Health expenditures and causa mortis: an investigation on municipalities of the State of Paraná, Brazil

The Brazilian government shares the responsibility of financing public health among federal, state, and municipal levels. Health expenditures are thus un-even across the country and cannot contribute equally to health outcomes across disease categories. This study aims to identify how the health expenditures of municipalities affect the mortality rate in the state of Paraná by causa mortis. We considered years of life lost for each municipality, the chapters of the International Classification of Diseases (10th revision), and the elasticity of this measure in relation to public health expenditure. Considering the possibility of endogeneity, this study follows the instrumental variable approach in a panel of generalized method of moments – instrumental variable (GMM-IV) with fixed effects. Our results show that a 1% increase in health expenditure could decrease the average number of years lost specifically for some causes from 0.176% to 1.56% at the municipal level. These findings could elucidate policy perspective within state finance.

Brazilian studies on health expenditures are strongly related to the literature on SUS funding. Since SUS is funded by taxes, the source and availability of resources are subject to scrutiny. The discussions inquire how much of the investment contributes to the access, equity, and comprehensiveness of treatment, which are the principles of SUS 3 . Overall, criticisms are aimed at the funding model, which was based on fixed proportions defined by the Tax Liability Law 4 up until 2015, when it started to suffer the effects of the Constitutional Amendment n. 85/2015. This law changed the rules of funding in the federal sphere, decreasing the investments in health care and resulting in the upward trend of municipalities injecting more of their own resources while the limitation of resources incentives the privatization of health services 5,6,24,25,26,27,28,29,30,31 .
The discussions on the health system also approaches the clear priority given to primary care compared to second and tertiary care 25 . The role of each level of care is examined within the principle of universality and the criticisms of its social design. As an example, Mendes & Marques 26 analyzed the Family Health Program (FHP), which is increasingly dependent on the municipality's health resources and therefore a liability for sustainable funding.
This study aimed to assess how the mortality rates in each group of diseases (according to the International Classification of Diseases, 10th revision -ICD-10) respond to the health expenditure applied by each municipality. This study considers the most recent publications on this research area, bringing theoretical and empirical novel methods of analysis 13,22,23 . To our knowledge, this is the first study to use this approach for Brazilian data, with important implications for local policymakers and policymakers from other similar developing countries. To approach the endogeneity issue, panel data was used from the GMM-IV method (generalized method of moments -instrumental variable). Data were analyzed from 2014 to 2018, collected from the Brazilian Ministry of Health 32 and from the Brazilian Information System for the Public Budgets in Health (SIOPS) 33 .
In this study, we follow the regional approach, choosing the state of Paraná as a case. The state has four macroregions and is subdivided into 22 health management regions 32 . Among Brazilian states, Paraná has one of the highest life expectancy, lowest infant mortality, and lowest poverty rate 32,34 , but still ranks the lowest within its region. The state's health expenditures per capita (BRL 373.73) are similar to the national average (BRL 376.85 per capita -USD 69.78, considering the exchange rate of USD 1.00 to BRL 5.40 on Nov 12, 2021), varying significantly across 399 municipalities. This indicates that other states perform better with almost the same budget and Paraná should further improve its health financing and achievement of health outcomes.

