How Good Are Livestock Statistics in Africa? Can Nudge and Direct Counting Improve the Quality of Livestock Asset Data?

There exists a major gap in measurement of livestock assets and inputs, particularly noticeable in countries like Ethiopia, which holds the largest livestock population in Africa but the sector contributes little to overall economic growth. In an effort to assess and improve the quality of livestock statistics, we introduce three main interventions in a large household survey targeting smallholder framers in the highlands of Ethiopia.

• Intervention one: Nudging
A specific nudge is implemented for a random subset of survey respondents. The nudging is intended to remind respondents on the purpose of data collection and hence build trust with respondents. The nudging information reminds respondents that the livestock data will be used solely for research purposes and not for identification of beneficiaries of social protection programs or tax purposes. In addition to the usual consent at the beginning of the survey, the following nudge is delivered midway through the survey and just before the starting livestock module.
“The purpose of this study is solely for research, involving the collection of data on the livestock population in the kebele, including cattle, goats, sheep, horses, and other livestock. This information will not be used for Productive Safety Net Program (PSNP) selection or any other related social assistance programs, nor will it be employed for tax purposes. The data will not be shared with any other organization or the government, and we will ensure the anonymity of farmers' identities. There are no additional benefits to participating in the survey, and there are no risks involved. Please be aware that providing inaccurate information may distort the study's results and hinder potential interventions for livestock in the kebele. We kindly request that you provide an accurate count of the cattle, goats, sheep, horses, and other livestock you own."

After providing this information, the enumerator will collect livestock data from the household head and spouse separately.
• Intervention two: Direct counting of livestock assets owned by households
In collaboration with local livestock experts, enumerators undertake a direct count of household livestock assets, particularly cattle, mules, horses, and donkeys. The initial step involves hiring a local animal health/veterinary extension worker to work with enumerators in facilitating this counting. Using local livestock experts allows us to capitalize on the experts' familiarity with village households and the trust that respondents may place in them. Furthermore, as an incentive for households participating in this process and for ethical purposes, the livestock experts are also asked to identify emaciated animals in need of deworming and determine the appropriate dosage. Thus, a randomly selected livestock-owning households were asked to participate in this process and after the direct observation and counting the livestock experts administer the deworming tablets (Albendazole) for those animals in need of it. The dosage is tailored to the type and number of livestock owned by each household. Before the observation and counting process, households are informed and asked to provide consent to participate.
Logistically, enumerators and the livestock expert in each village coordinate with survey respondents to facilitate the physical counting while the animals are in their homes. Early in the morning, before the cattle are let out, the enumerator and the livestock/veterinary extension worker will arrange a visit to the households. The extension agent will then request the household head to show them all the cattle, donkeys, mules, and horses to assess their body condition and determine which ones should receive deworming tablets. Simultaneously, the enumerator will count the number of cattle, donkeys, mules, and horses the household has and verify how many of them belong to the household, checking if the household has cattle elsewhere. Each household will receive deworming tablets if they have cattle with poor body conditions. If possible, the veterinary extension worker will administer the tablet with water to the cattle. We note that counting small ruminants can be infeasible in some instances, especially if the household owns a large number of small ruminants. Similarly, in pastoral settings, Ethiopian households own large stock of animals, and they usually live widely disbursed across large area of grazing land, which requires large amount of effort and logistics than we have in this experiment. Thus, the counting exercise in this project focused on large animals in highland areas.
• Intervention three: Administering livestock module to both spouses in the household
Traditionally, agricultural surveys in LMICs have been administered to head of the household, often men, as key representative of the of the households. This approach assumes that the head of the households are knowledgeable about agricultural practices and assets of the households. However, merging studies show important gender differences in livestock management roles and control of incomes derived from various livestock species kept by households. Consequently, it is expected that women and men members of households may possess different levels of information about the livestock species maintained by the household. For instance, in Ethiopia, women often serve as decision-makers for small animals (Mulema et al., 2017; Kinati and Mulema, 2019). This gender disparity could potentially result in underreporting or overreporting of different livestock species owned by households, depending on the individual's role in livestock production within the household. In this project, the livestock module is administered separately to both primary male and female respondents in each household. This approach aims to address potential variations in reported livestock ownership based on the respondent's role in livestock production. The project anticipates that these innovations will help identify and alleviate issues of under (over)-reporting. Through these interventions, we aim to provide important insights to improve livestock assets measurement.

2023-11-30

2024-01-31

Primary outcomes:
Total number of oxen, bulls, and calves
Total number of cows and heifer
Total number of large ruminants (cattle)
Total asset ownership TLU

Primary outcomes explanation:
Total number of oxen, bulls, and calves: This outcome considers the total number of local, cross-bread and exotic breed oxen, calves, and bulls the household owns.
Total number of cows and heifers. This will be total number of cows and heifer (including local, crossbreed and exotic/ Frisian breeds) households own.
The total number of large ruminants is calculated as the sum of total number of oxen, bulls, calves, cows, and heifer the household owns. This includes local breeds, cross-bread oxen, exotic breeds, and the household owners.
Tropical livestock Unit (TLU). Total asset ownership TLU is calculated as follows:
TLU=camels + (0.7*Total large remnant) +(0.8*horses)+(0.5*donkeys)+(0.5*mules)+(0.1*Sheep_no)+(0.1*goats_no)+(0.01*chicken)

In addition to the primary outcome variables, we will examine the following secondary outcome variables: the number of cows, oxen, donkeys, mules, horses, and indicator variables for each type of livestock ownership.

