Drivers of rice production: evidence from five Sub-Saharan African countries

Abstract

Background

In spite of considerable rice production gains over the past 50 years, Sub-Saharan Africa is becoming increasingly dependent on rice imports as demand is outpacing domestic supply. The serious economic and social strains caused by this have urged national leaders to address production deficits. The aim of this article is to analyse and discuss the drivers behind recent changes in rice production in Africa South of the Sahara, focusing on Ghana, Malawi, Nigeria, Tanzania and Mozambique. Considering the period 2002–2008, we model production performance and changes in production amongst 317 rice-growing households using multilevel and longitudinal data. We evaluate and discuss the role of three key processes: the role of commercial drivers, farm technology and macro-level conditions.

Results

We show that until 2002, production was driven by a combination of the three key processes considered, while during the period 2002–2008, production increases were primarily associated with area expansion and commercial drivers. This suggests that production lately has been more driven by processes of extensification than intensification. We also note that in none of the periods considered, the share of the state budget allocated to agriculture had a significant effect on production and that recent developments do not give any obvious support for an Asian-style state-driven Green Revolution in rice in Sub-Saharan Africa.

Conclusions

The role of commercialization in explaining changes in production suggests that policies strengthening food staple markets in the sub-continent hold great potential for driving rice production in the near future. Due to the scarcity of available land, the possibilities of further growth in the rice sector are limited without an intensification of production. Hence, farmers also need to access new farm technology, and positive development of rice production would in turn contribute to an improvement of food security.

Background

Global demand for agricultural products is expanding rapidly and the demand for food products is foreseen to continue to grow for several decades as a result of a combination of population growth, rising per capita incomes and urbanization [1]. In developing countries, approximately 60 % of total calories consumed are derived directly from cereals with values exceeding 80 % in the poorest countries [2]. Among the cereals, rice is the most important source of calories for humans. While per capita consumption is declining in parts of Asia, the demand for rice has increased considerably in Sub-Saharan Africa (SSA) since 1995 and is growing more rapidly there than in any other continent [3]. Increased rice consumption can be traced to a combination of population growth, urbanization,¹ changing consumer preferences and economic development [see e.g. 3, 5, 6]. Although rice production in SSA has increased substantially over the past 50 years², demand has outpaced domestic supply resulting in increasing imports [7–9]. The share of imports in African rice consumption has grown by 2 % per year over the past 50 years and reached 43 % in 2009 [10]. Indeed, almost one-third of the rice traded on the world market in 2011, or 11.8 million tons, was imported by African countries, compared to 0.5 million tons in 1961. Countries with international ports and populous cities, such as many West African countries, tend to be particularly large importers [10]. A recent estimate suggests that rice imports cost Africa more than US $4.3 billion annually [11]. The serious economic and social strains caused by the dependence on rice imports have urged national leaders, supported by international development partners, to address production deficits by strengthening the domestic rice sector [6, 12]. Expected positive effects, apart from improvement in balance of payments, include enhanced food security and reduced poverty among both producers and consumers.³

The aim of this article is to analyse and discuss the drivers of recent development in rice production in SSA, focusing on Ghana, Malawi, Nigeria, Tanzania and Mozambique. We model production performance and changes in production amongst 317 rice-growing households, considering the period 2002–2008. Specifically, we evaluate and discuss the role of three key processes: the role of commercial drivers, farm technology and state-specific conditions (including agrarian policies). Our findings can inform future strategies and policies for capacity development in the rice value chain.

In the following sections, after a brief presentation of the role of rice in the five countries, we introduce the data and the general modelling strategy. Thereafter, we outline key theoretical perspectives that inform this article. We then present descriptive statistics pertaining to select socio-demographic characteristics of the panel households as well as the variables that we use in our modelling. In the following section, we introduce the three models. Thereafter, we present and discuss the results, specifically considering the commercial, technology and policy drivers behind changes in rice production. The article concludes by outlining the implications of these findings for agricultural policy and practice in the region.

Rice production and consumption in the five case study countries

Rice is becoming increasingly important in the five countries considered in this study, with Ghana and Nigeria currently in the lead in terms of per capita consumption (Fig. 1). This is in line with reports that rice is much more important in West Africa than in the other regions of SSA [12]. However, Tanzania, as the country ranked second (after Madagascar) within Eastern, Central and Southern Africa in terms of rice production and consumption [13], has become a clear runner-up in these sub-regions over the past few years. In Fig. 1 we can see that the per capita consumption has been increasing steadily over the last three decades in Ghana, Nigeria, Tanzania and Mozambique. In these four countries, rice is today ranked between number three and five in terms of importance vis-à-vis other crops consumed,⁴ and represents approximately 10 % of daily caloric intake [14–20]. Only in Malawi is rice reported to be a crop of minor importance, which is also reflected by rice and wheat together accounting for only 4 % of daily caloric intake [21].

Fig. 1

Milled rice consumption per capita (kg) trends 1982–2013. Data source: [56]

Farmers in these countries, especially in Nigeria and Tanzania, have responded to the increasing demand for rice, as reflected in upward trends in aggregate production [22]. When adjusting for population increases (Fig. 2), the production trends are much less impressive, however, and as Fig. 3 shows, many of the countries are also becoming increasingly dependent on rice imports, fuelled by growing production-to-consumption gaps.

