Comments on Health Status Transitions

1
Comments on Health Status Transitions

• Byron G Spencer
• McMaster University

2
WPIII WPIV

• Andrew Bebbington and Judith Shapiro,
• Incidence of Poor Health and Long-Term Care
  Health Transitions in Europe Results from the
  European Community Household Panel Survey and
  Institutional Data, December 2006
• Maria Hofmarcher, Monika Riedel, Alexander
  Schnabl, and Gerald Sirlinger,
• Ageing, Health Status, and Determinants of
  Health Expenditure (AHEAD) Health Status
  Transitions, January 2007

3

• The concern the estimation of first order
  transitions -- only two periods of information
  required.
• To be done using a panel survey from 1994 to
  2001, initially involved 130,000 individuals
  living in 12 EU countries Austria, Finland and
  Sweden added later.

4

• Aside from sample attrition, more than 130,000
  observations each year for a period of up to
  eight years, and hence up to seven transitions
• However the information challenges proved to
  substantial

5
The problems

• survey instruments not standardized
• key questions differed from one country to
  another
• that was true even of SAH the variation was
  substantial (BS, p 4)
• even so, the results were usable among those who
  remained in the sample, less than 10 of
  transitions could not be calculated
• a major problem was sample attrition
• the concern here (as always) is that the
  attribution is not random
• it seems likely that attrition would be
  disproportionately among those with relatively
  poor health, and thus would be importantly
  associated with (unobserved) transitions to
  institutions or to death
• however, it appears that attrition rates differ
  little when classified by initial state of health
  and the second-year state includes death

6

• but the biggest problems arise from very
  considerable under reporting of (or failures to
  report) transitions to residence in long-term
  care facilities and to death
• furthermore the exercise requires that the count
  of deaths distinguish between those who died when
  normally residents of private dwellings or of LTC
  facilities
• data problems were so severe that it was not
  possible to draw comparative conclusions on
  mortality or institutionalization from any of
  these records (BS, p 8) even though that was
  a central purpose of the exercise!
• the problems of measuring the institutional
  population, and specifically the flows into and
  out of that population are well documented in BS

7
WPIII what is done?

• They consider two measures
• - SAH VG, G, F, B, VB
• - combined B and VB because of small
  numbers
• - HHC Hampering Health Condition
• combined No chronic condition with chronic
  condition but not hampered compared to chronic
  condition and hampered
• - (WP IV focussed only on SAH)
• in both studies the intention is to allow also
  transitions from each of these health states to
  LTC and to death, but not back

8

• within each country observations on annual health
  transitions were pooled
• poststratification adjustments to the weights of
  those who died, based on vital statistics records
  falling outside the sample
• however, similar adjustments are not possible for
  institutionalization
• partially ordered-probit functions estimated for
  persons living in the community (separately for
  those under 65 and those over) age and gender
  were entered as covariates and separate estimates
  were obtained for each initial health state
• allowance is made for possible transitions to
  each of the initially designated health states
  and also to death as an absorbing state, but not
  to residential care
• for those 65 and older, probit functions were
  estimated for admission to residency in a health
  care institution, using age-group admission rates
  which, under the assumptions, are the same as
  mortality rates for those institutions estimates
  are by gender age and age-squared are the
  covariates

9
Comments

• for those living in the community, why not pool
  those under 65 and 65 using a more flexible
  functional form if necessary
• it would be helpful to display illustrative
  results graphically
• for example, plots showing the probabilities
  of transition from each initial health state as a
  function of age
• authors note that the preferred functional form
  provided a suitable fit for all countries and
  that there are plausible regularities in the
  estimated results that are similar across
  countries
• that suggests that some cross-country pooling of
  observations might be possible with appropriate
  adaptation of the error specification something
  that could be explored further

10

• their finding that gender rarely has significant
  explanatory power in the transition equations is
  consistent with our results for Canada, once we
  controlled for income and education
• ignorance of the population in long-term care
  facilities remains a major problem
• that is true also in the Canadian context, where
  each province defines the term
• differences in definition rather than practice
  appear to account for much of the differences in
  reported rates of institutionalization across
  provinces

