The impact of health facility delivery on child mortality in Tanzania : an instrumental variable approach

(1)

The impact of health facility delivery on child mortality in

Tanzania: an instrumental variable approach

Master Thesis

Supervised by Prof. Pauline Rossi, PhD

David Serban

University of Amsterdam

July 15, 2017

Abstract

This paper examines the impact of health facility delivery on child mortality in Tanzania. I use an instrumental variable approach to estimate a causal effect and rid the coefficient on facility delivery of any adverse selection that biases results. By instrumenting facility delivery with distance and altitude and controlling for obvious covariates, I obtain estimates for facility delivery that differ substantially from OLS estimates. I find that facility delivery is consistently and significantly negatively related to child mortality, although the magnitude of the effect is hard to interpret. The results offer corroborating evidence for the hypothesis of an imperfectly informed patient that isn’t fully aware of the increased chances of survival of the child during a health facility delivery. In terms of policy advice, I interpret the results as implying a need for a stronger focus on information campaigns about the increased chances of survival by giving birth in a health facility.

(2)

1. Introduction

In the past, the lack of health infrastructure in developing countries compelled policy makers to advocate for the erection of new hospitals in remote areas, to provide access to healthcare to those who are hardest to reach and most in need. Nevertheless, cases of corruption, absenteeism and the lack of quality services have made for a notoriously bad record of such stand-alone policies in many developing countries (Chaudhury, Hammer, Kremer, Muralidharan, & Rogers, 2006). In fact, numerous researchers have observed that increasing the availability of healthcare services does not necessarily increase the use of these services (Thaddeus & Maine, 1994). As a result, it seems that the widespread consensus has moved towards more qualitative aspects of the provision of health, as health facilities remained empty due to absenteeism of staff or competing uses of time for patients. Indeed, there is a large body of literature confirming patient’s prevalent use of traditional medicine, traditional births attendants and non-modern methods of healthcare (Leonard, Mliga, & Mariam, 2002) (Klemick, Leonard, & Masatu, 2009). Leonard et al. (2002) have even found that patients simply bypass newly built or nearly located health facilities for other supposedly higher quality facilities further away. This puts health policymakers in a difficult position where the mere provision of facilities is not sufficient to guarantee their use. The absence of health facility usage is particularly intriguing during pregnancy and baby delivery. In fact, the millennium development goals address the staggering disparities in maternal and child mortality in developing world when compared to the developed countries. As the chances of survival for the baby are much higher in equipped and well-staffed health facilities than during a home delivery, it is puzzling to find that in much of the developing world, home births are still common (Terhi J. Lohela, 2012). Could it possibly be that the mother consciously decides to deliver the baby at home even though she could give birth in a health facility nearby? Said differently, could it be that the expectation of the quality of care received at a health facility is so low that it might compel women to deliver at home to raise chances of survival? Or is it rather that poor household decision making and lack of information or transparency are the problem? It is therefore warranted to investigate whether delivering a baby at home is ill-advised and misinformed or whether it is a conscious decision taken by a mother faced with facilities of inferior quality in the surrounding area.

In order to find an estimate for the difference in child survivorship between home deliveries and health facility deliveries, I will use an instrumental variable (IV) approach to see whether

(4)

2

child mortality can be explained by home delivery or whether the available health facilities make no difference in the survival of the baby. Ordinary least squares (OLS) estimates of the effect of facility delivery on child mortality might be biased because of unobserved factors in the error term that are likely to be correlated to the dummy variable for facility delivery. I will therefore attempt to use distance and altitude difference to the health facility as instruments for facility delivery and control for obvious correlates of distance and altitude identified in previous research.

The rest of this paper is structured as follows. Section 2 will present theoretical and empirical research in the field of access to health facilities as well as the main strands of research on the quality of health facilities in the developing world. A formal model will be presented that will allow for an interpretation of the results. In section 3 I will explain the methodology I used to get to my results and I will briefly lay out my regression models. Section 4 will present results for different outcome specifications and section 5 will attempt to provide corroborating evidence for the results. A discussion of the results will follow in section 6 and I will end with some concluding remarks in section 7.

2. Literature review

2.1. Access to health facility

The decision of giving birth to a baby in a health facility is related to distance in many ways. Usage of healthcare services is usually modeled either as an exponential function of distance or in some other nonlinear form (Stock, 1983). Intuitively, this means that the last kilometer walked makes a lesser difference for long travels of say 40km’s compared to a short one of 5km’s (Leonard, Mliga, & Mariam, 2002).The exact functional form for this distance decay is usually determined case by case by the best fit.

There are also other geographical dimensions that are expected to be considered before attending a health facility. Some of the most relevant factors include altitude and terrain, proper road infrastructure as well as having the proper means of transportation to go from point A to B. As an example, villages in the Philippines, Sri Lanka, Vietnam and Morocco participating in the construction of rural roads reported significantly shorter travel times and higher usage of modern health facilities compared to non-project villages after the roads were built (van de

(5)

3

Walle & Cratty, 2002). Although the health benefits are only part of the advantages of better road infrastructure, such policies must be kept in mind since they might be more effective in increasing health usage rather than health-specific infrastructure.

On top of the physical accessibility factors, costs in the form of foregone earnings of leaving the household unattended, being accompanied or transportation costs are major impediments to seeking care (Thaddeus & Maine, 1994). In Uganda for example, the mean expenditure for outpatient’s transportation was 35c compared to the price of the medication of 11c (Stock, 1983). The policy responses to such cost barriers that detract a woman from delivery in a health facility are likely to be more straight-forward than reforming a highly centralized and inefficient healthcare system.

In addition to that, unobservable and non-pecuniary factors such as safety of travel or emotional support during delivery might influence the decision to walk another mile to the health facility and hence be related to distance and other indicators of remoteness.

As has been presented in this paragraph, although there are many ways in which distance and remoteness influence facility delivery, if the effect of distance on child mortality is channeled only through facility delivery, this won’t bias my estimation technique. Section 3 will further delve into covariates of distance that could be problematic to the exogeneity assumption of two-stage least squares (2SLS).

2.2. Quality of care

The focus in increasing health facility usage in rural regions in Tanzania has gradually shifted from “mysterious cultural barriers” and distance to more qualitative measures (Klemick, Leonard, & Masatu, 2009). Moreover, it is argued that quality takes precedence if facilities are within reaching distance. By the same token, seeking care is not just a function of the distance to the nearest facility or the quality thereof, but rather a function of all health facilities accessible to the household and the quality thereof (Klemick, Leonard, & Masatu, 2009).

One of the reasons for not attending a closely located healthcare facility is simply that the personnel that ought to be there is absent during the visit. In fact, a study of absenteeism in six developing countries revealed that the absenteeism rate for medical staff was found to be 35% for primary health centers (Chaudhury, Hammer, Kremer, Muralidharan, & Rogers, 2006). It has also been found that if health facilities were judged by their ability to provide adequate

(6)

4

treatment by trained professionals, 83% were understaffed or underequipped (Thaddeus & Maine, 1994). Alleged institutional reasons for missing staff are corruption as well as too high of a degree of centralization in health care that causes a lack of oversight in facilities as well as rigid payment schemes that fail to incentivize staff performance (Leonard, Mliga, & Mariam, 2002).