Methodology
Following Martin et al. 22,23 and Claxton et al. 12 , health outcomes could be expressed from mortality rates, classified by cause of death considering the public health expenditures. Similarly to the authors who studied the British NHS (National Health Service), we analyzed a public health system funded by citizens' taxes -the SUS -using the dependent variable years of life lost (YLL) for each chapter of ICD-10.
Data were obtained from 2014 to 2018 for the 399 municipalities of the state of Paraná. In the Mortality Monitoring Panels, the Brazilian Ministry of Health 32 provides the number of deaths by location, cause of death (based on chapters of ICD-10), time, and age range. This classification groups death causes which have different mortality rates among each other. Regarding representativity, the causes of death that represented less than 1% of the total were excluded from the sample. Table 1 summarizes the occurrences and names of the classifications and the YLL for each cause.
The number of expected YLL was estimated considering the life expectancy at birth in each location as informed by the 2010 Population Census 34 . Marshall's 35 approach was used considering the population of a location i, where M deaths have occurred in a period of t years due to a particular cause j in a population of age group a; in which e is the measure of expected YLL for an individual that passes away at the age a. The number of expected YLL due to cause j in the location i and time t, summed by groups of age ranges, is then given by A ijt .
Cad. Saúde Pública 2022; 38 (7):e00272421 The empirical findings on this subject indicate a possible reverse causality between health expenditures and health outcomes such as life expectancy or YLL, resulting in endogeneity. According to Grossman 1 and Nixon & Ullman 2 , if a public health system helps reduce the burden of the diseases in a locality, more resources should be applied in this system to reduce YLL to these illnesses. However, health outcomes would also affect the direction of the resources in a particular area; that is, they affect and are affected by health expenditures. Instrumental variables were thus used while acknowledging the possibility of endogeneity between these two variables.
The dataset is constructed in the form of panel data, where constant non-observable or nonavailable characteristics are captured using fixed effects by locality.
According to the Tax Liability Law 4 , every Brazilian municipality must apply 15% of collected taxes on healthcare. Considering the recent trend of increased municipality funding in health activities 6,25,26 , municipalities have engaged in operational responsibilities and are using more of their own resources besides the mandatory 15%. Figure 1 shows municipalities' expenditures on the left axis and the amount applied in health care in 2014-2018 on the right. Health expenditure steadily increased for the whole period while the percentage applied dropped from 2017 onwards.
The data show that in 99.7% of all the annual investments within the considered period, municipalities of Paraná applied variable portions above 15% on healthcare. The variances of the percentage   Source: prepared by the authors based on information from Brazilian Information System for the Public Budgets in Health (SIOPS) 33 .
applied and its deviation are close to 21 whereas the standard deviations are around 4.5%, showing that the investments of the municipalities vary considerably. As a result, the percentages applied and their deviation from the minimum required express a proxy of the healthcare needs considered essential by the government of each municipality. Andrews et al. 36 and Claxton et al. 12 have previously used, in the United Kingdom, funding rules that determine the volume of resources each geographical area receives. The authors argue that these rules alone can influence the expenditures but not mortality rates. In Brazil, the revenue of municipalities is a complex combination of resources including taxes and transfers from other sources, such as the federal and state governments. Our parallel instrument would thus be the percentage applied in health and the deviation from the minimum stipulated by law, which are also possibly related to expenditure but not to mortality.
If the instrument was related to municipality revenue, it could result in endogeneity. However, since the linear correlation between the two variables is low (0.0033), these variables may be considered exogenous, kept as candidates for instrumental variables, and submitted to the appropriate tests.
Some studies 9,18,20,21 have indicated that the method of two-stage least squares (2SLS) with instrumental variables leads to biased results in the presence of heteroskedasticity 37 . We therefore tested the presence of heteroskedasticity in our estimations to verify this problem. Using the Pagan-Hall test, all causes of death returned heteroskedastic errors when regressed with our 2SLS instrumental variable method. We thus chose the GMM-IV method, robust for heteroskedasticity.
To control variables that may be indicators of the supply of health services and human capital, vaccinal coverage 19 and the number of health professionals for each 10,000 inhabitants were considered as proxies for SUS attendance in municipalities. The literature on the persistent anomaly 7 shows that the mortality-expenditure relationship could be affected by the number of health workers and lead to endogeneity. In Brazil, the expenditure of salaries and personnel costs might also be a large part of health expenditures, but the SIOPS data did not display those costs separately. The variable of health professionals was thus maintained as a control considering that if human capital leads to endogeneity, this variable must be significant. This information was obtained from the Brazilian Ministry of Health 32 .
To consider development variation among the municipalities, the included explanatory variables were demographic density, the gross value (GV) of agriculture, industry, services, public administra-Cad. Saúde Pública 2022; 38 (7):e00272421 tion, and GDP per capita. The population variable was obtained from the 2010 Population Census 34 estimations for 2014-2018. We tried to consider differences in education that may be related to health outcomes according to the literature 10,18,19,20 . The IDEB (Basic Education Development Index) score was used as a proxy for education outcomes by municipality.
Moreover, a dummy variable was included from 2016 onwards for two reasons. Firstly, this period comprises the possible effects of the Constitutional Amendment n. 85/2015 on the health funding of municipalities. Secondly, since a general election for mayors occurred in 2016, the changes in public expenditure and political scenarios could have affected our analysis. Formally, the model is estimated using fixed effects, with every variable in logarithmic form: Where: A ij are the YLL in year t, municipality i, and cause of death j, while G it is the health expenditure of the municipality i, X it is a vector of covariates, d t is our dummy from 2016 onwards, and t is a time trend.
To test the validity of the instruments, the estimations were submitted to underidentification tests, weak instruments tests, and overidentification tests (respectively Kleibergen-Paap rk LM, Kleibergen-Paap rk Wald F, and Hansen J). In every estimation, the outputs were within the values that allow accepting both instruments (expenditure percentage of tax collection and the difference to the minimum). The output of the tests is available under request. Furthermore, each estimation was submitted to the Hausman test for endogeneity robust to heteroskedasticity. The health expenditures were considered endogenous to mortality for some diseases.

