Our data come from a series of cross-sectional household surveys known as the National Household Sample Surveys (Pesquisa Nacional por Amostra de Domicilios
or PNAD in Portuguese) carried out by the Brazilian Institute for Geography and Statistics (IBGE). We use the three health-related supplements to the PNAD conducted in collaboration with the Ministry of Health in 1998, 2003, and 2008. The PNAD uses a three-stage complex probabilistic sample, and is representative of the national, regional, and state levels [11
]. A total of 1.12 million individuals are included in the three surveys, which obtained data by means of face-to-face interviews and rely on self-report.
We employ measures of horizontal equity developed by Van Doorslaer, Wagstaff, and others [12
]. These measures seeks to assess equity (fairness) in healthcare utilization by taking into account the fact that individuals have different health needs and that differences in health needs ought to translate into different demand for and use of health services. But once health needs are standardized across individuals, remaining utilization could be considered to be inequitable. The horizontal inequity index (HI) is used to operationalize this concept. It is defined as the difference between observed healthcare utilization and that which would be expected given the individual’s health needs [14
We base our set of control variables representing healthcare needs and other non-need factors on guidelines developed by the World Bank [14
]. Each individual’s need for healthcare is approximated by the following variables: sex/age categories (12 variables representing men and women aged 0–17, 18–34, 35–44, 45–64, 65–74, and 75
years and over, with males aged 0–17 as the reference category), self-rated health (measured as excellent/good/fair versus poor/very poor); any physical functioning limitation (any difficulty in toileting/feeding/bathing oneself, kneeling/stooping, walking up stairs, or walking 100 meters); previous medical diagnosis of any of the following conditions: arthritis, cancer, diabetes, bronchitis/asthma, hypertension, heart disease, kidney failure, depression, tuberculosis, cirrhosis, and/or tendinitis; and a measure of co-morbidity (two or more of the conditions listed previously). Additional determinants of healthcare utilization (also known as non-need factors) include literacy (whether the person can read and write), schooling (less than 3 completed years, 4–7
years, 8–10,11-14, and 15
years or more), log monthly family income, urban/rural location, geographic region (North, Northeast, South, Southeast, Central-West), affiliation with a private health plan, and coverage by the Family Health Strategy (available only in 2008). For all control variables (except income) dummy variables were created for all categories, using the lowest category as the reference group. Note that we do not adjust monthly family income for inflation, since in the statistical analyses, income is used to rank each individual at each year along the income distribution and is treated as a relative measure of social position in each time period.
Outcome variables are measures of access and utilization of healthcare services that are comparable across the three surveys. These include: any doctor visit in the past 12
months, any dental care visit in the past 12
months, any hospitalization in the past 12
months, and any health services sought in the previous 2
weeks. An additional variable captures whether the individual is able to identify a usual source of medical care (“Do you tend to seek healthcare services from the same place?”) and is used a proxy measure of continuity of care.
Analysis of equity requires a series of steps. First, it is necessary to regress medical care utilization, yi on a set of explanatory variables:
is use of the particular health care service by individual i,
) is the log of family income for individual i
is a vector of need determining variables, Zp
is a vector of non-need determining variables, α, β, γk
are parameters and ϵi
is the error term. This equation can be used to generate the predicted demand for any particular health care service,,
that is, the expected healthcare use of individual i
on the basis of his/her health needs. This quantity can be thought of as the amount of the health care the individual should consume, if s/he had been treated the same as others with the same healthcare needs.
After predicting demand we calculate indirectly standardized demand (
) by estimating the predicted y values by standardizing for the X (health need) variables while simultaneously controlling for the Z (non-need) variables.
We then calculate
is the indirectly standardized (predicted) demand, yi
is actual demand,
is the x-expected demand and
is the sample mean of actual demand (See equation
After completing the above standardization and graphing the concentration curves, we calculated the concentration index for both yi
, using the convenient regression method as outlined in O’Donnell et al [14
]. Once the concentration indices for actual (Cm
) and predicted demand (Cp
) are calculated, the Horizontal Inequity Index (HI) is calculated as follows:
where Lp(p) is the concentration curve of predicted demand and Lm(p) the concentration curve of actual demand. The HI ranges from −2 to 2 and is positive if there are inequities favoring the more advantaged members (richer) of society, which in these models is measured by family income.
Finally, we apply methods to decompose the concentration index in order to ascertain the contribution of each covariate (need and non-need factors as described above) to overall inequity in healthcare utilization. Because all outcomes are binary, we use techniques developed by Van Doorslaer et al [15
]. Decomposition is performed using a linear approximation of the model based on partial effects of each covariate evaluated at the sample means. This approach allows us to identify which factors are associated with pro-rich or pro-poor utilization and to approximate their contribution to the overall concentration index.
Analyses were carried out using Stata Version 12 [18
]. When appropriate, results are adjusted for the effect of the sample design and include individual probability weights.