Study effect of any gene on survival in dialysis
patients (e.g., ACE-polymorfism)

• Is blood pressure a confounder?

• Cardiovascular disease?
• Diabetes?
• Anything else?
•
•
•
 Cave!
• Previous knowledge necessary to decide whether:
– Something is an independent risk factor
– Something could be related to risk factor of interest
– Something is NOT in the causal pathway

• Can the data themselves tell you whether something is

a confounder?
 NO !!!
• Can a statistician tell you whether something is a
confounder?
 NO !!!
 How to deal with confounding?
(1) In study design
• Randomisation (i.e., allocate both known
and unknown confounders equally towards
both treatment groups)

• Restriction (i.e., study only old people)

• Matching (i.e., for each patient with grey

hair, select a patient without grey hair who is
of the same age)
 How to deal with confounding?
(2) In analysis
• Stratification (separate analysis for all
levels of confounder)

• Multivariate analysis:
– Linear regression
– Logistic regression
– Cox regression

Note: With the exception of randomisation,

all strategies only work for known confounders!
Examples from the literature (1)

Effect of BMI on risk to develop ESRD

What confounders to adjust for?

• age?
• sex?
• systolic blood pressure?
• proteinuria?
 Kidney International 2004

Adjusted for age, sex, SBP, proteinuria

 What if I run a multivariate model here to look for independent risk factors?

 Or to look for independent predictors? Age is no independent factor here?

 For death, stop breathing for 5 minutes is the only independent factor?
Examples from the literature (2)

Effect of haematologic malignancy on mortality in

patients with acute renal failure at ICU

Age
Sex
Apache
…

Haem. ?
malign. †
+/-
 NDT 2005

Crude HR haem. Malignancy = 1.52 (p=0.018)

HR adj for Apache-score = 1.28 (p=0.16)
NDT 2005
 Aetiology
• Study effect of a risk factor on an outcome
• Consider confounding:
– other risk factor for the outcome,
– also related to risk factor of interest,
– and not being in the causal pathway
 interference with relationship to be studied

• Adjust for confounding (stratification or

multivariate analysis)  real effect
 Prognosis (1)
• Try to predict outcome based on one or more risk factors

• What risk factors can be used as predictors?

 anything!

(e.g., blood pressure, use of medication, color of your

hair, even color of your socks)
 Prognosis (2)
• Put all possible risk factors in a model

• Let computer decide what is the strongest combination

• Weak predictors will be thrown out of the model

• Variables that add little on top of the other variables in

the model will be thrown out as well

• Stepwise modelling
 Prognosis (3)
• Example:

• GFR is strong predictor of mortality in dialysis

• MDRD and Cockcroft-Gault both estimate GFR

• Which one will be in the model?

• Other one is not important?

 Prognosis (4)
• Model will result in ‘best’ prediction

• Confounding is not an issue here

• Model might help to predict the future

• But not necessarily to understand why things will happen

 Cave interpretation!
 Examples (4): AJKD 2000
Aim: To determine which factors (albumin, CRP, …)
are most closely related to CV death

Both CRP and albumin are univariate predictors

in HD patients (p<0.005)

“Cox model: high CRP is predictive of CV death

(p=0.0001) while albumin is not (p=0.5)”

conclude that albumin is not prognostic?

conclude that albumin does not play a role?
 Cave!
• In prediction models, all ‘significant’ predictors are in the
model
• Intermediate variables or risk markers can be selected
in stead of true causes
• e.g., grey hair can be selected in stead of age
• Important confounders are sometimes not selected

 Consequently, variables in prediction model are not per se

important in the aetiology of the outcome

 Before model building: Define your specific aim:

either aetiology, or prediction
Carefully consider the purpose of your study:

Understanding = aetiology

Prediction: WHY would you like to predict?

1. To understand risk factors = aetiology
2. For intervention / to treat risk factors = aetiology
3. To estimate the relative risk of a variable = aetiology
4. To predict with modifyable risk factors = aetiology

5. To estimate the absolute risk for a practical decision = prediction

(start treatment yes/no, refer to nephrologist yes/no)
6. To estimate the absolute risk to inform a patient = prediction
7. Risk stratification / identify high risk group = prediction
8. Added value of a biomarker to a predictionmodel = prediction
(e..g., add CRP to Framingham risk score)
Predictors for the development of microalbuminuria and macroalbuminuria
in patients with type 1 diabetes: inception cohort study P

Conclusions Around one third of patients newly diagnosed with type 1 diabetes
develop persistent microalbuminuria within the first 20 years of diabetes. Several
potentially modifiable risk factors predict the development of persistent AP
microalbuminaria and persistent macroalbuminuria.

A Cox proportional hazards regression model was used to evaluate the relative A
contributions of covariates to the risk of developing persistent microalbuminuria,
correcting for duration of follow up. A stepwise backward selection was used. P

We have shown that poor

glycaemic control at onset of
diabetes (determined by first
assessment of haemoglobin A1c
concentration six months after the
onset of type 1 diabetes) is an
important predictor for the P
development of microalbuminuria.
Efforts should therefore be made
to normalise glycaemia when A
diabetes is first diagnosed.
P?? BMJ 2004….
 “DO” and “DON’T” (1)
• DO consider confounding in aetiology:
– independent risk factor for outcome
– associated with risk factor of primary
interest

• DON’T adjust for variable in the causal

pathway
 “DO” and “DON’T” (2)

• DO consider stepwise models in prognosis

• DON’T interpret variables in a prognostic

model as aetiology
You may confound thy enemies…

but do not get confounded yourself!