Health Systems for Outcomes Publication 53122 Health Worker Preferences for Job Attributes in Ethiopia: Results from a Discrete Choice Experiment Kara Hanson and William Jack April 2008 Health worker preferences for job attributes in Ethiopia: Results from a discrete choice experiment12 Kara Hanson London School of Hygiene and Tropical Medicine William Jack Georgetown University April 23, 2008 1 This is a preliminary draft, not for citation. 2 The survey for this study was ...nanced by a grant from the Bill and Melinda Gates Foundation and Norad, administered through the World Bank. We thank Dr Tedros, Dr Kebede and Dr Nejmudin of the Ministry of Health, and the Government of Ethiopia. We are grateful to Dr. Aklilu Kidanu for providing invaluable assistance to the authors, and for leading the survey team of the Miz-Hasab Reserach Center in Addis Ababa. Thanks also to Joost de Laat, Christopher H. Herbst, Magnus Lindlow, Gebreselassie Okubagzhi, Pieter Serneels, Agnes Soucat, and Kate Tulenko for comments and assistance. The opinions expressed in the paper are those of the ect authors and do not re the position of the Government of Ethiopia or the World Bank Group. Abstract This paper estimates the e¤ectiveness of a range of policy interventions aimed at improving the supply of health workers to rural areas in Ethiopia. Using data from a survey of 861 health workers, it employs stated preference techniques to predict labor market responses of doctors and nurses to changes in rural wages, working conditions, housing bene...ts, and training opportunities. Doubling wages in areas outside the capital would increase the share of doctors willing to work there from about 7 percent to more than 50 percent. Providing high quality housing would increase physician labor supply to about 27 percent, which is equivalent to paying a wage bonus of about 46 percent. Doubling wages paid to nurses for work in rural areas outside cities increases their labor supply from 4 percent to 27 percent, while the non-wage attribute that is most e¤ective in inducing them to relocate to rural areas is the quality of equipment and drugs. The same impact could be achieved by increasing rural nursing wages by about 57 percent for men and 69 percent for women. 1 Introduction The supply and geographic distribution of health workers are major constraints to improving health in low-income countries. A number of recent studies have highlighted the shortage of skilled health workers in many settings (WHO, 2006), the impact this has on health outcomes (Anand and Barnighausen, 2004), and the risk this poses for the achievement of the Millenium Development Goals (WHO, 2006; Joint Learning Initiative, 2004). However, there remains limited evidence about what sorts of policies will attract nurses and doctors into medical training, improve the retention of trained health workers, and encourage them to work in rural areas where problems of inaccessibility of services are most acute. A number of strategies have been employed to address the human resources challenge in low- and middle-income settings: these include compulsory rural service for new graduates (e.g. in South Africa); payment of incentives or "hard- ship allowances" for those posted to rural areas (as proposed in Rwanda); or encouraging self-selection by those with a commitment to rural service (as prac- ticed, for example, in Thailand, Wilbulpolpraser and Pengpaibon, 2003). Yet few of these strategies have been systematically evaluated, and the e¤ectiveness of each will likely re ect country-speci...c labour market conditions, political systems, and culture and tradition. The challenges of human development are particularly extreme in Ethiopia, a country with a population of over 70 million people, 85% of whom live in rural areas. It is one of the poorest countries in the world, with per capita income of about $150, and although the poverty rate has fallen by 8 percentage points over the last 10 years, it nonetheless remained at 37% in 2006. The country faces acute challenges in reaching all of the Millenium Development Goals, including the three goals relating to health - to reduce child mortality, improve maternal health, and combat HIV/AIDS, malaria, and other diseases. In 2005 the infant mortality rate was 77 per 1,000, the under-5 mortality rate was 123 per 1,000, and the maternal mortality rate was 673 per 100,000. In 2006 about half of all mothers received some kind of antenatal service, and 15% of deliveries were attended by a health worker. Ethiopia has escaped the ravages of HIV/AIDS compared with other countries in Africa, and had an adult prevalence of 2.1% in 2006. The WHO reports that in 2003 there were 1,936 physicians in Ethiopia, rep- resenting a population-physician ratio of approximately 38,000, or 0.03 physi- cians per thousand individuals.1 This is the ...fth highest population to physi- cian ratio among African countries, and compares pitifully with the ratio of 10,000 as recommended by the WHO. If anything near this ratio is to be at- tained, there will clearly need to be a sustained long term increase in the net supply of physicians to the Ethiopian market. The shortage of nurses is less acute,2 but similar expansions will be necessary. The recruitment and retention 1 See www.who.int/research/. 2 TheWHO estimated that there were 14,270 nurses in Ethiopia in 2003, or about one nurse for every 5,250 people. 1 of health workers in both the public and private sectors of the local market depends on the ...nancial and non-fanancial rewards that they expect to reap, relative to alternatives (such as non-health sector work, or migration). On the other hand, the allocation of a given supply of health workers across geographic regions, as well as to tasks and specialties, often depends on the relative rewards in the public and private sectors. Attracting health workers to remote areas is a particular challenge if the WHO-recommended ratios are to be met in a meaningful way. Rural and remote areas of Ethiopia are particularly underserved by health workers. We do not have nationally representative data on health worker lo- cation by rural and urban areas, but Ministry of Health data indicate that in 2004 about 20% of the approximately 1,000 doctors classi...ed as operating in the "public sector" worked in Addis Ababa, home to about 5% of the popula- tion (Ministry of Health, 2005). It is likely, of course, that physicians in other regions are also located in urban centers, so the share of public sector doctors in rural areas would be far less than 80%. To add further to this rural-urban disparity, Ministry of Health data suggest that a further 1,500 doctors work for NGOs, other governmental organizations (e.g., the military), the "central" government, and in the private sector. We do not have speci...c data on their location, but we believe that few of them serve the rural population. By some estimates, half of the physicians in Ethiopia serve the residents of the capital, Addis Ababa. During the 1980s and most of the 1990s, health workers trained in Ethiopia were typically assigned to their ...rst jobs by the central government. This assignment was by means of a lottery system, and the prevailing belief among o¢ cials we interviewed is that the control exercised by the state was such that most health workers accepted their assignments. Workers were required to serve a ...xed number of years before being "released" and permitted to apply for other positions. During the past ...ve years Ethiopia has embarked on a radical decentralization program across all areas of the public sector, with much of the responsibility for service delivery being devolved to lower levels of government. In each of the 10 regions, plus Addis Ababa, a regional health bureau has responsibility for the hiring and deployment of public sector health workers. As competition between regions for health workers has grown, anecdotal evidence suggests that the lottery system has become increasingly ine¤ective. It is believed that many new graduates do not register for the lottery, and those who do participate are prone to disregard their assignment if they so wish, knowing that they can apply directly to the regional health bureau for a job. The regions compete on salaries, time to release (which allows work in the burgeoning private sector, at least in Addis Ababa), and other attributes. Some regions, for example Oromia, have recently introduced explicit ...nancial incentives to attract individuals to remote areas within the region. Similarly, regions that are themselves remote (in terms of being far from Addis Ababa) have attempted to attract health workers by providing certain training options and ...nancial incentives. In this paper we estimate health worker preferences over di¤erent job at- 2 tributes in an attempt to identify the factors that are important to health work- ers in in uencing their labor supply