Fig. 2

Milled production per capita (kg) trends 1982–2013. Paddy rice was converted to its milled equivalent using a conversion ratio of 0.65 [see e.g. 6]. Data source: [22, 57]

Fig. 3

Milled production—consumption balance (kg/per capita) trends 1982–2013. Paddy rice was converted to its milled equivalent using a conversion ratio of 0.65 [see e.g. 8]. Data source: [22, 56, 57]

Methods

Introduction to the data set

The data derives from a longitudinal survey conducted in 2002 and 2008 within the framework of the Africa and Intensification (Afrint) project⁵, which focused on a group of countries located in what can be described as the African ‘maize and cassava belt’. Nine countries were purposively selected—Ethiopia, Ghana, Kenya, Malawi, Nigeria, Tanzania, Uganda, Zambia and Mozambique [23, 24]. The survey covered more than 4000 randomly sampled households in 103 purposively sampled villages. Informed consent was received from each survey participant. The project adhered to the Swedish Research Council’s ethical guidelines, and recommendations on good research practice [25].

In this article, we focus on the rice-growing households in the survey, defined as those that grew rice in both 2002 and 2008. For this reason, we have created a sub-panel of rice growers (referred to as ‘the panel’), comprising of 317 households interviewed in both 2002 and 2008, and for whom retrospective data for the year of household formation (referred to as ‘the reference year’) are also available. Only five of the nine countries, Ghana, Malawi, Nigeria, Tanzania and Mozambique, are represented in this panel, since there were no surveyed households growing rice in the other countries. The rice grower panel covers households in 25 villages (see Table 1). Household-level data are complemented by village- and country-level data.

Table 1 Distribution of households

Modelling strategy

The modelling strategy aims to capture the drivers of production changes by using multilevel and longitudinal data. Here, both countries and villages have been sampled purposively. In the model, we treat the effects from the villages and from the countries as random effects to reflect the variability in the results obtained due to localization and/or regionalization We thereby represent these as a sample of possible effects (as well as show the variability in these), instead of pointing out the effects from the specific villages and countries. In the longitudinal data analysis, our data refer to three different time periods:

• reference year (for which the average was 1984) to 2002 (t ₀ to t ₁; or period p ₁)

• 2002–2008 (t ₁ to t ₂; or period p ₂) and

• reference year to 2008 (t ₀ to t _2; or period p ₁ + p ₂, also denoted p ₁₊₂).

We model a characteristic of the individual household i in village j and country k at time t as an additive function of the possible effect of a number of explanatory variables. One such characteristic studied here is the logarithm of the amount of rice produced (in kg) in the individual households at different time points; another is the logarithm of the relation between the amounts of rice produced at two different time points.

Table 2 presents an overview of the explanatory variables involved and the reasons for including them in the analysis. These range from macro-level variables, being estimates of the effects of national policies related to the individual countries, to the most important determinants of production at household level, including technological drivers of intensification and commercialization variables. To investigate the distributional consequences of production changes, two distributional dimension variables, viz. gender of farm manager and village elite membership, are considered at the household level. The farm manager was self-defined as the person in charge of making general decisions regarding the farm on a daily basis. We acknowledge the shortcoming in this definition, which assumes that only one person in the household decides over all the plots and crops on the farm, when in fact women and men may control different plots and crops. The definition is used due to the limitations of the present data set.⁶ Following Andersson et al. [28], we use total cultivated area as proxy for elite membership,⁷ while being aware of its limitations. We also include household-specific control variables, namely age of farm under the present management in 2002 and whether the household has partitioned due to a generational shift during the period from 2002 to 2008. Finally, since extreme and unexpected weather events may severely affect the production performance, we also include two weather dependence control variables at the village level.

Table 2 Overview of variables and associated assumptions

Introduction to the rice grower panel

The majority of panel households in this study, or 88 %, are male-headed, with female-headed households making up the remaining 12 %.⁸ The head of household was almost invariably also the farm manager: 86 % of the households had a male farm manager and 14 % a female farm manager. In 2008, the average farm manager was 47 years old and had been to school for 5 years. The average household was made up of four members. In 2002, the mean age of the farm, i.e. the number of years it had been under its present management, was 18 years with a range from 2 to 75 years. Only a very small share (3 %) of the panel households consist of descendant households where the person who was the head of household in 2002 had deceased between the two rounds of data collection. Hence, the average age of the farms in the panel in 2008 is slightly lower than the 24 years that would have been the case if such partitioning of the original household had not occurred.

On average, a farm’s total cultivated area was 2.3 ha in 2008, of which 0.89 ha was devoted to rice. Individual farms had in some cases changed size between 2002 and 2008, and on average these farms had increased their cultivated area by 0.4 ha. Of the total number of farms, 7 % had increased their total cultivated area since the reference year. As a proxy for elite membership, we use the upper decile of farms in the village with respect to their total cultivated area in 2002. In 2002, 12 % of the total number of farms belonged to the elite membership group.

Rice production performance across countries and villages in 2002 and 2008

Table 3 shows that on average, panel households dedicated 0.89 ha of land to rice cultivation in 2008 compared to 0.76 ha in 2002, which represents an increase by 17 %. The overall rice production per household, however, remained almost constant during the period, only increasing by 20 kg, or 2 %, from 1075 to 1095 kg on average, while rice yields declined by 12 % between 2002 and 2008, from 1477 to 1296 kg/ha.⁹ These figures, however, obscure substantial spatial and temporal variation.

Table 3 Household rice production performance by country

If we consider area and total production, it can, for instance, be noted that in Nigeria, households dedicated on average 2.46 ha to rice in 2008 and harvested over 3143 kg, while in Mozambique the same year the corresponding figures were only 0.29 ha and 225 kg. If we consider yields in 2008, these ranged from an average of 402 kg/ha in Ghana to as much as 2237 kg/ha in Malawi, with higher yields not appearing to only be linked to smaller plot sizes. Patterns over the 2002–2008 period also vary among the countries: in Mozambique, falling acreages coupled with increasing production resulted in a 253 % yield increase. In Ghana, by contrast, a more extensive production pattern resulted in a 67 % yield decrease.