11
WP IV

• The purpose of this work is, again, to build a
  picture of the movements in health status of the
  whole population of each country by age (p 1),
  but to do so building on WP III, since not all
  the relevant transitions were considered there

12

• the major data problems are, of course, the same
• transitions into both LTC and death are greatly
  under reported
• not possible using the ECHP to estimate mortality
  rates separately for those in LTC and those not
• the solution adopted is to produce a set of
  demographic accounts using the best available
  information about population stocks and annual
  flows
• the general idea is to make the best use of all
  relevant information in order to complete the
  demographic accounting matrix

13

• much effort was put into obtaining data on
  residential care
• in the end remarkably very few countries were
  able to provide satisfactory (or even minimal)
  information
• only 8 of 14 countries could provide estimates of
  the population in residential care by age
• of those, only six could provide estimates also
  by sex
• only the Netherlands and Finland could provide
  estimates by a/s of the number of deaths among
  those in residential care even here, serious
  concerns about the quality of the numbers
• information is limited, but suggests that the
  prevalence of LTC varies enormously across the
  countries surveyed
• in consequence, it was concluded that transitions
  into (and out of) LTC would have to be country
  specific

14

• given the available information, it was decided
  to limit the development of the complete tables
  to three countries Belgium, Germany, and the UK
• Dutch mortality rates for those in LTC assumed to
  apply
• no correction is made for the under reporting of
  either deaths of number of residents, thereby
  implicitly assuming that the ratio of the two
  values is independent of such a correction

15
Comments

• why is the analysis is not based entirely on the
  proportions (or probabilities) rather than on the
  number or level of transitions?
• data smoothing could apply to the proportions

16

• the Stone algorithm was applied to levels rather
  than proportions, and major problems were
  encountered with implied negative transition
  probabilities, even after data smoothing
• perhaps such problems would have been reduced if
  proportions had been used throughout
• as an aside, the use of Sprague multipliers to
  adjust the age-specific numbers in LTC (fig 2)
  produces odd results why does the series not
  increase with age?
• because of measurement problems the sum of the
  rows and columns in the matrix are uncertain
  hence they are allowed to vary I note that
  working with proportions would avoid that problem
  and simplify the analysis

17

• the estimated transition probabilities appear
  generally plausible
• summarizing those probabilities through panel
  regression analysis is helpful, and allows us to
  see how patterns vary with age and sex, and
  across countries
• indeed, one might place greater confidence in the
  estimated probabilities based on the regressions
  it is entirely plausible that they would progress
  smoothly (though not linearly) with age, as
  required by imposing such a functional form

18

• calculations relating to life expectancy
  conditioned on LTC also help with the
  interpretation
• the results here are surprising for men aged 65
  it varies from 6 weeks (in Germany) to 13 weeks
  (Belgium) the average is somewhat more than
  twice as long for women, except in Belgium, where
  it is estimated to be more than three times as
  long at most ages
• one would like to get behind these estimates, to
  assess their reliability
• I note that they are not consistent with other
  evidence BS who state that the median time from
  admission to mortality in England is 20 months
• Inconsistent also with the inverse of the
  turnover rate, which suggests about 30 months
• the experiments designed to see what it would
  take to reduce weeks lived in residential care
  are interesting
• not surprisingly, one conclusion is that a lesser
  reduction in the transition probabilities would
  be required if people were in better initial
  health

19

• some four pages are spent discussing the possible
  of an EU benchmarking model based on the
  transitions, apparently in order to identify the
  most efficient health care provide
• however, given the huge data problems, such an
  application would seem premature at this time

20
Canadian SLID

• This is a repeated six-year panel.
• The first in 1994, with new ones starting in
  1997, 2000, and 2003
• We now have three complete panels, each with
  about 35,000 adults followed for six years.
• Information about health is limited, respondents
  are asked the standard SAH questions
• But Statistics Canada has generally been able to
  follow respondents who were institutionalized or
  died, partly by matching with vital statistics.
  Attrition rates were extremely low e.g., only
  0.8 of the 1996 respondents were not accounted
  for in 1998.