Combining farm and household surveys with modelling approaches to improve post-harvest loss estimates and reduce data collection costs

2.1Scope of the literature review

A literature review was conducted to identify whether on-farm loss models built using data from onfarm or household surveys have been estimated, how robust they were and what estimation strategy and set of explanatory factors were used. The review covered 126 journal publications that mentioned food loss models or post-harvest loss models in any form, of which 62 publications applied food loss models at the farm or household level. The nature of the models used largely depended on the objective of the modelling exercise and ranged from identifying the factors causing on-farm losses, to estimating losses indirectly or assessing the impact of losses or loss reduction on socio-economic factors, at the micro and macrolevel. For the purpose of this research, those papers estimating regression models to identify the drivers of losses were chosen to be the most relevant, since they were built on farm survey data and provided guidance for suitable sets of explanatory variables and the overall model structure.

The review shows that loss models estimated using farm survey data identify a wide range of significant explanatory factors. Most of these studies collected cross-sectional on-farm loss data on single commodities and sub-national regions, with only a few covering a larger number of crops. Africa is the most represented region (more than 25 papers), followed by Asia (15 papers), while fewer studies were found for the rest of the world. Great contributions to this research topic come from Ethiopia and Nigeria, followed by Bangladesh, Kenya, India, Nepal, Uganda and Ghana. Several studies build on each other, for instance Kumar et al. [12] and Basavaraja et al. [13] represent one of the first applications of regression analysis to determine on-farm loss drivers and were cited by various articles. The large majority of screened papers were produced between 2015 and 2020 and cover mainly grains (21 papers), roots and tubers (12 papers), and fruits and vegetables (19 papers). Some of the studies also conducted surveys and regressions for off-farm stages, but these were not further examined for the purpose of this research. In what follows, the main conclusions in terms of the estimation approach, the data used, and the dependent and explanatory variables are outlined.

2.2General model approaches used for farm loss modelling and survey data

The most commonly found estimation approach is a multiple linear regression, using a set of explanatory variables in order to explain the dependent variable of on-farm losses. This approach was used for instance by Kumar et al. [12] and Basavaraja et al. [13] on grains, roots and tubers in India; Begum [14] and Khatun et al. [15] on rice, wheat and tomato in Bangladesh; Arun et al. [16] and Paneru et al. [17] on a variety of commodities in Nepal; Adisa et al. [18] on yam; Babalola et al. [19] on tomato in Nigeria; Tadesse et al. [20] on potato in Ethiopia; and Ambler et al. [21] on cereals in Malawi. Some authors opted to use a double and semilogarithmic multiple regression analysis, as Folayan [22] on maize in Ethiopia, Aidoo et al. [23] on tomato in Ghana, and Huang et al. [24] on grains in China. Ansah et al. [25] used a fractional logistic regression model because of the proportional nature of the dependent variable, which in this case is post-harvest management as a way of assessing the inverse of on-farm losses. Hossain et al. [26] suggested a Cobb-Douglas production model to estimate the coefficients of the factors influencing potato storage losses in Bangladesh. Shee et al. [27] and Garikai [28] used an ordered probit model, employing on-farm loss categories that organize loss percentages into four loss categories. These categories were built from loss percentage data collected in the respective study survey. Amentae et al. [29] and Falola et al. [30] used a tobit regression model, applying only a binary category of on-farm losses (low and high losses; experience and do not experience losses). Kikulwe et al. [31] made use of a tobit censored regression model to solve the limitation in the dataset of a significant number of producers who reported zero losses.

Most of the surveys collected on-farm loss data by declaration together with other socio-economic characteristics of the producer or household, agronomic indicators of production or post-harvest management, and climatic and weather factors. Therefore, the data for the regression models was overly obtained from the same survey. In general terms, the surveys conducted for the loss studies had a sample size of about 100 to 300 households, depending on the target population, regional coverage and available resources.

