The Effects of Land Title Registration on Tenure Security, Access to Credit, Investment and Production: Evidence from Ghana

Transcription

1 The Effects of Land Title Registration on Tenure Security, Access to Credit, Investment and Production: Evidence from Ghana Niklas Buehren, Markus Goldstein, Robert Osei Isaac Osei-Akoto, Christopher Udry 1 Draft Report, March 217 PLEASE DO NOT CITE Abstract We use a regression discontinuity evaluation design combined with three rounds of household survey data collected over a period of six years to evaluate a pilot land titling intervention in the Awutu-Effutu-Senya area of the Central Region of Ghana. We find that there are strong markers that the land titling program was successful in reaching the target population through measuring, demarcating and ultimately titling land in the group. These first-order outcomes translate into a number of impacts of the titling program on asset holdings which vary substantially by the gender of the owner. We also find evidence of increased engagement in self-employment and business profits among women while agricultural production at the household level remains constant. 1 Buehren: World Bank, nbuehren@worldbank.org; Goldstein: World Bank, mgolstein@worldbank.org; Osei: Institute of Statistical, Social and Economic Research, rosei@ug.edu.gh; Osei-Akoto: Institute of Statistical, Social and Economic Research, ioseiak@ug.edu.gh; Udry: Yale University, christopher.udry@yale.edu. 1

2 1. Introduction The importance of well-defined property rights in stimulating access to credit and investment is well recognised in the literature (De Soto, 2). There are several avenues through which property rights achieve this. For example, property rights more generally and secure land tenure allows households to collateralize loans and thus obtain financing for investments. The additional capital obtained this way can drive both farm and non-farm investment and has been found to trigger labor productivity and income (Field & Torero, 25). This link has become a key argument for the role of land security in promoting development (Besley, 1995). Consequently, the link between land titling, access to credit and investment has been targeted as an intervening point for policies and programs by several governments and development organisations. However, several empirical studies have failed to confirm the positive impacts of land titling programs on access to formal credit (Deininger & Chamorro, 24; Galiani & Schargrodsky, 21; Zegarra et al, 28). In addition, those studies that have found empirical support for this relationship frequently qualify their findings in several ways. In particular, these evaluations of land titling programs highlight impact heterogeneity and the importance of the implementation approach (Mushinski, 1999; Dower & Potamites, 25). Taken together, the findings of these studies confirm that credit markets thrive within a plethora of enabling factors, of which land titling, and thus the ability to use real estate as collateral, is an important but not the sole driver to access to credit. Despite these mixed empirical results, theoretical arguments for the existence of a multitude of social, environmental and economic benefits stemming from secured property rights cannot be overlooked (Besley, 1995). In light of this, the Millennium Development Authority (MiDA) and the Government of Ghana (GoG) rolled out a pilot land titling program in the Awutu-Effutu-Senya area of the Central Region of Ghana in 29. The primary objective of this program was to encourage land users in the pilot district to register their claims to parcels they controlled either through deeds or titles if claims were existent for at least three years prior to the start of the intervention. By enhancing the bundle of informal rights possessed by land holders through formal land titles, it was anticipated that the valuations of these parcels would appreciate and increase owners access to credit and thus remove 2

3 one cause of the liquidity constraints faced by the target population. Ultimately, the program was intended to stimulate agricultural and non-agricultural investment in order to reduce poverty and spur economic growth in the long run. We use a regression discontinuity evaluation design (RDD) combined with three rounds of household survey data collected over a period of six years to evaluate this pilot intervention. We find that there are strong markers that the land titling program was successful in reaching the target population through measuring, demarcating and ultimately titling land in the group. These first-order outcomes translate into a number of impacts of the titling program on asset holdings which vary substantially by the gender of the owner. For example, while women decrease their amount of outstanding credit in the short run, men increase the amount of outstanding credit in the long run. This evaluation contributes to the literature linking property rights more generally and land titles in particular to investment and access to credit as well as household decision behaviour. In addition, while most studies investigating the effects of land titling on credit concentrate on either rural or urban households, this study builds on data collected in a peri-urban setting. In the study location there is considerable competition among alternative land uses: agricultural, commercial and housing. In addition, the dataset on which this study builds on has sex-disaggregated information on plot ownership and thus identifies which plots are controlled by men or women in the study households. This allows us to examine the sex-disaggregated impacts of land titling on investment and credit. The remainder of the report is organised into four sections. Section 2 presents background information on land tenure in the Ghanaian context and elaborates on the implementation arrangements of the MiDA Land Titling Project. In Section 3 we proceed to discuss our data and the analytical methods we employ for our estimations. Section 4 is organized into two parts. The first part provides descriptive statistics of variables of interest and balance tests. In the second part we estimate the impact of the land titling project on tenure security, investment and measures of household welfare. Section 5 discusses proposals for further work. 3