Faced with such uncertainty about the quality of care received, patients delay their hospital visit by trying to identify good from bad health facilities. As a result to this information asymmetry about quality of care at health facilities, patients are expected to make somewhat uninformed decisions, such as the decision to deliver a baby at home with a traditional birth attendant. This strand of literature emphasizes the role that health education and information campaigns can have on reducing health discrepancies (Gage, 2007) (Phoxay, Okumura, & Nakamura, 2001). In contrast to this, Leonard et al. (2002) challenges the assumption of the uninformed patient that is largely held in the literature by pointing towards evidence of a very informed assessment of facilities which are chosen according to severity of illness. To do this, Leonard et al. (2002) presents a framework that models change in expected utility ∆𝐸𝑈 as a function of travel costs 𝑇𝑖𝑗 between the facility 𝑗 and the patient 𝑖, fixed treatment costs 𝐹𝑗𝑘 at facility 𝑗 for illness 𝑘 and

value 𝑊𝑖 of treatment which in turn is a function of a vector of fixed inputs 𝑙 like staff and

equipment 𝑅_𝑗𝑘𝑙 which each have an illness specific beneficial return 𝑎𝑙𝑘.

∆𝐸𝑈 = 𝑊𝑖(∏ 𝑅𝑗𝑘𝑙 𝑎𝑙𝑘 𝐿

𝑙=1

) − 𝐹𝑗𝑘− 𝑇𝑖𝑗

Leonard et al. (2002) argues that viewing patients as informed consumers maximizing expected utility helps explain why the average patient travelled 2.8 kilometers further than necessary to seek healthcare and helps formulate more effective policies. Stock (1983) arrives at the same conclusion that although costs including travel and opportunity costs are certainly part of the decision-making process, expected quality of service received is the main determinant for care sought. With that being said, it is not surprising to find that there seems to be a relationship between mean quality of service received and length of travel journey (Stock, 1983). Said differently, people travel further than needed in hopes of better treatment and the numbers point towards their added travel time being worth the effort (Stock, 1983). These findings align with most of the findings on the nature of bypassing health facilities by patients as informed decision makers (Gauthier & Wane, 2011) (Leonard, Mliga, & Mariam, 2002). What’s more, it seems

(7)

5

that even relatively poor patients seem to seek better quality health facilities at higher cost (Gauthier & Wane, 2011).

Just like the quality assessment of health facilities has been subject to many studies, the relationship between distance and child mortality has been subject to much research (Kadobera et al, 2012). However, to the best of my knowledge there haven’t been many papers that have attempted to link distance, facility delivery and child mortality in a two-step approach. Lohela et al. (2012) attempt to empirically establish whether the likelihood of survival for babies is really higher at health facilities by looking at distance to faciltiy in Zambia and Malawi. They stratify their sample by faclity delivery into four strata and assume that facility delivery is a proxy for the confounding variable of complications related to pregnancy, which remains unobserved. They then seperately regress distance on facility delivery, care and neonatal mortality. Their measure for distance turns out to be a significant predictor of facility delivery for Zambia and Malawi, which confirms my expectations for the relevance of my instruments in the first stage of my IV estimation. Similar to the results obtained by Kadobera et al. (2012), they find that facility delivery was associated with lower neonatal mortality, although in their case the coefficients did not reach statistical significance.

The problem behind the Lohela et al. (2012) causal interpretations is that their stratification method implicitly assumes that low facility delivery is a symptom of complicated pregnancy cases, which I will choose not to assume in my models. The following section will seek to explain why IV methodology seems more warranted to out find out which interpretation for home deliveries is more pertinent in the case of rural Tanzania.

3. Data & Methodology

3.1. Data collection

I use DHS (Demographics and Health Survey) data from 2016 combined with SPA (Service Provision Assessment) data from 2015 to calculate straight-line distances between the village clusters and the health facilities included in the SPA data1. The DHS dataset reports

1_{My initial intention was using to calculate the actual distance traveled instead of using straight line distances as}

a proxy. To do this I used Google API’s to calculate the distance to travel by road network, only to reveal that for the Tanzanian sample, the lack of cartographical data created too many missing variables to be used.

(8)

6

geographical coordinates for each cluster (608 in total) at the cluster’s epicenter. This is the same as saying that every individual 𝑖 in cluster 𝑗 has the same geographical coordinate. This is due to the DHS program’s privacy policy which restrict household-level inference. In addition to this, the geographical coordinates are displaced from the epicenter of the cluster 𝑗 by a buffer of 0-5 km’s in 99%percent of the cases and by 5-10km’s in 1% of the cases for rural households (Perez-Heydrich et al,2013). Although the measurement error in my instrumental variables2_{undoubtedly biases the coefficient of the instruments towards zero, this error is}

random and only influences the relevance my instruments and not the interpretation thereof. The survey data includes questions pertaining to the health of the family, with some in-depth questions about maternity, female health and reproduction. The SPA data include a representative, yet non-exhaustive list of health facilities, their location and many other characteristics. Some facilities used will be missed in the analysis, yet the exclusion of facilities is another layer of random measurement error that shouldn’t bias results. In order to obtain my constructed measures, all the 608 DHS clusters have been matched with all 1496 health facilities listed in the SPA dataset. After matching the closest geographical points and calculating distances for each, I restricted my focus to the three shortest distances of health facilities to a specific cluster as has been used in previous literature (Leonard, Mliga, & Mariam, 2002). This gives each cluster 3 facilities that it is matched to. This matching process is hypothetical and a best guess, since there is no account given in the survey of the actual health facility used. The only information provided is the type of place of delivery (see Table 9 in Appendix). I construct my facility delivery variable by making it 0 in case of a delivery in the respondent’s home, other home or tradition birth attendant premise and 1 otherwise. Although the facility used for delivery might have been even further than the third closest facility, I further assume that the statistical effect of this is negligible and that the three closest facilities give a sufficiently accurate indicator of proximity to health facilities that are likely of having been used.

2_{Although the distance is certainly less precise because of the buffer, the altitude measured for the centre of the}

buffer radius should still reflect the altitude of the real location, unless there are substantial height differences in the buffer radius.

(9)

7

In addition to the distance measure, I use the difference in altitude between the three closest facilities and the village to include additional information about remoteness. The results of this matching procedure can be seen in Table 1 below.

Table 1: Descriptive statistics of the constructed instruments

Variable Obs Mean Std. Dev Min Max

Closest distance 37169 8.6446 8.2359 0.0793 55.8010

2nd closest distance 37169 14.3104 12.0778 0.0830 79.2457 3rd closest distance 37169 18.1780 13.5424 0.1276 84.6090 Altitude difference closest 37169 64.6614 119.7784 0.0000 1330.4 Altitude difference 2nd closest 37169 100.1134 172.8598 0.0000 1364.7 Altitude difference 3rd closest 37169 109.2967 166.8696 0.0000 1369.3 The dataset lists 37169 observations of which the exact place of delivery of the birth is only known for 10233 of the 37169 observations. This is because the focus of the survey was on child mortality and respondent women were only asked to report the births that occurred at most 5 years prior to the interview.