decisions. One approach to this problem is to study actual choices made by health workers. However, this method may su¤er from a range of selection and endogeneity problems, leading to biased parameter estimates. In addition, there is often limited variation in key job attributes (pay and non-pay), making it challenging to estimate the e¤ects of these parameters on labour market choices and to predict the e¤ects of changes in job attributes that lie outside the existing range over which these attributes vary in practice. Of course, the obvious downside to this approach is that we might have more con...dence in choices people actually make, not in the choices they might say they would make. Stated preference techniques have been widely used in health and environ- mental economics applications to study preferences for non-marketed commodi- ties. Discrete choice experiments (DCEs) have examined the valuation of dif- ferent attributes of health care service provision , dimensions of health bene...t beyond health outcomes, and quality of care attributes (for reviews see Ryan and Gerrard 2003a and 2003b). Studies of health worker valuations of job at- tributes have also adopted DCE methods, both in the UK (Scott 2001) and in a variety of low- and middle-income settings (Chomitz et al., 1998; Mangham and Hanson, 2007; Penn-Kekana et al., 2005). Chomitz et al. provide a useful review of the bene...ts and shortcomings of the approach. A study of recent medical and nursing school grauduates currently underway in Ethiopia (Lind- low and Serneels) elicits direct measures of the cost of taking a rural job, but does not employ the DCE technique. 2 Data and DCE methodology In this section we report out sampling strategy and the details of the discrete choice experiment we conducted. 2.1 Sampling Our sampling strategy aimed at obtaining representative samples of doctors and s nurses from three of Ethiopia' eleven regions ­the capital city of Addis Ababa, Tigray, and Southern Nations and Nationalities Peoples Republic (SNNPR). Addis is a city of about 3 million people and is located in the central highlands. Tigray has a population of about 4 million people and lies in the extreme north of the country, bordering Eritrea, while SNNPR, with a population of 14 million borders Kenya to the south. Our sample is representative within these geo- graphic areas.3 The design over-sampled doctors in SNNPR and Tigray due to the small number of doctors outside Addis Ababa: all doctors in these rural 3 Other regions, such as Oromia (which surrounds Addis Ababa) and Amhara (which is immediately north of Oromia) are larger (with 26 and 19 million residents respectively) and less remote, at least in terms of direct distance measures, but we have no reason to expect this to have introduced systematic biases in our estimates. 3 Addis Ababa SNNPR Tigray Total Facilities 40 39 18 97 Hospitals 6 12 11 29 Health centers and clinics 34 27 7 68 Health workers 362 206 293 861 Doctors 91 72 56 219 Nurses 271 221 150 642 Table 1: Facilities and health workers surveyed in Addis Ababa regions were sampled, while only about one third of doctors in Addis were. Our ...nal sample included 219 doctors and 645 nurses working in health centers and hospitals. A random sample of 1/3 of doctors was achieved in Addis Ababa by (a) randomly sampling facilities of the various types with sampling weights corre- sponding to the estimated proportion of doctors working across the di¤erent facilities; and (b) interviewing all doctors at the sampled facilities. In SNNPR and Tigray, all doctors were included in the sample. This was achieved by sampling all public hospitals in SNNPR and Tigray (there are generally no doc- tors in non-hospital health facilities in these regions and there were no private hospitals). A random sample of approximately 1/6 of all nurses was achieved in Addis Ababa by having the enumerators randomly select half of all nurses at the sam- pled facilities. In SNNPR and Tigray we (a) randomly selected 1/6 of all nurses working in government hospitals; (b) randomly selected 1/6 of the sub-regional districts or woredas which have a hospital, visited all health centers in these woredas, and interviewed all nurses in these health centers; and (c) randomly selecting 1/6 of the woredas without hospitals, visited all health centers in these woredas, and interviewed all nurses in these health centers. Although for logis- tical and budget reasons (to minimize transport costs) the sample was selected using a cluster approach (with the facility as the cluster), as there is no strong reason to expect health worker preferences within a facility to be correlated we have not adjusted for clustering in the analysis. A summary of our sample is provided in Table 1. Amongst doctors, the interview response rate varied widely across regions. In Tigray it was very high (88%), while in SNNPR and Addis Ababa it was lower ­ 58% and 66% respectively. In Addis, the response rates di¤ered in public and private facilities. At public facilities, all doctors present agreed to be interviewed, although 40% of sampled doctors were absent on the day of the interview (28% for unexplained reasons, and 12% for planned leave). However at private facilities, no unexplained absences were recorded, while 18% of doctors were absent on planned leave. In contrast to public facilities, the share of sampled doctors who were present but refused to be interviewed was 27%. In Tigray, non-response arose because one sampled facility no longer existed, and 4 one was inaccessible for security reasons, but at visited facilities absenteeism and refusal rates were very low. In SNNPR, 42% of doctors listed as being employed were absent at the time of the facility visit, although nine out of ten of them were reported as being absent for training purposes. For nurses, we do not have data on refusals to be interviewed, but we have calculated response rates as the ratio of the numbers of nurses interviewed to our initial target sample.4 These calculated rates varied by region: in Tigray, nurses at both hospitals and clinics appear to have been over-sampled, leading to an interview rate of 143% (i.e. 43% more nurses were interviewed than initially targeted), while in SNNPR about 70% of the target number were interviewed. Most of the under-sampling seems to have occurred at health centers, which may have been under-sta¤ed compared with our pre-survey estimates. In Addis there was a small degree of over-sampling ­ the sampling protocol appears to have been followed in hosptials (where 50 percent of nurses in sampled facilities were to be interviewed), with slight over-sampling in health centers. 2.2 The discrete choice experiment Each health worker interviewed was presented with a questionnaire with two s modules, the ...rst of which solicited factual data on the worker' circumstances, incomes, household characteristics, etc., and the second of which contained a series of hypothetical choices that the respondent was asked to make. The second module provides the underlying data for our discrete choice analysis. We characterized a job in the public sector by discrete values of each of six attributes. These attributes were chosen based on their perceived relevance to health worker decisions in Ethiopia, following discussions with o¢ cials from the Federal Ministry of Health and the heads of regional health bureaux in Addis Ababa, Mekele (the capital of Tigray) and Awasa (the capital of SNNPR). The choice of attributes was also informed by focus group discussions undertaken as part of a similar study in another low-income country in sub-Saharan Africa, Malawi (Mangham and Hanson, 2007). The attributes chosen are shown in Table 2. The attribute values or levels were chosen both to be realistic, and to provide a wide enough range of variation to enable predictions about relatively large policy changes to be made. The values of the location attribute di¤ered for doctors and nurses. In practice, very few doctors work outside towns, so for them we allowed the location attribute to be either "Addis Ababa" or "Regional Capital". For nurses however, this attribute took on the values "City" and "Rural".5 At the time of the study, public sector health workers were paid on the basis of a pay scale based on experience, quali...cations, etc. We used the 4 The total number of nurses interviewed in each region was determined by budgetary constraints. Following this, and based on pre-survey estimates of the number of nurses working at each facility, the data...rm was provided with an estimated proportion of nurses to be interviewed at the facilities (but could revise this in the ...eld if the pre-survey estimates did not match the actual size of the regional population of nurses). 5 A full description of the instructions given to respondents is in the appendix. 