 So what I want to talk about today are health measures that we use in household survey data that we use to make assessments about the need for care and that we base health inequalities on. And I want to talk about how our perceptions of health affect the reported health measures. I'm going to specifically speak about the differences in reporting behaviour between the poor and the non-poor, and I'm going to focus on South Africa. And just to give recognition to my co-author, Professor Adifan Duerslaw from the Timbergen Institute in Rotterdam. So I turn to South Africa, especially because I'm from South Africa, but secondly also because of the large income inequalities in South Africa and also unequal access to basic services. And this translates into difference in health outcomes between the poor and the non-poor, where the poor are severely disadvantaged. While a large port of 16% of the population is actually covered by medical schemes, this membership is largely concentrated amongst the affluent and leaving the poor having to rely on public health care, which is often undercapacitated and underresourced. So given the poor quality of public health care, a lot of the poor have to access out-of-pocket payments to access private health care in South Africa. So this is a stat from an article by Berger and her colleagues where they found that a fifth of health care utilization by the persons in the poorest quintile is from private providers. So if we're going to make assumptions about reporting behaviour in household surveys and the data that's usually used in household surveys is self-reported health measures. So they're usually not measured by medical professionals unless you're using anthropometric measures. These self-reported health measures can be anything from asking somebody to give an overall analysis of the health, so asking somebody to rate their health on a scale from one to five, where one is really good and five is really terrible. There can also be a question asking people to rate their ability to do certain daily activities, so the ADL index. And sometimes you also use self-reported chronic conditions such as hypertension or strokes. And especially amongst the poor, these can also go undiagnosed if you don't have access to good quality of health care, so they can also be underreported. So I'll be looking at the overall self-assist health question where there's sort of evaluating a health on a scale from one to five question. And when you're asking somebody to evaluate their health on a scale from one to five, people are going to use different reporting skills when analysing their own health. So saying that your health is excellent might mean one thing for one subgroup and something for another subgroup. And when one subgroup systematically over-estimate or under-estimates their health or systematically uses a different reporting skill, this can bias your health measures. So, and this is often referred to as reporting heterogeneity in certain papers. So one sort of often cited example of reporting heterogeneity is of the Aborigines in Australia where they found that when looking at self-reported health measures, the Aborigines reported having the best health in Australia, but when you looked at something more objective such as mortality rates, they performed the worst. And when you're looking at reporting behaviours, there's sort of this predictive pattern that you usually find that vulnerable subgroups are likely to have a different reporting skill, but also that they're likely to underreport their ill health. And this has been found for across provinces, income groups and different education levels. So why would we expect the vulnerable subgroups to under-estimate their ill health? The first reason, I mean there are a range of reasons, but one possible reason are differences in comparison groups. So when you take for instance the poor as a subgroup, they're comparing themselves to their peers and their environment. They're comparing themselves to sort of their neighbours and their family. And when you're comparing yourself to a group who starts off with a relatively low level of health, you might see yourself as better off than when you compare yourself to sort of somebody from a more affluent subgroup. So that's one possible reason. I mean another reason which is quite an interesting theory is that people sort of suppress or ignore their illnesses due to their inability to cope with the economic costs involved with being ill. So it's a coping strategy. When I'm speaking about economic costs, I'm not only talking about the healthcare costs, but also about the opportunity costs of not being able to work while you're ill. So Joachim Saabon and his colleagues differentiate between two types of coping strategies. The one is managing strategies. So once the costs are incurred, what do you do to cope with your health costs? So this would be something like taking an extra loan to deal with the health costs. The second one and the one that I'm sort of posing in this paper is the preventative strategy where you modify your illness perception or ignore your disease due to your inability to cope with the illness. So there's sort of this cognitive shift in what you think your health state is. And this is also a table from another table from Burger and her colleagues that show the reported health across income groups across years. And what you can see is that from the poorest quintiles, people report much lower illnesses than people from the higher quintiles. But also that once they do admit that they're all ill, they're also less likely to seek care. So if the poor and the non-poor use different reporting skills and the poor as a vulnerable subgroup underestimate their ill health, then there's lots of implications for estimating health disparities because you're underestimating the gap in the health outcomes between the poor and the non-poor. So yeah, I'll be focusing on reporting behavior according to wealth status where wealth is measured using the asset index and you're classified as poor when you fall within the bottom two wealth quintiles and non-poor if you're in the top three wealth quintiles. So establishing the possibility of health reporting, different health reporting behavior affecting health disparities. I take a two-step procedure. Firstly, I'll estimate whether or not the poor and the non-poor use different reporting skills. So do they have different reporting behavior? And often you establish this in what direction is this bias? So if the poor do use a different reporting skill, are they over-reporting or under-reporting their ill health? So do they think that they're better or worse than they actually are? The data I'll be using is the WHO Study on Global Aging and Adult Health, or SAGE, which was created in 2008. It consists about 3,200 observations and unfortunately it only focuses on the adult population. So it only looks at persons aged 50 years and older. The data, the self-reported data that I'll be using in the dataset is the self-assessed rating of your health in specific health domains. So here I've put an example of one of these questions. Participant is asked, over all in the last three days, how much difficulty did you have with moving around? Where one is none and five is extreme. So a higher value is a worse health state. And you have this question for the various domains. So mobility, appearance, anxiety, pain, vision, sleeping and energy, stuff like that. And then what