Related literature

Evidence from high-income countries suggest that the magnitude of attenuation bias from using self-reported health measures is large. Baker et al. documented errors in self-reported status for 13 disease categories, by linking data from a Canadian household survey to health administrative data in the province of Ontario [7]. In all categories, they recorded a false negative rate of over 50 percent. In comparison, rates of false positives were smaller, but relatively larger for conditions that individuals tend to self-diagnose, such as migraine.

Johnston et al. used the health survey for England to examine the gap between self-reported and clinically tested hypertension [6]. They found 28% under-reporting on average, and estimated an attenuation bias of 68 percent. Their estimates indicated that while household income had no effect on self-reports, a 1 log-point increase in household income lowered clinical hypertension by about 1%. They also found that the probability of false negative reporting declines with income.

Suziedelyte and Johar examined the consistency of self-reports on four invasive surgical procedures over the previous 5 years [13]. The procedures were knee replacement, removal of gall bladder, removal of prostrate and hysterectomy. They linked self-reports from survey data in the Australian state of New South Wales with hospital records on all admissions in the state. They also compared self-reported hypertension with claims data for prescription drugs for treating hypertension, and hypertension diagnoses in hospital admission records in the 12 months after the survey date. Misreporting was common in their data, with higher rates of over-reporting (false positives) for all procedures except prostrate surgery, ranging from 26% to 29%. For hypertension, over-reporting accounted for about 32% while the rate of under-reporting was about 28%. Respondent characteristics could explain little of the variation in reporting error.

Evidence on the degree of misreporting of health conditions in developing countries is meagre due to data limitations. An exception is [14] who compared self-reported rates of five common non-communicable diseases (NCDs) in India from the Study on Global Ageing and Adult Health, with age-adjusted prevalence rates of these diseases. They estimated disparities based on household wealth and education categories. They found evidence of significant under-reporting among low socioeconomic groups. They infer that this is due to poor access to high quality healthcare in India.

Cramm et al. used the same data source as this paper to examine the association between self-reported health and objective health based on tests conducted by the survey staff [15]. However, they did not compare subjective and objective indicators for the same disease. Their subjective health measures were self-rated health and dependence in activities of daily living, while their objective measures were test-diagnosed results for lung disease and grip strength. They found a weak correlation of at most 14%, between subjective and objective health indicators. They also found that respondents tend to overstate their health status, relative to what the objective measures indicate.

In contrast to [15], we compare subjective and diagnostic reports of two disease conditions—hypertension and lung disease—for the same individuals. This provides a validation exercise for the self-reported data. Our paper also differs from [14] in that we present estimates of bias that arise from using self-reports in place of objective measures. We examine characteristics associated with false negative reporting for both diseases.

Hypertension and lung disease in India

The two primary causes of death in India are coronary heart disease (CHD) and lung disease. A key risk factor for CHD is hypertension. In India, hypertension leads the list of NCD risks, estimated as a factor in about 10 percent of all deaths [16]. Moreover, prevalence rates among adults have risen dramatically over the past three decades and are expected to rise further [16–18].

Hypertension can often be asymptomatic, even at elevated levels of blood pressure [6]. Thus, though it is inexpensive and easy to detect, it is likely to remain undiagnosed if the individual is not accustomed to visiting a doctor periodically. This leads to preventable morbidity and deaths.

Lung disease is the second biggest cause of deaths in India. Lung disease refers to many conditions affecting the lungs, including asthma, chronic obstructive pulmonary disease (COPD), infections like pneumonia, influenza and tuberculosis, and lung cancer. The age-adjusted death rate due to lung disease, at 127 per 100,000 of population, places India at #1 in the world by this criterion [19]. While many conditions comprising lung disease such as respiratory illness come under NCDs, tuberculosis (TB) is an infectious disease. India also ranks first among the list of 22 countries constituting the highest TB burden in the world, and accounts for one-fifth of the global incidence of TB [19].

