4.1 Dependent variables

Our key dependent variable is farm net revenue, defined as the revenue from all crop typesFootnote ¹⁰ grown on a farm in a given year, net of the costs of fertilizer and manure. It is therefore an application of the Ricardian model where net revenues are used in place of land values, to capture the economic effect of climatic factors in agriculture (Deressa and Hassan, Reference Deressa and Hassan2010). The use of net revenue is particularly appropriate in this set-up as land markets are in fact not perfect.

The analysis also considers the adoption of soil and water conservation (SWC) as an additional dependent variable. Soil bund, terrace (local or modern) and tree planting are considered in the analysis. The purpose of these structures is to help collect surface runoff and increase water infiltration and prevent soil erosion. When used as a combination within a field, they help to increase the proportion of arable land by levelling steep land, reducing runoff and preventing erosion. Such measures are also said to be effective along with other sustainable land management practices, as promising adaptation strategies, especially in the high-rainfall areas (Kato et al., Reference Kato, Ringler, Yesuf and Bryan2009).

Farmers practice cut-off drains as a way of conserving moisture and preventing runoffs. In the study areas, cut-off drains are dug across a slope to intercept surface runoff and carry it safely to an outlet such as a canal or stream. They are used to protect cultivated land, compounds and roads from uncontrolled runoff, and to divert water from gully heads. Soil bunds are common types of SWC measures, but bunds are susceptible to heavy rainfalls and easily eroded by water and wind. As the bunds tend to be moisture- and nutrient-rich, some farmers destroy the soil bunds and disperse the accumulated soil over the whole field after a while. Other farmers also maintain their bunds and cultivate crops on them. The most dominant form of SWC activities in the sample is tree planting, which also both conserves moisture and prevents soil erosion.

4.2 Certification variables: kebele and household level

Implementation of the certification program was initiated simultaneously in all woredas. Although the simultaneous implementation was the initial plan, it was recognized that such action would stretch financial and manpower resources both at the kebele and the woreda levels, so a plan was set to cover all lands in the region within three years. The program reached some of the woredas within the Amhara region starting in 2004, and was eventually rolled out to the entire region by 2007 (Adenew and Abdi, Reference Adenew and Abdi2005). Furthermore, the slack agricultural seasons were chosen to initiate the program due to the availability of the famers then (as opposed to during the busy wet seasons) and the conduciveness of the dry-weather roads to carry out fieldwork by the officials. Preparation, completion and distribution of certificates were carried out by the officials during the peak agricultural season. The certification program was implemented in these villages sequentially.Footnote ¹¹ The discontinuity created by agricultural seasonality and road conditions along with the gradual roll-out created a geographic and a time gap in the implementation in the different kebeles (Deininger et al., Reference Deininger, Ali and Alemu2011; Bezabih et al., Reference Bezabih, Holden and Mannberg2016).Footnote ¹²

Hence, the construction of the certification variable was such that villages and households reached by the program earlier are categorized as certified while those enrolled in the program at later stages are considered control. The last round of the survey, conducted in 2007, was designed to gather information on the features of the certification program that enable analysis of the impact of the program on different variables of interest. We use two measures of certification to identify the program impacts: a kebele- and household-level certification. A kebele is considered certified if the initial demarcation of fields in the kebele by the LAC was completed 12 months before the last round of the survey in 2007 (Deininger et al., Reference Deininger, Ali and Alemu2011). The essence of this is to ensure that sufficient time has elapsed for the effects of the program to be realized.

The exact timing of the receipt of the land certificate depends not only on the time when the land demarcation activities started but also on the time when the land certificates were issued to households.Footnote ¹³ Since the date of issuance of the land certificate can be arguably exogenous to household characteristics, there will be variation across households on the date of receipt of certification. This variation could be the source of identification. In other words, household-level certification enables comparison of households that have received their land certificates on time to change their agricultural decisions measured during the last round of the panel survey to their peers that had not received their certificate on time. Hence, the appropriate level of certification that will be analyzed is household (instead of kebele). As shown in table 1, 43 and 58 per cent of the observations are certified at the kebele and household levels, respectively.