5 Doctors Attribute Possible levels 1 X Location Addis Ababa vs Regional Capital X2 Net Monthly Pay (Base = 2; 500) 1 Base; 1:5 Base; 2 Base X3 Housing None, Basic, Superior X4 Equipment and Drugs Inadequate vs Improved X5 Time Commitment 2 years vs 1 year X6 Private Sector Yes vs No Nurses Attribute Possible levels X1 Location City vs Rural X2 Net Monthly Pay (Base = 1; 250) Base; 1:5 Base; 2 Base X3 Housing None, Basic, Superior X4 Equipment and Drugs Inadequate vs Improved X5 Time Commitment 2 years vs 1 year X6 Supervision High vs Low Table 2: Job attributes and levels (unweighted) average monthly salary from these scales to determine a "base" salary for doctors and nurses separately, and let the pay attribute take on values each to 1, 1.5, and 2 times this value. The third (housing), fourth (equipment and drugs), and ...fth (time6 ) attributes in Table 2 took on the same values for doctors and nurses. For doctors, the ...nal attribute was permission to work in the private sector (taking the values yes and no). Since opportunities for providing nursing services outside regular hours are limited, the opportunity to work in the private sector is of limited use for nurses. However, experience from other countries has suggested that active and supportive supervision is an important job attribute for these health workers. This is the sixth attribute we included for nurses. Our questionnaire presented individuals with a series of pairs of jobs, and asked them to choose the one they preferred from each pair. There are in principle 144 (= 2 3 3 2 2 2) distinct jobs charactized by the 6 attributes, and hence 20,592 (= 144 143) distinct pairs. However, using SPSS software, we generated a main e¤ects fractional factorial design with just 16 job scenarios. These jobs are shown in Table 3. This design satis...es the criteria of orthogonality, minimum overlap and level balance (Huber and Zwerina, 1996). To simplify the cognitive task for respondents we elected to use a questionnaire with one job with "middling"attributes as a constant comparator and paired the remaining scenarios to it, giving 15 choices altogether for each respondent. This number of choices is consistent with practice in the health economics literature (Ryan and Gerard, the AHEHP paper). 6 Time refers to the number of years that an individual is required to work at an institution per year of further training sponsored by that institution, after the training is completed. 6 Equipment Pay-back Private sector/ Location Pay Housing and drugs time Supervision Job 1 Addis 1.5 Basic Inadequate 1 Yes/High Job 2 Addis 1.5 Superior Inadequate 2 Yes/High Job 3 Rural 1 Superior Improved 2 Yes/High Job 4 Rural 1 Basic Improved 1 Yes/High Job 5 Addis 1 None Improved 1 Yes/High Job 6 Rural 1.5 None Improved 2 No/Low Job 7 Rural 1.5 None Improved 1 No/Low Job 8 Addis 1 None Inadequate 2 No/Low Job 9 Rural 2 None Inadequate 2 Yes/High Job 10 Addis 2 Superior Improved 1 No/Low Job 11 Rural 1 Superior Inadequate 1 No/Low Job 12 Addis 1 None Improved 2 Yes/High Job 13 Rural 1 Basic Inadequate 2 No/Low Job 14 Addis 2 Basic Improved 2 No/Low Job 15 Rural 2 Basic Inadequate 1 No/Low Job 16 Addis 1 Basic Inadquate 1 Yes/High Table 3: The constant job, Job 1, and the 15 comparator jobs To examine whether the placement or ordering of scenarios a¤ected responses we administered the questionnaire in four formats. In two of these, the "con- stant" job was ...rst (on the left-hand-side) in each of the 15 choices, and in the other 2 it was the second (on the right-hand-side). Similarly, in two formats, the series of 15 pairs were presented in the reverse order. While we did not identify any major e¤ects of the questionnaire version on results, it did allow us to identify a small number of errors in the ...nal versions of the questionnaires. Rationality of responses was investigated by including one scenario that was clearly superior to the other, assuming that individuals prefer Addis Ababa (doctors) or urban (nurses) location, higher pay, better housing, more equip- ment and drugs, shorter time commitment, ability to work in the private sector (doctors), and more supervision (nurses). We found that over 95% of respon- dents chose the clearly superior job. As it is very possible that some doctors or nurses would have a preference against Addis or urban areas, we chose to retain the full sample of respondents in our analysis. 3 Speci...cation and estimation We label individuals by an index q = 1; :::; Q. Each potential job is characterized by a set of a = 6 attributes, and we label the jobs i = 1; :::; 16. A pair of possible jobs is called (i; j). To motivate our analysis of the data, let y(i;j)q be de...ned 7 by 1 if q chooses i over j y(i;j)q = : 0 otherwise Let X be an a-dimensional column vector of attribute levels with kth element X k , and let Xi be the vector of attribute levels that characterize option i. Similarly, let Zq be a c-dimensional column vector of personal characteristics l for individual q, with lth element Zq We hypothesize that individual q derives some utility from option i given by T T T Uiq = + Xi + Zq + Xi Zq + uiq (1) T T T = + Xi + Zq + Xi Zq + (ei + vq + "iq ) where ei is a job-speci...c shock, vq an individual speci...c shock, and "iq are uncorrelated shocks, independent of the Xs, the Zs, and ei and vq . We allow interactions between all pairs of Xs and Zs. is an (a c)-dimensional matrix of coe¢ cients with (k; l)th element kl , and we de...ne the operation Pa Pc by T kl k l Xi Zq k=1 l=1 Xi Zq . If we suppose ei = 0, then the di¤erence in e utility earned by individual q between options i and j, y(i;j)q , is e y(i;j)q = Uiq Ujq (2) T T = (Xi Xj ) + (Xi Xj )Zq + (i;j)q : where (i;j)q = "iq "jq . Notice there is no constant in this expression. We assume then that individual q chooses option i over j (when given this binary e choice) if and only if y(i;j)q > 0. This occurs with probability P(i;j)q y = prob(e(i;j)q > 0) T T = prob( (Xi Xj ) + (Xi Xj )Zq + (i;j)q > 0) (3) T T = prob( (i;j)q < (Xi Xj ) + (Xi Xj )Zq ) T T = F( (Xi Xj ) + (Xi Xj )Zq ) as long as F (t) = prob("iq "jq < t) is such that f (t) F 0 (t) is symmetric about zero. The parameters and are estimated using a random e¤ects probit estimator to capture the within-individual correlation among choices. Where functions of estimated parameters are interpreted, 95% con...dence intervals are estimated using the bootstrap method, which has been shown to produce accu- rate and robust estimates of willingness-to-pay measures (Hole 2007). 3.1 Interpretation of estimated coe¢ cients s With this speci...cation, individual q' marginal utility of the kth job attribute (which, due to our linearity assumption, is independent of the attribute levels associated with alternative i, Xi ) is @Uiq X k kl l = + Zq : @X k l 8 More meaningfully, the marginal rate of subsitution between the kth and hth attributes (which again, with this speci...cation is independent of the attribute levels associated with alternative i) is kh @Uiq @Uiq M RSiq = (4) @X h @X k h P hl ! l + l Zq = k P kl + l Zq l kh If X k is pay (i.e., k = 2), then the absolute value of M RSiq is the marginal s value of attribute h to individual q, or individual q' marginal willingess to pay for attribute h. In the special case where Zq is a binary scalar variable (e.g., sex) taking on the values 0 (male) and 1 (female), the marginal rate of substitution between the kth and hth attributes is h kh M RSiq = k if Zq = 0 (i.e., for men) and ! h h kh + M RSiq = k k + if Zq = 1 (i.e., for women). 3.2 Attribute interactions Under the speci...cation in (1), the marginal rate of substitution between di¤er- ent attributes is independent of the attribute levels (see 4) - that is, indi¤erence curves are straight lines, and the attributes are perfect substitutes. The mar- ginal rate of substitution can be allowed to vary with the mix of attributes by introducing non-linear terms in (1). The simplest way to do this is to include a complete set of interaction terms between the di¤erent components of Xi . (We assume that of the interaction terms between attribute levels and individual characteristics, only the linear ones are potentially signi...cant - i.e., there are no k h l terms of the form Xi Xi Zq .) This yields a utility level for person q in job i of T T T T Uiq = + Xi + Xi Xi + Zq + Xi Zq + uiq where is an (a a) upper triangular matrix of coe¢ cients, with (k; h)th element kh 7 . The di¤erence in utility levels obtained by individual q between jobs i and j is e y(i;j)q = Uiq Ujq (5) T T T T = (Xi Xj ) + Xi Xi Xj Xj + (Xi Xj )Zq + (i;j)q : 7 kh kh 0 for k h, = 0 for k < h. 9 Following (3), the probability that individual q will choose job i over j is P(i;j)q y = prob(e(i;j)q > 0) (6) T T T T = F( (Xi Xj ) + Xi Xi Xj Xj + (Xi Xj )Zq ); and the parameters are estimated using maximum likelihood methods for a given assumption about F . The marginal rate of subsitution between the kth and hth attributes is now kh @Uiq @Uiq M RSiq = @X h @X k 0 hP Pa i P 1 h h mh m hm m hl l + m=1 Xi + m=h Xi + l Zq = @ hP Pa i P A(7) k k mk m km m kl l + m=1 Xi + m=k Xi + l Zq As an example, consider the e¤ect that job location might have on the rela- tive valuation of private practice and money for doctors. The attribue X 1 can take on two values, X 1 = 1 if the job is in Addis and X 1 = 0 if it is in another city, and similarly the attribute X 6 can take on two values, X 6 = 1 if private sector work is permitted and X 6 = 0 if it is not. If we ...nd that the coe¢ cient 16 is positive, and that all other kh = 0, then the marginal rate of substitution between private sector work and money for a job in Addis is 6 P ! 