4 2. Land Tenure in Ghana and the MiDA Land Titling Project In Ghana land is categorized into four different types: Stool, Family, State-owned and Freehold. Stool and Family land constitute about 78 percent of all land while State-owned and Freehold land form the remaining 2 percent and 2 percent respectively (Deininger, 23; Kuntu-Mensah, 26; Awuah et al, 213). With regards to Stool and Family land, the Ghanaian legal framework allows for customary freehold, stranger usufruct rights, sharecropping and leasehold of less than 1 years to be held by individuals. The Ghanaian Land Title Registration Act of 1985 specifically permits the above rights to be formally registered so that any interests held by individuals on any parcel of land can be protected (Sittie, 26). However, poor record keeping (emanating from the oral nature of transactions associated with land controlled by Stools and Families), 2 process complexities and associated costs have inhibited title registration in the past. Additionally, cost of land titling is relatively high and often connected with extensive time lags (Awuah et al, 213) To streamline land titling in Ghana, the Land Title Registration Act 1985 was introduced. Although the new Act did not represent a dramatic deviation from prevailing practices, it was intended to address the weaknesses of land related laws (PNDC Law 152). Prior to the enactment of the Land Title Registration Act in 1985 (and the accompanying law: PNDC Law 152), there already existed legal instruments which supported deeds registration in Ghana (Zevenbergen, 1998). The operation of the deeds registry helped to identify transactions related to land, but failed to confer title on the individual who held the deed. Cadastral maps which accompanied such deeds were also frequently inaccurate or, in some instances, not required and thus missing (Kuntu-Mensah, 26). Therefore, the system failed to address the issues of multiple claims to the same parcel of land. To address these challenges, legislative reforms were initiated in 1987 and backed by the Land Title Registration Act of More specifically, these reforms were meant to introduce a system that allowed the registration of land titles across the country in a stepwise manner. 3 This system was designed to 2 To improve efficiency of record keeping within the customary system (stool and family lands) the Customary Land Secretariat was established in 24 with 38 branches throughout the country. Although potentially beneficial the state of operation of the various branches have been mixed some functioning fully whilst others are yet to take off. 3 That is after the declaration of a registration district by the lands commissioner at least 75% of parcels within that district were to be accurately registered before another district was declared. A t the same time Deed Registration was to continue operation in other parts of the country. 4

5 operate side by side with the deeds registration processes that were already in place. Naturally, the main objective of the title registration system was to confer title to the holders of the certificate and assure the holders that in times of any threat to their rights, the government will ensure that they are protected. Any title issued under this law could only be nullified by a court of law (Sittie, 26). Population growth, urbanisation and expansion of commercial agriculture over the past decades have increased scarcity of land in Ghana. These developments pose challenges to the traditional way in which land ownership and land use rights have been managed even in the face of the laws discussed above. Traditionally, Chiefs were in charge of the allocation of land to ensure equity in access to land (Udry, 21; Onoma, 21). In the face of the new dynamic presented by increased demand for land, some Chiefs have taken advantage of the situation and sold land multiple times to different buyers especially in the urban areas. Of course, such practices can create an array of conflicts, disputes and ownership insecurity, as for example shown in Kuntu-Mensah (26), which have resulted in numerous litigations in the courts (Aryeetey & Udry, 21). As noted by Jones-Casey & Knox (211), the Ghanaian courts were clogged with 35, land disputes in 26. Narrative evidence suggests that in urban centres the impact of insecure land tenure rights is felt by households through high prices of housing for residential uses and offices for businesses. Given the lack of clear land rights, investments in housing provision and mortgage markets have been inhibited, causing the rental costs of housing to rise rapidly as a result of insufficient supply. In 21 for example when the housing stock deficit stood at 1,2, houses, only 199, units of houses were built (Afrane et al, 216). In rural areas, insecure land tenure often manifests itself in lacking agricultural investment. That is, farmers appear to be less willing to make long-term investments, in for example, tree crops. Soil investments, such as extended fallow periods, are also curtailed in light of insecure land tenure (Goldstein & Udry, 28). In the wider international context, tenure insecurity particularly discourages investments by multinational companies in Ghana and thus the national economy forgoes potential positive effects from additional job creation and technology transfer (Barthel et al, 211). 5

6 It has become clear in Ghana that the existing customary tenure systems mainly administered by Chiefs and family heads is a crucial building block to resolve the challenges related to insecurity of land tenure. To bring clarity and transparency into the customary land rights institutions 38 Customary Land Secretariats have been set up as part of a strategy to help streamline and address some of these challenges being faced by the sector. These secretariats are intended to improve the efficiency of record keeping by managing land allocations and transactions within the customary setting (Biitir & Nara, 216). It is noteworthy that although registration of land titles has been enabled in Ghana for nearly three decades, very few land titles have been issued. As of 26, only 42, registration applications had been submitted and of these a mere 3 percent had been granted (Kuntu-Mensah, 26). This situation suggests that there are impediments preventing progress in Ghana s attempt to give titles to land owners and users. In the light of this, the Government of Ghana and other development agencies have undertaken several interventions in the last decade to remove some of the barriers which are preventing progress and improve the title registration processes (Jones-Casey, 211). Despite these reforms being undertaken by the government, the lack of transparency and lack of institutional commitment has remained and the system of land administration and registration is still relatively weak. Other private sector participants, non-governmental organizations (NGOs) and bilateral partners have also initiated programs to speed up and improve the titling process on pilot bases. Notable among these efforts is the MiDA program that targeted a comprehensive pilot titling program in the Central Region of Ghana (Jones-Casey & Knox, 211). The experiences and outcomes of these interventions especially those by MiDA, which is at the center of this evaluation, should guide future decisions on which approaches appear to be most beneficial, scalable and practical in the context of the existing legal framework and land administration system. As of 211 the MiDA Land Titling Project succeeded in issuing land titles to about 27 parcels (Jones-Casey & Knox, 211). The nature of the project and the innovations which accompanied this pilot program makes it quite attractive for scaling up. The fact that the program created a new registration district and an office with state of the art equipment for land data collection, processing 6

