2.1 Data

The Study on Global Ageing and Adult Health (SAGE) project compiles studies on the population aged 60 and older. Countries included are China, Ghana, India, Mexico, Russian Federation and South Africa. We are using SAGE wave 0, which for China includes information for years 2007/08 and 2009/10; Ghana 2007/08; India 2007/08; México 2009/10; Russia 2007/08 and 2009/10; and South Africa 2007/08 [22]. Unfortunately, at the time of writing this study, we were not able to get access to the second wave studies collected at 2014/15. These surveys measure the determinants of active ageing through SES data, the physical and social environment, behaviour, cognition and affect, functionality, mental well-being, medical and health conditions, as well as use and access to health services.

For Colombia, the Healthcare, Welfare and Ageing Survey (SABE, for its acronym in Spanish) was collected in 2015, and its design is comparable to the other ageing population surveys [23]. Appendix A in S1 Appendix presents the main characteristics of the health systems of these countries.

One of the main characteristics of those surveys among older individuals is that they involve clinical measures such as blood pressure. Following a standardised procedure, nurses measured three times systolic and diastolic BP. After discarding the first BP measurement, we take as our objective measure of HBP whether the average of the measurements is above 140 mmHg for systolic BP (SysBP), or above 90 mmHg for diastolic BP (DiasBP). This is a standard procedure as the first BP measurement is usually higher than average because respondents tend to be nervous, a phenomenon known as the white-coat effect. In SABE, SysBP was measured in both arms (six measurements), while in SAGE it only was in one arm (three measurements). In Mexico, there were only two measurements, so we take the second one. Figure A1 in the S1 Appendix presents the densities of the final SysBP measure for each country. In this set of six countries, we observe an ample range of BP patterns: from the low levels of India to the high ones of South Africa.

We include in our analysis all observations of individuals aged 60 and older for whom there is available information in the following characteristics: SAGE studies include information for individuals 50 to 59 as well. We exclude them as SABE does not cover this age group. Nevertheless, online Appendix D in S1 Appendix presents the exercise using this sample, which are qualitatively the same as those presented in the main text. valid measurements of blood pressure, self-reported diagnosis of high blood pressure, gender, age, education, weight and height, smoking behaviour, assets, and health insurance (in all countries but Mexico). As a common limitation for using biomarkers data, the resulting sample is likely to be more educated, wealthier and care more for their health than the regular population [24, 25]. See Appendix B in S1 Appendix for further details, and the limitations under the discussion section for the implications of the selection pattern. Table 1 presents the means of the variables included, and sample sizes. In the rest of the document we explain the table and how variables are used. Individual sample weights are used in all exercises. These weights account for sample selection.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Means by country.

https://doi.org/10.1371/journal.pone.0269118.t001

2.2 High blood pressure classification

Following Falaschetti et al. [26], we classify respondents in four groups, based on whether they are aware of having HBP and objective BP measures. Someone is considered aware if he or she reports having been diagnosed with HBP by a doctor or nurse, excluding women diagnosed during pregnancy. The groups are defined as follows, and a diagram of these definitions presented in Fig 1. (i) No HBP: those respondents who declare not having being diagnosed with HBP and do not present objectively measured BP levels that indicate otherwise (SysBP < 140 and DiasBP < 90). (ii) Undetected HBP: those respondents who report not having HBP, but according to their BP measurements (SysBP ≥ 140 or DiasBP ≥ 90), they likely have the condition. (iii) Uncontrolled HBP: The person is aware of being diagnosed with HBP, but the BP measurements suggest that current treatment (if any) is ineffective. And (iv) Controlled HBP: those respondents who are aware of HBP, and whose BP levels are under control.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Definition of high blood pressure classification groups.

Notes: Diagnosis of HBP by a physician correspond to the self-response of individuals. SysBP and DiasBP correspond to the blood pressure measurement by a nurse at the time of the survey.

https://doi.org/10.1371/journal.pone.0269118.g001

In this study, the denominator of our analysis are individuals who either report having ever been with HBP, or those whose BP is above the standard diagnosis threshold (groups ii to iv). Henceforth, we refer to them as the at-risk population (AR). While group (i) is essential for public health interventions to prevent them from developing HBP, in this work, we concentrate on those who already require attention by the health system. In this sub-population, it is possible to discuss the false negatives of the system (group ii). False-positives, those diagnosed with HBP but who did not have the condition, cannot be detected with this survey-based approach. This group is important as it indicates the inefficient use of resources of the system. The other leading group of interest is (iii), as it involves those for whom the health system has not provided effective treatment. Ideally, these two groups should be zero. Another element of interest is the proportion of those who are aware (groups iii and iv) and are receiving treatment. We concentrate on an outcomes perspective, rather than on an input perspective. Therefore we make emphasis in the controlled/uncontrolled partition. Thus, our main dependent variables will be:

• to have undetected HBP for those individuals in the population AR;
• to have uncontrolled HBP conditional on being aware of the HBP condition.