Lung disease is often associated with some typical symptoms such as cough, breathlessness and fever. However, these symptoms can be confused with those of milder conditions such as influenza, and ignored. A visit to a specialist is required to diagnose lung disease. Problems of access to specialist medical care and financial considerations prevent the prompt diagnosis of this very serious disease.

Measurement error in self-reported health and associated bias

We attempt to quantify the ‘measurement error’ in the self-reported hypertension and lung disease status. We assume that the disease status as reported in the biomarker module is the ‘truth’ and we measure the error as the difference between the self-reported status (S) and the true status (T). The analysis is based on [24] and [7].

Suppose we want to estimate the following model: (1)

However, one of the variables in the X vector, T, is not observed. Instead, we only observe S where S_i = T_i + u. That is, T is measured with error, but we assume that cov(ϵ, u) = 0. In the case of classical measurement error (CME), u is also uncorrelated with T. Since both S and T are dichotomous variables, the measurement error is not CME in our case. If T = 1, then u ≤ 0, and if T = 0, then u ≥ 0. This means that the errors are mean-reverting, implying that cov(T, u) < 0. For this case, the proportional bias in estimating β is equal to the regression coefficient from a hypothetical regression of u = S − T on X. When S is the only explanatory variable ([X] = [S]), this is equivalent to . We report b_uS for both diseases.

We next follow [7] for testing the hypothesis that self-reports of disease status are unbiased estimators of the true disease status. The tests are based on ordinary least squares (OLS) and bivariate probit regressions. We estimate OLS regressions for hypertension and lung disease with the error (S − T) as our dependent variable, a vector of control variables and a constant. The control variables comprise individual and household characteristics that reflect risk factors for the disease, including socioeconomic characteristics of the household. The test statistic is an F-test for the joint significance of the control variables. In the bivariate probit regressions, our dependent variable is S for one equation and T for the other, with the same set of control variables used for the OLS regressions. We test the hypothesis that E[S − T|X] = 0, based on an F-test that the estimated parameters on X in the two equations are equal.

The medical literature identifies age and lifestyle factors as risk factors for hypertension [25]. Lifestyle factors include excessive consumption of dietary salt, smoking, alcohol consumption, high BMI, physical inactivity and a diet deficient in fruits and vegetables. There is evidence suggesting that education [26, 27] and household income [6] lower the propensity for hypertension. The LASI does not contain information on dietary habits of respondents. But we are able to control for the other risk factors and moderating characteristics.

Risk factors for chronic lung disease include smoking, indoor air pollution caused by biomass fuel used for cooking, outdoor air pollution, occupational exposure to dust/gas, respiratory-tract infections during childhood and chronic asthma [28]. Many of these factors are in turn correlated with poor socioeconomic status. While we do not have information on childhood disease history or occupational exposure to pollutants, we do control for type of fuel used for cooking, urban living (that is likely to be correlated with outdoor air pollution), smoking behavior and socioeconomic variables in estimating reporting error regressions for lung disease. Following [20], we define cooking fuel to be of good quality if the household uses any of the following fuels for cooking: coal, charcoal, natural gas, petroleum, kerosene, or electric. In our statistical analysis, we use this definition of good cooking fuel. However, guidelines issued by the World Health Organization (WHO) consider coal and charcoal to be highly polluting and do not include these among recommended fuels for cooking [29]. We therefore use an alternative definition of good quality cooking fuel that excludes coal and charcoal. We test all our results for robustness to the use of this alternate definition.

False negative reporting

We also estimate models of false negative reporting (S = 0|T = 1), to identify characteristics that might be associated with such errors. There is a methodological issue with modeling false negative reporting, which arises from the fact that we only observe such errors for individuals diagnosed with the corresponding disease (T = 1). For those without the disease (T = 0), we do not observe what they would report in the counterfactual scenario. We follow the approach outlined in Johnston et al., of using a censored probit regression model to address this sample selection issue [6]. Van de Ven and Van Praag describe the censored method for addressing sample selection in detail [30]. We outline the model below.