7 and storage removes some of the bottlenecks associated with previous attempts. The program also created an incentive structures to nudge officials to maintain a constant workflow. Finally, the program facilitated negotiations with Chiefs and family heads who hold allodial titles to land in order to ensure consent for the issuance of titles. The process of land titling has ten steps which include: filling out land registry forms; lodging the application; issuing acknowledgement notes; entering certificate numbers; drafting certificates; typing and checking land registers; preparing files for signing by the appropriate Land Administration Director; plotting parcels and binding certificates. The MiDA intervention made it possible to undertake all these processes within 31 days a benchmark which was hoped to be adopted elsewhere in the country. The aim of this evaluation is to assess how effective this intervention was during the early stages of implementation in terms of benefitting households in the targeted areas measured by a wide array of outcomes related to investment, asset ownership, agricultural production and welfare. 3. Data and estimation 3.1 Data This impact evaluation mainly builds on household-level panel data that was collected over a period of five years from both households that were targeted by the MiDA land titling pilot intervention as well as those households located just outside the intervention area. In total, there were three survey waves conducted in 2 communities located around a main road that divides many of the sample communities into two halves. This road forms a loop (see Figure 1). Households within the loop (area shaded blue in Figure 1) were eligible to participate in the MiDA land titling pilot while the households outside of the loop were not eligible in the first phase of the program. The physical demarcation of the road dividing communities into two, forms the bases for our empirical evaluation strategy. In a nutshell, households located within the loop of the road and not more than 2 meters away from the road (marked as A in Figure 1) qualify as the group in our study. On the other hand, we 7

8 sampled from two sets of control groups: a short term control group and a long term control group. At the time of project initiation, it was anticipated that households just outside the loop would be the most likely future recipients of the land titling assistance. We refer to households outside of the loop but within 2 meters of the road as the short term control group (marked as B in Figure 1). Households outside the loop and located more than 5 meters away from the road constitute the long-term control group as these household were unlikely to benefit from the MiDA pilot programme during the evaluation phase (marked as D in Figure 1). Consequently, there was a buffer of about 3 metres between the short-term and long-term control groups (marked as C in Figure 1). From each of these three groups all households were chosen for the first round of data collection. The locations of sample household around the round used to distinguish between and control households is also shown in Appendix Figure 1. Figure 1: Sampling and data collection design 8

9 Box 1: The survey instruments consisted of seven modules Module 1: Seeks the household s consent to participate in the data collection exercise and also captures information on the household roster and members. Module 2: This module is administered to individuals and is designed to gather basic demographic information on household members, on employment and the different sources of income for the male spouse only. Module 3: These modules are designed to collect data on assets owned by the household and individuals within the household. The main sub-sections include tools, durable goods, farm assets, as well as financial assets. Module 4: The purpose of this section is to collect data on the household's agricultural activities. It covers agricultural assets such as land, livestock and equipment. Furthermore, it provides data on agricultural production technology and processing, marketing, input use, output and incomes. Module 5: This module is designed to gather information on employment, time use and different sources of income for the household head and spouse either as individual owners or jointly owned businesses which are not farm based. Module 6: In this module the goal is to appraise the level of financial knowledge among respondents. Module 7: The marital module aims to understand the relations between husband and wife/wives and to study how households in Ghana function. The first of the three surveys was conducted in 21; the second in 211; and the third in Therefore, all three survey waves were collected after the land titling intervention was initiated in 29. The first two survey waves allow us to understand the short-term effects of the land titling program immediately after program initiation and the third survey wave is informative about the medium-term program impacts. Each of the three survey waves covered seven specific areas of interest. These areas included demographic characteristics of sampled households; paid employment engaged in by household members; individual and household assets; agricultural production and land titling; non-farm 4 Although the entire survey in the third round were stipulated to be undertaken in the year 214, not all the targeted households could be contacted for interviews in 214. Thus in the first half of 215, a tracking exercise was undertaken to mob-up as many households as possible which were missed in the 214 episode of data collection. 9