3.2. Estimation technique

I use 2SLS estimation since there might be potential confounders with facility delivery in the error term of a naïve OLS regression of facility delivery on child mortality that might bias results. In fact, Lohela et al. (2012) mentions that it is difficult to get accurate and comparable data on complications related to pregnancy as a means of adjusting to this estimation problem. Lohela et al. (2012) clearly layed out the problem or running OLS regression of facility delivery on child mortality:

“However. it can be difficult to demonstrate the beneficial impact of facility delivery on early neonatal survival due to confounding by complications during pregnancy or childbirth. In contexts

where most deliveries occur at home . those seeking care at facilities may well be complicated cases . with a higher risk of early neonatal” (Terhi J. Lohela, 2012,p.1)

To get around this problem, I use two instrumental variables, namely straight-line distance to facility and the difference in altitude as instruments for facility delivery, while controlling for obvious covariates in my first stage regression (1) seen below.

(10)

8 First stage (𝟏) 𝑓𝑎𝑐𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦 = 𝛼0+ 𝛼1∗ 𝑙𝑛𝑎𝑣𝑔𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑗+ 𝛼2∗ 𝑙𝑛𝑑𝑖𝑓𝑓𝑎𝑙𝑡1𝑗+ ∑ 𝜃ℎ∗ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖 𝐻 ℎ=1 + ∑ 𝛾𝑘𝑟𝑒𝑔𝑖𝑜𝑛𝐹𝐸𝑘 𝐾 𝑘=1 + 𝜀𝑖 Second stage (𝟐) 𝑐ℎ𝑖𝑙𝑑𝑑𝑒𝑎𝑑𝑖= 𝛽0+ 𝛽1∗ 𝑝𝑓𝑎𝑐𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑖+ ∑ 𝛿ℎ∗ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖 𝐻 ℎ=1 + ∑ 𝜌𝑘𝑟𝑒𝑔𝑖𝑜𝑛𝐹𝐸𝑘 𝐾 𝑘=1 + 𝑢𝑖

The choice of the functional form of my instruments is done by choosing the highest correlation or best fit as is shown in Table 11 and Table 12 in the Appendix. The unbiasedness of my instruments rests on (a) the relevance of the instruments in my first stage regression and (b) the exogeneity of my instruments, that is that:

𝐶𝑜𝑣(𝑙𝑛𝑎𝑣𝑔𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑖, 𝑢) = 0 and 𝐶𝑜𝑣(𝑙𝑛𝑑𝑖𝑓𝑓𝑎𝑙𝑡1𝑖,𝑢𝑖) = 0

I use 𝐻 controls at the individual level 𝑖, instruments 𝑙𝑛𝑎𝑣𝑔𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑗 and 𝑙𝑛𝑑𝑖𝑓𝑓𝑎𝑙𝑡1𝑗at the cluster

level 𝑗 and fixed regional effects at level 𝑘. The second stage regression (2) then regresses a dummy for child survivorship on the predicted values from the first stage regression for facility delivery while including the same identical controls and fixed effects as in (1).

3.3. Relevance and exogeneity

The relevance of the instruments will be tested by checking for individual and joint significance in the first stage regression. I expect my instruments to be highly significant even after controlling for other factors influencing facility delivery, as has been found in previous empirical results (Terhi J. Lohela, 2012).

However, the exogeneity of my instruments is impossible to prove. Nevertheless, I strongly believe that by construction, the possibility that my instruments are correlated with the error term 𝑢 in the second stage equation is quite limited. Since my instruments do not document the facility used for delivery, but rather a proxy for proximity of health facilities, I believe that qualitative aspects of health facilities are only remotely related to my instruments. The quality of health facilities is by construction averaged out in my instruments, so that every individual 𝑖 in cluster 𝑗 is matched to the same quality of facility. Said differently, by not using the actual distance traveled to the health facility used for delivery, I avoid problems pertaining to biased sample selection through bypassing behavior or through the adverse selection of only risky

(11)

9

deliveries in facilities. Thereby, it is less likely that after controlling for many confounders, the average distance to the three closest facilities is still related to complications during pregnancy. In other words, I strongly believe that through my choice of instruments, the exogeneity assumption is likely to hold, since I don’t think that quality of health facility used or women’s complications during pregnancy and care seeking behavior are systematically related to average distance to cluster 𝑗 after controlling for confounders. On the one hand, the constructed instruments are less precise in predicting facility delivery and thereby less relevant than the actual unknown figure of distance traveled to access facility. On the other hand, this construction limits variation between my instruments and unobservable factors affecting childhood mortality. Again, there is no problem if uncontrolled or unobservable factors influence facility delivery through distance or altitude. yet I believe that there is very little influence of average distance and altitude on childhood mortality other than through the channel of facility delivery.

In Figure 1 copied from Leonard et al. (2002). my set up assumes that my instruments 𝑙𝑛𝑎𝑣𝑔𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 and 𝑙𝑛𝑑𝑖𝑓𝑓𝑎𝑙𝑡1 only influence the outcome variable of mortality through the pathway of facility delivery after controlling for confounders as indicated in Figure 1 through the arrows with the negative sign.

(12)

10

3.3. Controls of cofounding variables

In a very influential article response to Thaddeus & Maine’s (1994) literature review of the determinants of delivery service use, Gabrysch & Campbell (2009) identify the most obvious confounders of distance and facility delivery that have been used in previous research.

They identify four distinct groups, namely sociocultural factors, economic accessibility, physiscal accessibilityand perceived need. I focus on the first three groups as controls, as the perceived need measures aren’t as easy to observe and might cause problems of “bad controls”. For instance. the inclusion of a dummy reporting antenatal care might render my instruments insignificant or biased in the first stage regression, since antenatal care is strongly correlated with facility delviery and could be considered an outcome variable in the first stage equation instead of a control.

With that being said. I drop the set of perceived need controls and add a greater emphasis on proxies for autonomy and empowerment of the pregnant woman. For those controls where the expected sign of the control variable on facility delivery is clear, I indicate it in brackets. A detailed account of the possible interpretations of some controls I use can be found in (Gabrysch & Campbell, 2009). I conclude this section with a list of controls that will be used throughout the rest of this paper.