2.3Set of dependent and independent variables used for the farm loss modelling

Almost all surveys estimated total post-harvest losses aggregated for all post-harvest activities, while only some studies disaggregated losses by or concentrated on one specific activity. Storage losses were one of the more specific areas of study (Kimenju et al. [32] on maize; Falola et al. [30] on yam; Hossain et al. [26] on potato). Ambler et al. [21] conducted the regression model on aggregated post-harvest losses as well as disaggregated by post-harvest operation for maize, soya and groundnuts. The main differences are the choice of independent variables, which are more specific if the study focuses on a single post-harvest activity. Two studies, namely Kikulwe et al. [31] on banana in Uganda and Qu et al. [32] on grains in China, also included harvest losses in addition to post-harvest losses. The most frequently used dependent variable is loss quantity in kilograms, as applied by Aidoo et al. [23] and Aidu et al. [34] in Ghana on tomato, Tadesse et al. [20] on potato in Ethiopia, or Folayan [22] on maize in Nigeria. Four studies considered on-farm loss in kilograms per hectare, as a way to relate the loss quantity to the size of the farm, as Kumar et al. [12] and Basavaraja et al. [13] in India and Begum [14] or Khatun et al. [15] in Bangladesh. On the other hand, six studies used loss percentage as the dependent variable, relating loss quantities to the total quantity produced, as Mebratie et al. [35] and Amentae et al. [29] on Ethiopia or Paneru et al. [17] and Arun et al. [16] on Nepal. Each of these three approaches can have different implications in terms of the relevance of loss drivers. Loss quantities are likely to be positively related to production volume and factors influenced by the size of the farm. Loss percentages are more commonly used to identify structural losses usually caused by the type of production system, climate, and agronomic practices. Loss quantities per hectare could be interpreted similarly to loss percentages, although it does not take into account the differences in productivity that are counted in when using loss percentages. Kikulwe et al. [31] applied a regression analysis on the drivers of loss quantities and loss percentages, which provides the possibility of comparing the changes in the significance of explanatory factors between both types of dependent variables. On the other hand, three studies (Shee et al. [27]; Maziku [36]; Garikai [28]) used on-farm loss categories of minimum, low, medium and/or high losses. These were either reported directly as categorical variables by the producer or based on the loss percentages the producer declared. Some studies (e.g., Kwami et al. [37] and Falola et al. [30] on food losses in yam and plantain in Nigeria) need to be analyzed separately as they build the regression analysis on the adaptation or use of technologies that are directly linked to higher or lower losses.

Apart from the general mapping, special emphasis was put on systematizing the explanatory variables used in these studies and their overall significance, which can be accessed in detail in the report of the 50x2030 Initiative literature review [38]. The studies show certain similarities in terms of the set of variables chosen. These cover a set of variables that can be grouped according to the household’s socio-economic characteristics, production characteristics, post-harvest management and market relations, and agro-ecological and weather conditions. Most studies cover each of these groups with at least one variable. In general terms, the group of production characteristics and agro-ecological and weather factors seem to have the most significant relation to on-farm losses, followed by post-harvest management and market relations, while socio-economic characteristics have a less significant impact. Within the production characteristics, production or farm size, the experience in farming and access to extension services and training, as well as the time or days of harvesting were those most commonly used and which showed to be significant for losses. Within the group of agro-ecological and weather factors, weather conditions during harvesting (rainfall during harvesting, good weather conditions during harvesting), the agro-ecological zones and general weather conditions (annual mean temperature) showed to be significant, as well as general geographical indicators (district, altitude). In terms of the group of post-harvest characteristics, the type and use of storage facilities (cereals and pulses), and the distance or time to the nearest market are relevant indicators, while those variables relating to other post-harvest operations like packaging, threshing and transportation provided less significant results. In terms of socio-economic variables, education, age and sex were the most commonly used, but resulted in limited explanatory power for on-farm losses. Family size was another widely used variable and had mixed results but with a tendency towards being significant.