10 enterprises; marital history of household heads and spouse(s); and financial literacy training. The survey team also collected data on household and plot locations using Global Positioning Systems (GPS). During the first survey round in 21, a sample of 2,45 households was interviewed which represents households from the (79), short-term control (862) and long-term control (798) groups. The second survey round in 211 reached a total of 2,99 households in the (693), short-term control (724) and long term control (682) groups. Finally, the third and final survey round was collected in 214 and a total of 1,714 households were traced in the (553), short term control (619) and long term control (542) groups. Box 2: Field Work During the first round of field survey in 21, a total of 65 enumerators were trained over a period of six days. Out of this number, 54 enumerators were selected for field work. The selection was based on the outcome of a test conducted to examine enumerator competence with regards to the questionnaire administration. In addition, a language fluency examination was also undertaken to ensure that enumerators had command over the Twi Language which was going to be the main medium of communication with respondents. From the set of 54 enumerators, three working teams of 18 members were created. Each team consisted of a Supervisor, a Field Editor, 2 Plot Mapping Experts and 14 Enumerators. The Supervisor was the team leader and was responsible for overseeing, monitoring and, where necessary, correcting the work of the interviewers and the Field Editor. The Enumerators conduct daily interviews with the head and spouse of sampled households. The Plot Mapping Experts were responsible for demarcating boundaries within which enumeration should be conducted based on the three terms and also to map or take waypoints of plot location (, short term and long Term). A similar strategy was adopted in the second and third rounds of survey. This means that between and Round 2, 351 households representing about 14 percent of the original sample could not be reached. Furthermore, between and Round 3, a total of 736 households which represented about 3 percent of the initial sample could not be tracked. This resulted in a one-year and four-year tracking rate of above 85 percent and 7 percent respectively 5. 5 A randomized control trial in Uganda which sought to examine women empowerment over a period of four years and with an initial sample of 5,966 adolescent girls, achieved a two-year tracking rate of 82 percent and a four- 1

11 The first panel of Table 1 presents the sample frequencies by and round of survey. The second panel presents the balanced sample across the three rounds of survey by status. A similar presentation is made for plots in Table 2. Next, we examine whether specific household characteristics as well as the status are important determinants of attrition. To do this we create a dummy variable for attrition which equals 1 if the household form a panel or does not attrit (between and Round 2/3) and if they attrit. We regress this dummy variable on combinations of the household characteristics and the status. In Table 3 and Table 4, we present estimates from OLS regressions in columns 1 to 5 and estimates from a Probit regression in column 6. Based on the results of the OLS and Probit regressions shown in Table 3, we find that between survey and Round 2, household size, age of household head and affected a household s likelihood of forming part of the panel significantly. Treatment households have a 1.1 percentage point (pp) higher probability to remain in the panel between survey Rounds 1 and 2. A unit increase in household size also increases the probability of the household remaining in the panel by close to 1pp and an increase of the household head s age decreases the same probability by.1pp. Similarly, comparing the group only to the short-term control group, Column (3) indicates that the probability of panel inclusion is 1.5pp higher for households in the group. However, there is no difference in the likelihood of panel inclusion between the short-term and the long-term control group. Table 4 shows the results from the equivalent regressions for panel inclusion focussing on Rounds 1 and 3 only. The results indicate that between and Round 3, the age of the household head, household size and an indicator for whether the household head has ever attended school affect the likelihood of panel inclusion. Contrary to the analysis between and Round 2, a year increase year tracking rate of percent (Bandiera et al, 215). Similarly, in a study investigating the impacts of education a sample of female youth were followed-up after a period of four years and 81.6 percent were tracked (Friedman, et al., 211). Also, when the impact of education on health outcomes were followed up after seven years in Kenya, 55 percent of the initial sample of 19,289 boys and girls contacted in 23 were tracked in 21 (Duflo et al, 214). These studies suggest that our own tracking rate of 86 percent after one year and 7 percent after four years compares well. 11

12 in the household head s age now increases the likelihood of a household to form part of the panel. Consistent with the findings from before, larger households increase are more likely to form part of the panel between and Round 3 (by nearly 2pp per unit increase). Households whose heads have ever attended school were also about 3pp more likely to remain in the panel compared to those whose heads have never attended school. Treatment negatively affects the likelihood of panel inclusion when the group is compared to only the short term control group. In contrast, there is no statistical difference between the and the long term control group as can be seen from column 2 in Table Empirical Approach Our empirical work is based on the natural experiment generated by the allocation of households into and control based on their location on one side or the other of the road which divides the communities in our sample into two halves. We focus on impacts using intention-to-treat (ITT) estimates. That is, although the program targeted all households in the geographic that were eligible for land title issuance, participation was entirely of voluntary nature households had to decide whether to participate or not. Not all the households with parcels in the area participated. In addition, not all households or individuals who possessed land in the area were able to negotiate approval for the titling process with local authorities such as chiefs. There were others whose interest in the land did not exceed the minimum of three years required before parcels could be issued. Given that these issues are likely to involve significant degrees of endogeneity, we will rely on estimating the impact of the intention to title in this report. The particular nature of program implementation, whereby a major road separates and control groups allows for a rigorous impact assessment an RDD methodology. The key identifying assumption that needs to be maintained is that conditional on community fixed effects, unobserved determinants of the outcomes we measure are on average the same for households on either side of the road. Of course it can be argued whether this assumption is valid. Therefore, it is worth noting that the road was used as a convenient and apparently arbitrary boundary to delineate an area in which to pilot 12