Table 4 shows that there are also substantial variations between villages within countries in relation to production patterns over time. This suggests that the overall production increase observed for the panel as such, although minimal, is due to a mixed pattern of extensification and intensification. That the relative change in overall rice area was larger than the relative change in production reflects the fact that in many areas yields are falling, which is also reflected in the overall average yield decrease of 12 %.

Table 4 Household rice production performance by village

When considering weather-related shocks, it can be noted that 34 % of households were affected by the drought in Eastern and Southern Africa in 2002, while 7 % were affected by the floods in West Africa in 2008.

Technology adoption, management and commercialization

Table 5 shows that three-quarters of the households reported to be planting rice in lowland conditions in 2008, which is a somewhat lower share than in 2002. The table also suggests that there was greater upland–lowland mobility in rice production between 2002 and 2008 than between the reference year and 2002. This could in part be due to the new and high-yielding upland rice varieties developed by the Africa Rice Center and the International Rice Research Institute (IRRI), which have been successfully introduced in many countries in Africa over the past 10 years. This has made upland rice production a more productive and profitable enterprise than it used to be [30].

Table 5 Rice technology adoption, management and commercialization

In relation to the adoption of improved varieties, we can also see a substantial increase. About 25 % of farmers reported having started using such varieties between 2002 and 2008 and in total 42 % of the respondents use improved varieties in 2008 compared to only 24 % in 2002. But only 14 % of households cultivated improved rice varieties in both 2002 and 2008. We can also see that only about 20 % of the households have used fertilizer at both points in time, and that the use of fertilizer has become slightly more common over time. For ploughing by tractor, there are modest changes over time.

The commercialization variables, referring to whether a farmer is marketing some of the rice crop and whether s/he has entered or exited the rice market or changed sales volumes over time, show that while the majority of farmers sell rice, there has been a slight withdrawal from markets, from 66 % in 2002 to 61 % in 2008. Out of these farmers, as many as 47 % reported that they had sold rice both in 2002 and in 2008. It can also be noted that mobility in and out of markets has increased over time: The share of households that stopped or reduced their sale of rice between 2002 and 2008 is similar to the share that started or increased the sale of rice during this time period. Over longer time periods, which span from the reference year to 2002 and from the reference year to 2008, respectively, very few stopped or reduced their sales, suggesting that market exit was temporary.

Macro-level conditions

To capture the influence of government policy, three macro-level variables are specified in Table 6, disaggregated by country. As the table shows, government expenditure on agriculture and rural development as a share of total government expenditure varies significantly between the countries, even if some improvements over time can be noted in all countries but Tanzania, likely reflecting a more conducive policy environment following the 2003 Maputo declaration. In 2008, only Malawi surpassed the 10 % share of public spending on rural development agreed to in this declaration. However, the explanation for the relatively high expenditure share in Malawi is principally due to the agricultural input subsidy programmes run in the country [31]. While Ghana and Mozambique were close to reaching the Maputo target, the share of government expenditure devoted to rural development was still considerably lower than in Malawi. The variable that we use to trace economic growth and development in general, i.e. gross domestic product (GDP) per capita, has improved over time in all countries.

Table 6 Macro-level variables for models

Finally, we consider rice imports. Table 6 (see also Fig. 3) shows that Malawi and Tanzania are almost self-sufficient in rice. For Malawi, this is probably due to the low share that rice constitutes in national diets, together with the fact that, as pointed out by Demont [10], it is a landlocked country, which makes large-scale rice imports prohibitively expensive. For Tanzania, consumer preferences are geared heavily towards local rice varieties, explaining the relatively low share of imports [10]. In contrast, the remaining three countries are increasingly dependent on rice imports. Although consumers are known to prefer local rice varieties in Mozambique, this is the country where import of rice in relation to gross domestic production is the highest [10]. This paradox shows the inability of domestic production—which in the case of Mozambique is still very small—to keep up with increasing consumer demand for rice. In Nigeria and Ghana, by contrast, Demont [10] reports dominant consumer preferences for imported rice varieties, likely explaining the dependence on imports in these countries. Nigeria, as the second largest importer of rice in the world, spends $1.56–2.2 billion per year on rice imports [32]. Moreover, it seems that in countries where consumers have already acquired a taste for imported rice it is difficult to reverse this trend, not the least because imported rice has established quality standards against which local rice (particularly, due to poor post-harvesting and milling practices) fails to compete [10]. Hence, it is of little surprise that the table shows that countries starting off with a big import share are even more import dependent today, by around 2.5 times so. Efforts to alter this have proven difficult. For instance, Nigeria initiated a rice transformation agenda in 2011 aiming for rice self-sufficiency [33]. Inter alia, the government announced that it was planning to ban rice imports from late 2015 to effectively end the country’s extreme dependence on rice imports [34], suggesting that Nigerian rice farmers could soon face a totally restructured market for their produce. As a first step, a 110 % duty on rice imports was introduced in January 2013 to encourage local rice production, although it did not have the desired effect as informal cross border trade rather than local production boomed [34, 35]. In 2014, a revised rice sector policy lowered the import duty to between 30 and 70 % up to 2017 in what appears to be an attempt to discourage smuggling and make investments in the rice sub-sector more profitable again [32, 35–37]. But as the ambitious target of achieving self-sufficiency together with the introduction of an import ban is approaching, it can be noted that Nigeria is projected to produce less rice in 2015 than it did in 2014, while imports are projected to reach a record 3.5 million tons, or up 17 % from 2014 [38].