26 + 16 + l hl Zq l M RSAddis;q = 2 P ; + l kl Zql and the M RS between private sector work and pay for a job outside Addis is 6 P ! 26 + l hl Zql M RSN on Addis;q = 2 P : + l kl Zql Similarly, the valuation of housing could well depend on the location of the job under consideration. The marginal rate of substitution between housing and money being ! 3 P 23 + 13 + l hl Zq l M RSAddis;q = 2 P ; + l kl Zql in Addis, and ! 3 P hl l 23 + l Zq M RSN on Addis;q = 2 P kl : + l Zq l for a job outside the capital. 10 3.3 Wage equivalents We will ...nd it useful to measure the supply response to changes in non-wage attributes in terms of equivalent changes in wage rates. If a change in say rural housing is estimated to have a certain impact on rural labor supply, we calculate the change in the rural wage that would have the same quantitative e¤ect on the willingness of health workers to take rural jobs. This sub-section outlines the methodology we employ. Suppose that a standard or typical job in Addis Ababa is described by a certain bundle of characteristics, XA , and that the typical rural job is described by a vector XR . A policy intervention that is aimed at attracting workers to rural areas might improve one or more of the attributes typically found in a rural job, such as improved housing, etc. The vector of attributes de...ning the average rural job under this policy is denoted XP . A particular example of a policy intervention involves a change in just the wage earned in rural areas, keeping other attributes at the levels typically found in rural jobs. We think of this policy intervention as an equivalent wage policy. Denote such a vector of attributes by XE - each component of XE is the same as the corresponding component of XR , except for the wage. These bundles are represented by the following vectors, with the values of the numerical components derived from the survey: 0 1 0 1 0 1 0 1 0 1 0 1 1 0 0 0 0 0 B wA C B wR C B wR C B wP C B wE C B wE C B C B H C B C B H C B H C B C B 0 C B C B C B C B C B C XA = B C ; XR = B XR C = B h C ; XP = B XP C ; XE = B XR C=B h C: B 1 C B XE C B 0 C B XE C B XE C B 0 C B C B R C B C B P C B R C B C @ 2 A @ XR A @ 2 A T @ XP A T @ XR T A @ 2 A P P P 1 XR s XP XR s Note that both h and s will typically vary between doctors and nurses: for doctors, s = 0 as there is e¤ectively no private practice in rural areas, but for nurses s represents the prevailing level of supervision that nurses enjoy in rural jobs, which might be non-zero. Similarly, about 40% of doctors in our sample outside Addis report receiving a housing allowance, but less than 10% of nurses in these regions do so (see Table 5 below). The attribute di¤erences between the rural job under the policy intervention and the Addis job are 0 1 1 B wP wA C B C B XP H C dXP A = XP XA = B B C B XP 1 C E C @ XP 2 A T P XP 1 while the di¤erences between the rural job with the equivalent wage policy (XE ) 11 and the Addis job are 0 1 0 1 0 1 B wE wA C B wE wA C B C B C B XRH C B h C dXEA = XE XA = B B C=B C B XR 1 C B E C B 1 C C @ XR 2 A @ T 0 A P XR 1 s 1 In the base attribute-only model, the shares of respondents taking job P over job A, and job E over job A, are respectively T T PP A = F ( (dXP A )) and PEA = F ( (dXEA )): T T If these are set equal, then (dXP A ) = (dXEA ), or L W H H E E T T P P + (wP wA ) + XP + (XP 1) + (XP 2) + (XP 1) L W H E P = + (wE wA ) + h + (s 1) or W H H E E T T P P W (wP wA )+ (XP h)+ XP + (XP 2)+ (XP s) = (wE wA ): Alternatively, w = wE wP 0 H 1 0 H 1 XP h 1 B B E C B XE C B P C C: = W @ A @ XP T T 2 A P P XP s This is the extent to which a simple wage bonus would need to exceed the wage in the rural bundle to have the same impact on labor supply. Note that in general, the wage equivalent of a policy change that improves a single non-wage attribute will not be the same as the marginal value of, or marginal willingness to pay for, that attribute. This is because the MRS between two attributes is calculated holding all other attributes constant, while the wage equivalent compares two jobs with di¤erent attribute levels. In a model with characteristic interactions, the same attribute vectors are used, but to ...nd the wage equivalent for a given type of person we now equate the di¤erences in mean latent utilities for that person type. The shares of respondents with characteristics Z taking job P over job A, and job E over job A, are now respectively T T PP A = F ( (dXP A )+ dXP A Z T ) and PEA = F ( (dXEA )+ dXEA Z T ): These are equal if T T (dXP A ) + dXP A Z T = (dXEA ) + dXEA Z T 12 or T (dXP A dXEA ) + [dXP A dXEA ] Z T = 0: In the case where Z has just one component, sex, (0 = male; 1 = f emale), this reduces to 0 H 1 0 H 1 XP h 1 B E C B XE C wmale = wEmale wR = W B T C B P @ A @ XP 2 A T C P P XP s for men, and 0 H H 1 0 H 1 + XP h 1 B E + E C B XE C f wf emale = wEemale wR = B C B P C W W @ T + T A @ XP T 2 A + P P P + XP s where W is the coe¢ cient on the wage-sex interaction, H is the housing-sex interaction, E is equipment, T is time, and P is private/supervision. In general, the wage equivalent for individuals with a given characteristic vector Z is 0 H P Hl l 1 0 H 1 + l Z XP h P El l 1 B E+ CB E C w(Z) = wE (Z) wR = B T P l T l Z C B XP C W P Wl l @ + l Z l A @ XT 2 A + l Z P P P P + l P lZ l XP s where Xl is the coe¢ cient on the interaction term between attribute X and characteristic l. If we want to know the wage equivalents for men and women separately, just substitute Z sex = 0 or 1, and use the mean values of the other characteristics in the Z-vector. 4 Results 4.1 Summary statistics A summary of facility-level information is provided in Table 4. This table includes information provided by a facility administrator in response to a the facility survey, as well as information provided by individual health workers in response to questions about the quality of the facility. Both data sources in- dicate that workers and patients operate in facilities of generally poor quality, and that on some dimensions at least rural facilities face particular challenges.8 8 The facility administrators paint a somewhat rosier picture of conditions than workers. Our interviewers did not independently verify conditions as reported by the administrators: but we could speculate that their relatively positive evaluations might have been due to strategic mis-reporting (due to a sense of pride perhaps), or due to incomplete information (if the administrators did not face the realities of poor working conditions on a daily basis). 13 Private facilities in Addis Ababa appear to be ranked consistently better qual- ity than public facilities there and in rural areas. Within the public sector, di¤erences between Addis and rural areas are not large, and indeed sometimes favor the rural areas. As well as physical infrastructure, the work environment is conditioned by underlying work practices. One indicator of this is the level of supportive supervision that health workers reported, which at less than 50 percent, is rather low.9 Descriptive statistics regarding health workers, and indicators of their labor market status, are reported in Table 5. In economic terms, doctors in Addis do better than those in the regions. As reported in panel II of Table 5, asset ownership is higher in Addis, with one half and one quarter the doctors working in private and public facilities respectively reporting ownership of a car, com- pared with less than two and ...ve percent, respectively, in SNNPR and Tigray. House ownership is higher among private sector physicians in Addis (35%), but the rates among other doctors are similar (10-16%). These patterns of asset ownership naturally match the patterns of earned incomes. Doctors working in the public sector in Addis earn salaries about 50% more than the average doctor in the regions, while salaries of private sector doctors are three times as much. Part of this di¤erential likely re ects the return to experience (Addis doctors are older) and specialization (they are more likely to be specialized). However, we ...nd that the rates of specialization in the public and private sectors in Addis are virtually identical, suggesting that training is not the sole driver of observed income di¤erentials. Nurses in Addis earn signi...cantly smaller premiums over regional salaries ­about 14 percent if they work in the public sector and 36 percent in the private sector. The gap between private sector salaries in Addis and those of other doctors is partly o¤set by additional sources of income: public sector doctors in Addis earn additional income equal to 21% of their salaries, while the ...gures in SNNPR and Tigray are 17% and 33% respectively, and between a third and a half of doctors in the regions outside Addis report receiving housing allowances (although we do not have data on the monetary value of these allowances). Indeed, signi...cant shares of doctors working outside the Addis private sector report holding more than one job ­ from 23% in the Addis public sector, to 12% in Tigray. On the other hand, private sector doctors in Addis supplement their (much higher) salaries by only 3 percent. Although 20% report holding more than one job, we expect that these multiple jobs are in some sense considered together to s make up the worker' primary occupation, which accounts for the small amount of supplemental income. Finally, physician household incomes are higher in Addis than elsewhere. 