Sociocultural factors

- Age

- Marital status (married or in partnership =1) - Husband’s age

- Education (+)

- Husband’s education (+) - Number of children born

Economic accessibility

- Standardized z-score for wealth (+)

Physiscal accessibility

- Owns a car or truck (+) - Owns a motorcycle (+) - Owns a bike (+)

- Rural/Urban (Rural = 1) (-) - Regional dummies

(13)

11

Empowerment

- Beating justified if out without permission(-) - Beating justified if wife neglects children(-) - Beating justified if wife argues with husband(-)

- Beating justified if wife refuses to have sex with husband(-) - Beating justified if wife burns the food(-)

- Person who takes health decision ( wife or husband together with wife = 1) (+)

4. Results

4.1. First stage regression

As can be seen in Table 2 below, the instruments are insignificant in the first stage specification without regional fixed effects in column (1). Moreover, they have the wrong sign, indicating that I might be omitting factors that determine facility delivery. The reason for this is the heterogeneity of facility delivery across regions in Tanzania which can be seen in further detail in Table 10 in the Appendix. Intuitively, it makes sense that the distance decay varies according to unobserved regional factors such as terrain, road networks and other unobserved regional level characteristics. In contrast to pooling all regions in specification (1), regional fixed effects (2) estimate the way that distance and altitude affect facility delivery for each region. After adding regional fixed effects to the regression, both instruments reach significance in column (2). For the most part, the sign of the controls remains unchanged, yet some controls win or lose their significance in the second specification. Most importantly, the instruments are relevant and significant individually as well as jointly significant (F=38.57>10). Although it is not essential to 2SLS, it is reassuring to see that the Adj. R squared is high at 0.2269 since more precise predictions in the first stage generate tighter standard errors for the 2SLS estimates.

(14)

12

Table 2: First stage regression with and without regional fixed effects

Exp. sign Variable OLS Fixed effects

(1) (2) - lnaverageds 0.0038 (0.52) -0.0645*** (-7.75) - lndiffalt1 0.0018 (0.5) -0.0099** (-2.51) - problemdistance -0.0635*** (-6.97) -0.0675*** (-7.49) +- age 0.0111*** (10.12) 0.0065*** (5.9) +- husband age -0.0017** (-2.49) -0.0008 (-1.21) + education 0.0165*** (11.56) 0.0158*** (11.22) + husband education 0.0018** (2.42) 0.0018** (2.49) + wealth (z-score) 0.0684*** (9.41) 0.0654*** (8.73) +- married or partner 0.0018 (0.06) -0.0077 (-0.28) +- number of children -0.0442*** (-14.89) -0.0294*** (-9.73) - rural -0.1231*** (-9.23) -0.0431*** (-2.99) + bike 0.0218** (2.42) 0.0319*** (3.49) + motorcycle 0.0279* (1.84) 0.0264* (1.79) + cartruck -0.0577** (-2.02) -0.0398 (-1.43) - beatingout -0.0423*** (-3.21) -0.019 (-1.47) - beatingneglect 0.0119 (0.89) 0.0146 (1.11) - beatingargue 0.008 (0.6) 0.0005 (0.04) - beatingsex 0.0184 (1.42) 0.0151 (1.18) - beatingfood -0.0239* (-1.78) -0.0279** (-2.11) + healthempower 0.0283*** (2.7) 0.0236** (2.26) _cons 0.5205*** (15.28) 0.8386*** (18.48) F(2. 10212) 0.38 38.57 Prob > F 0.6844 0 Number of obs = 10233 10233 Adj R-squared = 0.1785 0.2269 * p < 0.1, ** p < 0.05, *** p < 0.01. t-values in parentheses.

(15)

13

4.2. Two-stage least squares versus ordinary least squares

Since the previous Table 2 confirmed the relevance of my instruments, I proceeded in running 2SLS regressions. Column (1) and (2) in Table 3 indicate naive OLS and 2SLS estimates for the outcome variable childead12, where childdead12 restricts the dummy to children that were older than 12 months or have died before the date of the interview. Similarly, childdead24 restricts the sample to either those survivors that are older than 24 months at the date of the interview or those children that have died previously to reaching 24 months of age. This reduces the bias that could exist by including children younger than 12 or 24 months respectively at the date of interview that might still be alive but die after the date of the interview.

As can be seen in Table 3 below, the naïve OLS estimates for facility delivery have the wrong sign and are statistically indistinguishable from zero. The TSLS estimates in column (2) and (4) indicate a strong impact of the instrumented facility delivery variable on child mortality. A closer look at the magnitude of these coefficients however raises some fundamental questions. The childhood mortality figure for the whole sample was 0.042 for childdead12 and 0.055 for childdead24, yet the magnitude of the coefficient on facility delivery in column (2) and (4) is about -0.1659 and -0.2834 depending on the outcome variable used. IV regressions measure the local average treatment effect (LATE) which in this case is the effect of facility delivery on child mortality for the subpopulation of compliers, that is, the part of the population which is affected by distance in their decision to deliver at a facility. I do not pick up the effect of those pregnant women who always prefer to deliver at home or always choose the facility, regardless of the amount of time needed to travel to the facility. Nevertheless, even if we assume that childhood mortality would be higher in this subpopulation, the magnitude of the coefficient in column (2) and (4) still doesn’t make sense.

(16)

14

Table 3: Naïve OLS results versus TSLS results on childdead12 and childead24

OLS TSLS OLS TSLS

(1) (2) (3) (4)

childdead12 childdead12 childdead24 childdead24 faciltydelivery 0.0031 (0.6) -0.1659** (-2.49) 0.0051 (0.73) -0.2834*** (-2.76) age -0.0015*** (-2.66) -0.0003 (-0.42) -0.0029*** (-3.8) -0.001 (-0.98) husband age -0.0006* (-1.7) -0.0007** (-2.11) -0.0008* (-1.8) -0.0011** (-2.28) education 0.0012* (1.65) 0.0041*** (2.96) 0.0019* (1.95) 0.0068*** (3.38) husband education -0.0003 (-0.74) 0.0000 (0) -0.0005 (-0.91) 0.0000 (0.06) wealth (z-score) 0.0090** (2.34) 0.0207*** (3.36) 0.0003 (0.06) 0.0239** (2.38) married or partner 0.0088 (0.59) 0.0064 (0.44) 0.0219 (1.1) 0.0218 (1.02) number of children 0.0085*** (5.39) 0.0031 (1.12) 0.0108*** (5.21) 0.0021 (0.57) rural -0.0052 (-0.72) -0.0237** (-2.28) -0.0145 (-1.51) -0.0458*** (-2.96) bike -0.0027 (-0.56) 0.002 (0.35) -0.008 (-1.23) 0.0006 (0.07) motorcycle 0.0075 (0.95) 0.0111 (1.23) 0.0119 (1.13) 0.017 (1.33) cartruck 0.0007 (0.05) -0.0074 (-0.44) 0.0063 (0.33) -0.009 (-0.39) beatingout 0.007 (1.02) 0.0047 (0.59) 0.01 (1.08) 0.005 (0.46) beatingneglect -0.0067 (-0.97) -0.0047 (-0.61) -0.0201 (-2.16) -0.016 (-1.43) beatingargue 0.0095 (1.38) 0.0087 (1.15) 0.0073 (0.79) 0.007 (0.63) beatingsex 0.0081 (1.2) 0.0093 (1.21) -0.0018 (-0.2) -0.0002 (-0.02) beatingfood -0.0066 (-0.95) -0.0112 (-1.44) 0.0205** (2.2) 0.0123 (1.07) healthempower 0.0031 (0.57) 0.0063 (1.1) 0.0084 (1.15) 0.0133 (1.6) problemdistance 0.0007 (0.15) -0.013* (-1.75) -0.0016 (-0.25) -0.0251** (-2.24) _cons 0.0663*** (3.01) 0.181** (3.38) 0.117*** (3.94) 0.3168*** (3.79)

Region fixed effects Yes Yes Yes Yes

Observations 8178 8178 5937 5937

Adj R-squared 0.0044 0.21218 0.0075 0.25689

* p < 0.1, ** p < 0.05, *** p < 0.01. t- values in parentheses.