The censored probit model comprises two latent variables: , an individual’s likelihood of having the disease j, {j = hypertension, lung disease}, defined as follows: and measuring the likelihood of false-negative reporting: We then have if , and if and

Thus, for hypertension, y_i1 equals 1 if individual i has hypertension and zero otherwise, and y_i2 equals 1 if y_i1 = 1 and S_i = 0, and zero otherwise. The vectors z₁ and z₂ comprise observed socio-demographic characteristics of the individuals in the sample, while ϵ₁ and ϵ₂ are distributed bivariate standard normal with covariance ρ: . The covariance parameter ρ captures the impact of unobserved characteristics that might affect both the propensity of having the disease and the propensity to report not having it. We need valid exclusion restrictions to identify the model. The LASI data has no information to indicate whether individuals have a genetic pre-disposition to either disease. Instead, we use other variables that could potentially serve as exclusion restrictions.

The C-reactive protein (CRP) concentration in blood and the pulse reading (heart rate) are two independent predictors of cardiovascular disease (CVD). CRP is a stable and sensitive biomarker for systemic inflammation; an elevated level of CRP in the blood, above the medically-accepted threshold of 3 mg/L, is indicative of infections in the past and is considered an independent risk factor for CVD [31–33]. Pulse/heart rate is yet another independent risk factor for CVD [34]. Since both these variables are risk factors for CVD, they are likely to be correlated with hypertension. At the same time, there is no reason to expect these variables to influence the self-reports, one way or the other. Hence we use these two variables as instrumental variables (IVs) for hypertension.

Exposure to passive smoking and measured Epstein-Barr virus (EBV) antibody activity are our chosen IVs for lung disease. The LASI asks each respondent, “Does any usual member in your household smoke usually inside the home?”, with a valid response being a Yes/No. Those exposed to passive smoking are likely to be at greater risk of lung disease. The Epstein-Barr Virus (EBV) antibody titer is a reliable measure of cellular immune function. EBV is a member of the herpes virus family. Over 90% of the world’s population carry this virus, however, it doesn’t affect most individuals over their lifetimes. For the EBV to remain in a latent state, a well-functioning immune system is crucial [35]. While there are no clinically accepted thresholds for this measure, elevated levels of EBV antibodies indicate a compromised immune system, making individuals vulnerable to many diseases including certain types of lung diseases [36, 37].

Measurement error in self-reported health and associated bias

The sample means in columns (1) and (2) in Table 3 are the same as those reported in Table 2. Column (3) is the magnitude of error in S. Estimates of the corresponding proportional bias, as reported in column 4, are substantial and similar in magnitude for the two diseases. The bias estimates imply that if we use self-reported hypertension (lung disease) as the control variable instead of the true disease status in any regression analysis, the estimated coefficient will be 83% (87%) smaller than the true coefficient. When we use T1 and T2, the alternate definitions for true hypertension, the associated bias estimates are 91% and 83% respectively, with standard errors of 0.03 in each case. These estimates are very large, and call into question any inference based on using self-reports in place of the true disease status.

Download:

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

Table 3. Error decomposition.

https://doi.org/10.1371/journal.pone.0202786.t003

Baker et al. report estimates of b_us for 13 health conditions including hypertension, asthma and bronchitis [7]. Their estimates are 0.36 for hypertension, 0.57 for asthma and 0.9 for bronchitis. In comparison, our bias estimate for hypertension is notably bigger, following from the large error in the self-reports. Our b_us estimate for lung disease is very similar to their estimate for bronchitis. Our bias estimate is also significantly bigger than the 0.68 estimate of Johnston et al. for hypertension [6].

Table 4 reports results from OLS and bivariate probit models of reporting errors, to test the hypothesis that self-reports of disease status are unbiased estimators of the true disease status, controlling for individual and household characteristics.