2.3 Socio-economic status / wealth index

In order to understand the role of SES on our HBP variables, we constructed a wealth index based on assets’ ownership. This is a composite measure of a household’s cumulative living standard. This index measures wealth based on household’s ownership of selected assets and is built using factor analysis. Details of the weights given to each variable are presented in the online Appendix B in S1 Appendix. The index was constructed with the following variables: the household appliances that counted the home, type of dwelling, housing rights (ex: own house, rent), physical characteristics of the dwelling (roofing, walls and floors material). In the Colombian case, we also included a socio-economic level indicator known as estrato (according to a utility subsidies and taxes classification). Fig 2 presents estimated densities of the index for each country. The value of the index is relative for each country: our objective is to understand the concentration of wealth, rather than it’s level. As shown in the graph, India and Ghana are in one extreme, while Russia is in the other. All the other countries are relatively similar in that respect.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Densities of the wealth index by country.

Notes: Epanechnikov Kernel densities using a bandwidth of 0.03 (asset index ranges from 0 to 1). Graph constructed over the sample of SABE and SAGE of respondents with a valid BP measure and a self-report of HBP. The asset index was derived using factor analysis for all individuals who answered the assets section. See online Appendix B.2 in S1 Appendix for further details.

https://doi.org/10.1371/journal.pone.0269118.g002

2.4 Discrete choice models

In order to approximate the relationship between and individual being classified as with undetected HBP with household wealth, and between belonging to the group uncontrolled HBP with household wealth, we estimate logistic regression models given that in both cases the dependent variable is binary. The probability of observing that for individual i outcome Y_i take the value of 1 instead of 0, conditional on a set of covariates X_i, country R_i, and wealth index wealth_i is given by the following equation: (1) where Λ represents the logistic function. For the case of uncontrolled HBP, the sample is restricted to those diagnosed with the condition. In this regression, on top of the countries fixed effects and a dummy for respondents being surveyed in 2009/10, we include the interaction between the country dummies and the independent variables. As a result, the interaction of the wealth index and the country dummy provides a measure of the gradient (γ_r) for each country. The comparison group is being surveyed on 2007/08. The Colombian survey was the only one in 2015, so its information is already part of the specific survey fixed effect. These models were estimated using Stata 16.

From the results of these models, we compute the average marginal effects (AME) of wealth on the respective probabilities: . This expectation can be computed conditional on the country. The resulting figures can be interpreted as increases on percentage points (pp) over the probability of Y = 1 as opposite to Y = 0. The odds ratio, an alternative for interpreting the coefficients of the model, are presented in the Table C1 in S1 Appendix.

A generalisation of the model is the multinomial logistic model. This model fits the context as our at-risk group is composed of three categories. The model considers the probability of occurrence of two of the three events relative to a basis. We present in the Table C1 in S1 Appendix the relative risk ratios (rrr), similar to the odds ratio in the binary case.

2.5 Blinder Oaxaca decomposition

The Blinder Oaxaca decomposition technique is a common tool for explaining inequality of health outcomes [27]. We use it to study mean outcome differences in uncontrolled and undetected High blood pressure (HBP) by poverty status. We considered as poor to 20% of people with the lower income. Sensitivity to alternative cutoffs is presented in the online Appendix G in S1 Appendix. This method divides the uncontrolled or undetected HBP differential between two groups (poor and non-poor) into a part that could be explained by group observable characteristic differences and a residual part that cannot be accounted for by such differences in HBP determinants. The unexplained part subsumes the effects of group differences in unobserved predictors.

