Spatial distribution of healthcare access and utilization: do they affect health outcomes in Turkey?

ABSTRACT

This paper examines the link between healthcare access/utilization and health outcomes in Turkey within a spatial framework. Our initial set of findings highlight an overall duality in health indicators which is getting stronger once a spatial dimension is included. Specifically we find wider spatial dichotomy for health outcomes relative to access and utilization measures. Finally once we consider unobserved heterogeneity, spatial spillovers and spatial variability; our results pinpoint a non-robust link between healthcare access/utilization measures and health outcomes which works better among the already developed regions of Turkey. Overall, our combined results indicate an ongoing polarization of health-based human capital development which coincides with local variations of the relationship between healthcare access/utilization and outcomes in Turkey.

KEYWORDS:

• Healthcare
• inequalities
• spatial dependence
• Turkey

JEL CODES:

• I11
• I15
• R11

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Fırat Bilgel http://orcid.org/0000-0002-2585-5975

Notes

1 Our data set covers the 81 provinces for the 2009–2014 period. Only for the primary healthcare visits we are unable to obtain regional data for the pre 2012 period.

2 See Redding and Schott (2003) and Redding and Venables (2004) for the formal NEG model.

3 We use motorway distances to measure the physical proximity. Data is obtained from Ministry of Transportation. Another way to consider the distance is to use the travel time distances. However we do not have reliable data on travel time; moreover it is also less likely to control for the quality of infrastructure and road-networks at the regional scale.

4 Note that we also consider the accessibility within the same province. In case i=j, we use the Head and Mayer (2006) approximation for internal distance as $D_i=0.66/\sqrt{Are{a}_i/\prod )}$.

5 Moran's I lie between −1 and +1, where − and + values represent negative and positive spatial autocorrelation respectively; 0 represents spatial randomness. For Geary's C, values lower than 1 represent increasing positive spatial autocorrelation values higher than 1 represents increasing negative spatial autocorrelation and 1 represents spatial randomness. Moran's I is more consistent especially in large samples (Anselin & Getis, 1992). In our spatial concentration analyses we use both to check for the robustness of our results.

6 We use different orders to determine the threshold level (fourth, sixth, eight etc.). These results are available upon request.

7 Note that the type of hospital (i.e. public, private, university) may be used as an additional control. However we fail to use this variable due to high number of missing observations. Moreover for the period of our analysis, the number of hospitals is quite stable. As a result, we used the share of university hospitals in our empirical analysis, resulting in a significantly lower sample size without any significant change in the results. Therefore, we keep these analysis out of the text, however they are available upon request.

8 Akaike Information Criteria (AIC), Bayesian Information Criterion (BIC) and Cross Validation (CV) are used for bandwidth selection. Moreover following Fotheringham et al. (2002) we implement the adaptive Kernel approach.

9 In our spatial auto-correlation analysis, we take the order of contiguity as 1 for contiguity weight matrix construction. For the threshold distance weight matrices the number of neighbors is set to 5. We have tried different order levels and number of neighbors for contiguity and k-nearest weight matrices. These results are available upon request.

10 We use inverse distance weighting for LISA calculations. Local spatial analysis yields similar results upon the use of different weight matrices despite the sensitivity of the strength of the global measure to weight matrix specification. These results are available upon request.

11 The results of the LISA analysis for the healthcare accessibility and utilization indicators are given in Figures A1–A8 of the online appendix.

12 See Gezici and Hewings (2004) and Gezici and Hewings (2007) for spatial dimension of regional income inequalities in Turkey

13 See online appendix Figures A9–A12.

14 See Section 4.3 and the online appendix for details.

15 See Longhi, Nijkamp, and Poot (2006) for year by-year application of GWR models in panel structures. Here we only report results for the beginning and ending years of our sample. Results of all individual years are available upon request.

16 Frohlich, Ross, and Richmond (2006) discuss the Canadian experience and underline that policy tools directed to channels that affect the overall inequalities can also have sizable influence to combat against health disparities among different segments of the society.

17 See Figures A13–A16 on maternal care non-utilization rates; Figures A17 and A18 on consanguinity, Figures A19 and A20 on under-20 birth ratio and Figures A21 and A22 on female illiteracy rates.

Additional information

Funding

This work benefited from a financial grant from the Economic Research Forum (ERF) under grant no. 2015-145 of Global Development Network-Research Competition on the Economics of Healthcare in the ERF Region. The content and the recommendations of this research do not necessarily reflect the views of the ERF.