This section outlines the experimental design of the intervention, which is embedded into a large follow-up household survey in Ethiopia. This follow-up survey builds on a 2019 survey on Feed the Future (FTF) Ethiopia Zone of Influence (ZOI). The 2019 FTF survey followed a multi-stage random sampling and included 264 enumeration areas (EAs) selected from 132 woredas/districts within the ZOI (two EAs from each woreda). Out of the 264 EAs in the baseline, 180 of them were identified to be currently safe and accessible for a survey. Thus, our full sample for the follow-up survey contains 180 enumeration areas (EA) with a random sample of 20 households in each EA or village. For the purpose of this experiment, we specifically target regions and households engaged in mixed farming, excluding pastoral regions such as Somali and Afar (29 villages). The exclusion is mainly due to the challenges associated with implementing the observational counting exercise in pastoral settings. Counting herds, which are usually disbursed across a wide area of grazing lands in pastoral settings, can be infeasible. Thus, we focus on focus on the remaining regions in Ethiopia, utilizing the 151 EAs or villages (and associated households) for this experiment. The randomization of households into the different methods of measuring livestock assets is conducted at the household level. To ensure balance and comparability, we stratify the randomization as follows. In each village, we assign 8 households to self-reporting (control group); 8 households to nudges and 4 households to direct observation and counting. Figure 1 shows the distribution of households in each village into the three measurement arms.

Not available

The randomization was done at the household level using the baseline list of households. We initially selected those villages that are accessible for the survey. We focus on those households practicing mixed farming and hence living in the six highland regions of Ethiopia. On average, the baseline data includes 20 households in each village. We stratify the randomization and hence assign 8 households in each village into the nudge, 4 households into direct counting, 8 households into control group. A reserve list was also prepared in case some of the villages are not accessible because of conflict.

Household level

No

About 2,567 households and 4,492 respondents. This sample is computed based on the following assumptions and the power calculation we report below. In our sample, 75 of the households have two respondents: primary male and primary female respondent. The remaining 25 percent of households have only one primary respondent, most of which are female headed households with no adult male in the household. We also assume an attrition rate of 15 percent.

About 2,567 households and 4,492 respondents. This sample is computed based on the following assumptions and the power calculation we report below. In our sample, 75 of the households have two respondents: primary male and primary female respondent. The remaining 25 percent of households have only one primary respondent, most of which are female headed households with no adult male in the household. We also assume an attrition rate of 15 percent.

Number of households by treatment arm:
C (Control): 1,026 households, about 1,795 respondents
T1 (Nudge): 1,026 households, about 1,795 respondents
T2 (Direct count): 510 households, about 890 respondents

The statistical power calculation considers several outcomes of interest, and we compute the number of households and respondents needed for the primary outcomes described above. In the baseline sample there were an average of 20 households in each village, and we anticipate being able to trace about 85 percent of them, 17 households per village. Our power calculations aim to achieve the standard and widely adopted 80 percent power at a significance level of 5 percent. We note that the power calculations are computed and reported only for the primary outcomes. We compute the number of respondents needed for each primary outcome separately, and then selected the maximum sample needed to detect impacts across these outcomes. We use the baseline data to inform our power calculation. Table 1 reports the mean and standard deviation of the four livestock asset indicators we aim to use for this study. We compile empirical evidence on the impact of nudges in various settings, including those meant to reduce recall and telescoping bias (e.g., Abate et al., 2022). Following these studies, we assume a 15 percent impact of the nudges, compared to the control and hence business as usual practice based on self-reporting. Previous studies report comparable impacts of similar interventions to reduce recall bias (e.g., Abate et al., 2022). As shown in Table 1, with these features and assumptions, the smallest sample needed is for detecting impacts on TLU (2,090) while the largest sample is for detecting corresponding impacts on the number of cows and heifers (3,012). The number of respondents in our sample is larger than this, about 3,580 respondents. Thus, effectively, our sample enables us to detect even slightly smaller impacts. Table 1 summarizes our power calculations involving several primary outcomes meant to test the impact of nudges. Table 2 summarizes our power calculations involving several primary outcomes meant to test the impact of direct counting. We follow similar assumptions and features of the baseline data to compute the sample size needed to detect the impact of the direct counting. Two important features of direct counting are worth noting. First, we anticipate that direct counting will be more impactful than nudges because direct counting can address both unintentional recall errors as well as intentional underreporting. Thus, we assume a slightly higher Minimum Detectable Effect (MDE), 20 percent. Many studies comparing self-reported and direct observation of various outcomes show even larger impacts. For example, most of the recent studies comparing self-reported harvest with crop-cut harvest show much more impacts of the latter (Abay et al., 2019; Kosmowski et al., 2021; Ayalew et al., 2023). Second, implementing the direct counting requires substantial logistical arrangements, including hiring local livestock experts to accompany and facilitate the counting by enumerators. For these reasons, we reduced the share of households exposed to direct counting, to be about half of the control group. Table 2 shows that the largest sample needed is for detecting corresponding impacts on the number of cows and heifers (1,908) while the smallest sample needed is for detecting impacts on TLU (1,325). The number of households and respondents we assigned for direct counting is larger than those reported Table 2 (510 households and 890 respondents. This implies that our sample allows us to detect even slightly smaller impact.