Note: Column (1) and (2) indicate naive OLS and 2SLS estimates for the outcome variable childead12, where childdead12 restricts the dummy to children that were older than 12 months or have died before the date of the interview. Similarly, childdead24 restricts the sample to either those survivors that are older than 24 months at the date of the interview or those children that have died previously before 24 months of age. This reduces the bias that could exist by including children younger than 12 or 24 months respectively at the date of interview that might still be alive but die after the date of the interview while still being younger than 12 or 24 months.

(17)

15

Another reason for the magnitude of the coefficient might be that the linear predictions of the binary variable facility delivery in the first stage are just too imprecise. To the best of my knowledge, the statistical software used doesn’t allow for nonlinear predictions in the first stage while still adjusting standard errors for robustness in the second stage. Since this would offer an alternative explanation for the magnitude of the coefficient, I manually run predicted values for facility delivery by linear estimation (OLS) and non-linear estimation (Logit & Probit). Logit and Probit estimates gives predications that are bounded between 0 and 1 while OLS gives values that exceed the bounds of the binary variable. Some descriptive statistics of the distributions of the predictions obtained through these different estimation techniques are tabulated in Table 4 below.

Table 4: Predictions by using linear(OLS) versus nonlinear predictions for facility delivery

Variable Observations Mean Std. Dev Minimum Maximum

facility delivery 10233 0.6329 0.4820 0 1

pprobit 10233 0.6327 0.2372 0.0518 0.9997

plogit 10233 0.6329 0.2387 0.0624 0.9979

pols 10233 0.6329 0.2315 -0.0033 1.383

Table 5 runs the second stage regression for the predicted values obtained by OLS, Probit and Logit on the two different outcome variables. It must be repeated that the magnitude of the coefficients are similar to those that would be obtained through 2SLS that the statistical software would run, yet the significance of the coefficients is overestimated in Table 5 since manually running 2SLS doesn’t allow to adjust standard errors for robustness. As expected, the magnitude of the coefficients in the nonlinear specifications in column (2),(3),(5) and (6) is smaller, although some of this effect could be attributed to the loss of significance in the coefficients. Although the coefficient in column (3) might marginally lose its significance if the regression would adjust standard errors for 2SLS, the magnitude of the coefficient -0.0825 in column (3) qualitatively offers the same LATE interpretation, yet makes more sense quantitively compared to the coefficient -0.251 in column (1).

(18)

16

Table 5: Manual 2SLS for linear and nonlinear predictions of facility delivery

tchilddead24 tchilddead12

(1) (2) (3) (4) (5) (6)

pOLS pLogit pProbit pOLS pLogit Probit

pfacdelivery -0.251*** (-3.12) -0.0713* (-1.65) -0.0825* (-1.79) -0.1571*** (-2.62) -0.0209 (-0.65) -0.0288 (-0.84) age -0.0011 (-1.22) -0.0024*** (-2.89) -0.0022*** (-2.72) -0.0004 (-0.63) -0.0014** (-2.22) -0.0013** (-2.09) husband age -0.0011** (-2.27) -0.0009* (-1.94) -0.0009** (-1.98) -0.0007** (-2.06) -0.0006* (-1.75) -0.0006* (-1.78) education 0.006*** (3.73) 0.0032* (2.63) 0.0034* (2.72) 0.0038*** (3.13) 0.0016* (1.8) 0.0018* (1.89) husband education 0 (-0.01) -0.0003 (-0.63) -0.0003 (-0.6) 0 (0.01) -0.0002 (-0.62) -0.0002 (-0.6) wealth (z-score) 0.0218** (2.58) 0.0067 (1.07) 0.0076 (1.19) 0.0224*** (3.54) 0.011** (2.35) 0.0116** (2.44) married or partner 0.0205 (1.02) 0.0218 (1.09) 0.0223 (1.12) 0.0074 (0.5) 0.0087 (0.58) 0.0088 (0.59) number of children 0.003 (0.96) 0.0085*** (3.49) 0.0081*** (3.25) 0.0037 (1.55) 0.0078*** (4.22) 0.0076*** (3.99) rural -0.0368*** (-3.1) -0.0214** (-2.07) -0.0223** (-2.14) -0.0193* (-2.17) -0.0074 (-0.96) -0.0081 (-1.04) bike -0.0006 (-0.09) -0.0057 (-0.88) -0.0055 (-0.84) 0.0019 (0.37) -0.002 (-0.41) -0.0018 (-0.37) motorcycle 0.0156 (1.48) 0.0129 (1.23) 0.013 (1.24) 0.0101 (1.28) 0.0079 (1) 0.008 (1.02) cartruck -0.0083 (-0.42) 0.0015 (0.08) 0.0005 (0.03) -0.0084 (-0.56) -0.0006 (-0.04) -0.0012 (-0.08) beatingout 0.0051 (0.55) 0.0084 (0.92) 0.0083 (0.91) 0.004 (0.58) 0.0065 (0.95) 0.0064 (0.94) beatingneglect -0.0162* (-1.72) -0.0189** (-2.03) -0.0189 (-2.02) -0.0044 (-0.63) -0.0064 (-0.92) -0.0063 (-0.91) beatingargue 0.0075 (0.81) 0.0073 (0.78) 0.0073 (0.79) 0.0097 (1.41) 0.0095 (1.38) 0.0095 (1.38) beatingsex 0.0018 (0.2) -0.0008 (-0.09) -0.0006 (-0.07) 0.0103 (1.52) 0.0084 (1.24) 0.0085 (1.26) beatingfood 0.0126 (1.31) 0.0182* (1.93) 0.0176* (1.87) -0.0116 (-1.61) -0.0074 (-1.05) -0.0077 (-1.09) healthempower 0.0143* (1.9) 0.01 (1.37) 0.0103 (1.4) 0.0068 (1.2) 0.0037 (0.66) 0.0039 (0.7) problemdistance -0.0206** (-2.38) -0.0074 (-1.04) -0.0083 (-1.15) -0.0112* (-1.73) -0.0011 (-0.21) -0.0017 (-0.31) _cons 0.2854*** (4.73) 0.1669*** (4.11) 0.1735*** (4.16) 0.1721*** (3.81) 0.0821*** (2.71) 0.0871*** (2.8) Region fixed effects

Yes Yes Yes Yes Yes Yes

Observations 5937 5937 5937 8178 8178 8178

Adj R-squared 0.0090 0.0079 0.0079 0.0052 0.0044 0.0044 * p < 0.1, ** p < 0.05, *** p < 0.01. T-values in parentheses.

