By Suzie Carson & Brian Easton
New Zealand Journal of Social Policy December 2000, p.121-128. Based on a paper presented to the 1999 Conference of the New Zealand Statistical Association, Wellington, July 5-7.
The increasing use of the Household Economic Survey for policy purposes raises issues about the assumptions which are used for transforming the unit records into aggregates which underpin the social policy analysis. This paper reports upon an HRC funded project to investigate the relationship between personal health status and economic status (especially location in the household distribution, but also in relation to other measures). The project uses unit records of the Household Economic Survey for 1994/5-6/7 years when personal health status was recorded, using both objective and subjective measures. The paper explores some of the processing issues which the analysis is addressing.
In the past research has been limited by access to the data base, which has been at either a high level of aggregation or the processing has been based on predetermined assumptions, without much opportunity to interact with the data to improve the estimates.(2) As a result a number of problems with the method have hardly been addressed. A research grant from the Health Research Council is funding extensive use of the Statistics New Zealand Data Laboratory (SNZDL) and so give the researchers direct access to HES data.(3) This offers the opportunity to deal with some of the past assumptions.
The Standard Model(4)
The HES collects a variety of information on household status and economic activity, including household composition and before-tax market plus benefit incomes. Each record contains both household-wide information (such as household spending and housing status), and individual information on each member of the household such as personal characteristics and income received (which can be aggregated up to household characteristics). Access to the SNZDL means that the project will be able to work with both sets of records.
The processing occurs as follows:
(1) The after tax (or disposable) income for each household can be calculated by applying known tax and abatement rates to individual records, and aggregating.
(2) Household needs vary with household composition, including the number in the household, and their ages. Aggregate household income is scaled to reflect this composition. Rather than use a simple per capita measure, a household equivalence index allows for economies of scale and the lower relative needs of children.
(3) The resulting ratio is called “household equivalent income” (HEI). The households are ranked in order of their HEI, and either divided into quantiles or partitioned by a poverty line (or lines).
The resulting estimates have been widely used. They are the primary data base for the debate on whether poverty has increased or decreased, and for the current discussion on income shares, which acknowledges that while the top decile has experienced a rise in its standard of living over the last fifteen years, the bottom eight deciles have not.(5)
Some Statistical Issues
Various conceptual issues complicate the standard model, including to what extent income is a measure of welfare? There is also a debate about the correct poverty line. Here we focus on the statistical ones and steps taken to resolve them, together with a comment in italics for the implications for social policy analysis.(6)
1. A minor, but frequently overlooked problem is that the data is often reported on a household basis. But since households have varying numbers of members, and because large households tend to be poorer, the proportion of household below any poverty line is less than the proportion of people in poverty. The analyst needs to check whether the data is presented by households or people.
2. Note also that the HES data is often reported in terms of years to March. This covers all the respondents in the period April of the preceding year to the march-identified year. Because respondents are asked to report retrospectively on their preceding year’s expenditure, the income data actually reflects more closely the period for September year preceding the March year. Thus, March 1999 year income data would be better attributed to the September 1998 year. This becomes important where there are income comparisons or where price deflation occurs. The data published in the recent edition of “New Zealand: Incomes” by SNZ allows for this,(7) but some of the earlier researchers do not. Unless there is an explicit mention of the time period issue it is probably sensible to assume the adjustment has not been made. This does not affect distributional shares over time, but without adjustment real income changes and timing of changes in real and nominal income levels may be wrong, although the long-run comparisons may be relatively reliable.