4.2 Direct e¤ects model We ...rst estimated a model containing only the direct e¤ects of the job attributes, running this separately on the data for doctors and nurses. The results, shown 9 In our analysis we include the level of supervision as a job attribute for nurses, but not for doctors. 14 All regions Addis Ababa SNNPR Tigray Facility survey Public Private Number of facilities 77 8 31 21 17 Reliable elec./phone (%) 92 100 100 97.3 97.6 Functioning laboratory (%) 100 100 100 100.0 100.0 Su¢ cient water supply (%) 74.2 20.2 96.3 87.2 85.7 Su¢ cient medicine (%) 78.6 92.5 71.5 88.1 50.0 Su¢ cient equipment (%) 86.3 87.3 82.6 100.0 69.1 Individual survey (%) Doc Nurse Doc Nurse Doc Nurse Doc Nurse Doc Nurse Availability of supplies Soap 75.0 69.0 68.7 69.1 100 100 63.8 59.7 53.5 67.1 Water 75.0 75.2 82.5 79.9 98.0 100 59.0 61.8 44.2 77.2 Plastic gloves 88.7 85.7 84.3 84.8 100 100 92.2 84.3 68.6 82.8 Facial mask 58.7 43.0 57.8 51.8 88.9 92.5 49.1 32.1 16.2 23.5 Sterile syringes 93.7 91.8 91.1 92.1 100 100 94.7 92.1 84.4 87.2 Medicines 73.9 70.9 61.3 76.1 97.8 91.3 79.3 73.0 42.2 50.8 Workload Often not time to do tasks 55.1 48.2 67.3 58.2 22.0 20.3 82.1 61.2 61.6 31.5 Usually time to do tasks 43.0 51.1 32.7 40.4 72.0 79.8 18.0 38.8 38.4 67.1 Idle time common 2.0 0.6 0.0 1.0 6.0 0.0 0.0 0.0 0.0 1.3 Condition of facility Good 43.4 40.9 30.3 24.2 58.0 79.8 39.3 37.0 40.7 46.3 Fair 42.1 45.6 48.5 53.2 38.0 18.6 38.5 51.6 45.4 41.6 Bad 14.5 13.5 21.2 22.6 4.0 1.6 22.2 11.4 14.0 12.1 Supervision Supervisor reprimands 31.1 40.3 34.7 39.5 36.0 49.0 34.2 38.8 12.8 38.9 Supervisor supportive 45.3 46.1 32.0 38.3 62.0 68.8 50.4 45.2 26.7 45.0 Table 4: Facility level information, based on interviews with an administrator ("facility survey") and individual health workers * Includes for-pro...t and non- pro...t NGO and missionary facilities ** Includes 3 private facilities 15 Doctors Nurses All Addis SNNPR Tigray All Addis SNNPR Tigray Demographics Public Private Public Private Female (%) 18.2 30.0 16.0 2.6 26.7 64.3 73.8 84.4 53.0 61.8 Married (%) 56.6 61.3 74.0 33.3 45.2 63.3 65.3 65.5 50.2 79.3 Age (years) 36.1 39.2 41.2 29.3 31.5 33.4 34.5 35.3 31.0 34.7 (0.88) (1.64) (1.78) (1.16) (1.61) (0.49) (0.73) (0.86) (1.25) (0.71) Primary job priv. (%) 34.9 * * 9.4 0.0 14.0 * * 5.4 0.0 Specialist (%) 27.8 40.4 38.0 6.8 19.8 * * * * * Income 16 Salary (US$) 284.5 244.6 480.5 156.4 176.6 100.9 106.8 128.3 87.7 100.8 (17.4) (10.5) (39.0) (14.8) (13.9) (2.0) (2.1) (9.6) (2.7) (1.96) Other compensation 52.7 29.3 46.0 85.5 53.5 47.0 15.5 35.9 73.3 48.7 with job (%) Housing (%) 18.9 0 0 52.1 34.8 5.9 0 0 11.7 6.7 Total health worker 320.9 297.0 496.8 181.4 233.1 102.6 109.3 130.1 87.7 103.7 income (US$) (24.8) (24.8) (40.1) (29.7) (38.2) (2.1) (1.7) (9.5) (2.70) (3.7) Total household 443.8 509.2 696.9 196.3 264.3 201.2 298.8 263.9 139.4 157.5 income (US$) (28.1) (49.1) (55.7) 30.0 (46.8) (12.8) (22.1) (25.6) (10.9) (10.0) Table 5: Demographic characteristics and incomes of sampled health workers Doctors Nurses III. Value as % Variable I Coef II. S.E. I Coef II. SE of base salary Pay 1000 0.620 0.029 0.992 0.033 Doctors Nurses Location 0.415 0.052 0.895 0.031 26.8 72.1 Housing 0.501 0.036 0.582 0.020 32.4 46.9 Equipment 0.409 0.056 0.619 0.033 26.4 49.9 Payback Time -0.282 0.053 0.144 0.030 -18.2 -11.6 Private/Super 0.743 0.059 0.404 0.033 48.0 32.6 % correctly predicted 79% 81% Log likelihood -1383.52 -4038.84 LRT p < 0:001 p < 0:001 n 216 640 216 640 Table 6: The direct e¤ects model, for doctors and nurses in Table 6, con...rm that all job attributes signi...cantly inuence job choice in the expected directions. Columns I and II of the table report the coe¢ cients and their standard errors respectively. Column III reports the marginal value of each non-wage attribute, as a percentage of average base public sector wages (2,500 Birr, or $275, per month for doctors and 1,250 Birr, or $140, per month for nurses). These values are equal to the marginal rates of substitution between the corresponding attribute and pay, as calculated in (4) (with each = 0). These results suggest that on average, the extra value of a job in Addis relative to one in a regional city for doctors amounts to about one quarter (27%) of the base public sector physician salary, the value of improved housing is about one-third (32%), the value of equipment is about one quarter (26%), and the value of reduced time commitment is about one ...fth (18%). The most highly prized attribute for doctors is however, the ability to work in the private sector, which has a value of about half (48%) the base salary. For nurses the most valuable job attribute is location. Indeed, location appears to be valued more by nurses than by doctors, especially when the value is measured as a share of the base salary. This partly re ects the fact that "location" means something di¤erent in the questions nurses were presented with than it does for doctors - switching a job from a rural area, which in principle can be very remote, to a regional capital, increases its value by 72% of s the base public sector nurse' salary. (The other factor is of course the fact that the base nurse salary is only half the base doctor salary, 1,250 Birr, or about $140.) The least valued attribute for nurses appears to be payback time, as it is for doctors - having to pay back an extra year after receiving training is equivalent to a pay-cut of about 12% of base salary. Improved supervision is valued, but not as highly as the other non-time attributes. 17 4.3 Full model with characeristic and attribute interac- tions We extend the direct e¤ects-only model by incorporating interactions with char- acteristics that we expect might be correlated with marginal attribute valua- tions, and by including attribute-attribute interactions to assess non-linearities and synergies between attributes. We are particularly interested in explor- ing which attribute changes are likely to induce individuals to move to a rural posting, and which types of people are more likely to respond to a particular policy intervention. The demographic characteristics of greatest interest mari- tal status, number of children, and sex. We are also interested in the e¤ects on s attribute valuation of characteristics of the respondent' current job, including its location, housing bene...ts, and for nurses the level of supervision provided. We adopt a data-driven approach to model construction, in which we ...rst sequentially add interactions of a particular characteristic with pay and non-pay attributes; we then estimate a full model including all the interactions that are individually signi...cant (at the p = 0:10 level); ...nally we remove from this full model all the interactions that are no longer statistically signi...cant (at the p=.05 level). Note, however, that because the M RS and other key model outputs are functions of multiple parameters, some otherwise insigni...cant interactions are retained in the model in order to calculate standard errors around these functions. Table 7 present the results for doctors and nurses. Our estimates in the full model allow us to examine the heterogeneity of attribute valuations across health workers with di¤erent demographic charac- teristics. Table 8 reports selected marginal valuations for doctors, expressed as a percentage of the average base salary of all doctors (2,500 Birr per month). Note that the table includes only marginal valuations of those attributes with statistically signi...cant characteristic interactions. We ...nd that married doctors value a job in Addis twice as highly as single doctors (38% versus 19% of base salary). The most natural explanation for this e¤ect is a combination of joint-career issues, and