PROTOCOL: Agricultural Input Subsidies for improving Productivity,Farm Income,Consumer Welfare and Wider Growth in Low- and Middle-Income Countries: A Systematic Review

Access to agricultural inputs

With 12.5 per cent of the world's population currently undernourished (FAO, 2013) there is an urgent need to improve food security. Increased agricultural productivity has been identified as a potential means for improving the food supply aspect of food insecurity, and agricultural inputs can help to increase productivity greatly (Gordon, 2000). Adequate inputs are important to achieving agricultural productivity in developing countries (Buringh & Dudal, 1987; Hazell et al., 2007, Ajah & Nmadu, 2012). There is a strong concern that in developing countries, the inputs and technologies needed to achieve increased productivity are financially unaffordable or unattractive to many poor farmers (e.g. Wiggins & Brooks, 2010). Agricultural input subsidies are a potential way of incentivising farmers to purchase inputs that they are unable or unwilling to obtain at market rates; for example, because they lack access to credit or find or expect the inputs to be unprofitable at market prices given existing knowledge about their benefits, and so on. Input subsidies could thus provide a means for achieving higher agricultural productivity, improved food security and, through lower food prices, pro-poor economic growth.

There is, however, considerable contention among policy makers and analysts regarding the effectiveness and efficiency of investments in agricultural input subsidies and the conditions under which they may or may not work (Wiggins & Brooks, 2010; Kilic et al., 2013; Pauw & Thurlow, 2014).

Agricultural input subsidy interventions

Agricultural input subsidies were common in poor rural economies in the 1960s and 70s, but conventional wisdom deemed them ineffective by the 1980s and 90s (Dorward, 2009). However, in recent years, there has been a resurgence of interest and investment, mainly in Africa, in so-called ‘smart subsidies’, which seek to maximise the multiple benefits of subsidies to different stakeholders while minimising their distortionary effects on inter alia efficient commercial market operation and development (Morris et al., 2007). This resurgence was the result of a need to boost food security in the short-term while also implementing longer-term investments to raise productivity. Such programmes use vouchers for targeting, rationing and private-sector delivery.

How subsidies might work

Impacts on agricultural production and productivity are the direct impacts that are expected from input subsidy programmes. Historically, the economic justification for agricultural input subsidies involved reducing temporary knowledge and risk constraints to farmers' adoption of improved technologies and practices, and increasing productivity with farmer benefits and/or consumer gains, from produce and labour market and price effects.

However, wider potential subsidy impacts associated with the introduction of new and innovative agricultural input subsidy programmes are increasingly recognised and investigated, including private market development, soil fertility replenishment, social protection, national and household food security, and growth (Dorward, 2009). These indirect (or dynamic) environmental, income and consumer welfare effects are only expected when there are direct (or static) production effects.

These wider impacts, still critically dependent on links from productivity impacts, are affected by a variety of factors, including:

Subsidy design and implementation: inputs subsidised, targeting and access, beneficiary household characteristics, complementary inputs and investments including credit, subsidy scale and rationing (for example, value per beneficiary), market effects and technical fit (for example, marginal productivity);

This is illustrated in the impact pathway attached (Figure 1). This figure has been developed from reviews of theory and literature (e.g. Dorward, 2009 and intensive evaluation by Dorward and Chirwa of the Malawi Agricultural Input Subsidy from 2006/7). It shows the main outcomes and impacts flowing from input subsidies and the causal links between them. On the right hand side are the main contextual influences that affect impacts and the links between them.

OPEN IN VIEWER

1

Impact Pathway of Agricultural Input Subsidies

A policy-relevant systematic review of agricultural input subsidy impacts therefore requires:

examination of direct (static) effects such as impacts on beneficiary farmers' incomes;

Why it is important to do the review

While subsidies sometimes played a key but time-limited development role (e.g. Timmer, 2004), empirical studies generally revealed negative impacts and difficulties in cost control, diversion (inputs being stolen or used by others other than the intended recipients), overuse of inputs and capital (discouraging more labour-intensive methods), regressive benefits, and market distortions inhibiting private investment in agricultural services (e.g. Ellis, 1992; Morris et al., 2007; Timmer et al., 2009). Dominant analyst and donor thinking therefore considered agricultural subsidies to be generally ineffective and inefficient.

This viewpoint has been challenged by re-examination of successes and failures of both state- and market-led agricultural development (e.g. Fan et al., 2004; Djurfeldt et al., 2005; Dorward, 2009), and by calls from African governments, NGOs and some donors for input subsidies to address agricultural stagnation in Africa. Furthermore, wider potential subsidy impacts associated with different delivery mechanisms are, in turn, generating new impact studies (e.g. Ricker-Gilbert et al., 2009; Liverpool-Tasie et al., 2010; Ricker-Gilbert et al., 2010; Ward & Santos, 2010; Chirwa et al., 2011; Pan & Christiaensen, 2011; Liverpool-Tasie, 2012).

