In this study, we collected personal and environmental information of patients with ESRD across the years to investigate the spatial and temporal association between risk factors and the incidence of ESRD. In the spatial analysis, we found spatial clustering effects for ESRD incidence rate, and a high-high association (hot spot) was observed in southern Taiwan every year. Considering time-dependent and spatial effects, we applied a spatial panel data model to estimate the marginal effects of personal and environmental factors on the incidence of ESRD. We found that environmental factors such as population composition (proportion of old adults and level of education), healthcare resources, and unemployment rate were associated with the incidence of ESRD. Personal risk factors, including diabetes, hypertension, NSAIDs, and aminoglycosides, were independent risk factors for ESRD after adjustment for spatial and temporal effects. Furthermore, the spatial spillover effect was not significant, implying that personal and local environmental factors were substantial after adjusting for all covariates, rather than the effects in neighboring areas.
The residential location in the registry of the NHID is not always equal to the residential area. Therefore, we applied the algorithm of residence area proposed by Lin et al. [9] and linked the NHID to the NHIS to obtain the actual residence area and evaluated the accuracy using this algorithm. The average PPV was approximately 84.3%. Therefore, the algorithm of residence area is suitable for estimating a person’s residence area in the NHID registry.
The spatial analysis showed a high clustering effect of ESRD incidence in the southern and southeastern townships. These townships are the most remote areas with insufficient healthcare resources, such as the number of hospitals, medical staff, and beds, compared to the northern areas in Taiwan. A previous study indicated that different sizes of hospital and medical manpower resources were associated with the deterioration of renal function. [10] It may be one of the reasons for the higher ESRD incidence in remote areas. There were some urban townships in the southern areas, such as Tainan City and Kaohsiung City, which also showed a high clustering effect of ESRD incidence, and these areas have sufficient healthcare resources. The reason for the higher ESRD incidence is different from that in other remote areas. A previous study reported that sufficient healthcare resources could improve accessibility. Patients would accept dialysis because they were able to access healthcare resources. [11]
Spatial analysis is a univariate analysis that cannot be used to investigate the association between multiple factors and outcomes. Therefore, a spatial panel data model can be used to evaluate the effect of a risk factor after adjusting for spatial and time-dependent effects and other risk factors. In the spatial panel-data model analysis, a higher proportion of older adults would increase the risk of ESRD, which was similar to a previous study in which renal function would be progressing with age. [12] A study in the United States of America reported that there was a higher rate of rapid deterioration of renal function in towns with a high proportion of ethnic minorities or immigrants (OR = 1.25, 95% Cl = 1.2–1.31), [11] and this finding was similar to our result that higher proportion of aboriginal people was associated with higher rate of ESRD incidence. Most aboriginal people lived in remote townships in Taiwan, which had insufficient healthcare resources and were associated with an increased ESRD risk. The area with a higher proportion of bachelor's degree was significantly associated with lower ESRD incidence, which was similar to a previous study that found that the higher the level of education, the lower the incidence of diseases and deaths from diseases. [11, 12] People with higher education are more likely to realize and accept health suggestions, [13] have a higher socioeconomic status, and have affordable medical expenses. [14] We also found that a higher unemployment rate was associated with an increased risk of ESRD, which could be attributed to socioeconomic status. [14]
NSAIDs and aminoglycosides are well-known nephrotoxic agents. However, in clinical observational studies, the detrimental effect of NSAIDs in chronic kidney disease (CKD) progression was inconsistent. In a prospective observational study of 10,184 individuals older than 65 years, the increased risk of CKD progression was only seen in NSAIDs users with glomerular filtration rate (GFR) > 60 mL/min/1.73 m2, but not in NSAID users with GFR < 60 mL/min/1.73 m2. This study could not exclude the possible bias that, in clinical situations, greater NSAIDs use may be avoidable in patients with less renal function. [15, 16] In another prospective cohort study of 4,101 patients with rheumatoid arthritis, a significant deleterious effect of renal function decline was only seen in patients with CKD stage 4–5, but not in patients with CKD stage 1–3. [17] In our study, through spatial analysis, we found that the use of NSAIDs or aminoglycosides within 90 days prior to the index day increased the incidence of ESRD. Therefore, we reinforced the importance of avoiding exposure to nephrotoxic agents in patients with CKD. Interestingly, the use of aminoglycoside within 180–365 days prior to the index date had protective effects against renal function decline in ESRD. The reason for this was not clear, and it might imply that early recognition of CKD and knowledge of further avoiding exposure to nephrotoxic agents are practical strategies to preserve renal function in patients with CKD.
Ignoring the spatial dependence on the response and explanatory variables may reduce the efficiency of the results and produce biased and inconsistent model estimations. In this study, we consider spatial panel models to study the association of risk factors with ESRD, which include lags of the dependent variable and of the independent variables in both space and time, and provide a useful tool to quantify the magnitude of effects from the variables, both in the short- and long-term. In the model, the direct effects of the variable on the dependent variable can be investigated in its own ESRD incidence rate rather than the coefficient estimate of that variable. In addition, the indirect effects of spatial spillovers exist rather than the coefficient estimate of the spatially lagged dependent variable, and/or the coefficient estimates of the spatially lagged independent variables can be examined. We applied the models to empirical data and ESRD incidence rate data from townships in Taiwan from 2004 to 2011. In this model, we found that NSAIDs or aminoglycosides seem to have an upward and then downward significant effects on ESRD incidence rates (Table 2).
Limitations and strengths
This study had several limitations. First, the patient’s residence may have been misclassified because of our uniform insurance system. Patients may undergo dialysis or have medical visits away from their place of residence. Our validation study showed that the lowest PPV was for Kinmen, an offshore island with limited healthcare resources. Second, the information on particulate matter (PM2.5 and PM10) from county monitoring stations, not townships, may result in some loss of precision concerning intra-township variability risk. Another limitation was the absence of data on renal function in patients who received NSAIDs or aminoglycosides. The risk of nephrotoxicity due to NSAIDs or aminoglycosides occurs in patients with an underlying renal disease.
The primary strength of this study was that it examined a confirmed ESRD population with minimal financial barriers and readily accessible healthcare under Taiwan's full-coverage NHI program. We estimated the risk factors for ESRD after adjusting for geographic influence, socio-environmental factors, comorbidities, and medications. This will provide useful information for policymaking in different regions by targeting potential risk factors.
Spatial panel data models provide a useful tool for capturing time-dependent and spatial clustering in responses and covariates. However, there is a big issue in the specification of spatial weights matrix W and m. Consequently, empirical studies follow a statistical approach driven by data analysis and are limited to one or a few pre-specified W matrices. Often, spatial weight matrices are specified in terms of a well-founded background for certain spatial interaction effects and the frequency of their use, resulting in criticism in empirical studies. In view of this, it should be clear that the way of thinking and model selection strategies are used in most empirical studies to determine the structure of the spatial weights’ matrix W.