Results
Overall, the results suggest significant and negative elasticities, indicating that health investments can reduce the mortality rates of some diseases. In turn, population density was almost unanimously positively related to mortality and YLL but significant only in some specifications -possibly because of its association with urbanization, industrialization, and population growth. These phenomena are also related to higher exposure to pollution, lack of access to sanitation, and other factors which may be negatively affecting mortality rates.
The results of immunization efforts may seem contradictory: in most cases, better vaccine coverage seems to increase mortality rates instead of decreasing them. However, using ICD chapters directly related to immunization (i.e., those which contains diseases preventable by vaccines), the relation between vaccine coverage and mortality is negative, as expected, and more significant with higher magnitudes, indicating the fitness of the model. The number of health professionals variable was not statistically significant in any specification.
Variables aimed at measuring the degree of development conditions and education were not considered significant in the model. We hypothesized that, in the short term, their variation might not directly affect health outcomes. The gross value of public administration, however, was found significant and had high coefficients in many specifications.
Interestingly, the dummy variable for 2016 onwards showed different patterns among the causes of death, being negatively related to YLL in most cases. This could indicate that, overall, health investments were relatively more effective for some diseases during this period. Tables 2 and 3 present results by ICD-10 chapters. The analysis for Diseases of the blood (Chapter III), Mental and behavioral disorders (Chapter V), Diseases of the nervous system (Chapter VI), Diseases of the circulatory system (Chapter IX), Diseases of the respiratory system (Chapter X), Diseases of the skin and subcutaneous tissue (Chapter XII), Diseases of the genitourinary system (Chapter XIV), Pregnancy, childbirth and the puerperium (Chapter XV), Symptoms, signs and abnormal clinical and laboratory findings (Chapter XVIII), and External causes of morbidity and mortality (Chapter XX) were not found endogenous and are statistically insignificant. These analyses therefore were not analyzed in the text.
The results for the remaining causa mortis are detailed as follows.

Certain infectious and parasitic diseases (Chapter I)
About 38% of deaths in this chapter were caused by bacterial diseases and 25% were caused by HIV 32 . Expenditure was found endogenous to mortality and its elasticity was negative and significant at around -0.176. This implies that for every 1% increase in health expenditure YLL are expected to decrease in 0.176% for this cause of death. Population density was significant only at 10% while immunization and health professionals were not statistically significant. The gross value of public administration is highly significant at 1% significance.

Neoplasms (Chapter II)
Almost all (98%) deaths from neoplasms are caused by localized malignant tumors 32 . Expenditure was found endogenous to mortality and its elasticity was negative and significant at around -0.520. This implies that for each 1% increase in health expenditures, YLL are expected to decrease in 0.520% for this cause of death. Population density was positively related to mortality at a 5% level of statistical significance. Immunization showed a statistically significant and negative coefficient, but with no obvious reasons. The dummy variable for 2016 was found negatively related to YLL for this cause of death. The number of health professionals and GDP are not significant in this setting. GV of the public administration was found to be negative and statistically significant.