children (although see below on the latter). Also, younger doctors are seen to value shorter pay- back periods following training ­ at ...rst this seems surprising, as the young have "time on their hands," and we might expect them to be willing and able to pay back more time after training. The result likely arises from the fact that age is confounded with experience and training, so that older doctors, relatively many of whom are already specialized, do not place a high value on the training o¤ered. Alternatively, younger doctors might feel more able to take advantage of their training, say by entering the private sector, or seeking future promotions. The future stream of bene...ts associated with training, even accounting for the length of career over which those bene...ts will accrue, may be greater for younger, more adaptable, doctors, than for older generations. The impact of children seems perhaps surprisingly small, particularly the im- pact of the ...rst child: Doctors with one child value an Addis job (presumably with better schools etc.) just 2 percentage points of base salary more than doc- tors without children, and there is virtually no di¤erence in the value of housing 18 Doctors I. Coef II. S.E. Nurses I. Coef II. S.E. Direct e¤ects Direct e¤ects Pay 1000 0.623 0.154 Pay 1000 0.816 0.090 Location -0.026 0.209 Location 0.082 0.103 Housing 0.611 0.064 Housing 0.246 0.048 Equipment 0.448 0.116 Equipment 0.890 0.066 Time -0.822 0.285 Time -0.120 0.045 Private/Supervion 0.626 0.115 Private 0.207 0.063 Characteristic interactions Characteristic interactions Income 1000 * location 0.030 0.017 Income 1000 * location 0.053 0.019 Married * location 0.402 0.160 Married * location 0.011 0.071 Sex * location 0.082 0.182 Sex * location 0.236 0.065 Child * location -0.022 0.082 Child * location 0.050 0.029 Addisnow * location -0.181 0.007 Citynow * location 0.343 0.074 Sex * pay 1000 -0.016 0.095 Supervision * location 0.173 0.096 Child * pay 1000 -0.074 0.042 Sex * pay 1000 -0.179 0.069 Married * pay 1000 0.066 0.083 Child * pay 1000 -0.023 0.031 Age * pay 1000 -0.002 0.004 Married * pay 1000 0.187 0.076 Sex * housing 0.016 0.120 Citynow * pay 1000 0.268 0.071 Child * housing -0.061 0.046 Housenow * pay 1000 -0.398 0.130 Sex * equipment 0.190 0.189 Married * housing 0.088 0.042 Child * equipment 0.035 0.073 Citynow * housing 0.305 0.043 Sex * time 0.130 0.177 Housenow * housing -0.145 0.080 Age * time 0.016 0.008 Married * equipment -0.091 0.067 Child * time 0.058 0.075 Married * time 0.036 0.062 Child * private -0.004 0.077 Married * supervision 0.069 0.070 Sex * private -0.026 0.200 Attribute interaction Housing * location 0.356 0.065 Attribute interaction Equipment * location -0.241 0.075 Private * location 0.442 0.205 Supervision * location 0.291 0.078 Table 7: Full model with interactions for doctors and nurses 19 Marginal attribute value (percent of average base salary) Characteristic Location Housing Payback time 24 -23.2 Age * 34.6 -13.0 40 -8.0 Single 19.1 Marital status Married 38.6 0 28.6 33.7 -13.0 Number of children 1 30.6 33.8 -11.0 2 33.0 33.9 -8.4 Male 28.3 33.4 -14.1 Sex Female 33.5 35.1 -7.1 * Ages are the 10th percentile, mean, and 90th percentile. Table 8: Heterogeneity in attribute valuations: doctors by number of children. Having a child reduces the value of reduced payback time following training, but from a relatively small base (13 percent). These results on the e¤ects of children should however be interpreted with caution as the median number of children in our sample of doctors was zero, so extrapo- lations to larger numbers of children are likely to be imprecise. Nonetheless, taken together with the e¤ect of marriage, they suggest that joint career con- cerns (and perhaps the prospect of children) are more important barriers to rural labor supply than parenthood. Finally, in terms of di¤erences by sex, women are observed to value work in the capital more than men (33% versus 28% of base salary), while men value reduced payback time about twice as much as women. In contrast, as reported in Table 9, married nurses value urban work, and housing, less than single nurses. We do not know why marriage should a¤ect nurses'valuations di¤erently to those of doctors. One di¤erence is, of course, that "location" means something di¤erent in our estimation of the preferences of nurses and doctors.10 Married nurses value reduced payback time half as much as singles, suggesting mobility might be more valuable to the latter. The impact of children seems somewhat larger for nurses than doctors (in terms of the percentage of base salary), but again, having children does not seem to be an especially impenetrable barrier to rural work. 5 Policy experiments Our basic policy concern is over what the government can do to induce health workers to take jobs in underserved locations. The speci...cation in (1) allows 1 0 Perhaps it is more important for single nurses to be in a city "marriage market" than it is for single doctors. 20 Marginal attribute value (percent of average base salary) Characteristic Location Housing Payback time Single 74.7 56.6 -12.2 Marital status Married 61.3 52.9 -6.9 0 60.4 Number of children 1 66.5 2 72.9 25th percentile 65.4 Income Mean 66.5 95th percentile 72.4 No 61.4 30.2 Lives in a city now Yes 70.8 45.3 No 63.4 Works in private sector now Yes 73.7 No 53.8 Receives housing allowance now Yes 72.7 Table 9: Heterogeneity in attribute valuations: nurses us to infer the estimated probability of accepting one job (i) over another (j). We use this information in section 5.1 to calculate the probability that a worker will accept a rural job over an urban job, and how this probability varies both with other job attributes and across di¤erent people, for example by sex. In section 5.2 we convert the e¤ects of job attributes on labor supply to equivalent wage changes by asking what wage change would have the same e¤ect on the probability of accepting a rural job as a given attribute change. Finally in section 5.3 we illustrate graphically the impact of job attribute improvements on the e¤ect of rural wage bonuses in increasing rural labor supply. Note that in our simulations we do not report the impact of allowing doctors to engage in private sector work outside of Addis Ababa. This is due to the fact that while respondents reported a high valuation of working in the private sector, there are few opportunities to do so outside Addis at this time (although the situation is likely to be changing rapidly), so application of the corresponding coe¢ cients to rural jobs would be misleading. 5.1 Impact of attribute changes on rural labor supply First we estimate the impacts of changes in job attributes on the probability that an individual will accept a job in a rural area over a job in Addis Ababa (for doctors) or in a zonal capital (for nurses). For doctors we de...ne job j to be in Addis Ababa, with the prevailing attributes of an average job there set at levels approximating those reported by health workers in the ...rst part of the survey instrument (and similarly for nurses in zonal capitals). Holding public sector wages constant (i.e., without introducing wage bonuses), we calculate the 21 Doctors Nurses p 95% CI Increase p 95% CI Increase Baseline 0.074 (0.029,0.122) ­ 0.046 (0.034,0.058) ­ Basic housing 0.109 (0.046,0.173) 47% 0.097 (0.080,0.115) 112% Superior housing 0.269 (0.137,0.400) 262% 0.192 (0.152,0.233) 319% Equipment 0.167 (0.105,0.229) 125% 0.198 (0.165,0.231) 332% Pay-back time 0.114 (0.047,0.180) 53% 0.056 (0.041,0.072) 22% Equip & housing 0.226 (0.144,0.308) 204% 0.323 (0.284,0.362) 605% Supervision - - - 0.075 (0.055,0.095) 64% Table 10: Impact of non-wage attribute improvement on probability of accepting a rural job, for doctors change in the estimated probability of an individual accepting a rural job when one non-wage attribute is improved. The results of this exercise are reported in Table 10. Our point estimates indicate, for example, that about 7.5 percent of doctors would be willing to take a rural job over a job in Addis under pre- vailing conditions, if they had the choice. Providing incentives in the form of superior housing increases the chance of accepting a rural job to more than one- in-four, while provision of basic housing, and training incentives (measured by a reduction in time commitment to one year) have relatively small e¤ects, each increasing the likelihood from 7.5 percent to about 11 percent. The e¤ect of improving the availability of equipment is in the middle of the range, increasing the probability of choosing a rural job to 17%. For nurses, the non-wage attribute with the single biggest impact on the share of workers willing to take a rural job is the provision of adequate equip- ment. At baseline levels, only 4.4 percent of nurses would choose a rural job over a city job, but this jumps to 20 percent if they can be guaranteed ade- quate levels of equipment. The provision of basic housing, reducing pay-back time and providing better supervision have substantially smaller e¤ects on the probability of choosing a rural job, increasing it to levels in the range of 5-8%. 