Followup and confounding followup study theme 23052006

1
Follow-up and confoundingfollow-up study
theme23-05-2006

• Volkert Siersma

2
A study
exposure
outcome
?
3
Mediation
The mediating variable should not be in the
analysis as we are interested in the overall
effect of the exposure
covariate
exposure
outcome
4
Confounding
Part of the outcome, but also the exposure is
influenced caused by the covariate. This
effect needs to be counted in the analysis.
covariate
exposure
outcome
Randomisation of the exposure removes the
confounding effect
5
Follow-up study diagram a schematic example
6
HbA1c development and GP characteristics
Configuration of GP characteristics
The development of HbA1c
exposure
outcome
Hansen LJ, Olivarius NdF, Siersma V, Andersen JS.
(2003) Doctors characteristics do not predict
long-term glycaemic control in type 2 diabetic
patients. Br. J. Gen. Prac., 53, 47-49
7
HbA1c development and GP characteristics
(only) patient age influences slope, plt0.0001, in
a random effect linear regression model with
various GP characteristics as covariates.
Age, Sex, Practice form, Experience, Help in
practice, Nr. of patients, Location, etc.
Hansen LJ, Olivarius NdF, Siersma V, Andersen JS.
(2003) Doctors characteristics do not predict
long-term glycaemic control in type 2 diabetic
patients. Br. J. Gen. Prac., 53, 47-49
8
HbA1c development and GP course attendance
Baseline covariates known to influence HbA1c
development
Not confounders, just to get smaller variance.
Could just as well have been confounders though
covariate
Speed of HbA1c increase
exposure
outcome
Number of courses the GP went to
Olivarius NdF, Siersma V. GP project course
attendance and its relation to the patients
treatment results.
9
HbA1c development and GP course attendance
The influence of GP course attendance modelled as
a time-varying covariate in a random effect
model. P0.0103
Curious the course dosage effect gives first
higher Hba1c and only after attendance of 5 or
more courses a lower HbA1c
Project courses
Olivarius NdF, Siersma V. GP project course
attendance and its relation to the patients
treatment results.
10
Longitudinal models
The development of one outcome modelled in detail
through a function where time is explictly
mentioned. The development itself is the outcome.
11
Longitudinal models
Detailed modelling of the outcome. No
equidistant observations needed. Random
effects, correlation structures.
Interpretation of time-varying covariates
difficult. No feedback, i.e. disease
development dependent on treatment AND
treatment dependent on disease history.
12
HbA1c and symptoms at diagnosis
Plt0.001 in a multivariate logistic regression of
log(HbA1c) on the precense of symptoms and 9
other covariates.
Age Sex Systolic BP BMI Macrovascular
complications Retinopathy Renal
involvement Peripheral neuropathy Antihypertensive
medicine
HbA1c
Symp
Drivsholm T, Olivarius NdF, Nielsen ABS, Siersma
V. Symptoms and signs in newly diagnosed type2
diabetic patients and their relation to
glycaemia, lood pressure and weight.
13
HbA1c and symptoms one year after diagnosis
P0.1141 in a multivariate logistic regression of
the Precense of symptoms on HbA1c and 12
other covariates.
Age Sex Systolic BP BMI Macrovascular
complications Retinopathy Renal
involvement Peripheral neuropathy Antihypertensive
medicine
HbA1c
Symp
Cohabitation Smoking Phys. activity
Nielsen ABS, Olivarius NdF, Gannik D, Siersma V.
Symptoms, self-rated health and glycaemic control
one year after diagnosis in patients with type 2
diabetes mellitus.
14
Multivariate models
The dependencies between the various variables
can be modelled in detail Time can be incorporated
15
Multivariate models
Relationships between developments of multiple
variables. Visualisation of correlation structure
s.
Observations at fixed time points. Large models.
16
Dynamic treatment
treatment
treatment
Were interested in the final disease status.
A
B
disease
disease
Mediating variable for A Confounder for B
17
Modelling of transitions
treatment
treatment
disease
disease
But the effect of treatment in this transition
model is the short-term effect. That is not too
interesting
18
Effects in dynamic treatments
The effect of a sequence of treatments is wanted,
or rather the effect of a strategy. The aim is to
find an optimal strategy. By careful modelling
one can evaluate strategies and even try to
optimise them.
19
Causal dynamics
All other possible confounders
treatment
treatment
disease
disease
20
Advice

• Formulate the aim of the study in non-statistical
  terms
• Talk with a statistician before you start
• This stuff is difficult