We used a twofold decomposition using the coefficients from a pooled linear probability model over both groups as the reference coefficients. Then, the outcome difference is divided into two components: (2) where the first component, (3) is the part of the outcome differential that is explained by group differences in the predictors. As predictors, we used behavioural risks which include the variables of age and being male, demographic risks, which have smoking history and obesity status, and a variable that indicates if the individual lives in an urban area. The second component in this decomposition, (4) is the unexplained part, which in addition to capture all the potential effects of differences in unobserved variables can be a proxy of inequality in HBP by poverty status.

Fairlie [28] consider an extension of the original decomposition but using non-linear probability models such as the logit and probit. Under this framework, the contribution of each variable to the gap is equivalent to the change in the average predicted probability if the distribution of the poor is replaced with the distribution of the non-poor. According to the authors, the linear probability model may produce notoriously different covariates that have strongly non-linear relationships with the outcome. Therefore, we compute the decomposition under this extension in the Appendix C in S1 Appendix. These models were estimated using the command Oaxaca in Stata 16 [29].

3.1 High blood pressure groups and wealth

Panel A of Table 1 presents the means of the relevant variables conditional on different samples (for which there is no missing data in all covariates). Columns marked with BPS (Blood Pressure Sample), corresponds to those individuals for whom there is information on their blood pressure. Given that, we can determine the proportion of total respondents who are AR, which is presented in the first row. In Russia and South Africa, at the time of the SAGE interviews between 2007–10, around 81% of respondents aged 60 or older were AR, followed by Mexico (77%) and Colombia (69%). Below was China (67%), and far from them Ghana (57%) and India (39%). The second row presents those who were aware of having HBP. If the denominator is the entire population (columns BPS), we have the detected prevalence, which was as high as 61% in Russia and as low as 15% in Ghana around 2007–10.

The other set of columns of Table 1, present the proportions for the population AR. For instance, China had a detected prevalence of 33% on the overall 60 and older population. These individuals were nearly half (49%) of those at risk, the AR population. Thus, the undetected BP rate was 51%. Relative to the AR population, 12% had their BP levels controlled (SysBP/DiasBP BP below 140/90 mmHg). Thus, the proportion of individuals aware, but with their BP being uncontrolled, corresponded to 39% (= 51%—12%) of the AR population at the time (around 2007–10). This is the uncontrolled HBP rate. Panel B of Table 1 presents the total number of observations for each of these groups.

Concerning undetected HBP there were four groups. First, in Colombia and Russia, undetected HBP rates were 18% and 25%, respectively. Second, for Mexico the corresponding rate was 45%. Third, for China, India, and South Africa it was around 53% at the time. And fourth, in Ghana, this figure was just above 74%. Once we consider control of the condition, the best performing country was India: 20% of Indians AR had their SysBP below 140 mmHg and their DiasBP below 90 mmHg at the time. The worst performing country was Russia, where 61% of the AR were known cases of HBP which are not under control at the time.

Table 1 presents as well socio-economic characteristics of the respondents, corresponding to the years 2007–10 for all countries except Colombia (for which data correspond to 2015). Age and gender produce ‘mechanical’ differences between countries. However, has fewer men in the sample (33%), yet BP levels were one of the highest at the time, in comparison with the other countries included in the sample; the South African population was on average as old as the Indian (68 years), but there were more than ten mmHg of difference between their average BP levels. Differences were more likely to be related to habits and lifestyle: obesity in South Africa was nearly 50%, while in India it was 4% at the time. Yet, China obesity rate was as well low, 7%, but their average SysBP was the same in South Africa. Mexico and Colombia had the highest rate of people living in an urban area, close to 80%. However, BP in Mexico was higher than in Colombia. Although Mexicohad the highest wealth index also had one of the highest BP in our sample of countries. Another consideration is that there were no substantial systematic differences between the BPS and AR population because AR represented a large proportion of all individuals older than 60 in these countries.

Another set of variables presented in Table 1 is related to the health system. As described in the context section, there are noteworthy differences. Russia had no VHI at all at the time, while in Ghana, 41% of the individuals had it. Almost the rest of the population is uninsured in Ghana, only worse to India, where nine out ten did not have such protection at the time. The opposite is again Russia, where everyone was insured, similar to Colombia, where people without health insurance represented only just 2% in 2015. A very similar pattern occurs for living in rural areas, where more barriers have been found to expand universal coverage in the world. Unfortunately, this information is not available for Mexico.