Note: When running 2SLS in Stata, the standard errors in the second stage regression output are adjusted fort the imprecise predictions made in the first stage. Since the predicted observations are manufactured in stata and not observed as for example the controls, the standard errors of running a 2SLS estimation manually are too small. Although the magnitude and sign of the coefficient don’t change whether or not one runs 2SLS regressions manually, the standard errors are underestimated by running manually. That means that the significance of the pfacdelivery variable obtained by pOlS, pLogit or pProbit might become insignificant. To the best of my knowledge, there is no command allowing for 2SLS with nonlinear predictions in the first stage in stata

(19)

17

5. Robustness checks

5.1. Self-reported perception of distance as a problem

I used the variation present in the self-reported perception of distance being a problem to access care to divide my sample into two distinct groups. The group that reportedly identified distance as a problem is on average close to 5 kilometers further away and 28 meters further away in altitude (see Figure 2 and Figure 3 in Appendix), both of which warrant their response to the survey question. Generally, these women and their husbands are much less educated, poorer and have more kids than those that didn’t report a problem as can be seen in Table 6 below.

Table 6: Mean differences between the group that reported that distance was a problem to access health facilities and those that didn’t consider it a problem

Mean Std. Dev.

no problem big problem no problem big problem

averageds 11.8579 16.6406 10.6137 11.3742 diffalt1 51.5285 79.4612 95.7919 138.1764 wealth (z-score) 0.2325 -0.2794 1.0884 0.8524 education 6.17144 4.9813 3.5900 3.5146 heducation 5.7258 4.9984 6.4382 6.4609 age 29.2137 29.4894 7.1945 7.3219 number of children 3.8191 4.2821 2.5346 2.6618 beatingfood 0.1842 0.2467 0.3877 0.4312 healthempo~r 0.6223 0.5577 0.4848 0.4967 Observations 5722 4511 5722 4511

The LATE obtained by 2SLS is largest for the population that is most likely to respond to a change in treatment, which in my case answers the question of whether distance and altitude make a difference in the decision to access a health facility. We would therefore expect the problematic grouping to exhibit higher and more significant coefficients, yet we find the opposite in Table 7. As puzzling as it may seem, Table 6 showed that the average distance was 5 kilometers longer for the problematic group. One possible explanation for this could be that the problematic group lives so far away, that the impact of the reduction of distance or difference in altitude marginally makes a bigger difference for those that are not affected by distance and live closer. This interpretation would be in line with the literature on distance decay which postulates that there is a bigger benefit to reducing distance to health facility if the distance is small than if it is prohibitively big (Stock, 1983).

(20)

18

Only the coefficients in column (2) and (4) reach significance, yet this can be attributed to the reduced number of observations obtained by truncating the sample. Nevertheless, the coefficients have the expected sign for each of the specifications. The magnitude of the coefficients is again overestimated, yet the cause of this is the same as the results for the whole sample in Table 3.

Table 7: 2SLS regressions for the grouping that reported that distance was a problem to access health facilities and those that didn’t consider it a problem

tchilddead12 tchilddead24

(1) (2) (3) (4)

Problem distance No problem Problem distance No problem facility delivery -0.1205 (-1.43) -0.1924* (-1.79) -0.153 (-1.27) -0.3573** (-2.13) age 0.0006 (0.53) -0.0013 (-1.26) -0.0011 (-0.77) -0.0012 (-0.76) husband age -0.0012*** (-2.77) -0.0001 (-0.2) -0.0018*** (-3.03) -0.0001 (-0.2) education 0.0035* (1.75) 0.0039** (2.14) 0.0037 (1.39) 0.0082*** (2.91) husband education -0.0003 (-0.84) 0.0002 (0.46) -0.0004 (-0.79) 0.0004 (0.49) wealth (z-score) 0.0179** (1.99) 0.0234** (2.51) 0.0119 (1) 0.0298* (1.82) married or partner 0.0168 (0.83) -0.0066 (-0.3) 0.0586** (2.07) -0.0232 (-0.7) number of children 0.0017 (0.45) 0.0041 (1.03) 0.0047 (1.02) 0.0002 (0.03) rural -0.0273 (-1.42) -0.0198 (-1.6) -0.0462* (-1.78) -0.0417** (-2.18) bike 0.0077 (0.97) -0.0036 (-0.49) 0.0076 (0.73) -0.0058 (-0.52) motorcycle 0.0061 (0.41) 0.0127 (1.1) 0.0161 (0.8) 0.0153 (0.93) cartruck -0.0093 (-0.26) -0.0032 (-0.17) 0.0008 (0.02) -0.0099 (-0.39) beatingout -0.0107 (-1.02) 0.0216 (1.85) -0.0086 (-0.59) 0.0248 (1.46) beatingneglect 0.0098 (0.89) -0.0183 (-1.78) -0.0112 (-0.76) -0.0265* (-1.71) beatingargue 0.0141 (1.3) 0.004 (0.38) 0.0162 (1.06) 0.0018 (0.12) beatingsex -0.0025 (-0.23) 0.0189* (1.76) -0.0106 (-0.74) 0.0082 (0.53) beatingfood -0.0089 (-0.84) -0.013 (-1.15) 0.0217 (1.46) 0.0089 (0.52) healthempower 0.01 (1.28) 0.0007 (0.08) 0.0152 (1.42) 0.0099 (0.76) _cons 0.1074* (1.87) 0.2434** (2.56) 0.1752** (2.08) 0.4264*** (2.92) Observations 3645 4533 2652 3285 * p < 0.1, ** p < 0.05, *** p < 0.01. T-values in parentheses.

(21)

19

5.2. Cluster-level estimates

Some of the high magnitude of the coefficients obtained could also be explained by the difficulty in predicting a binary variable not only in the first stage, but also in the second stage. Therefore, I collapse my data to the cluster lever and run the first stage regression with linear and nonlinear predictions. The regression in the second stage (3) seen below is the same as before, yet calculated for each of the 𝐽 =608 clusters instead of the individual 𝑖.

(𝟑) 𝑐ℎ𝑖𝑙𝑑𝑚𝑜𝑟𝑡𝑎𝑙𝑖𝑡𝑦𝑗= 𝛽0+ 𝛽1∗ 𝑝𝑓𝑎𝑐𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑗+ ∑ 𝛽ℎ∗ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑗 𝐻 ℎ=3 + ∑ 𝜌𝑘𝑟𝑒𝑔𝑖𝑜𝑛𝐹𝐸𝑘 𝐾 𝑘=1 + 𝑢𝑗

After looking at Table 8, the results are qualitatively the same as for the whole sample with the naïve OLS regression revealing an insignificant coefficient on facility delivery and 2SLS yielding coefficients with the expected sign and significant coefficients. Again, the magnitude of the coefficient becomes smaller when I use nonlinear predictions in the first stage. Although the econometrics of using such a nonlinear two-step procedure are not straight-forward, the intuition behind these results is the same as the for the results obtained through regular 2SLS with linear predictions.