Modelling production for the period p ₁ (reference year to 2002) ¹⁰

The first model we consider, model 1, deals with the period from the reference year (t ₀) to 2002 (t ₁), i.e. period p ₁. Model 1 is defined as a production function where $\ln (P_{{ijkt_{1} }} )$ is the logged production of rice (in kg) of household i in village j and country k at time t ₁.¹¹ The logged production, $\ln (P_{{ijkt_{1} }} )$, is a function of a number of explanatory variables that refer to circumstances either at the reference year (time t ₀), at the year 2002 (time t ₁), or during the period from the reference year to 2002 (period p ₁) (N.B. t ₁(L) indicates variables referring to lagged values with respect to t ₁).

Model 1 is specified as follows:

$$\begin{aligned} \ln (P_{{ijkt_{1} }} ) & = \delta_{{kt_{1} }} + \gamma_{{jkt_{1} }} + {\text{Farmage}}_{{ijkt_{1} }} \beta_{{1t_{1} }} + {\text{DH}}_{ijk} \beta_{{2t_{1} }} + {\text{Area}}_{{ijkt_{1} }} \beta_{{3t_{1} }} + {\text{IRV}}_{{ijkt_{1} }} \beta_{{4t_{1} }} \\ & \quad + \;{\text{AULL}}_{{ijkp_{1} }} \beta_{{5t_{1} }} + {\text{StoppedULL}}_{{ijkp_{1} }} \beta_{{6t_{1} }} + {\text{StartedULL}}_{{ijkp_{1} }} \beta_{{7t_{1} }} \\ & \quad + \;{\text{AUF}}_{{ijkp_{1} }} \beta_{{8t_{1} }} + {\text{StoppedUF}}_{{ijkp_{1} }} \beta_{{9t_{1} }} + {\text{StartedUF}}_{{ijkp_{1} }} \beta_{{10t_{1} }} \\ & \quad + \;{\text{AUP}}_{{ijkp_{1} }} \beta_{{11t_{1} }} + {\text{StoppedUP}}_{{ijkp_{1} }} \beta_{{12t_{1} }} + {\text{StartedUP}}_{{ijkp_{1} }} \beta_{{13t_{1} }} \\ & \quad + \;{\text{ASellingR}}_{{ijkp_{1} }} \beta_{{14t_{1} }} + {\text{StoppedSellingR}}_{{ijkp_{1} }} \beta_{{15t_{1} }} \\ & \quad + \; {\text{StartedSellingR}}_{{ijkp_{1} }} \beta_{{16t_{1} }} + {\text{FarmS}}_{ijk} \beta_{{17t_{1} }} \\ & \quad + \;{\text{Gender}}_{ijk} \beta_{{18t_{1} }} + {\text{Drought}}_{{jkt_{1} }} \beta_{{19t_{1} }} \\ & \quad + \;{\text{SGBudget}}_{{kt_{1} (L)}} \beta_{{20t_{1} }} + {\text{ImportR}}_{{kt_{1} (L)}} \beta_{{21t_{1} }} \\ & \quad + \;{\text{GDP}}_{{kt_{1} (L)}} \beta_{{22t_{1} }} + {\text{u}}_{{ijkt_{1} }} \\ \end{aligned}$$

where variables and notations (with suppressed subscripts when clarity is not standing the risk of being lost) are as described below:

ln(P):

logarithm of rice production (in kg)

Farmage:

age of the farm under the present management in 2002 (logged number of years)

DH:

dummy variable for descendant household

Area:

logarithm of area under rice (in ha)

IRV:

dummy variable for the use of improved variety

AULL:

dummy variable for having used lowland at both time points

StoppedULL:

dummy variable for having stopped using lowland

StartedULL:

dummy variable for having started using lowland

AUF:

dummy variable for having used fertilizer at both time points

StoppedUF:

dummy variable for having stopped using fertilizer

StartedUF:

dummy variable for having started using fertilizer

AUP:

dummy variable for having used tractor plough at both time points

StoppedUP:

dummy variable for having stopped using tractor plough

StartedUP:

dummy variable for having started using tractor plough

AsellingR:

dummy variable for having sold rice at both time points

StoppedsellingR:

dummy variable for having stopped or decreased selling of rice

StartedsellingR:

dummy variable for having started or increased selling of rice

FarmS:

dummy proxy variable for elite membership (i.e. among the upper decile of farms in the village with respect to their total cultivated area in 2002 and that had increased their cultivated area since the reference year)¹²

Gender:

gender of farm manager in 2002

${\text{Drought}}_{{jkt_{1} }}$ :

dummy variable for drought in E and S. Africa, for village j in country k at time t ₁

${\text{SGBudget}}_{{kt_{1} (L)}}$ :

logarithm of government expenditure share on agriculture and rural development in country k at time t ₁ , lagged by 3 years (i.e. in 1999)

${\text{ImportR}}_{{kt_{1} (L)}}$ :

logarithm of import of rice in relation to gross domestic production in country k at time t ₁, 5-year average, lagged by 3 years (i.e. 1995–1999)

${\text{GDP}}_{{kt_{1} (L)}}$ :

logarithm of GDP per capita in country k at time t ₁ lagged by 1 year (i.e. 2001) (constant year 2000 USD)

Finally, $\delta_{{kt_{1} }}$ is a latent characteristic (variable) for country k (a random intercept) at time t ₁; $\gamma_{{jkt_{1} }}$ is a random factor for village j in country k at time t ₁; and $u_{{ijkt_{1} }}$ is a residual, a random factor at the household level, i.e. for household i in village j and country k at time t ₁.