Endocrine, nutritional, and metabolic diseases (Chapter IV)
More than 80% of the victims of this ICD-10 chapter were diagnosed with diabetes mellitus 32 . Expenditure was found endogenous to mortality and its elasticity was negative and significant around -0.234 at the 5% level. This implies that for each 1% increase in health expenditures, YLL to this cause of death are expected to decrease in 0.234%. Population density was not significant in this setting. This could be positively related to mortality since diabetes is prevalent in more industrialized or urbanized regions 38 . Other variables considered proxies for regional development -such as GDP or GV of services/industry/public administration -and immunization were also not significant.

Diseases of the digestive system (Chapter XI)
The victims of this ICD-10 chapter were, in 75% of the cases, patients with diseases of the liver, gallbladder, pancreas, or intestine 32 . Expenditure found endogenous and significant, with an elasticity of -0.272. This implies that for each 1% increase in health expenditures, YLL due to this cause of death are expected to decrease in 0.272%. Both immunization and the number of health professionals are not statistically significant. GV of public investment was statistically significant and negatively related to this cause of death at the 5% level of significance.

Certain conditions originating in the perinatal period (Chapter XVI)
In 51% of the cases, the victims of this condition underwent complications of pregnancy and delivery 32 . Expenditure was found endogenous to mortality and significant, reaching an elasticity of -1.563 to 1% of significance. That is, for each 1% increase in health expenditure, YLL to this cause of death are expected to decrease in 1.563%. Immunization was significant at the 1% level and negative around -0.577. Although this result cannot be verified directly from the database used, it may be related to the application of BCG and hepatitis B vaccines, which must be applied in infants within the first month of their life. GV of public administration was found statistically significant, having a negative relationship with our dependent variable, while our dummy variable for 2016 was significant and positive. This positive relationship is an exception, suggesting that the changes in public funding and political environment that occurred in 2016 onwards somehow positively affected the numbers of this ICD-10 chapter. The coefficient of health professionals was not statistically significant for this cause of death.

Congenital malformations, deformations, and chromosomal abnormalities (Chapter XVII)
Out of the victims of this chapter, 52% are patients with malformations in the circulatory or nervous systems 32 . Expenditure was found endogenous and significant at 1%, with a negative coefficient of -0.702. This implies that for each 1% increase in health expenditure, YLL due to this cause of death are expected to decrease in 0.702%. This result may be surprising at first since many malformations are fatal and cannot be mitigated by the health system. However, since 1993, the European community has classified the mortality of this chapter as deaths preventable by the health system via surgery and medication 39 . The health professionals and immunization variables had no statistically significant coefficients. GV of public administration was significant and negatively related to YLL. Our dummy for 2016 was found statistically significant at 5%, being negatively related to our dependent variable.