3. It turns out that incomes reported in the HES are inaccurate, probably to the extent of being 20 percent lower than accurate measures of income. SNZ is reviewing this inaccuracy. An important issue is whether the error has drifted over time (or even suffered a major change in the mid 1980s when benefit incomes were grossed up). Absolute levels may be incorrect, although changes over time may be more reliable.(8)
4. The results are sensitive to the choice of Household Equivalence Scale. A number are available, some based on a priori arguments, some on econometric estimation (at various levels of sophistication). The most popularly used, the Jensen 1988 scale (based on a priori assumptions), may be extreme compared to the other available ones, giving lowest poverty levels (especially among children).(9) One issue is the extent to which the scales change over time, perhaps as a consequence of relative price changes, especially the increased use of user charging for health services, education and housing which affects different household compositions and different positions in the income distribution (as when the user charging is income tested). The study has done some preliminary work (in draft publication) which suggests that the distribution in general, and the location of some social groups in particular, may be very sensitive to the choice of equivalence scale, among the ones currently available. Claudio Michelin (with Srikanta Chatterjee) has been econometrically estimating new equivalence scales.(10) While not yet having been used in the standard household model, preliminary indications are that they are of higher quality than those currently available.
5. Housing circumstances matter. The standard model treats those who own their own home (with or without a mortgage) and one who is renting (at market or subsidised rents) as all having exactly the same spending power with their reported income. One approach has been to deduct household spending on housing from disposable income. Not only does the resulting hybrid of income and expenditure indicate this method is conceptually wrong, but there also needs to be a resulting adjustment to the equivalence scales. A rigorous way may be to impute “normal” household spending on housing (based on housing characteristics and market income), and treat the net difference between imputed and actual housing spending as imputed income. This is yet to be done systematically. At this stage all the social policy analyst can do is be cautious. It seems likely that not adjusting for housing on average raises the incomes of households with children compared to those without, and lowers the relative incomes of the elderly who tend to own their own housing without mortgage.
6. A related problem exists for spending on education and health. For instance, a household with disability or illness may have medical outlays which a well household does not. The problem is not as large on average as for housing, and therefore probably not as acute across all households. It may be very important, though, to the sick. Again the only current counsel is caution in choice of equivalence scale and through time, especially where policy changes have affected private outlays on educational and health services.
There is, however, a deeper problem behind these statistics for the social policy analyst. What do they actually mean? So what, if we are told that such and such a percentage of the population are in poverty?
What we really want to do is be able to relate certain sorts of behavioural consequences – like the extent to which sickness is caused by poverty. While there are limitations from the essentially cross-sectional data of the HES, the inclusion of health status questions in some surveys means some progress is possible. A basic research technique is to contrast those who are well with those who are unwell. Suppose there is enough information on personal characteristics to make it is possible to predict each well person’s income and expenditure. The personal characteristics of each sick person can then be used to predict their income and expenditure as if they were well. The differences can be used to indicate to what extent the unhealthy’s living standards are depressed by poor health. The remainder of this paper describes how this might be done with the existing data.
Health Status and the HES Questionnaire
For the three years between 1994/95 and 1996/97 a set of health questions was asked in addition to the standard HES questions. Some of the information collected for all household members, including children, about their use of the following health services over the preceding twelve-month period, is essentially “objective”:
* Accident and emergency at a hospital;
* Other hospital services such as outpatient clinic, hospital pharmacies, laboratories or day wards;
* An ambulance;
* Nights spent in a hospital as a patient;
* Nights spent in a nursing home or similar;
* Length of time since visiting a GP;
* Number of visits to a GP or family doctor, nurse, medical specialist or consultant, chemist or pharmacist, optician or optometrist or other medical personnel;
* Medical support services such as laboratories, x-ray clinics or health caravans.
The health supplement also included a “subjective question”:
* In general, how would you rate your health?
with response options of “Excellent”, “Good”, “Not so good”, “Poor”.
Additionally, household members were also asked whether they had medical insurance, a Community Service Card or a High Use Health Card.
Constructing Health Status Indexes
The first step in exploring the relationship between health status and socioeconomic status will be to construct an overall measure of objective health status for each individual. This will be by way of a regression procedure with the objective measures of health status as independent variables and the subjective health measure as the dependent variable. Currently principle component analysis is being used to reduce the data set of health usage to a few variables.