Considering the β coefficients, we can use the $\beta_{{3t_{1} }}$ coefficient to exemplify how these coefficients should be interpreted: By taking logged value of area, the $\beta_{{3t_{1} }}$ coefficient will directly reflect the scale effect of relative change of production with respect to relative change in area at time t₁. The interpretation of β coefficients for dummy variables is also straightforward, since the antilog of the β value (i.e. exp(β)) indicates the percentage change in production between households having the value of 1 for the dummy, as compared to those having the value 0.

Modelling production for period p ₁₊₂ (reference year to 2008)

The second model, model 2, is also a production function, referring to the period from the reference year (t ₀) to 2008 (t ₂) where ln($P_{{ijkt_{2} }}$) denotes the logged rice production of household i in village j and country k at time t ₂. Model 2 reconsiders the role of the explanatory variables as explanation of production of rice by covering a longer time period (p ₁₊₂). Hence, this second model replicates the results of the first model and thereby provides an indication of robustness and reliability of the models. Moreover, changes in regression coefficients may also be significant in terms of our assumptions.

Model 2 is specified as follows:

$$\begin{aligned} \ln (P_{{ijkt_{2} }} ) & = \delta_{{kt_{2} }} + \gamma_{{jkt_{2} }} + {\text{Farmage}}_{{ijkt_{1} }} \beta_{{1t_{2} }} + {\text{DH}}_{ijk} \beta_{{2t_{2} }} + {\text{Area}}_{{ijkt_{2} }} \beta_{{3t_{2} }} + {\text{IRV}}_{{ijkt_{2} }} \beta_{{4t_{2} }} \\ & \quad + \;{\text{AULL}}_{{ijk(p_{1} + p_{2} )}} \beta_{{5t_{2} }} + {\text{StoppedULL}}_{{ijk(p_{1} + p_{2} )}} \beta_{{6t_{2} }} + {\text{StartedULL}}_{{ijk(p_{1} + p_{2} )}} \beta_{{7t_{2} }} \\ & \quad + \;{\text{AUF}}_{{ijk(p_{1} + p_{2} )}} \beta_{{8t_{2} }} + {\text{StoppedUF}}_{{ijk(p_{1} + p_{2} )}} \beta_{{9t_{2} }} + {\text{StartedUF}}_{{ijk(p_{1} + p_{2} )}} \beta_{{10t_{2} }} \\ & \quad + \;{\text{AUP}}_{{ijk(p_{1} + p_{2} )}} \beta_{{11t_{2} }} + {\text{StoppedUP}}_{{ijk(p_{1} + p_{2} )}} \beta_{12t2} + {\text{StartedUP}}_{{ijk(p_{1} + p_{2} )}} \beta_{{13t_{2} }} \\ & \quad + \;{\text{AsellingR}}_{{ijk(p_{1} + p_{2} )}} \beta_{{14t_{2} }} + {\text{StoppedSellingR}}_{{ijk(p_{1} + p_{2} )}} \beta_{{15t_{2} }} \\ & \quad + \; {\text{StartedSellingR}}_{{ijk(p_{1} + p_{2} )}} \beta_{{16t_{1} }} + {\text{FarmS}}_{ijk} \beta_{{17t_{2} }} \\ & \quad + \;{\text{Gender}}_{ijk} \beta_{{18t_{2} }} + {\text{Floods}}_{{jkt_{2} }} \beta_{{19t_{2} }} \\ & \quad + \;{\text{SGBudget}}_{{kt_{2} (L)}} \beta_{{20t_{2} }} + {\text{ImportR}}_{{kt_{2} (L)}} \beta_{{21t_{2} }} \\ & \quad + \;{\text{GDP}}_{{kt_{2} (L)}} \beta_{{22t_{2} }} + {\text{u}}_{{ijkt_{2} }} . \\ \end{aligned}$$

The variables are all defined as before, when not considering the subscripts specifying the spatial and temporal aspects. As the subscripts in the model above show, the following variables (here with suppressed subscripts) refer to the same time period t ₁ as in the first model: Farmage, DH, FarmS and Gender; the following refer to time t ₂: Area and IRV; the following refer to time t ₂ (L) (i.e. time t ₂ with lagged values): SGBudget, ImportR and GDP;¹³ and the following refer to the behaviour over the longer time period from the reference year to 2008, period p ₁₊₂: AULL, StoppedULL, StartedULL, AUF, StoppedUF, StartedUF, AUP, StoppedUP, StartedUP, AsellingR, StoppedsellingR and StartedsellingR. One new variable is introduced, namely:Floods _jkt2 dummy variable for floods in W. Africa, for village j in country k at time t ₂.

Again,$\delta_{{kt_{2} }}$, $\gamma_{{jkt_{2} }}$ and $u_{{ijkt_{2} }}$ refer to random factors on the respective levels of the country, village and household, but at time t ₂. Also the interpretation of β coefficients follows the same logic as in model 1, but here referring to time t ₂.

Modelling changes in production

In the third model (model 3), the variable under study is the change in production represented by the logged ratio of production of rice in 2008 over production in 2002. By using panel data and a reduced form model as per Glewwe and Hall [39], this model was developed to identify the drivers behind changes in production through the combination of two separate models for the two panel rounds 2002 and 2008, respectively, and one model for the change in production between 2002 and 2008.

Also here, explanatory variables refer to either a state in 2002 (time t ₁) (alternatively to the reference year t ₀), or to the period from 2002 to 2008 (period p ₂). The influence from the explanatory variables on production at the various points in time may be different. This is then reflected in different β coefficients in the models at the two points in time. Δβ denotes the change in the β coefficient between 2008 (t ₂) and 2002 (t ₁) (i.e. $\beta_{{t_{2} }} - \beta_{{t_{1} }}$). Such a change in the influence from an explanatory variable can be estimated from a model of the logarithm of the ratio between production at the two time points (or equivalent: the difference of the logarithm of the production at the two time points) as seen by subtracting model 1 from model 2.