13 this systematic land title registration intervention. Indeed, as program implementation started, it was discovered that a significant number of chiefs of the study villages actually lived outside of the area. This observation provides us with confidence that the boundary choice (conditional on community) can be treated as exogenous. We concentrate on estimating program impacts for the sample of households located closest to the road, i.e. households located within 2 meters on either side of the road. Consequently, all households are located close to the boundary. We include community fixed effects in all specifications, thus limiting comparisons of outcomes to households on either side of the road within a single community. Appendix Figure A1 makes the identification assumption clear. Almost each cluster of points around the project boundary represents a single community. We compare outcomes of households within one of those communities and inside the boundary with those of other households in the same community but outside the boundary. These FE OLS estimates present the most robust impact estimates in our view. However, it is possible that even within a single community, households which are closer to one another in geographical terms face similar geographically determined conditions and shocks. Thus clusters of households are potentially affected by the same (unobservable) influence factors. In order to address these local correlations, and improve estimation efficiency we use a spatial autoregressive (SAR) estimation specification as a robustness check for our main impact estimates. Moreover, there may be omitted variables which are correlated within neighborhoods. Hence, we also plan to use spatial fixed effects estimates to control in a further iteration of this report. Ordinary Least Squares (OLS) Our basic OLS specification to examine the impact on the variable y it for household i at time t is: y it = α + βx i + τ 1 treat i time 1 + τ 2 treat i time 2 + τ 3 treat i time 3 +μ 1 time 2 + μ 2 time 3 + θv i + ε it (1) 13

14 Where X i denotes the characteristics of the household at. 6 treat i is an indicator variable that equals one if the household is located in the are which is demarcated by the main road dividing each community. time represents a dummy variable capturing the year of the survey round in which the observation is measured. Finally, V is a vector of village dummy variables that will control for local, time-invariant effects. At the centre of our interest, of course, is τ t which denotes the coefficient that measures the ITT impacts of the MiDA Land Titling Pilot for each survey round separately. The standard errors are clustered by village. Spatial Autoregressive Regression (SAR) The first step of our second approach is to define the nature of expected spatial interdependencies among households. There can be three different interaction effects: (i) endogenous - where the dependent variable of one household depends on the dependent variable of another and vice versa; (ii) exogenous - where the dependent variable of a unit depends on the independent variable of other units; and (iii) interactions are among the error terms of nearby units. In our present analyses we concentrate on the first type of interactions that takes into account potential spill overs or correlated shocks in outcome variables as a result of proximity between households. The SAR is based on the OLS specification in equation (1) and introduces an additional term which accounts for endogeneity among outcomes of spatially distributed households based on their proximities with the use of a spatial weights matrix. Equation (1) becomes: N y it = α + j=1 ρw ij y jt + βx ij + τ 1 treat i time 1 + τ 2 treat i time 2 + τ 3 treat i time 3 +μ 1 time 2 + μ 2 time 3 + θv i + ε it (2) 6 More specifically, the baseline control variables include household size, number of female household members, number of householder member aged 5 or younger, an indicator for male household head, the age of the household head and whether the household has ever attended any school. 14

15 Where all variables are defined as before except w ij which are elements of a spatial weight matrix W, which forms together with the y jt the endogenous interaction effects. W is an N by N inverse distance matrix or an inverse distance matrix with a certain cut-off point. 7 ρ is referred to as the spatial autoregressive coefficient and, essentially, our model would reduce to an OLS if ρ were to equal zero. 8 In estimating equations (1) and (2) we limit the sample to households to those located in a 2 meter band to the main road. This sub-sample is further refined by mostly focusing on panel households. Panel households are those households which we observe in all three survey rounds. To account for the possible imbalances caused by differential attrition between and control households we also employ inverse probability weighting (IPW) as a robustness check. 4. Results and discussions In this section we present and discuss our main results for outcomes measured at the household level. Several outcomes, however, are reported by different individuals within households which allows us to explore gender-differentiated impacts of land titling. We begin by providing descriptive statistics of all variables which are used in our empirical estimations Descriptive statistics In Appendix Table 1 we present basic descriptive statistics for each variable of interest. The table is too extensive to be discussed in detail. The control variables used in estimating Equation (1) and (2) are listed in Table 5 partitioned by the status and the difference between and control for. Statistical significance of the difference is indicated where *** denotes significance at 1 percent, ** at 5 percent, and * at 1 percent. We do not expect the program to have impacts on these variables (especially in ). However, the significant difference in the number of female 7 The distance decay can also be formulated as a power function and the elements of W matrix can be inverse distances with a distance decay factor, γ, = 1 d ij, where d ij denotes the distance between units i and j. 8 Eventually, we plan on adding a specification that builds on spatial fixed effects (Conley & Udry, 21). 15