The objective measures that best predict subjective health status should be identifiable. It may be that a combination of several simple objective health measures is a good predictor of subjective health status. The predictors are likely to differ by gender and age.
The objective and subjective health status of each household member will be combined into a household health status index, which can be compared with the household economic status using such measures as household equivalent income, housing status, material consumption, and employment status of household members.
Evaluating Health Status and Economic Status
The subjective and aggregate objective index will enable the researchers to investigate such questions as:
* To what extent do poorer households have poorer health, and those with poorer health live in poorer households?
* Does health related spending differ between different income groups when controlled for health status?
* What is the relationship between housing status and health status, when controlled for income?
* Are there any specific issues relating to children=s health and income?
* Do the unemployed have different health status from other groups, when income is controlled for?
* What is the effect of non-household health funding (such as medical insurance or the community service card) on health spending?
* How effective is the community services card? How successfully is it targeted?
* What is the impact of medical insurance on private health service spending and on health service utilisation?
* Are there differences in spending patterns on other commodities (such as food) between households with different health statuses?
Inevitably there are limitations to interpreting the data. Suppose we observe a concentration of those with poor health among those with lowest incomes. That does not tell causality. It may be the unhealthy become poor, or it may be that the poor become unhealthy. Probably it is a bit of both. Even so, the project will add to knowledge of where the unhealthy are located in the income distribution, in the housing tenure spectrum, and in the source of income spectrum. The study cannot resolve all the questions about health status and economic status. Its more modest objective is to use the HES to make some progress by providing some of answers.
While social policy analysts, especially those concerned with health and socioeconomic status, may eagerly await the research outcomes, they should also be cautious using the existing research, given the problems of data transformation identified here, but still unresolved.
1. The authors would like to thank the following for their assistance in the project thus far: Rob Bowie, Paul Brown, Denise Brown, Len Cook, the HRC, Dean Hyslop, Sandra McDonald, Diane Macaskill, Claudio Michelini, Clare Salmond, John Scott, Helen Stott, and Alistair Woodward.
2. Quasi-unit Records, which are averages of three observations stratified by household type, tenure and income are available for the 1995 year. Indications are they behave sufficiently like unit records for many purposes, and some of the results reported below use them. They are available from Brian Easton, and while there are some restrictions upon their use, which will be no complication for a serious researcher.
3. Any results presented in this study are the work of the authors, not Statistics New Zealand. Access to the data used in this study was provided by Statistics New Zealand in a secure environment designed to give effect to the confidentiality provisions of the Statistics Act 1975.
4. For a more extensive account of the model see B.H. Easton (1991) Updating the Economic Model of the Household, Paper to the Conference of the Social Policy Research Centre, University of NSW, Sydney, July 1991, Economic And Social Trust On New Zealand, Wellington.
5. See B.H. Easton (1999) AWhat Has Happened to the New Zealand Income Distribution and Poverty?, Social Policy for the 21st Century, Proceedings of the National Social Policy Conference Sydney, 21-23 July 1999, SPRC University of NSW, . Vol 2, p.55-66.
6. B.H. Easton, (1997) “Measuring Poverty: Some Problems” Social Policy Journal of New Zealand, 9, Nov 1997, p.171-180.
7. e.g. in New Zealand Now: Incomes 1998, SNZ, Wellington, 1999.
8. B.H. Easton (1997) How Accurate are the Incomes Reported in the Household Economic Survey? (Revised), Internal Paper, Economic and Social Trust on New Zealand.
9. B.H. Easton (1997) Household Equivalence Scales and the Household Survey Internal Paper, Economic and Social Trust on New Zealand.
10. C. Michelini (1998) The Estimation of Some Rank 3 Demand Systems from Quasi-unit Record Data of New Zealand Household Consumption, Discussion Paper 98.12, Massey University Department of Applied And International Economics.