Changes in production between the two points in time may also result from changes occurring in the explanatory variables during the same time period. Especially, the macro-level variables have been subject to major changes between the two points in time, and we will therefore model the effects of these on changes in production levels.

Model 3 is specified as follows:

$$\begin{aligned} \ln \left( {\frac{{P_{{ijkt_{2} }} }}{{P_{{ijkt_{1} }} }}} \right) & = (\delta_{{kt_{2} }} - \delta_{{kt_{1} }} ) + (\gamma_{{jkt_{2} }} - \gamma_{{jkt_{1} }} ) + {\text{Farmage}}_{{ijkt_{1} }} \varDelta \beta_{1} + {\text{DH}}_{ijk} \varDelta \beta_{2} \\ & \quad + \;\varDelta {\text{Area}}_{{ijkp_{2} }} \beta_{3} + {\text{AUIRV}}_{{ijkp_{2} }} \varDelta_{4a} + {\text{StoppedUIRV}}_{{ijkp_{2} }} \varDelta_{4b} \\ & \quad + \;{\text{StartedUIRV}}_{{ijkp_{2} }} \varDelta_{4c} + {\text{AULL}}_{{ijkp_{2} }} \varDelta \beta_{5} \\ & \quad + \;{\text{StoppedULL}}_{{ijkp_{2} }} \varDelta \beta_{6} + {\text{StartedULL}}_{{ijkp_{2} }} \varDelta \beta_{7} \\ & \quad + \;{\text{AUF}}_{{ijkp_{2} }} \varDelta \beta_{8} + {\text{StoppedUF}}_{ijkp2} \varDelta \beta_{9} \\ & \quad + \;{\text{StartedUF}}_{{ijkp_{2} }} \varDelta \beta_{10} + {\text{AUP}}_{{ijkp_{2} }} \varDelta \beta_{11} \\ & \quad + \;{\text{StoppedUP}}_{{ijkp_{2} }} \varDelta \beta_{12} + {\text{StartedUP}}_{{ijkp_{2} }} \varDelta \beta_{13} \\ & \quad + \;{\text{ASellingR}}_{{ijkp_{2} }} \varDelta \beta_{14} + {\text{StoppedSellingR}}_{{ijkp_{2} }} \varDelta \beta_{15} \\ & \quad + \;{\text{StartedSellingR}}_{{ijkp_{2} }} \varDelta \beta_{16} + {\text{FarmS}}_{ijk} \varDelta \beta_{17} \\ & \quad + \;{\text{Gender}}_{ijk} \varDelta \beta_{18} + \varDelta {\text{SGBudget}}_{k} \beta_{20} \\ & \quad + \;\varDelta {\text{ImportR}}_{k} \beta_{21} + \varDelta {\text{GDP}}_{k} \beta_{22} \\ & \quad + \;({\text{u}}_{{ijkt_{2} }} - {\text{u}}_{{ijkt_{1} }} ) \\ \end{aligned}$$

Again, Farmage, DH, and Gender (with suppressed subscripts) are defined as before. ΔArea _ijkp2 refers to the change in logarithm of rice area of the household i in village j and country k during time period p ₂. The specific use of improved rice varieties during period p ₂ is described by introducing three new dummy variables (here with suppressed subscripts):

AUIRV:

dummy variable used for improved rice variety at both time points

StoppedUIRV:

dummy variable for having stopped using improved rice variety

StartedUIRV:

dummy variable for having started using improved variety

Elite membership is here (with suppressed subscripts) defined as:

FarmS:

dummy proxy variable for elite membership (i.e. among the upper decile of farms in the village with respect to their total cultivated area in 2002).

The variables AULL, StoppedULL, StartedULL, AUF, StoppedUF, StartedUF, AUP, StoppedUP, StartedUP, AsellingR, StoppedsellingR and StartedsellingR, all with subscripts i j, k, all refer to household i in village j and country k during period p ₂.

The macro-variable changes are: ΔSGBudget _k (change in logarithm of budget allocations to agriculture in country k between 1999 and 2005), ΔImportR_k [change in logarithm of import of rice in relation to gross domestic production in country k between the two lagged five-year time periods (1995–99 and 2001–05)] and ΔGDP_k (change in logarithm of GDP in country k from 2001 to 2007).

The latent variables that are constant over time will be eliminated through the formation of differences of log production. Likewise, latent household effects that are variable over time will contribute to the difference of log production as an effect from the difference of values of the latent variable. Hence, any systematic effect (the time-averaged effect) will be removed from the difference of log production and the remaining is only the variation in this effect, i.e. $u_{{ijkt_{2} }} - u_{{ijkt_{1} }}$. The same applies to the latent effects on country level, $\varDelta \delta_{k} = \delta_{{kt_{2} }} - \delta_{{kt_{1} }}$, as well as on village level, $\varDelta \gamma_{jk} = \gamma_{{jkt_{2} }} - \gamma_{{jkt_{1} }}$; only the difference between the effects in time between t ₁ (2002) and t ₂ (2008) contribute to the model. The particular effect of ageing (in time) is captured in the Farmage and DH variables. In this third model, the β coefficients reflect elasticities of production. Hence the β ₃ parameter, for instance, reflects the elasticity of production with respect to area, with values over unity reflecting intensification, i.e. increased production stemming from increased yields, while values below unity by contrast reflect extensification, i.e. expanded area with lower than expected proportional correspondence in yield.