16 household members between the and the control group indicates that the two groups may differ in basic observable characteristics and underlines the importance of controlling for these variables in the regression Impact analysis In Table 6 and 7 we present estimation results of the Land Titling Program in Ghana using the natural experiment. All tables follow a similar format. Column (1) provides the mean of the dependent or outcome variable of interest in the control group. The number of observations used in the regression is reported in Column (2). Columns (3) through (5) report the impact estimates for each survey round separately which correspond the coefficients on the interaction terms in Equation (1). Columns (6) through (8) report the equivalent estimates using the SAR specification from Equation (2). We consider these estimates as a robustness check. The stated coefficients capture the ITT impact estimates for the program. Again, *** denotes significance at 1 percent, ** at 5 percent, and * at 1 percent level. In the discussion, we focus our attention on outcome variables that are intended to directly capture land investments and to investigate whether the land titling intervention had any impacts on asset ownership. First, we concentrate on outcomes related to land owned by women. A first-order condition for program impacts to materialize and to validate the evaluation strategy is that the program reached the target population. Table 6a shows that on various metrics, the intervention reached a significantly higher proportion of households with plots owned by women in the group relative to the control group. Across all three survey rounds, there is a consistent and significant impact of the program on all three main indicators of program implementation: land registered, measured and demarcated. These strong indications for successful program implementation and targeting, however, do not appear to translate into immediate measures of land investment such as the construction of buildings or fallowing of land as can be seen from Table 6b. However, women are more likely to have obtained the land through purchase in the short run () which may be seen as improved female land market participation. 16

17 In terms of outcomes capturing asset holdings shown in Table 6c, women appear to benefit from the land titling intervention through accumulating higher stocks of durable assets at least in the long run, i.e. in Round 3. This effect is particularly strong, in terms of the level effect, when concentrating on the winsorized monetary amount in cases where a female respondent was present which constitutes our preferred outcome measure. 9 However, other asset holdings do not seem to be affected. Instead, however, women appear to decrease the value of their outstanding loan in response to the program as shown in Table 6d. This effect is restricted to the intensive margin, i.e. women reduce the amount of currently outstanding loans they need to repay while they do not appear to be less likely to have an outstanding loan in the first place. These impacts can only be observed in the short-run, i.e. in and in Round 2. The data also hints at an increase on current lending in, as shown in Table 6e, even though this impact disappears when the outcome variable is winsorized. Finally, there is weak evidence that women decrease their saving stock in the longer term. Table 6f indicates that total savings decrease in Round 2 for the un-winsorized variable. If anything, this effects seems to stem from a decrease in savings held in more formal financial institutions. Turning to men, we see equivalent strong impacts on land registration, demarcation and measurement in Table 7a even if the impacts on land registration in Round 2 statistically do not differ from zero. Similarly, we also do not observe any strong marker for direct investments into the land such as planting trees or construction. However, there is an interesting pattern of impacts on the self-assessed total value of land as can be seen from Table 7b 1. While these self-evaluations decrease in, there is a sizable increase in Round 3. Contrary to women, men are not more likely to have obtained the land through purchases but instead are more likely to have accessed land through renting and free allocations in the short run and inheritance in the long run. There is a stark increase in livestock ownership for men which can be observed for both the extensive and the intensive margin in Round 3 (Table 7c). This finding stands in contrast to the decrease in the 9 Winsorizing refers to the limiting of extreme values at the top and at the bottom of the distribution. In general, we winsorize at the 1 and 99 percentile. 1 We had a surveyor value a small number of parcels in round 3. We are in the process of comparing those values with reported values of the land holders. 17

18 value of livestock in the short run in. Both the decrease as well as the increase in the total value of livestock is substantial. In contrast to the women, men appear to increase current borrowing (Table 7d). This is true for the likelihood of currently having outstanding credit (in Round 2) as well as for the value of current outstanding credit (in Round 3). Hence, both of these effects appear to be at play in the longer term. Interestingly, this finding is paired with an enormous increase in outstanding lending in the short run (shown in Table 7e) and in increase in savings in the long run (shown in Table 7f). 5. Future work The analysis so far has revealed a number of interesting patterns with substantial gender differences in the data. In the next iterations of the analysis, it will be important to expand the breadth as well as the depth of dataset where possible especially in terms of outcomes. In particular, preliminary results show that women are more likely to be engaged in self-employment activities and are able to increase business profits in response to. These impacts are achieved despite the fact that households agricultural production remains constant. These findings are not yet reflect in the analysis presented in this draft. In addition we plan on refining some of the estimation methodologies such as using inverse probability weighting (IPW) in order to account for the possible imbalances caused by differential attrition between and control households as a robustness check as well as dealing with spatial correlations between outcomes by using, for example, spatial fixed effects regressions.. 18

19 Table 1: Unbalanced and balanced household samples by and wave of survey Round Treatment Short Term Long Term Total All Households 1 (21) ,45 2 (211) ,99 3 (214/15) ,714 Total observations 2,36 2,25 2,22 6,263 Panel Households 1 (21) ,76 2 (211) ,76 3 (214/15) ,76 Total observations 1,647 1,845 1,626 5,118 Source: Generated from the Ghana Land Titling Datasets, 21, 211 and 214 Table 2: Household plots and wave of survey (Frequency) Time Treatment Short Term Control Long Term Control Total All Plots 1 (21) 1,13 1,299 1,149 3,578 2 (211) 1,185 1,314 1,247 3,746 3 (214/15) 984 1, ,124 Total observations 3,299 3,763 3,386 1,448 Source: Generated from the Ghana Land Titling Datasets, 21, 211 and 214/15 19

