Title: Project Overview
1Issues in estimating incidence from linked
hospital admissions data
Tim Churches and Baohui Yang Centre for
Epidemiology and ResearchNSW Department of
Health October 2008
2Introduction why estimate incidence
- Health authorities traditionally report annual
population-based counts and rates of hospital
admissions for particular diseases and conditions - Such metrics reflect both the underlying disease
incidence rates in the community, and variations
in hospital utilisation for those diseases and
conditions - Cases of diseases are using defined as being
incident when they are first diagnosed or
detected - We are interested in variations in both hospital
utilisation and incidence, but it is much better
if they can be disentangled. - Estimation of disease incidence is required to
properly assess the impact of disease prevention
efforts, and to estimate prevalence and burden of
disease
3CVD hospital separation rates
4Estimating incidence from hospitalisations
- If the disease is acute and requires only one or
few hospitalisations, then hospital admission or
separation rates may approximate incidence rates - eg community-acquired pneumonia
- For chronic diseases which require multiple
hospitalisations, annual hospital admission or
separation rates bear little or no relationship
to underlying disease incidence rates - However, for some diseases, first hospitalisation
for each person for the disease may be used as a
proxy for incidence, if - almost all cases of the disease require hospital
treatment as an admitted patient - hospital treatment is typically required soon
after diagnosis - Linked data allows us to identify the first
hospital admission for each person for the
disease of interest, and thus estimate underlying
incidence of diseases which meet these criteria
5The Prevalent Pool Effect
2
2
1
2
1
1
1
0
Time ?
6The Prevalent Pool Effect
2
2
1
2
1
1
1
0
Time ?
7The Prevalent Pool Effect
5
2
1
1
1
0
Time ?
8Apparent incidence of lung cancer
9Actual incidence of lung cancer
10Solutions to the Prevalent Pool Effect
- Use a clearance period
- initial years of linked data are ignored
- how many years depends on condition
- clearance period derived by visual inspection of
trends - but no way to distinguish a true downward trend
from the prevalent pool effect - wastes a large amount of data
- only practical if many years of linked data are
available
11Backcasting methods
- Developed by Brameld and Holman in WA
- Kate J Brameld, C DArcy J Holman, David M
Lawrence, and Michael ST Hobbs (2003). Improved
methods for estimating incidence from linked
hospital morbidity data. International Journal of
Epidemiology 2003 32 617-624 - Retrograde actuarial survival model to derive
correction factors for over-ascertainment of
first-time events - described two methods
- a less accurate annual method
- a method which provides individual-level
correction factors - per-individual method is more accurate and quite
easy to implement - based on the estimated probability that an
apparently first-time admission in an individual
may have actually been preceded by unobserved
admissions
12Brameld-Holman backcasting method theory and
assumptions
- a is date of admission
- b is date of previous admission
- c is the start of the observation period
- retrograde follow-up time l a b (for a
previous admission), - or l a c (if censored)
- use life-table to calculate the hazard of a
previous admission at reverse time ?(t) and the
corresponding survival function S(t) - Assumption is that for a chronic disease for
which the retrograde hazard function is
monotonically decreasing in reverse time, there
is a point tf at which ?(tf ) 0 ie there is no
further hazard of a previous admission prior to
reverse-time tf
13Brameld-Holman backcasting method theory and
assumptions
- calculate a correction factor for
over-ascertainment of incident cases due to the
prevalent pool effect for an apparently first
admission at time tj - for reverse times lt tf , Cj S(tf ) / S(tj )
- for reverse-times gt tf, , Cj 1
- The correction factors for each case, Cj, are
summed to give an adjusted count by year of
incident cases
14Brameld-Holman backcasting method
implementation
- fractional polynomial regression is used to fit a
smooth curve to the life-table hazard function in
order to extend it into the past - this curve is then inspected to determine the
reverse time at which it approximates zero - a fractional polynomial curve is fitted to the
survival and this is used to evaluate S(tf ) and,
for each individual, S(tj ), and Cj is then
calculated - We did all this in SAS using a custom fractional
polynomial fitting macro written by Baohui Yang - Fractional polynomials are a convenient way of
fitting non-linear regression curves, but other
curve-fitting methods can be used
15Issues in implementation of the method
- Quite a long period of follow-up is need for many
chronic diseases in order to determine when the
hazard approaches zero. - We used 6 years of linked data
- Sufficient for relatively acute chronic diseases
such as lung cancer, where diagnosis is quickly
followed by treatment and then either recovery of
death - Insufficient of many other disease with a
chronic, relapsing course, including breast
cancer and even AMI - Hazard function for previous admission for these
condition does not approximate zero at 6 years in
the past - Some conditions have a pattern of remission and
relapse, such as many (treated) leukaemias, and
in these, the hazard function may be slightly U
shaped
16Hazard of previous admission for leukaemias
17Hazard of previous admission for leukaemias
18Modified backcasting method
- Because of the difficulties in fitting a
monotonically decreasing curve to hazard and
survival functions that, at 6 years of
reverse-time was not yet flat, we decided to omit
this step and just use 6 years for tf - We found that the correction technique was robust
to the selection of tf and this simplification
made little difference to the results - There were still problems with conditions for
which the survival function S(t) was still not
very flat at the 6 years of reverse-time
available to us. - The fractional polynomial regression fitting
procedures needed to be constrained to only
select monotonically decreasing polynomial
expressions.
19Results Survivorship for lung cancer
20Incidence estimates for lung cancer
21Incidence estimates for COPD
22Incidence estimates for schizophrenia
23Incidence estimates for ischaemic heart disease
24Conclusions
- Some form of adjustment (or use of a clearance
period) is essential the prevalent pool effect
is real and substantial - The Brameld-Holman backcasting incidence
adjustment method works well if there is a
sufficiently length of follow-up available - If not, great care needs to be taken
- Anomalous corrections are possible
- The method can be simplified by omitting the
identification of tf - Other curve-fitting methods may give better
results - Planners of record linkage facilities Important
to have the longest times series of linked data
as possible
25References
- Kate J Brameld, C DArcy J Holman, David M
Lawrence, and Michael ST Hobbs. Improved methods
for estimating incidence from linked hospital
morbidity data, International Journal of
Epidemiology 2003 32 617-624 - Patrick Royston, Gareth Ambler and Willi
Sauerbrei. The use of fractional polynomials to
model continuous risk variables in epidemiology.
International Journal of Epidemiology 1999 28
964-974.