3.1The CART method for post-stratification to specify and identify the loss model

There are different uses of generalized linear models. One common use is to establish the relationships of selected independent variables as determinants of losses, while another is to predict mean responses, where the inclusion of a wide variety of independent variables are used as determinants focusing on the prediction capabilities and statistical efficiency of loss estimates. While the literature review highlights mainly on-farm loss models to identify on-farm loss drivers, this research seeks on formulating on-farm loss models for prediction purposes. Therefore the above-mentioned results from the literature review provide a key orientation point for specifying the models but need further discussion to generate models for prediction purposes. In this paper, the objective is to test one possible way of specifying on-farm loss models for prediction purposes making use of a classification method derived from a Classification and Regression Tree (CART) post-stratification [39, 40]. It is expected that this modelling approach not only provides good and reliable on-farm loss estimates, moreover it simplifies the modeling procedure and helps to improve the efficiency of the mean estimate by a reduction of its standard error. These gains, in turn, can support data collection on relatively small sub-samples. Data collection costs could be optimized without considerably compromising data quality, resulting in a complementing strategy to be used to design national data collection of losses within national farm and household surveys.

As a first step, a better understanding of the available survey datasets is needed, especially on the overall availability of significant variables from the surveys to explain on-farm losses. Based on the insights obtained from the literature review multiple linear regression models were tested, choosing a set of independent variables related to harvest and post-harvest known to be relevant for losses. On the other hand, different dependent variables proposed in the literature were thereby assessed to revise whether to use quantity losses or percentage losses, total post-harvest losses or losses disaggregated by operation (harvest, cleaning, drying, storage, etc.) As a first conclusion, on-farm loss percentages seem to be a better suited dependent variable than quantity on-farm losses for the given survey datasets. Percentage losses seem to better indicate the structural problems causing losses, the efficiency of handling the grains, while the quantity of losses is to some extent driven by the production volume. Since the recorded percentage losses show a positive skewed distribution, the use of the natural log transformation of the percentage losses is suggested. A linear regression could be used to generate a model to predict mean percentage losses, but in that case, predictions from the model are in log scale and the reverse transformation results in a bias of the estimated mean losses. These could be corrected by including a function of the variance of the errors in the estimated mean losses. Nevertheless, the use of a Poisson regression model can be a better alternative to log-linear regression, because the link function of this model is the natural log of the response. Additionally, a Poisson regression handles outcomes that are true zeros, while a log regression does not consider zeros because of ln(0) is -∞. Poisson distribution assumes that the expected response equals the variance of the response, so the use of robust standard errors is useful to handle these assumptions.

As a second step, post-stratification is prepared with the main idea to improve the efficiency of the parameter estimates obtained from the sample survey and respective sub-samples. As stated by Smith [41], it can be a useful method to reduce variance and correct for possible bias, and in this case Classification and Regression Trees (CART) are used to generate the post-stratification. The output of the CART is a decision tree where each end node represents a stratum with a final prediction for the outcome variable, in this case on-farm losses. The algorithm selects the relevant independent variables and their respective cutting point where the difference of the mean response for the resulting groups are maximized. Thereby, part of the variance in the sample survey is explained by the mean differences between the resulting groups. In order to make use of post-stratification in the modelling approach, the results of the classification and regression tree are used to set the estimation model, where the classification variable is used as predictor and the mean prediction is used as the estimator of the mean on-farm loss. In what follows, the models are tested to determine whether they (i) are sufficiently well-specified to provide reliable estimates, and (ii) reduce the standard error as an assumed effect of the post-stratification procedure.

To test if the model is well suited, the linktest [42] is used to detect any specification error in the proposed models. This uses the linear predictor value X⁢b^ and linear predictor value squared (X⁢b^)2 as the predictors to rebuild the model (X represents de predictor variables and b^ the estimated model coefficients). The variable (X⁢b^)2 should have no predictive power and the estimated parameter should be zero. On the contrary, if (X⁢b^)2 is significant, the linktest is significant, meaning that we have omitted relevant variables, or our link function is not correctly specified. In this case, it implies a model with lack of fit which is of limited use for prediction purposes. On the other hand, X⁢b^ should be close to 1, which is considered a good linear predictor.