One should note that even if model 3 shows similarities to what would have been the case if subtracting model 1 from model 2, model 3 is not the equivalent of the latter, since model 3 models the effects of changes in some of the explanatory variables, while for other variables it models the effect of changes in the coefficients.

The analysis of the data in all three models has been carried out through `mixed models’ using the statistical software program IBM SPSS Statistics 20. Furthermore, we have carried out the same analysis by using MLwiN program (developed especially for multilevel modelling). However, the results were similar to the SPSS results. Therefore, the presented results described below are from the SPSS program.

Results and discussion

Household control variables

As suggested by Table 7, the household control variables are not significant in any of the models. This means that neither the number of years the farm has been under its current management nor whether there has been a generational shift has any significant effects on production. As can be expected, the cultivated area is one of the most important determinants of production in all three models. Model 1 shows that that an increase in cultivated area of 100 % in the first period (p ₁) would result in a relative increase in production of 63 % on average. The effect in model 2, covering the longer period from the reference year to 2008 (p ₁₊₂), is slightly bigger, and in model 3, related to production change in period (p ₂), it is slightly smaller. These results indicate that about half of the change in area will be carried over to the change of production.

Table 7 Modelling rice production an d production change over time

Whether farmers plant rice in upland or lowland conditions does not seem to affect production results. This picture is consistent in all three models: the production outcome for those who grew rice in lowland conditions both in 2002 and 2008 did not differ significantly from the results in rain-fed uplands. Similarly, the level of production of farmers who planted lowland rice at both points in time did not differ from those households that either stopped or entered this type of production during the same period. This is somewhat surprising, given that paddy varieties are usually higher yielding than upland varieties [40]. The fact that a majority of the households that grew rice in lowland conditions did not use irrigation could shed some light on this issue, since lowland varieties perform best under irrigated conditions. Moreover, with the more recently developed high-yielding upland rice varieties, the differences between upland and lowland conditions may have been somewhat reduced [30].

Farm technology indicators

The effects of the farm technology indicators, namely improved seed, fertilizer and ploughing varied¹⁴. For all three models, the use of improved rice varieties did not significantly affect production. There could be various reasons for this. The farmers may not be able to distinguish effectively between improved varieties and traditional,¹⁵ undermining the reliability of the variable as such. This has been reported in relation to for instance pigeon peas in Tanzania [41] and wheat in Pakistan [42]. Moreover, many improved varieties need to be complemented by proper management practices, including fertilizer application [40, 43]. The fact that only a minority of the households in the panel reported the use of inorganic fertilizer could partly explain why the use of improved varieties did not affect production.

In the first period (p ₁), there is a strong relationship between higher production and both the usage of fertilizer and of tractor ploughing: Those who used fertilizer both in the reference year and in 2002 had 63 % higher production levels than those households that did not. Also, those who started using fertilizer during this period had levels of production that were 62 % higher than those who never used fertilizer. Similarly, those who ploughed with a tractor both in the reference year and in 2002 had levels of production that were 51 % higher than those who did not use this type of soil preparation. However, neither fertilizer use nor tractor ploughing had any significant effect on production during the other two periods.

The lack of statistical significance for the fertilizer and tractor ploughing variables for the latter period indicates that the relative dynamism in production observed after 2002 has not been significantly driven by fertilizer technology or tractorization. A possible explanation is that farmers applying fertilizer in the latter period used too small quantities to have any impact on production [44].

It could also partly be due to the floods in West Africa in 2008, which had an overall negative (−59 %) impact on rice production, while the 2002 drought in Eastern and Southern Africa did not seem to have had any significant negative effect on rice production. The lack of association between tractor ploughing and rice production might also be related to farmers using other types of techniques for land preparation, such as oxen ploughs. Both these findings warrant further research to explore the specific reasons.

Commercialization variables

The commercialization variables have strong impact on the production dynamics in all three models. Increases in production are linked to having sold rice as well as to having entered the market or to increased market participation during the considered periods. For instance, the selling of rice both in the reference year and in 2008 had 62 % higher production compared to those households that did not sell at those points in time. Farmers who sold rice both in 2002 and 2008 had about 50 % higher production than those who did not. For households who entered the market or increased their sale of rice in period p ₂, this had a particularly strong impact on production, which was then 62 % higher than in 2002. Moreover, the withdrawal from market activities between 2002 and 2008 or selling smaller quantities of rice in 2008 as compared to 2002 had a significantly negative effect on production, which decreased by 56 %. Hence, while household market integration has been important throughout these years in explaining rice production performance, it seems to become even more important towards the end of the period.

Macro-level indicators

We now consider the macro-level policy indicators for all three models. The share of state budget devoted to agriculture and rural development has not had a significant effect on production. This is somewhat surprising, given that, as noted in Table 6, all the countries except for Tanzania have substantially increased their agricultural budgets over time. Our data do not indicate that the new political commitment after the 2003 Maputo declaration has translated into any significant effects benefitting rice farmers. Imports of rice, on the other hand, have a significantly negative effect on rice production in the period p ₁, supporting the hypothesis that domestic production is undermined by cheap imports. However, in relation to the other two models, this effect was not significant. Yet, imports became more important vis-à-vis domestic production for all countries except Tanzania. The reasons for this need to be explored, particularly regarding the effects of higher import tariffs on rice, introduced in the wake of the collapse of the Doha round in the fall of 2006, as well as to rising world market prices for food staples after 2007. These developments may have made domestically produced rice more competitive and thereby provided incentives for farmers to produce more.

