578 Panel Data Methods
observations, and are comprehensive, often providing observations on a complete
population rather than a random sample. They tend to be less prone to unit and
item non-response than survey data and may give better coverage of hard-to-reach
groups of the population and the socially disadvantaged. Also they tend to be less
affected by reporting bias, but are still vulnerable to data input and coding errors.
Given their primary purpose, administrative datasets are not designed by and for
researchers. This means they may not contain all of the variables that are of interest
to researchers, such as socioeconomic characteristics, and many different sources
may have to be combined to produce a usable dataset. In some cases, sources of
administrative data may be combined and made available with researchers in mind.
For example, the Oxford Record Linkage Study (ORLS), used by Seshamani and Gray
(2004), is a longitudinal dataset that links statistical abstracts for hospital inpatient
and day cases to birth and death certificates for people living in the Oxford region
of England. It provides 10 million records for over 5 million people between 1963
and 1999.
Dusheikoet al.’s (2004) study of the impact of practice budgets for GPs on hospi-
tal waiting times in the English National Health Service (NHS) provides an example
of the complex and painstaking process that is often required to link administra-
tive data. Information on waiting times was obtained from the Hospital Episode
Statistics (HES) for 1997/98 to 2000/01. HES is an annual database of hospital
inpatient activity, including day cases, with more than 10 million records per
year. Dusheikoet al.extracted information on the waiting times for over 5 mil-
lion finished consultant episodes and linked average waiting times to GP practices.
Information on practice populations was obtained from the Primary Care Trust
(PCT) database at the National Primary Care Research and Development Cen-
tre (NPCRDC: http://www.primary-care-db.org.uk)..) Practice characteristics, such
as the GP’s age and sex, qualifications, size of practice, etc., were obtained from the
Prescription Pricing Authority, the Department of Health’s Organisational Codes
Service and their General Medical Statistics, along with the NPCRDC database.
Patient characteristics for each practice were obtained from the 1991 Census and
components of the Index of Multiple Deprivation, with these small area data
mapped to GP practices. Finally, supply side factors, such as distances to hospi-
tals, were obtained from the Department of Health’s Allocation of Resources to
English Areas (AREA) project.
Some of the key administrative datasets that have been used in health
economics are summarized in Table 12.1.
12.3.1.1 Non-response and attrition
Non-response and attrition are a common feature of longitudinal survey data.
Nicoletti and Peracchi (2005) list possible reasons for non-response: these include
demographic events, such as death; movement out of the scope of the survey,
such as institutionalization or emigration; refusal to respond at subsequent waves;
absence of the person at the address, along with other types of non-contact. Jones
et al.(2006) investigate health-related non-response in the first 11 waves of the
BHPS and the full eight waves of the European Community Household Panel