Table 8: Naïve OLS, 2SLS and manual 2SLS estimates for the cluster level

Naive OLS 2sls pOLS pLogit pProbit

(1) (2) (3) (4) (5) facility delivery -0.0103 (-0.65) -0.2518* (-1.7) -0.1918** (-2.49) -0.085* (-1.76) -0.095* (-1.87) age 0.0003 (0.14) 0.0039 (1.18) 0.0018 (0.88) 0.0009 (0.46) 0.001 (0.51) husband age -0.0005 (-0.34) -0.0012 (-0.76) -0.0008 (-0.56) -0.0006 (-0.42) -0.0006 (-0.44) education 0.003 (1.29) 0.0108* (1.89) 0.0054** (2.14) 0.0043* (1.73) 0.0044* (1.79) husband education -0.0009 (-0.6) 0.0013 (0.71) -0.0006 (-0.39) -0.0006 (-0.44) -0.0006 (-0.43) wealth (z-score) 0.0031 (0.41) 0.0148 (1.33) 0.0216** (2,02) 0,009 (1,08) 0,0099 (1,18) married or partner -0,0172 (-0,31) -0,0162 (-0,25) -0,0161 (-0,29) -0,0183 (-0,33) -0,0178 (-0,32) number of children 0,0125** (2,29) 0,0027 (0,33) 0,0063 (1,04) 0,0098* (1,73) 0,0094* (1,65) rural -0,0127 (-1,29) -0,0212* (-1,85) -0,0246** (-2,25) -0,018* (-1,75) -0,0185* (-1,79) beatingout 0,0021 (0,08) -0,0038 (-0,14) -0,0017 (-0,07) -0,0017 (-0,07) -0,0021 (-0,08) beatingneglect 0,0113 (0,44) 0,0205 (0,63) 0,0137 (0,54) 0,0127 (0,5) 0,0129 (0,5)

(22)

20 beatingargue -0,0116 (-0,44) 0,0074 (0,25) -0,0065 (-0,25) -0,0095 (-0,36) -0,0092 (-0,35) beatingsex 0,051** (2,08) 0,0377 (1,11) 0,0497** (2,04) 0,0524** (2,14) 0,0524 (2,15) beatingfood -0,0185 (-0,71) -0,0479 (-1,34) -0,028 (-1,06) -0,0213 (-0,81) -0,022 (-0,84) healthempower -0,0232 (-1,23) -0,0175 (-0,8) -0,0201 (-1,07) -0,0213 (-1,13) -0,0212 (-1,13) bike -0,0152 (-1,14) -0,0006 (-0,03) -0,0103 (-0,77) -0,0134 (-1) -0,0132 (-0,98) motorcycle -0,011 (-0,47) 0,0151 (0,54) -0,0107 (-0,46) -0,0088 (-0,38) -0,0087 (-0,37) cartruck 0,0019 (0,06) -0,0422 (-0,92) -0,0102 (-0,31) -0,0076 (-0,23) -0,0088 (-0,27) problemdistance -0,0355*** (-2,85) -0,08** (-2,5) -0,0479*** (-3,59) -0,0405*** (-3,19) -0,0414*** (-3,24) _cons 0,0666 (1,22) 0,1611* (1,74) 0,1837** (2,52) 0,1128* (1,84) 0,1185* (1,91) Outcome variable Childdead12 Childdead12 Childdead12 Childdead12 Childdead12

Observations 608 608 608 608 608

Adjusted R squared 0,0304 0 0,0403 0,035 0,0357

6. Discussion

Even though the exact magnitude of the coefficients in my previous section are hard to interpret, facility delivery is consistently and significantly negatively related with childhood mortality. Put differently, after controlling for obvious confounders, the mere fact of attending a facility during delivery increases chances of survival of the baby. In each of the specifications, the quality of care received during delivery at the facility is superior to the quality at home as measured by survivorship of the child. These results align with the hypothesis of the misinformed healthcare consumer, which has incomplete health knowledge of the benefits of facility delivery. This however does not mean that the whole story is about the lack of information. There are certainly financial, physical and sociocultural constraints that make it prohibitively difficult for a pregnant woman to realize the gains of delivering a baby in a facility, yet the evidence clearly points towards a record of better survivorship for the average woman in the sample delivering in a facility. To which degree this information is known and how this knowledge is incorporated in their decision making cannot be observed. What’s more, the valuation of the life of the baby and effort exerted to guarantee a slightly higher chance of survival almost certainly varies from woman to woman.

Conceptually, there are two alternative explanation to why the potential benefits of facility delivery are not realized and women still deliver at home. On the one hand, when using the

(23)

21

Leonard et al. (2002) model for the expected utility of care and incorporating the finding that survival likelihood is higher in facilities deliveries than at home, this would mean that the well-informed patient would value the added benefit of survivorship of the child 𝑊_𝑖(∏_𝐿 𝑅_𝑗𝑘𝑙 𝑎𝑙𝑘

𝑙=1 ) less than the treatment expenses 𝐹𝑗𝑘 and travel expenses 𝑇𝑖𝑗 incurred so that

this results in ∆𝐸𝑈 ≤ 0 and the baby is delivered at home. In other words, if we assume that patients respond rationally to an increase in expected utility by delivering the baby at a health facility, the gain realized through increased survival chances must be less than the total costs incurred. On the other hand, incomplete knowledge of (∏𝐿_𝑙=1𝑅_𝑗𝑘𝑙 𝑎𝑙𝑘) would simply underestimate the value of facility delivery and compel women to deliver at home through incomplete information.

I would personally side with the literature on misinformed consumers of healthcare since the latter interpretation offers more plausible reasons for home deliveries in my view. Furthermore, policies aimed at increasing facility deliveries have more possible directions. According to this model, reductions of 𝐹𝑗𝑘through healthcare reform or reductions in 𝑇𝑖𝑗 through the building of

roads or hospitals should be accompanied with information campaigns raising the valuation 𝑊_𝑖(∏_𝐿 𝑅_𝑗𝑘𝑙 𝑎𝑙𝑘

𝑙=1 ) of facility deliveries.

Due to the high absenteeism rates in health facilities and the compromised quality that is often found, prospective patients and pregnant mothers have certainly revised their expectations downward, yet information campaigns in schools should stress that even though the provision of services in health facilities is often far from perfect, the child still has a higher chance of surviving. Jensen (2010) reported impressive gains from schooling by just informing pupils of current differences in average earnings between high-school graduates and non-graduates. Similarly, I believe that informing women about the actual differences in child mortality at home compared to at a facility could have a potential impact at very low cost. What’s more, if perceptions don’t change fast enough, the benefits of increasing the mere quality ∏_𝐿 𝑅_𝑗𝑘𝑙 𝑎𝑙𝑘

𝑙=1

could have very limited benefits if this information is not fully integrated in household decision making. Nonetheless, the difficulty remains in presenting an order of magnitude for the benefit of delivering at a facility that is not as hard to interpret as the results obtained through the 2SLS coefficients presented in this paper, yet doesn’t suffer from omitted variable bias of a naïve OLS regression.