The third macro-level variable, GDP per capita, a proxy for economic development in general, had a significant positive effect on rice production in the first period p ₁ as indicated by a 45 % volume increase. The powerful role that agriculture can play in fostering national economic growth during the early stages of development has been well established [45, 46], and more recent literature points to the two-way linkage here, i.e. agriculture can both stimulate and be stimulated by national economic growth, reflecting the interdependence of the agricultural and non-agricultural sectors in the economy [47]. In the specific case of rice, we believe that it is mainly due to the increases in consumer demand as countries develop, caused by factors including higher incomes, changing consumption patterns and urbanization, which in turn creates a production incentive for the rice farmers. However, when considering the two time periods covered by models 2 and 3, variations in GDP per capita do not have any significant effect on production suggesting that the links between production and macro-level developments have weakened over time. Also in the first period, they were more a reflection of global processes and overall national economic development than the state prioritizing the agricultural sector¹⁶. Hence, our case study does not give any obvious evidence for an ongoing state-driven Green Revolution in rice in SSA.

Distributional variables

When considering the distributional variables, we see that the gender of the farm manager is not statistically significant in any of the models. This suggests that women farm mangers are not discriminated in relation to rice production, and that their farms perform as well as those of their male counterparts. This is consistent with findings that women and men farmers are equally efficient farm managers [48–52]. It can also be noted that households defined as belonging to the village elite, with other conditions equal, had 36 % higher production than the other households in the first period, p ₁. However, the role of elite membership seems to have dwindled in significance over time and did not affect production outcomes for the other periods. This could be an indication that the earlier elite bias is disappearing and that smallholder farmers are not discriminated against the way they used to be. In addition, it suggests that the recent dynamism in rice markets have brought more smallholder farmers into commercial production, which is also supported by the market entry coefficient in the third model, which is positive and significant at the 1 % level. Similar findings are found also for maize and cereal crops in general for the wider data set [28, 53, 54]. Hence, contrary to the Asian experience, the development during the first decade of the new millennium seems to have been driven more by the market than the state.

Table 8 considers the variance by country, village and household. As can be noted, the largest share of variance is found at the household level and is significant over all the time periods pointing to the importance of variables endogenous to households when explaining production performance.

Table 8 Estimates of variance component effects by country, village and farm/household effects

Although the country effect variation in period p ₁₊₂ (0.62) may indicate general differences between countries, this is not significant due to the small number of countries. In period p ₁, we do not even see such an indication. In period p ₁₊₂, there is significant variation in between-village effects on the average rice production, indicating that different villages have different conditions for the production of rice. Such a variation is, however, not observed in period p ₁. This could point to more recently introduced factors that affect production at the village level, e.g. improved water control techniques or improved access to market outlets and processing facilities. In period p ₂, the between-village effect variation is very low and not significant, indicating that the village-dependent effects are rather unchanged over time (from 2002 to 2008) or altered in the same direction and magnitude for a majority of the villages from 2002 to 2008.

Conclusions and implications for policy

In this article, we have reported on the production performance amongst rice-growing households in five Sub-Saharan countries: Ghana, Malawi, Nigeria, Tanzania and Mozambique. Using panel data for 317 households interviewed in 2002 and 2008, we have shown that while the mean rice area among the surveyed households has increased by 17 % between 2002 and 2008, rice yields have decreased by 12 % and the mean rice production has remained almost constant with only a 2 % increase. But we also noted substantial cross-country and cross-village variations, obscured by these figures, suggesting that a mixed pattern of extensification and intensification explain production increases.

We have also identified and evaluated the role of three main types of production drivers: the role of commercial drivers, farm technology and macro-level conditions, studying three time periods: one lasting from on average 1984 (i.e. the reference year) to 2002, one lasting from the reference year to 2008 and one lasting from 2002 to 2008. Of particular interest has been the last period, given increasing political attention to the welfare of smallholders across the continent in recent years, but also to the role of smallholders as key drivers of food production. More generally, the agricultural sector has been identified as the core sector for achieving broad-based development.

Using statistical modelling, we have shown that until 2002, the key drivers of production were a combination of area expansion, market integration, farm technology in terms of fertilizer and tractor ploughing, village elite membership and key macro-level conditions including rice imports and overall economic development in the countries as operationalized by GDP per capita. In relation to the period 2002–2008, we found that production increases were primarily associated with area expansion and commercial drivers. In none of the periods considered did the share of the state budget allocated to agriculture have a significant effect on production.

The fact that there is a strong association between area and production throughout the models, which is not matched by an association between new technology (improved varieties, fertilizer, tractor ploughing) and production in relation to the years 2002 and 2008, suggests that production in later years is more driven by processes of extensification than intensification.

Rice in SSA is still primarily a staple food preferred by urban consumers for whom rice is much more important than national consumption averages suggest. Considering that urbanization levels continue to increase, and that the preference for rice is increasingly spreading to rural areas, the demand for rice may well grow even more rapidly over the coming years, compared to the past few decades [see also 3]. Concerted efforts to increase production of rice are therefore urgently needed.

The role of commercialization in explaining changes in production in the three models suggests that policies that effectively combat the typical negative characteristics of food staple markets in SSA—including uncertainty, discouraging prices, atomism and high transaction costs [55]—hold great potential for driving rice production in the continent in the near future. However, even with improved staple food markets, the possibilities of further growth in the rice sector in SSA are limited without an intensification of production due to the scarcity of available and suitable land for rice production. Eliminating barriers restricting the access to new farm technology for small rice farmers would considerably improve the possibilities of a positive development of rice production in SSA, which in turn would contribute to an improvement of food security and to a reduction of both rural and urban poverty.

Availability of supporting data

The data set supporting the results of this article is available in the Afrint repository at http://www.keg.lu.se/en/research/research-projects/current-research-projects/afrint.