3.2Evaluate the gains of the modelling approach for sample size reduction

The main idea of this research is to make use of the improvements in the mean estimate obtained from post-stratification in order to reduce on-farm loss data collection to a sub-sample. The models will be assessed on the basis of their capacity to produce estimates, which should not deviate considerably from the estimates obtained from the full sample. Additionally, a measure of efficiency is defined in order to evaluate whether the model-based estimates provide considerable gains that can be used for sample reduction. Here, the ratio of the model-based variance to the actual full sample variance of an estimator is built to express the efficiency gains in terms of a reduction in the variance. The relative efficiency of a model-based estimate compared to the full sample-based estimate is then:

Where V⁢(Lm^) represents the variance for the loss estimate based on the specified model (implying a post-stratification effect), and V⁢(L^) represents the variance for the survey-based loss estimate from the full original sample RE is then the percentage of variance reduction.

In order to test the possibility for sample reduction, a simulation will be run on given survey datasets. To create sub-samples, a progressive random elimination of 10% of the sample to a maximum of 50% reduction is conducted, generating five sub-samples. For the full sample and for each of the sub-samples, sampling theory-based and model-based mean loss estimates are obtained and compared:

Compared with:

In addition to the model-based estimates, the specified loss models can also be used to impute missing values. To simulate the imputation of possible data gaps, the on-farm loss model is used here to impute losses from the sub-sample to the full sample. This exercise has some known limitations especially when imputation techniques are applied to a larger proportion of the sample and lead to an artificial reduction of the standard error of the mean estimates. Therefore, the exercise is presented in the results section but highlighting the limiting interpretation of the standard error.

For the purposes of comparison, we define:

4.1GSARS66 farm loss surveys in Malawi and Zimbabwe

The first set of available farm survey loss data comes from the field tests conducted for the “Guidelines on the measurement of harvest and post-harvest losses” [8] in Malawi [43] and Zimbabwe [44].

These farm loss surveys were implemented in 2017 and 2018 on a sub-national level covering the Salima and Lilongwe districts in Malawi for maize, rice and groundnuts, and the Makonde district in Zimbabwe on maize. The analysis focuses on maize only, the most important staple food in all the surveyed countries, and the crop for which the most data is available in the analyzed surveys. The sample size in Malawi achieved 447 observations for maize crops, in the case of Zimbabwe 307 observations are obtained for maize. In each of these regions, agricultural production is the main source of livelihoods. Average area harvested is 1.2–2.7 hectares per household in Zimbabwe and 0.4–0.6 hectares in Malawi, predominantly rainfed and based on manual harvesting methods (see Table 1 for a summary of descriptive statistics of the main variables).

Although data was collected for local, hybrid and composite maize, the modelling approach will be applied to the aggregation of all varieties of maize. This is due to the fact that, on the one hand, recent empirical evidence highlights problems of misclassification with respect to farmersdeclaring seed varieties [45, 46]. On the other hand, the sample sizes of the GSARS surveys are relatively small for testing modelling approaches. The main variable of interest – the percentage of losses over production – has been calculated as the total quantity of maize losses (from harvest to storage) divided by the quantity of land cultivated with maize.

4.2Living Standard Measurement Study – Integrated Survey on Agriculture (LSMS-ISA) in Malawi and Nigeria

The study uses two datasets from nationally representative household surveys in Malawi and Nigeria: the Fourth Malawi Integrated Household Survey 2016/17 (IHS4)77 and the Nigeria General Household Survey 2015/16 (GHS 2015/16).88 The surveys are part of the Living Standards Measurement Study – Integrated Survey on Agriculture (LSMS-ISA) and contain an integrated household and agricultural component. The household survey component collects detailed socioeconomic information, including household-level data on consumption, income, assets and housing, and individual-level data on demographics, education, and health. The agricultural component collects detailed information, among other items, on agricultural inputs used and outputs produced, as well as output disposition, at the plot-level. Particularly important for the analysis are that information on inputs are collected at plot level; that information on agricultural output is collected at crop/plot level; and that output disposition is collected at crop level. In addition, both the IHS4 2016/17 and GHS 2015/16 datasets include a number of exogenous climatological and geospatial variables. These include measures of distance, climatology, soil and terrain and other environmental factors. Time-series data on rainfall and vegetation has also been used to describe the survey agricultural season relative to normal conditions.