20 Table 3: OLS and Probit regression for attrition between 21 and 211 (dependent variable: panel inclusion) (1) (2) (3) (4) (5) (6) OLS 1 OLS 2 OLS 3 OLS 4 OLS 5 dprobit Treatment.11*.11*.11* (.1) (.6) (.6) Number of household members.7**.7** (.3) (.3) Number of female household members (.4) (.4) Number of household members (below years of age) (.5) (.5) Male household head.1.1 (.7) (.7) Age of household head -.1* -.1* (.) (.) Household head attended any school (.7) (.7) Treatment v. long term control.8 (.8) Treatment v. short term control.15** (.7) Short term control v. long term control -.7 (.7) Constant.939***.943***.936***.943***.939*** (.3) (.5) (.5) (.5) (.13) Observations 6,262 4,57 4,24 4,227 6,253 6,253 R-squared Notes: *** denotes significance at 1% level; ** at 5%; and * at 1%. The dependent variable is a dummy that is equal to 1 if a household attrit between the baseline and endline surveys and otherwise. The standard errors in parenthesis are clustered by community Source: Generated from the Ghana Land Titling Datasets, 21, 211 and 214 2

21 Table 4: OLS and Probit regression for attrition between 21 and 214 (dependent variable: panel inclusion) (1) (2) (3) (4) (5) (6) VARIABLES OLS 1 OLS 2 OLS 3 OLS 4 OLS 5 dprobit Treatment (.15) (.1) (.1) Number of household members.18***.19*** (.5) (.5) Number of female household members (.7) (.7) Number of household members (below 5years of age) (.8) (.8) Male household head (.12) (.12) Age of household head.1***.1*** (.) (.) Household head attended any school.29***.3*** (.11) (.11) Treatment vs. long term control.8 (.12) Treatment vs. short term control -.29** (.12) Short term control vs. long term control.36*** (.12) Constant.823***.84***.84***.84***.694*** (.6) (.9) (.8) (.9) (.21) Observations 6,262 4,57 4,24 4,227 6,253 6,253 R-squared Notes: *** denotes significance at 1% level; ** at 5%; and * at 1%. The dependent variable is a dummy that is equal to 1 if a household attrited between the baseline and endline surveys and otherwise. The standard errors in parenthesis are clustered by community Source: Generated from the Ghana Land Titling Datasets, 21, 211 and

22 Table 5: Mean comparisons of variables of control variables in All households within 2m bandwidth (panel) Topic Variable N Treatment mean Control mean Difference Household 1, characteristics Number of household members (.158) Number of female household members 1, * (.8) Number of household members (below 5years of age) 1, (.41) Male household head 1, (.31) Age of household head 1, (.892) Household head attended any school 1, (.29) 22

23 Table 6a: Program impacts on female respondents Topic Variable Round 1 mean (control group) N Plot/land characteristics (female) Any land registered Any land registered, conditional on female plot existing Proportion of plots registered Proportion of plots registered, conditional on female plot existing Land measured Land measured, conditional on female plot existing Proportion of plots measured Proportion of plots measured, conditional on female plot existing Plot demarcated Plot demarcated, conditional on female plot existing Proportion of plots demarcated Proportion of plots demarcated, conditional on female plot existing All households within 2m bandwidth (panel) OLS Round 2 Round 3 SAR Round 2 Round 3 (1) (2) (3) (4) (5) (6) (7) (8). 1, ***.176***.27***.135***.175***.26*** (.39) (.43) (.48) (.335) (.34) (.349). 1, ***.176***.27***.136***.175***.26*** (.39) (.43) (.48) (.335) (.34) (.349).54 1,299.13***.147***.179***.129***.147***.178*** (.38) (.33) (.47) (.38) (.312) (.321).54 1,298.13***.147***.179***.129***.147***.178*** (.38) (.33) (.47) (.38) (.312) (.321).5 1,6.46***.246**.165**.383***.263***.186*** (.96) (.93) (.73) (.467) (.477) (.42).5 1,6.46***.246**.165**.383***.263***.186*** (.96) (.93) (.73) (.467) (.477) (.42).47 1,6.338***.28**.173**.322***.225***.186*** (.82) (.9) (.62) (.431) (.442) (.388).47 1,6.338***.28**.173**.322***.225***.186*** (.82) (.9) (.62) (.431) (.442) (.388) ***.292**.193**.552***.295***.197*** (.6) (.18) (.69) (.567) (.548) (.671) ***.292**.193**.552***.295***.197*** (.6) (.18) (.69) (.567) (.548) (.671) ***.271**.174**.522***.272***.175*** (.6) (.13) (.7) (.5) (.54) (.661) ***.271**.174**.522***.272***.175*** (.6) (.13) (.7) (.5) (.54) (.661) 23