I'm also using, which also in this dataset, and that's why I'm using this SAGE, is our anchoring venets. And our anchoring venet is this hypothetical person or this hypothetical scenario that a person, which the respondent is asked to evaluate the health of. So just to give an example, I've written this down, Alan is able to walk a distance of up to 200 meters without any trouble, but gets tired after taking a flight of stairs or walking for one kilometer. And then the respondent is asked to rate Alan's mobility on a scale from one to five. And what makes the anchoring venets great are that they represent a fixed health state. So the health state doesn't vary across individuals, rather the way that the individual evaluates the health of the venet varies. So any sort of variation in the evaluation of the venet's health can be appropriated to reporting heterogeneity. So within each domain, there are five different venets ranging in different fixed health states, from good to average to poor health state. And then unfortunately, because it's such a sort of a data-hungry exercise, it's only asked for a subset of individuals. So you have these venet ratings for about 500 individuals, while you have the overall rating for all the respondents. So this is just a summary of the covariates that I'm including in the analysis, and it differentiates between the poor and the non-poor. The sample is about 55% female and 62 years of age. Persons in the non-poor group are more likely to be married and have higher levels of more years of education than the poor. The poor group is almost 80% African black, while the Asian Indian and the white population group's form is completely in the non-poor. Then if you look at the sort of hot apple cider colour picture at the bottom, this is just an illustration of what a venet rating would look like. So this is a venet rating in the dim health domain of mobility moving around, and this is a health venet with a sort of an average health state. So the column on the left-hand side is for the non-poor, and on the right-hand side is for the poor, and the non-poor are more likely to say that the venet has an extremely poor health state, and are less likely to say that it has no or mild health difficulty. So if you look at this, just as an example, but this is just one of many, then it looks like the non-poor or more pessimistic in the analysis of the venet's health. So for estimation, I'm using the hierarchical ordered probate model, which is a sort of two-part estimation model where in the first part of the model called the reporting behavior equation, what you do is you use the sample that have the venet ratings, and your dependent variable is the venet rating on a scale from 1 to 5, where 5 is the worst health set, and you regress this onto your various individual characteristics where one of the individual characteristics is your wealth variable, which is equal to 1 if you're poor and 0 if you're not poor, and also it also includes a constant which is equal to 0 for your first venet. And so this is basically like a generalized ordered probate because you allow the cut points of the ordinal variable to shift with the individual characteristics. What you do is you use the cut points that you estimate from this equation and you apply it to the next part of the model, which is the health equation. And then the second part of the model, you use the entire sample and you use a person's evaluation of their own health in a specific health domain, and you take the cut points that you estimated from the previous regression, you fix it onto this regression and you regress your overall self-assessed health onto once again the individual characteristic. And what you basically end up with is a type of almost like an interval regression, which is an ordered probate, but with known cut points. And you estimate these jointly. So what you end up with at the end from the second equation is basically a self-assessed health variable that has been purged of differences in reporting behavior. So since the cut points are dependent on wealth status and the other individual characteristics you remove reporting heterogeneity from your variable. To test for reporting heterogeneity, you use the reporting behavior equation. So you use this equation and then you test for the joint significance of the poor variable across the cut points of this equation. And what I report here are the results from this test for reporting heterogeneity. I now see that the font looks okay when you're working on it at home, but when it's big on screen, you see that you make font errors. Okay, so these are the p-values from this test of joint significance. And in all the health domains that I tested it for, reporting homogeneity is rejected. So there's evidence of the poor using different reporting scales than the non-poor. So there is reporting heterogeneity, but it's not clear whether the poor are underestimating or over-estimating their health. And that's the next step. So what you can do is you take the results from the second part of the model from the hope of estimation and you compare it to... you compare it to the purge figures and you compare it to the ordered probate results. And by doing this you can see how the purge figures compare to the non-purge figures. So in the ordered probate, the poor... So just to say that... just to remember that the health variable ranging from one to five or five indicates a worse health state and a one indicates a better health state. That means a high probability of reporting poor health. And in the ordered probate, the results were... the coefficients were mostly negative. So the poor were more likely to report a better health state. And of the control for reporting heterogeneity, some of the... or most of the coefficients become positive. And you... so the poor are more likely to report a poor health state. So the direction of the change is the same for all variables, but unfortunately the variables... the coefficients aren't statistically significant. But I think you can draw some sort of conclusion of the direction of the bias, but not necessarily of the magnitude. Yeah, so just to conclude... a bit early, but... So like I said, there's indication that using self-reported health measures and drawing conclusions about the need for care and health inequalities could be biased when you're doing this across income group. But it's not... and it's likely that the poor are underestimating their health state or underestimating their ill health. But apart from that, apart from sort of the measurement implications, it's also important to note that if the poor are underestimating their ill health, there is a sort of suppression of health needs that are going unrealized. And what has happened in a lot of countries, the countries that have implemented universal health coverage has been a large increase in health care utilization especially amongst poor. And a large part of this can be attributed to... obviously tomorrow has it, but a lot of it can also be attributed to unmet health needs. And if this is true for South Africa, then we need to think of ways to remove the access barriers, supply access barriers to improve the access to good quality public health care. So improving possibly home-based care or maybe better quality mobile health plans, or even the national health insurance as a start which aims at the first step to improve basic care. Primary basic care could be a move in the right direction.