Discussion and conclusion
The studies by Martin et al. 22,23 and Claxton et al. 12 used data on specific health expenditures per health care program, that is, the financing earmarked for that disease or group of diseases. This study did not use this approach due to data unavailability. Health expenditure aggregates all expenses recorded for health at SIOPS, including elements of the intensive margin (such as the payment for surgeries and treatments) and the extensive margin (such as the performance of health surveillance). Our results may also differ from the international literature for several other reasons. As an example, Cad. Saúde Pública 2022; 38 (7):e00272421 our data allowed estimating expenditure with different controls and our dependent variable was estimated by age ranges and not by the exact age of death. Moreover, we used a specific life expectancy for each municipality based on the 2010 Population Census while the international literature fixed life expectancy at 75 years. Life expectancy surpasses this age only in 30% of Paraná's municipalities, which may affect result interpretation. Finally, international literature used standardized mortality rates for their population. Considering these limitations, Box 1 summarizes the comparison between our results and the researched literature. Some of our results corroborate those found in the British literature for cancer, diabetes, neurological diseases, and gastrointestinal diseases, but our coefficients are considerably different in magnitude. The most diverse results concern cardiovascular diseases, which were not significant in this analysis despite the previous evidence from other countries. Moreover, respiratory diseases were not significant for expenditure despite the impact of immunization. This group was found to be negatively elastic with expenditure and significant in Claxton et al. 12 . In Martin et al. 22,23 , the same diseases were also found negatively related and significant, but since they did not pass the test of weak instruments, the authors recommended caution in using this estimate.
To our knowledge, Martin et al. 23 is the only other study in the literature which also mentions the analysis of diseases of the genitourinary system. The authors, however, did not report the analysis results since they considered them contradictory and counterintuitive. Since we found no significance for this ICD-10 chapter, overall, this classification may not react as expected to health expenditure.
Comparison of the other variables used in the model is less direct since most studies which adopted the fixed-effects panel regression conducted investigations between countries and therefore adopted different variables 2,17,18,20 . Regarding provincial data 9 , the lag of time was greater in our study. In Bilgel & Tran 21 , the dynamic panel considers lags of time not directly comparable with our results. Martin et al. 22,23 and Claxton et al. 12 do not use other control variables for British provinces.
Regarding human capital proxy, the number of health professionals was not significant in any of our specifications, indicating that the persistent anomaly presented by Cochrane et al. 14 may not occur for the state of Paraná.
The elasticity of public health expenditure for each cause reaches a magnitude below one, which is expected considering the context of whole-budget elasticities instead of by-program. That is, for per cent increases in health expenditures, the accompanying per cent decreases in YLL are proportion-

Box 1
Comparison between our results and the international literature. ally lower. Among causes of death, the conditions originating in the perinatal period stand out. In this setting, elasticity above one indicates that the YLL for this cause is elastic and improvements in health expenditure may decrease the average YLL proportionally more. This information is significant to improve public policies related to the care of mothers and infants. However, this cause of mortality represents only 1% of deaths in Paraná in the analyzed period. Neoplasms are the second most elastic death, in which YLL decrease in 0.502% for every 1% increase in expenditure in a municipality. They cause 19% of deaths in Paraná, indicating a potential to offer better treatment to numerous patients and thus effectively extend longevity.

DISEASE
Our results indicate reformulating funding by budget programs which allocate funding by diseases or group of diseases could offer better adjust each need according to each elasticity.
In fact, chronic diseases that increase mortality (i.e., cancer, diabetes) respond well to health expenditure. As expected, the same care, however, is not effective with chronic incurable diseases with historic lower mortality, such as neurological diseases, which do not achieve a similar result to other diseases despite relying on palliative care.
Death caused by cardiovascular diseases, the major cause of death, was not considered elastic to health expenditure. This result indicates at least two perspectives. On one hand, this suggests that patients affected by a heart attack or stroke could inevitably die despite medical care. In turn, it indicates that the population suffering from this these diseases may not be receiving or seeking adequate medical treatment in time or before the attack occurs. Unfortunately, the data available do not allow further assessment in this regard.
Overall, acute and deadly conditions such as violence, suicide, disorders caused by psychoactive substances, and death from undefined causes are not statistically significant relationship to health expenditure. This result indicates how the health system often cannot achieve comprehensive needs since, in many cases, nothing can be done.
This study has several limitations. Considering that life expectancy cannot increase unlimitedly, it may be associated with diminishing marginal returns. Unfortunately, due to multicollinearity in the endogenous variable, a square coefficient could not be tested for health expenditures. Another limitation is the absence of individual-level data for our population since many other factors, such as health behavior, may have influenced mortality rates.
Some diseases which can be treated and mitigated by the health system showed relevant results. Further studies should contribute to expand information to promote equity in health considering this evidence on how SUS financing has affected health outcomes.

Contributors
T. K. Feldens and K. B. Souza conceptualized the study and contributed to the literature search. T. K. Feldens compiled and analyzed data, produced the figures and tables, and wrote the original draft of the manuscript. K. B. Souza provided statistical guidance and input on the analysis and edited early drafts of the manuscript. All authors contributed to subsequent drafts of the manuscript, provided critical interpretation of the results, and approved the manuscript for publication.

Conflict of interests
None.