(24)

22

7. Conclusion

This paper has attempted to determine whether giving birth to a baby in a health facility in Tanzania makes a difference for the child’s survival. Using 2SLS estimation, I estimated the effect of facility delivery on whether the child has survived or not. The coefficient on facility delivery was consistently statistically significant and negatively related to child mortality. Although a precise order of magnitude could not be established due to imprecise predictions, the sign and significance are robust to changes in outcome variables and estimation technique. This result stands in stark contrast to the naïve ordinary least squares estimates that underestimate the effect of facility delivery due to adverse selection of risky cases at health facilities.

My results offer corroborating evidence for the general view of imperfectly informed pregnancies, since the gains from reduced child mortality in facility deliveries are not fully utilized. In my view, a better informed pregnant woman and her household could take better decisions concerning her and her babies’ health if information about improved survival chances of facility delivery were regularly presented to her. Further research could attempt to find a precise order of magnitude for the effect studied, while still controlling for the adverse selection of risky pregnancies happening.

(25)

23

8. References

Chaudhury, N., Hammer, J., Kremer, M., Muralidharan, K., & Rogers, F. H. (2006). Missing in Action: Teacher and Health Worker Absence in Developing Countries. The Journal

of Economic Perspectives, 91-116a.

Gabrysch, S., & Campbell, O. M. (2009). Still too far to walk: Literature review of the determinants of delivery service use. BMC Pregnancy and Childbirth.

Gage, A. J. (2007). Barriers to the utilization of maternal health care in rural Mali. Social

Science & Medicine, 1666-1682.

Gauthier, B., & Wane, W. (2011). Bypassing health providers: The quest for better price and quality of health care in Chad. Social Science & Medicine, 540-549.

Jensen, R. (2010). The perceived returns to educaiton and the demand for schooling. Journal

of Economic Perspectives, 515-548.

Kadobera, D., Sartorius, B., Masanja, H., Mathew, A., & Waiswa, P. (2012). The effect of distance to formal health facility on childhood mortality in Tanzania 2005-2007. Global

Health Action, 2012.

Klemick, H., Leonard, K. L., & Masatu, M. C. (2009). Defining Access to Health Care: Evidence on the importance of quality and distance in rural Tanzania. Economics,

American Journal of Agricultural, 347-358.

Leonard, K. L., Mliga, G. R., & Mariam, D. H. (2002). Bypassing Health Centres in Tanzania: Revealed Preferences for Quality. Journal of African Economies, 441-471.

Perez-Heydrich, Carolina, Joshua L. Warren, Clara R. Burgert, & Michael E. Emch. (2013)

. Guidelines on the Use of DHS GPS Data. Spatial Analysis Reports No. 8.

Phoxay, C., Okumura, J., & Nakamura, Y. (2001). Influence of Women's Knowledge on Maternal Health Care Utilization in Southern Laos. Asia Pacific Journal of Public

Health, 13-19.

Stock, R. (1983). Distance and the utilization of health facilities in rural Nigeria. Social Science

Medicine.

Terhi J. Lohela, O. M. (2012). Distance to Care, Facility Delivery and Early Neonatal Mortality in Malawi and Zambia. PLOS ONE. doi:https://doi.org/10.1371/journal.pone.0052110 Thaddeus, S., & Maine, D. (1994). Too far to walk: Maternal mortality in context. Social

Science & Medicine, 1091-1110.

van de Walle, D., & Cratty, D. (2002). Impact Evaluation of a Rural Road Rehabilitation

Project. World Bank. Siche gaangen

(26)

24

9. Appendix

Table 9: Descriptive statistics on place of delivery

Place of delivery Freq. Percent Cum. Percent

Respondent's home 3,118 30.47 30.47

other home 549 5.36 35.84

tba premises 89 0.87 36.70

national/zonal referral/ spec. hospital 386 3.77 40.48 regional referral hospital(government/p 485 4.74 45.22 regional hospital(government/parastatal 393 3.84 49.06 district hospital(government/parastatal 1,326 12.96 62.02 health center(government/parastatal) 1,242 12.14 74.15 dispensary(government/parastatal) 1,418 13.86 88.01

clinic (government/parastatal) 6 0.06 88.07

referal/ spec. hospital (religious/volu 51 0.50 88.57 district hospital (religious/voluntary) 316 3.09 91.65

hospital (religious/voluntary) 261 2.55 94.21

health center (religious/voluntary) 159 1.55 95.76

dispensary (religious/voluntary) 72 0.70 96.46

specialized hospital (private) 16 0.16 96.62

hospital (private) 93 0.91 97.53

health centre (private) 45 0.44 97.97

dispensary(private) 52 0.51 98.48

clinic (private) 7 0.07 98.54

other 149 1.46 100.00

(27)

25 Table 10: Heterogeneity of home delivery by region

Region Home delivery Total

dodoma 0.31 258 arusha 0.43 291 kilimanjaro 0.09 176 tanga 0.38 293 morogoro 0.21 248 pwani 0.18 239 dar es salaam 0.06 418 lindi 0.19 222 mtwara 0.15 181 ruvuma 0.11 262 iringa 0.06 225 mbeya 0.29 242 singida 0.36 358 tabora 0.44 535 rukwa 0.34 401 kigoma 0.48 431 shinyanga 0.38 490 kagera 0.51 368 mwanza 0.43 395 mara 0.46 494 manyara 0.48 371 njombe 0.11 233 katavi 0.55 486 simiyu 0.58 620 geita 0.50 502 kaskazini unguja 0.46 289 kusini unguja 0.24 256 mjini magharibi 0.14 325 kaskazini pemba 0.53 316 kusini pemba 0.51 308 Total 0.37 10.233

(28)

26

Table 11: Correlations of constructed distance measures with facility delivery

Correlogram Facility delivery

Facility delivery 1 First distance -0.1778 First distance^2 -0.1199 Ln(First distance+1) -0.2162 Exp(First distance*-1) 0.1892 Average distance -0.2002 Average distance^2 -0.147 Ln(Average distance+1) -0.2247 Exp(Average distance*-1) 0.1618

Table 12: Correlations of constructed difference in altitude measures with facility delivery

Correlogram Facility delivery

Facility delivery 1

First Difference altitude -0.0942 ln( First Difference altitude+1) -0.101 exp( First Difference altitude*-1) 0.0277 Average difference altitude -0.069 ln( Average difference altitude+1) -0.0989 exp( Average difference altitude*-1) 0.0592

(29)

27

Figure 2: Differences in average distance to the three closest facilities by self-reported problematic impact of distance on accessing health facilities

Figure 3: Differences in diffalt1 to the three closest facilities by self-reported problematic impact of distance on accessing health facilities

The impact of health facility delivery on child mortality in Tanzania : an instrumental variable approach