Socio-economic characteristics, lifestyle, and health system differences shape the relationship between wealth and health in each country. The unconditional relationship between wealth and our HBP quality of attention indicators are presented in Fig 3. Each point of the line presents the average rate of undetected and uncontrolled HBP for each level of wealth around the years 2007–2010 (and 2015 for Colombia). First, it shows a negative slope between wealth and probability of non-detection of HBP, which depends on the country. Second, concerning uncontrolled HBP, there are both positive and negative slopes, showing that the role of wealth is context-dependent. With the AME derived from the parametric model presented above, we can determine whether the observed slopes hold after conditioning of different sets of controls.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Undetected and uncontrolled HBP and wealth.

Notes: Kernel-weighted local polynomials using the samples of SABE and SAGE. For each country/region, the domain corresponds to the range covering 5%-95% of the total density of the wealth variable.

https://doi.org/10.1371/journal.pone.0269118.g003

3.2 Results from discrete choice models

Table 2 presents the AME after logistic regressions. It considers three models for each outcome which differ on the set of controls. Columns (1) and (5) only consider country fixed effects, age, and a gender dummy. Columns (2) and (6) expand them to lifestyle: whether the individual is obese and if she has ever smoked. These are the main risk factors, and there are not so many differences between both sets of columns. Columns (3) and (7) include education and whether the respondent resides in an urban area. Finally, columns (4) and (8) consider health insurance schemes. These are characteristics that might be correlated with wealth. Thus the resulting coefficients correspond to the remaining disparities not associated with these variables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Average marginal effects of wealth on undetected and uncontrolled HBP by country.

https://doi.org/10.1371/journal.pone.0269118.t002

If we consider only variables not related to wealth, we find that concerning undetected HBP, there has a wealth gradient in all countries (column 2). Ghana and Russia are the most extreme cases: if wealth has ten levels, moving up in one level is associated with a reduction of 7.27 pp the probability that someone above the age of 60 had undetected hypertension in Ghana, and of 7.64 pp in Russia, conditional on risk factors such as obesity and smoking antecedents. If we consider education level on top of these controls (column 3), the gradient drops to 6.64 pp in Ghana and 7.45pp in Russia. South Africa is on the opposite scenario, where the gradient was around 2.05 pp without education and 1.66 pp (not significant at 90% level) when these controls are added. Health system variables reduce the gradient slightly in most countries, especially in China. The considerable reduction on the gradient for Russia has to be taken carefully, as less than 1% of the individuals at risk has VHI, and none had no insurance. Therefore the change is such point estimate obeys more to the high imprecision of the estimate (a 95% confidence interval between -6.28 pp and -83.11 pp).

Regarding uncontrolled HBP, Table 2 presents a similar analysis but conditional on the diagnosis. Here we are exploring the probability of having HBP in comparison with a mix between those people with controlled HBP and those who are not aware of the condition. In the Appendix F in S1 Appendix, we present results without conditioning on the diagnosis. The table shows a significant negative gradient for all countries but for Mexico and South Africa. For these countries, the gradient (around the years 2007–10) was negative as well, but the precision of the estimates cannot rule out the non-presence of an association. The largest figure is for Russia (-0.41 in column 6) and Mexico (-0.42, but not significant at 95% level), and the smallest for South Africa (-0.11). Notice that the dispersion between countries is smaller than for detection. In most cases, the addition of health insurance variables reduces the magnitude of the gradient, especially for China where the gradient disappears.

Estimated odd ratios of the models (3) and (6) are presented in Table C1 in the S1 Appendix, using the same controls as columns 3 and 7 of Table 2, as well as regressions for SysBP and an alternative multinomial logistic model. Regardless of the model specification, and even with the continuous SysBP variable, similar patterns to the ones presented above are obtained. This is also the case when alternative definitions of the age range (Appendix D in S1 Appendix) and the assets index (Appendix E in S1 Appendix); and when individual regressions per country and set of controls are considered (Appendix H in S1 Appendix).

3.3 Decomposition results

We further explore the gradient by comparing differences between those at the bottom of the wealth distribution (the lowest quintile) with the rest of the sample. The black diamonds in Fig 4 present the difference in the proportion of undetected HBP (on the left) and uncontrolled HBP (on the right). Concerning the previous section, the pattern for the overall gap differs in the case of Russia. For this country, there exists a small gap in this exercise while a large wealth gradient. This indicates that the gradient at the time was probably linked to a specific upper interval of the domain of the wealth index (where most of the distribution is located as shown in Fig 2), and not a feature that is constant across the wealth distribution. For Ghana, the large gap goes in line with the steep gradient.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Contribution to gap in undetected and uncontrolled HBP conditional on being aware o hypertension.