24 Table 6b: Program impacts on female respondents Topic Variable Round 1 mean (control group) N Plot/land characteristics (female) Number of female plots, conditional on female respondent present Any female owned plot, conditional on female respondent present Worried to lose plot if left empty, conditional on female plot existing Any disagreement ever over this plot, conditional on female plot existing Any part of plot fallowed, conditional on female plot existing Trees planted in past year, conditional on female plot existing Any structures on land, conditional on female plot existing Any improvements to structures Any improvements to structures, conditional on structures existing Any improvements to structures, conditional on female plot and structures existing Any chemicals applied in past major season Any fertilizer applied in past major season, conditional on any chemicals answered All households within 2m bandwidth (panel) OLS Round 2 Round 3 SAR Round 2 Round 3 (1) (2) (3) (4) (5) (6) (7) (8).896 2, (.61) (.56) (.68) (.583) (.583) (.561).653 2, (.34) (.21) (.26) (.275) (.274) (.264).326 1, (.74) (.35) (.54) (.45) (.412) (.441).42 1, (.31) (.37) (.33) (.249) (.251) (.269).195 1, (.4) (.44) (.37) (.339) (.339) (.363).184 1, ** (.49) (.47) (.58) (.37) (.368) (.491).631 1, (.) (.48) (.39) (.424) (.429) (.46) (.78) (.66) (.113) (.512) (.496) (.549) (.82) (.66) (.113) (.524) (.497) (.551) (.82) (.66) (.114) (.524) (.497) (.551).178 1, (.48) (.22) (.5) (.397) (.32) (.345) , ** *** (.31) (.19) (.3) (.223) (.179) (.192) Any herbicides applied in past major season, conditional on any.75 1, chemicals answered (.34) (.2) (.47) (.3) (.286) (.38) Any other chemicals applied in past major season, conditional on any chemicals answered.6 1,

25 Topic Variable Round 1 mean (control group) N Value of land [in cedis] Value of land [in cedis, winsorized] Value of land, conditional on female plot existing [in cedis, winsorized] Plot was obtained through purchase, conditional on female plot existing Plot was obtained through inheritance, conditional on female plot existing Plot was obtained through renting, conditional on female plot existing Plot was obtained through sharecropping, conditional on female plot existing Plot was obtained through (free) allocation, conditional on female plot existing All households within 2m bandwidth (panel) OLS Round 2 Round 3 SAR Round 2 Round 3 (1) (2) (3) (4) (5) (6) (7) (8) (.1) (.7) (.6) (.886) (.77) (.763) 2,23 1, , ** 2.5 (671) (2,36) (1,452) (1349.1) (1394.1) (1488.1) 2,175 1, , * (553) (1,65) (1,51) (175.2) (1111.2) (1185.9) 2,215 1, , * (541) (1,9) (1,493) (185.1) (1113.9) (1188.9).289 1,228.78* * (.39) (.47) (.79) (.447) (.464) (.534).357 1, (.67) (.52) (.8) (.445) (.462) (.531).188 1, (.32) (.33) (.23) (.322) (.335) (.385).12 1, (.27) (.35) (.2) (.237) (.246) (.283).244 1, (.41) (.86) (.5) (.387) (.42) (.464) 25

26 Table 6c: Program impacts on female respondents Topic Variable Round 1 mean (control group) N Individual assets (female) Owns any livestock, conditional on female respondent present Value of livestock, conditional on screening question asked [in cedis] Value of livestock, conditional on screening question asked [in cedis, winsorized] Value of livestock, conditional on screening question asked and female respondent present [in cedis] Value of livestock, conditional on screening question asked and female respondent present [in cedis, winsorized] Value of tools, conditional on screening question asked [in cedis] Value of tools, conditional on screening question asked [in cedis, winsorized] Value of tools, conditional on screening question asked and female respondent present [in cedis] Value of tools, conditional on screening question asked and female respondent present [in cedis, winsorized] Value of durable goods, conditional on screening question asked [in cedis] Value of durable goods, conditional on screening question asked [in cedis, winsorized] Value of durable goods, conditional on screening question asked and female respondent present [in cedis] Value of durable goods, conditional on screening question asked and female respondent present [in cedis, winsorized] All households within 2m bandwidth (panel) OLS Round 2 Round 3 SAR Round 2 Round 3 (1) (2) (3) (4) (5) (6) (7) (8).364 2, ** (.55) (.55) (.66) (.356) (.353) (.345) , ** (11.2) (27.8) (69.4) (38.85) (38.74) (37.21).1 2, (8.82) (19.6) (4.6) (21.53) (21.49) (2.62) , ** (11.5) (27.5) (67.8) (39.22) (38.87) (38.16).9 2, (9.13) (19.3) (38.) (21.4) (21.23) (2.82) , *** (1.14) (.923) (5.67) (2.972) (2.874) (2.761) , *** (.963) (.569) (3.5) (1.797) (1.738) (1.67) , *** (.98) (.569) (3.26) (1.845) (1.775) (1.742) , *** (1.15) (.921) (5.95) (3.16) (2.92) (2.848) , *** *** (14.4) (19.6) (6.4) (43.65) (43.5) (41.95) 272 2, * 393* ** (226) (36.8) (195) (176.5) (175.8) (168.8) 76. 2, *** *** (14.4) (19.1) (62.4) (46.28) (45.89) (45.24) 277 2, * 47* ** (228) (37.1) (21) (179.3) (177.8) (174.4) 26