5.2 Wage equivalents Without information on the costs of making attribute improvements it is di¢ cult to use this information for decision-making purposes. One step in that direction is to calculate the rural wage increases that would yield equivalent labor supply responses for each attribute improvement considered. For example, we ask how much rural wages would need to be increased, holding current non-wage attributes ...xed, to induce the same increase in the number of doctors willing to take a rural job that we found for each attribute improvement in Table 10. Using the full model estimates reported in Table 7, we calculate these wage equivalents (as percentages of the base salary) for men and women separately, as discussed in sub-section 3.3. The 95 percent con...dence intervals around the estimated wage equivalents (again measured in percent of base salary) are 22 150% equivalent Wage 100% Nurses Nurses Nurses Nurses 50% Doctors Doctors Doctors Doctors Nurses Nurses Doctors 0% M F M F M F M F M F M F M F M F M F M F M F Basic Superior Equipment Payback Housing Super- housing housing time and vision equipment -50% Figure 1: Estimated wage equivalents for each attribute, by doctor/nurse and by sex (M/F), as a percentage of the base wage. bootstrapped, in light of the fact that the quantities of interest are non-linear functions of the underlying parameter estimates. The results are presented in Table ?? for doctors and nurses separately. A striking result is that while the point estimates of wage equivalents for most attributes tend to be higher for women, the di¤erence is rarely statistically signi...cant. Only for pay-back time amongst doctors does the con...dence interval around the estimated wage equivalent for men (25:7 46:8) not overlap that for women (31:4 64:6). This is illustrated in Figure 1. 5.3 Rural wage bonuses and attribute incentives Finally, we investigate the impact of increases in rural pay on predicted health worker labor supply based on our regression results above. In particular, we estimate the probability that an individual will accept a rural job as a function of the excess of pay over the base pay rate. This is the most obvious way to induce greater labor supply, but we also calculate the impact of such wage in- creases when coupled with attribute improvements. The results, for doctors and nurses respectively, are presented graphically in Figures 2 and 3, respectively. For doctors, doubling pay while keeping other attributes constant increases the probability of accepting a rural job from 7% to 57%. Alternatively, to induce half of doctors to locate in rural areas under current conditions, a rural bonus of approximately 89% (2,225 Birr) is required. Providing basic housing does not a¤ect the impact of wages to a large extent, probably because most doc- tors already have at least basic housing. On the other hand, providing superior 23 Doctors Male Female W.E. (% base) p (rural) W.E. (% base) p (rural) Est. 95% CI Est. 95% CI Est. 95% CI Est. 95% CI Baseline 0.080 (0.029,0.131) 0.050 (0.000,0.119) Basic house 11.7 (9.3,14.2) 0.116 (0.046,0.187) 12.3 (7.2,17.3) 0.077 (0.000,0.119) Superior house 45.2 (35.8,54.6) 0.280 (0.139,0.421) 47.3 (27.9,66.8) 0.213 (0.000,0.443) Equipment 24.6 (14.2,34.9) 0.169 (0.100,0.237) 35.7 (19.6,51.9) 0.157 (0.036,0.279) Pay-back time 14.1 (4.2,24.0) 0.125 (0.053,0.198) 7.1 (-12.8,27.1) 0.065 (0.000,0.166) Equip & house 36.2 (25.7,46.8) 0.228 (0.139,0.317) 48.0 (31.4,64.6) 0.216 (0.060,0.372) Nurses Male Female 24 W.E. (% base) p (rural) W.E. (% base) p (rural) Est. 95% CI Est. 95% CI Est. 95% CI Est. 95% CI Baseline 0.063 (0.047,0.078) 0.038 (0.027,0.050) Basic house 44.1 (37.4,50.7) 0.126 (0.100,0.152) 53.7 (46.7,60.6) 0.084 (0.064,0.104) Superior house 92.6 (78.6,106.4) 0.237 (0.189,0.285) 112.7 (93.5,131.9) 0.171 (0.131,0.211) Equipment 57.4 (48.2,66.6) 0.244 (0.207,0.281) 69.9 (59.4,80.5) 0.176 (0.142,0.210) Pay-back time 8.0 (3.0,13.0) 0.076 (0.056,0.095) 9.8 (3.8,15.7) 0.048 (0.033,0.062) Equip & house 101.5 (86.5,116.5) 0.380 (0.337,0.423) 123.6 (105.0,142.3) 0.294 (0.251,0.337) Supervision 31.3 (25.0,37.7) 0.099 (0.074,0.124) 38.2 (30.2,46.1) 0.064 (0.045,0.083) * Con...dence intervals in these cells to be recalculated. Table 11: Demographic characteristics and incomes of sampled health workers 1 Probability Superior housing of taking a rural job 0.8 Basic housing and equipment 0.6 Equipment Time 0.4 Basic housing 0.2 Baseline 0 -0.5 0 0.5 1 1.5 2 2.5 3 Wage bonus (as a multiple of base salary) Figure 2: Share of doctors willing to accept a rural job as a function of the rural wage bonus (horizontal axis), with alternative in-kind attribute incentives. housing means that doubling wages increases the probability of accepting a rural job from 27% to 84%. Our results suggest that nurses are much less responsive to proportionate wage bonuses than doctors ­ a doubling of pay increases the probability of accepting a rural job from 4% to only 27%, and inducing half of the nursing workforce to locate in rural areas would require a wage bonus of about 155% of the base salary. This bonus amounts to 1,937 Birr, and is only marginally smaller than that needed to induce a similar proportion of doctors to take jobs in rural areas. The impact of adequate equipment, both on willingness of nurses to take a rural job in itselt, and on the impact of higher pay on such willingness, ect is of particular interest, especially since this attribute does not re personal consumption as such. Indeed, the impact of equipment is not only greater than that of basic housing, but it exceeds that of superior housing also. By itself, adequate equipment increases the likelihood of accepting a rural job from 4% to 21%, while coupled with a doubling of rural pay, this probability increases to 61%. 6 Conclusions Our analysis provides evidence that the locational labor supply decisions of health workers are responsive to both wage and non-wage factors, and that for some of these attributes the responses can be large. Proportionate wage bonuses for rural service increase labor supply, but these e¤ects seem to be larger 25 1 Probability of taking Basic housing a rural job and equipment 0.8 Equipment 0.6 Superior housing 0.4 Basic housing Supervision 0.2 Time Baseline 0 -0.5 0 0.5 1 1.5 2 2.5 3 Wage bonus (as a multiple of base salary) Figure 3: Share of nurses willing to accept a rural job as a function of the rural wage bonus (horizontal axis), with alternative in-kind attribute incentives. for doctors than for nurses. For example, under current working conditions in urban and rural facilities, attracting 50% of doctors to work in regional towns would require a wage bonus of approximately 89% over the base salary. Inducing the same rural labor supply from nurses to rural settings (which are on average more remote than regional towns) would require a bonus of about 155%. For both doctors and nurses, the joint provision of superior housing and equipment induce signi...cant increases in the probability of accepting a rural job. Provision of basic housing, reduced pay-back time and (for nurses) improved supervision all have positive but smaller e¤ects on the likelihood of choosing a rural post. Our broad-brush interpretation of these results is that health work- ers want to be paid more and to be better able to do their jobs, but that in-kind inducements in the form of accelerated training and improved supervision are valued somewhat less. The fact that superior housing is an e¤ective inducement likely re ects the fact that this would represent a large cash equivalent. These results can be usefully compared with two similar studies of stated preferences of health workers in developing countries. Chomitz et al. (1998) found that the promise of specialist training was an e¤ective, if expensive and ine¢ cient, way of inducing doctors from Java to relocate to the remote islands of Indonesia. They found that individuals from remote areas were more likely to take jobs in remote areas, and that modest ...nancial incentives could induce relocation to moderately (but not extremely) remote areas. Serneels et al (2005) report results from a survey of nursing and medical students in their ...nal year of study in Ethiopia. They ...nd that at the then prevailing starting wage of 700 Birr, fully one-third of nursing students reported 26 that they would choose to work in a rural area (de...ned as 500km from Addis), and that a rural bonus of just 31% would be su¢ cient to induce all student nurses to take such jobs. To get all graduating doctors to move to the rural areas requires a bonus of just 39% of the starting salary. These results stand in contrast to those of this paper: we found above that doctors in our sample would need to be paid a bonus of about two and a half times the base salary in order to induce (nearly) all of them to work in a rural area, while the corresponding ...gure for nurses is about three time. The di¤erence may stem from the fact that our samples were very di¤erent: Serneels et al. interviewed students, while we surveyed health workers at various stages of their careers. Finally, Serneels s et al. report that the availability of children' educational opportunities was one of the main attractions of work in Addis Ababa. Somewhat surprisingly, we ...nd only weak evidence of this e¤ect in our data: the number of children a health worker has does not appear to have an economically signi...cant in uence on her/his valuation of alternative job attributes (including location). We speculate that this may be due to the widespread practice of sending children s to boarding school amongst Ethiopia' upper classes. These results can provide guidance to policy-makers about the potential trade-o¤s between alternative policies to encourage health workers to accept rural jobs. However, without detailed information on the costs of altering the speci...ed attributes, it is impossible to rank the alternative policy interventions in terms of any cost-e¤ectiveness measure.11 However, our wage-equivalent analysis is a ...rst step towards allowing such a comparison. In addition, the limitations of stated preference studies should be kept in mind, and ideally we would seek to validate our results by comparing them with evidence from revealed preference analyses. 1 1 Note that cost-e¤ectiveness is a useful measure if the objective of increasing labor supply is taken as given. It does not inform the question of whether such changes in labor supply are warranted - we take this as self-evident in this case. 27 7 References Anand and Barnighausen (2004): Lancet Chomitz, Kenneth, Gunawan Setiadi, Azrul Azwar, Nusye Ismail, and Widi- yarti (1998): "What do doctors want? Developing Incentives for Doctors to s Serve in Indonesia' Rural and Remote Areas," World Bank Policy Research Working Paper 1888, World Bank, Washington DC. Hole AR (2007): "A comparison of approaches to estimating con...dence intervals for willingness to pay measures," Health Economics 16(8): 827-40. Huber J and Zwerina K. (1996): The importance of utility balance in e¢ cient choice designs. Journal of Marketing Research 33: 307-317. Joint Learning Initiative (2004). Human resources for health: overcoming the crisis. Boston MA: Joint Learning Initiative Mangham L and Hanson K (2007): "Eliciting the employment preferences of public sector nurses: results from a discrete choice experiment in Malawi," Unpublished mimeo. Ministry of Health, Government of Ethiopia (2005): Health and health- related indicators: 1997. Penn-Kekana L, Blaauw D, Tint KS, Monareng D, Chege J (2005): "Nursing sta¤ dynamics and implications for maternal health provision in public health facilities in the context of HIV/AIDS," Johannesburg: Centre for Health Policy, University of the Witswatersrand. Ryan M and Gerard K (2003): "Using discrete choice experiments in health economics: moving forward," In Scott A, Maynard A and Elliott R, eds. Ad- vances in Health Economics. John Wiley and Sons Ryan M and Gerard K (2003): "Using discrete choice experiments to value health care programmes: current practice and future research re ections," Ap- plied Health Economics and Health Policy, 2(1): 55-64 Scott A (2001): "Eliciting GPs preferences for pecuniary and non-pecuniary job characteristics," Journal of Health Economics 20: 329-347. Serneels, Pieter, ose Garcia-Montalvo, Magnus Lindelow, and Abigail Barr (2005): "For Public Service or for Money: Understanding Geographical Imbal- ances in the Health Workforce," World Bank Policy Research Working Paper 3686, World Bank, Washington DC. WHO (2006): World Health Report 2006: Working together for health. Geneva: World Health Organization Wilbulpolprasert S, and Pengpaibon P. (2003): "Integrated strategies to tackle the inequitable distribution of doctors in Thailand: four decades of ex- perience," Human Resources for Health, 1: 12. 28 8 Appendix: Description of job attributes for doctors and nurses Here we report the descriptions of job attributes and possible levels that were presented to respondents. 8.1 Doctors For doctors, the job attributes and possible levels were described to the respon- dents as follows: Geographic Location This attribute speci...es whether your place of work is in Addis Ababa or in a zonal capital of one of the zones. If the latter, you should think of the job as being randomly situated in one of the zonal capitals in Ethiopia, or alternatively, in "an average zonal capital" . Net Monthly Pay (including regular allowances) This attribute takes on di¤erent Birr levels. The ...rst represents the base salary for a physician at an "average"grade in the civil service pay scale, while higher levels are multiples of this average base level. Note that the base salary ect does not necessarily re your current actual salary. Government-provided Housing This attribute measures the existence, and quality, of government-provided housing, and has three possible levels. "None" means there is no housing pro- vided by the government as part of the conditions of employment. "Basic" housing means the government provides housing for the health worker, but that it is rudimentary, having no electricity or running water, and with at best an outside toilet. "Superior" housing means the government provides housing of higher quality, including the presence of electricity and running water, including an inside ush toilet. Availability of Equipment and Drugs This attribute simply takes on two values ­ "inadequate" and "improved" . "Inadequate" is the standard of equipment and availability of drugs that you might expect in a poorly equipped public facility in the given location. "Im- proved"is that level of supplies that would result from a doubling of the budget currently spent on equipment and drugs. Time Commitment following Training Suppose your employer provides or sponsors training on your behalf. This attribute measures the number of years you are required to work for the sponsor for each year of training provided. It can take on two values: 1 and 2. 29 Permission to hold a Second Job in the Private Sector This attribute is 1 if you are permitted work in the private sector (either using the public facility or not), and 0 if you are not permitted to do so. 8.2 Nurses For nurses, some of the job attributes and possible levels di¤er to those o¤ered to doctors: Geographic Location This attribute speci...es whether your place of work is in a City (i.e., a zonal or regional capital, or Addis Ababa), or in a Rural area. If the job is a "City" job, you should think of it as being randomly situated in one of the zonal capitals . or larger cities in Ethiopia, or alternatively, in "an average city" If the job is a "Rural" job, you should think of it as being randomly situated in a town or village outside of the zonal capitals and larger cities. Net Monthly Pay (including regular allowances) This attribute takes on di¤erent Birr levels. The ...rst represents the base salary for a nurse at an "average" grade in the civil service pay scale, while higher levels are multiples of this average base level. Note that the base salary ect does not necessarily re your current actual salary. Government-provided Housing Same as for doctors. Availability of Equipment and Drugs Same as for doctors. Time Commitment following Training Same as for doctors. Level of supervision This attribute attempts to measure the degree of professional interaction you have with your superiors, and takes on two values ­ high and low. A high level of supervision could result from regular and productive interaction with a supervisor who works in the same facility as you, or from regular visits (say every one or two weeks) from a more senior health worker from another facility, such as a zonal hospital. A low level of supervision could arise due to lack of interaction by more senior health workers who work at your facility, or because of infrequent visits (say once every six months or less) by such superiors from other institutions. 30 This paper estimates the effectiveness of a range of policy interventions aimed at improving the supply of health workers to rural areas in Ethiopia. Using data from a survey of 861 health workers, it employs stated preference techniques to predict labor market responses of doctors and nurses to changes in rural wages, working conditions, housing benefits, and training opportunities. This paper was produced by the World Bank's Africa Region Human Resources for Health team, with funding from the Government of Norway and the Gates Foundation. 2009 © All Rights Reserved. Health Systems for Outcomes Publication THE WORLD BANK