Notes: Contribution of covariate groups to the gap on undetected and uncotrolled HBP between individuals in the lowest wealth quintile agains the other quintiles. These numbers are computed with the Blinder-Oaxaca decomposition after linear probability models.

https://doi.org/10.1371/journal.pone.0269118.g004

For this exercise, we focus on decomposition analysis. For example, for China the gap between poor and rich on undetected HBP was of -23.3 pp. This number is obtained from taking the undetected HBP rate of 44.9% for the 80% of the sample with higher wealth and subtracting the 68.1% rate for the lower quintile (see Table C2 in the S1 Appendix). For this alternative view of the gradient, we can decompose the contribution of each of the observed factors -behaviours, demographics, education and urban/rural division- using the Blinder-Oaxaca method. For the case of China, 56.2%(= -0.131/-0.233) of the gap at the time can be explained with the difference in the level of the observed variables. This is roughly the same proportion for most countries (from 45% to 58.4%), apart from Mexico and South Africa. In Mexico, it was 34.4% of a gap of 7.32 pp which is not significant at the 95% level, and in South Africa, it is 86.32% of the gap of 8.41 pp can be explained with the observed variables.

Almost all the explained differences are due to the urban/rural division in the first place and the education level in the second. Again, for the case of China, differences in the proportion of people living in urban areas represented 80.15%(= -0.105/-0.131) of the explained difference at the time; this figure is also high for South Africa (83.5%), and very small in Mexico (3.4%). Differences in education level represent another 11.8% in China, and just an 8% for the rest of the variables (socio-demographics, smoking prevalence and obesity). In Mexico, 91.3% of all the explained gap can be associated with education. In general, these two factors accounted for most of the observed gap (more than 80% of it). However, in Ghana, age and gender differences represented nearly 11.3% of the explained difference. In Figure C1 in the S1 Appendix, we perform the decomposition separately by urban and rural areas. It shows that all the other observed characteristics played a minor role in explaining wealth gaps within these areas 2007–10. For Ghana, the number of observations with a positive diagnosis of HBP in the rural setting is too small for the analysis.

Tables C5 in the S1 Appendix includes the health insurance variables into the decomposition. The proportion of the gap explained is similar to the results above. These variables explain 14% of the difference in Ghana, 11.6% in South Africa, 10.6% in India, 9.4% in Colombia, and 2.5% in Russia at the time. In contrast, for China (-3.5%) the level of observable variables suggest that the gap should run in the opposite direction.

The absolute difference in uncontrolled HBP, conditional on the diagnosis, between poor and non-poor people is statistically significant for China, Colombia and South Africa (see Table C3 in the S1 Appendix). For the other countries, the total difference cannot be rejected to be equal to zero. For Ghana, India, Mexico, and Russia, there is no evidence of a gap though there was evidence of a gradient. Once again, this might be related to the point of comparison of the wealth distribution. Ghana presents interesting results, as the point estimate is small, but the explained component suggests a gap similar to the Colombian or the Indian. For this country, any analysis should be careful as the number of individuals diagnosed with HBP is small, hence results are very imprecise.

For the three countries for which there is evidence of a gap, the explained component is roughly between 50% and 60%. For China and Colombia, more than 90% was explained by rurality, while this factor is only 36% in South Africa. In the last country, education accounts for 77% of the explained gap, while it is 6% in China and Colombia. When health insurance variables are included (Table C5 in the S1 Appendix), there are almost no changes at all, and these variables explain around 4% to 6% of the explained gap.

As discussed in the methods section, the BO decomposition with linear probability models might be biased for covariates with a strong non-linear relationship with the outcome. In the S1 Appendix, Tables C6 and C7 show the results of the Fairlie extension based on a logit model [28]. Results are almost the same as those presented above. Moreover, as for the results of the marginal effects, we consider alternative definitions of the age range (Appendix D in S1 Appendix), the assets index (Appendix E in S1 Appendix), and the uncontrolled HBP results conditional on being aware of the conditions (Appendix E in S1 Appendix). We also consider the sensitivity of the results to alternative definitions of the non-poor sample (Appendix G in S1 Appendix). Qualitative results are stable across these alternatives.