4.3 Rainfall and temperature variables

The monthly rainfall and daily temperature data were obtained from the Ethiopian Meteorology Authority, in eight stations close to the study villages (kebeles) for the years 1976 to 2006. To convert the station level observations into farm-level observations, an inverse distance weighting interpolation method, which combines latitude, longitude and station-level rainfall and temperature observations, was employed to impute farm-specific information. The interpolation leads to varying observations at the household level. While a farm is typically comprised of different plots located in different places, the plots are also typically closer to the house as the house will be built in a manner that it is close to the plots. Therefore, we have taken the location of the farm to be the location of the household. While this could be a strong assumption in some instances, the GPS information is taken for the household and not for the different plots and, therefore, the interpolation is based on information on the household.Footnote ¹⁴

The interpolated observations are used to compute annual mean and standard deviations from which rainfall and temperature anomalies are computed We use the measurement of anomalies-deviations of current observations from the long-term mean (O'Loughlin et al., Reference O'Loughlin, Witmer, Linke, Laing, Gettelman and Dudhia2012). The rainfall anomaly values are then computed as the ratio of the difference between the short-term (annual) and the long-term mean rainfall to the standard deviation of the long-term rainfall. In terms of range, three cut-off points are used to construct the anomalies: −1.5 to 1.5, −1 to 1, and −0.5 to 0.5. Accordingly, for the first cut-off, the negative rainfall anomaly dummies are constructed as anomalies below −1.5 and the corresponding positive rainfall anomaly dummies are constructed as anomaly values above 1.5. These dummies represent adverse climatic conditions. The same procedure is followed for the temperature anomaly dummies. Based on anomalies, we also compute the censored continuous measures, which are constructed as 0 to positive and from 0 to negative values.

4.4 Socioeconomic variables

The key socioeconomic characteristics used in the analysis are presented in table 1. Around 17 per cent of the surveyed households have a female head. The average age of the household head is 46. The proportion of household heads that can read and write is 58 per cent. An average household has two male adults of working age. Livestock ownership, measured in tropical livestock units, is around six per household. The average farm size is 1.64 ha, slightly higher than the national average of 1.1 ha.

Table 1 also presents a mean difference test for the variables used in the analysis. The test results show that the difference between the treatment and control groups is insignificant for many of the variables. The temperature anomaly variables, however, show significant mean difference between the two groups.

5.1 Generating comparable groups based on observables – PSM

The main purpose of PSM is to choose a sample of households in the treatment and control groups that are as similar as possible to each other in important, observed, aspects. More specifically, PSM balances the observed distribution of covariates across households in the two groups such that the conditional probability of participation in the program (i.e., receiving a certificate) should be equal in the treatment and control groups. The difference in this conditional probability is called the ‘propensity score’ and is the relevant summary statistics to ensure that the samples are sufficiently balanced (Rosenbaum and Rubin, Reference Rosenbaum and Rubin1983).Footnote ¹⁶ The matched sample so formed is then used in the DID analysis that is described below.

5.2 The difference-in-difference analysis

Our empirical strategy on the interplay between rainfall and land certification on farmers’ welfare relies on: (i) the gradual rollout of the land certification program to farm households at the village level, (ii) the exogenous variation in rainfall and temperature anomalies, and (iii) the panel nature of our dataset. We estimate a model where the identification of the effect of tenure security on welfare comes from comparing reported net revenues in year 2005 (before the implementation of the land certification) and in year 2007 (after the implementation of the land certification) in certified and non-certified villages and its interaction with rainfall anomalies. The measurement of anomalies of both temperature and rainfall enable us to capture non-linearities by creating dummies representing positive and negative deviations of short-term observations from the long-term temperature mean. Accordingly, the relationship between a farm level revenue, temperature and rainfall anomalies is given by:

(1)\begin{equation}\ln {y_{ht}} = {\theta _{ht}} + \beta {x_{ht}} + \gamma {r_{ht}} + \vartheta {w_{ht}} + \phi {l_t} + \vartheta {p_k} + \delta {l_t}{p_k} + {\varepsilon _{ht}}.\end{equation}

The log of farm revenue from farm household $h\;$in year $t\;$is denoted by ${y_{ht}}$. ${r_{ht}}\;$ represents a vector of temperature anomaly dummies for each household h in year $t$. Similarly, the rainfall anomaly dummies are represented by ${w_{ht}}$ while ${x_{ht}}$ are the socioeconomic and physical farm characteristic variables. The treatment variable ${p_k}$ is a dummy variable representing whether the household is in a treatment or control village. The variable ${l_t}$ stands for the year dummy. The coefficient of the post-certification interaction variable ${p_k}{l_t}$ thus corresponds to the effect of certification. ${\theta _{ht}}$ is the intercept term. Incorporating the effect of interaction between the certification and the anomaly dummies extends equation (1) to:

(2)\begin{align}\ln {y_{ht}} & = {\theta _{ht}} + \beta {x_{ht}} + \gamma {r_{ht}} + \vartheta {w_{ht}} + \phi {l_t} + \vartheta {p_k} + \delta {l_t}{p_k} + \rho {l_t}{p_k} \times {r_{ht}}\notag\\ & \quad + \partial {l_t}{p_k} \times {w_{ht}} + {\varepsilon _{ht}}.\end{align}

The error term ${\varepsilon _{ht}} = {\alpha _h} + {u_{ht}}$ is composed of a normally-distributed random error term, u _ht ~ N(0, σ _u²u_ij), and an unobserved household-specific effect, ${\alpha _{ih}}$. Under the assumption that ${\alpha _h}$ is orthogonal to the observable covariates, a random effects estimator can be employed (Wooldridge, Reference Wooldridge2002). However, allowing arbitrary correlation between ${\alpha _h}$ and the regressors/observed covariates requires a fixed effect, as it takes ${\alpha _h}$ to be a group-specific constant term and uses a transformation to remove this effect prior to estimation (Wooldridge, Reference Wooldridge2002). For a robustness check, we added a battery of controls to the estimating equation. Many important controls, however, do not change much in the period of interest. Since it may be presumed that the observed covariates are correlated with the unobserved individual effects, our estimation procedure involves the inclusion of fixed effect in the estimation following the Mundlak approach (Mundlak, Reference Mundlak1978; Wooldridge, Reference Wooldridge2002).Footnote ¹⁷ The results are very consistent and reported in column (4) of table 3.

5.3 Testing for the common trend assumption

In order to correctly identify the impact of certification, the treatment and control villages must be randomly assigned such that the change in the agricultural revenue, the outcome variable of interest, is attributed solely to the effects of the treatment. Closely related to the random choice of villages is the common trend assumption, which holds if the unobserved influences on farm level revenue, conditional on weather shocks, do not change in different ways between the control and treatment groups. We analyze the extent to which it is possible to establish that the trend in revenue, given observable variations that can be controlled for, and conditional on weather shocks, is the same for the treatment and control groups (see table 1).

As the common trend assumption is fundamentally untestable after the introduction of the program, we base the test on data prior to, and including, 2005. Accordingly, we estimate a panel data equation that consists of different intercept terms for the treatment and control villages for each year; weather shock variables; the interaction between these intercept terms and weather shocks; and a set of control variables. The coefficients from the estimation are then used as inputs in the common trend assumption test. The test investigates whether, prior to the land certification, households that have eventually received their land certificate used to be more resilient to weather shocks. Accordingly, the null hypothesis of the common trend assumption test is that the coefficients of interaction between certification and experience of weather shocks in 2000, 2002 or 2005 are not statistically different from zero. The pattern of coefficients during the pre-policy change period in the years 2000, 2002 and 2005 show no significant difference. This indicates that there is no statistical evidence of violation of the parallel trend assumption. Estimation and test results are reported in the online appendix, table B4.