The IHS4 2016/17 is the fourth wave of the Integrated Household Survey and includes 12,480 households surveyed in 780 enumeration areas. Households were visited once throughout the 12 months of fieldwork between April 2016 and April 2017. The GHS 2015/16 is the third wave of the Nigeria General Household Panel and includes 4,581 households surveyed in 500 enumeration areas and visited twice between September 2015 and April 2016, once right after the end of the planting activities and once right after the end of the harvest activities relative to the 2015/16 rainy season. Given this two-visit setup, the length of the recall period of the GHS 2015/16 is shorter than length of the recall period of the IHS4 2016/17. As with the GSARS surveys in Malawi and Zimbabwe, the samples for the IHS4 2016/17 and GHS 2015/16 were restricted to maize (all varieties of maize aggregated).

Both the Malawi IHS4 and Nigeria GHS 2015/16 samples have been restricted to observations reporting positive values for losses, thus excluding zero observations. The very high percentage of zero losses shed doubts on the accuracy of such zero reported losses, particularly given the very low percentage of zero losses in the GSARS surveys. The assumption here is that in LSMS-ISA surveys farmers, in some cases interviewed several months after the harvest, might have reported only substantial losses and omitted marginal losses (see Section 4.3 for an explanation of the differences in the scope and methodologies between GSARS and LSMS-ISA surveys). Including “false” zero losses could lead to downwardbiased estimates. Although we recognize that excluding zero losses could lead to biased (in the case that a certain percentage of reported zero losses are “true” zeros) or partial estimates (which, by definition, apply to the specific case of farmers reporting non-zero losses), after cross-checking loss variables with other strongly correlated variables (such as whether the crop was stored or not) we decided to restrict the sample to positive loss observations for the estimation of models and the simulation of subsampling scenarios, with the clear statement that the results from LSMS-ISA surveys in the study refer to farmers reporting non-zero losses. In order to assess the potential bias introduced by the sample restriction, the probabilities of non-zero and zero reporting were tested using propensity scores technique, as well as the significant difference in characteristics between those reporting zero losses and those reporting positive losses. Both tests show that the two groups are not systematically and significantly different (not shown here, available upon request) (see Table 2 for a summary of descriptive statistics of main variables).

4.3Differences in the survey design regarding farm losses

The on-farm loss data obtained from the GSARS harvest and post-harvest loss surveys and the LSMS-ISA surveys differ from each other in terms of their design and data collection method. First of all, the GSARS data stems from a survey specifically designed to measure on-farm losses. With all attention paid to the loss indicators, the questionnaire includes loss questions disaggregated by on-farm activities (harvest, threshing, winnowing/cleaning, and storage) as well as complementary data on socio-economic, production and post-harvest characteristics. Additionally, on-farm loss data was collected by farmers’ declarations as well as by physical measurement. On the other hand, the scope of the survey is the local level, where it is probable that the population is less diverse than at the national level.

The LSMS-ISA surveys, on the contrary, are designed to estimate agricultural production and productivity of the rural households, where losses are covered as one complementary indicator among various in the crop disposition section. It is collected by one sub-question on the destination of the harvested production, where farmers declare total post-harvest losses among quantities self-consumed, sold, given away as a gift, and used as seeds and as animal feed, without detailing the post-harvest activities. Harvest losses are not considered. The set of variables collected in the survey are less tailored to on-farm losses, but cover a broader range of socio-economic, production and environmental characteristics. The survey is nationally representative, whereby it is assumed to cover a more heterogeneous population compared to the GSARS surveys. All surveys allowed for reporting loss and production quantities in local non-standard units, converted into kilograms using correspondence tables specific to each survey.

For the 50x2030 Initiative, a similar questionnaire structure to the GSARS harvest and post-harvest loss survey was developed for the corresponding loss module recommending a more detailed assessment of harvest and post-harvest losses. On the other hand, the 50x2030 Initiative seeks nationally representative surveys, whereby the LSMS-ISA survey will help to better understand the implications of nationwide data in the modelling approach.

Figure 1.

Regression trees.

5.1Food loss models estimated for Malawi, Zimbabwe and Nigeria

5.2Results to reduce the sample size to a sub-sample and use the models to improve loss estimates