Download:

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

Table 4. Tests of hypothesis that error in reporting = 0.

https://doi.org/10.1371/journal.pone.0202786.t004

For both hypertension and lung disease, tests based both on the OLS and bivariate probit specifications unequivocally reject the hypothesis that the self-reports of the disease status are identical to the true disease status in expectation, conditional on the control variables. We therefore estimate models of false negative reporting, to identify characteristics that might be associated with such errors.

Estimates of false negative reporting

Table 5 presents the marginal effects from censored probit models of false negative reporting for hypertension. We present estimates from two specifications—one controlling for basic socio-demographic characteristics including household income, caste, religion, urban status and state of residence. The second specification also controls for educational attainment, some additional proxies for household income, lifestyle variables, health insurance status and physical risk factors. For comparison purposes, we also present estimates from standard probit models that do not correct for censoring.

Download:

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

Table 5. Estimates of false negative reporting: Hypertension.

https://doi.org/10.1371/journal.pone.0202786.t005

For the censored probit estimations, the two instrumental variables (IVs) included in the hypertension equation—pulse rate and an indicator for cardiovascular disease (CVD) risk, denoted as 1 if the C-reactive protein (CRP) concentration in blood is over 3 mg/litre—are statistically significant in both specifications, as measured by the Wald test. However, in both specifications, ρ, the estimated correlation between the error terms of the two equations is imprecisely estimated. Moreover, the estimated correlation declines significantly in the extended specification, relative to the basic one. This implies that selection is not a serious concern in our data. In view of this, our preferred specification is the one based on the probit model.

According to the probit estimates in the third column, the propensity to misreport hypertension goes down significantly with age. This is robust to the inclusion of controls for literacy, educational attainment, lifestyle variables and physical risk factors (column 7). Moreover, women have a lower propensity to misreport hypertension. Literacy lowers the propensity to under-report by about 7 percent. However, schooling appears to have no relationship with the dependent variable. Physical risk factors—specifically, being overweight and having a bigger waist-to-hip ratio—also lower the propensity of false negative reporting. This is presumably due to overall poorer health associated with these characteristics, which might require frequent visits to a doctor. The negative effect associated with former smokers is probably a selection effect, reflecting greater awareness and/or self-control—factors that are positively correlated with health in general. Notably, the use of good cooking fuel and availability of indoor plumbing are also associated with a lower probability of under-reporting. These are household characteristics reflecting higher socioeconomic status.

Our finding that women have a lower propensity to misreport is consistent with that of [6]. In contrast to our estimates, however, they find that controlling for censoring in the misreporting of hypertension is important. They also find a small, negative and statistically significant effect of household income on under-reporting. We do not find any evidence of an income gradient.

Table 6 presents the results for lung disease. Since valid spirometry test results are available for a much smaller sample compared to the sample that had blood pressure readings taken, we use a more parsimonious set of control variables than in Table 5. The IVs used in the selection equation—exposure to passive smoking and the level of Epstein-Barr antibody activity in the blood stream—are statistically significant in both specifications at the 5% level, as measured by the Wald test. We again find no evidence of selection in false negative reporting of lung disease. This is even less surprising since there is almost no variation in self-reports among those identified as having lung disease (see Table 2). Our preferred specifications are therefore the probit specifications.

Download:

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

Table 6. Estimates of false negative reporting: Lung disease.

https://doi.org/10.1371/journal.pone.0202786.t006

As with hypertension, the probit estimations reveal a small, negative and statistically significant effect of age on the dependent variable. There is no evidence of an income gradient or an education gradient. Current and former smokers are about 3% less likely to misreport. This is plausible given the physical symptoms that appear with regular smoking as well as the strong public campaigns to increase awareness of the detrimental effects of smoking. In contrast to the estimates in Table 5, physical risk factors appear not to influence misreporting.