While the literature has been reviewed (e.g. Chirwa & Dorward, 2013), no systematic review of agricultural input subsidies has, to our knowledge, been published in the past, using systematic data collection, critical appraisal and statistical synthesis using meta-analysis, and neither have more general reviews been sufficiently theoretically rigorous in addressing the range of possible programmes, outcomes and impacts discussed above.

This systematic review will therefore not only provide the first systematic review of this topic, it will also address a major gap existing in general literature reviews by taking a more holistic approach in addressing direct and indirect effects.

Population

Eligible populations are people for whom data are collected at all levels (e.g. country, region, community, household or individual) living in a low- or lower-middle-income country (as defined by the World Bank) at the time the intervention was carried out. ‘Low- and lower-middle-income countries’ as defined in March 2012 by the World Bank are divided according to 2008 GNI per capita, calculated using the World Bank Atlas method ² . The group of countries included consists of the 36 low-income economies ($1,025 or less); and 48 lower-middle-income economies ($1,026 to $4,035) at the time of the search. Studies that relate to countries that were developing countries by such definition during the experimental study period but are no longer on the World Bank list at the time the search is conducted, will not be explicitly sought, but will be included where identified. The populations considered include direct beneficiaries of the intervention (farmers and farm households) and those who may be indirectly affected (wage labourers and food consumers).

Intervention

Interventions studied will be limited to direct agricultural producer subsidies for inputs. ‘Agriculture’ is defined as animal or crop production (i.e. excluding forestry and fisheries). ‘Agricultural input subsidies’ are defined as grants (or loans, if repaid at below the market price) given to a farmer as a means of reducing the market price of a specific input used in agricultural production (Harman, in prep). We will distinguish between full (100%) and partial subsidies as appropriate. Credit and loans for unspecified purposes or inputs are not included in this study as they have distinct economic effects, as covered in a much broader literature. We will also exclude early-stage agricultural research station field trials and humanitarian relief programmes, as the adoption of these trial inputs and such emergency interventions are unrepresentative of impacts of input subsidies in normal agricultural practice.

Comparator

Eligible comparisons include no active agricultural input subsidy intervention, wait-list, alternate input subsidy intervention, or other interventions providing access to inputs. We will collect data on comparison conditions and test for systematic differences in effects accordingly in moderator analysis.

Outcomes

Study Designs

The different elements of the review question pose different challenges in empirical determination of subsidies' impacts. This can be seen with reference to the impact pathway in Figure 1: some impacts affect subsidy beneficiaries directly (for example, changes in beneficiaries‘ productivity and incomes) while others affect beneficiaries and non-beneficiaries indirectly (farm incomes, wages rates, consumer welfare and wider growth). While direct impacts are normally amenable to experimental and quasi-experimental investigation comparing beneficiaries and non-beneficiaries, indirect impacts are not amenable to such investigation unless the subsidies are administered in restricted areas within which market impacts are also restricted, using clustered designs at appropriate levels. This situation is not common and does not usually apply to large-scale subsidies, which are of major policy interest.

In view of these differences between potential methods for investigating direct and indirect subsidy impacts, we define different types of study design that will be included to answer questions on direct and indirect effects (Table 1). We will only include studies that involve some counterfactual comparison of results with and without subsidy treatments.

Direct subsidy beneficiary impacts: methods include randomised control trials and studies that use some formal methods for removing likely biases from non-random assignment of subsidy receipt. Such methods include regression studies using difference in differences (or fixed effects models), instrumental variables regression, regression discontinuity, and propensity score matching methods, as appropriate for analysing panel or cross-sectional household data with randomised or quasi-randomised beneficiary selection or beneficiary selection by programme planners.

Indirect subsidy impacts: methods include studies with appropriate cluster-level assignment (e.g. randomisation at the level of the market), quasi-experimental methods such as interrupted time series and fixed effects regression analysis across different regions or states (e.g. Fan et al., 2004), as well as simulation models that allow comparison of with and without subsidy situations (for example, partial equilibrium models (PEM), CGE, and other statistical simulation models that link direct subsidy impacts into wider labour and produce markets).

It is recognised that studies vary in their methodological rigour, in the quality of data, and in their contextualisation. Clear recording criteria will be utilised for coding each study on these parameters.

Critical appraisal

In assessing the studies, the clarity of aims, clarity and appropriateness of methodology, isolation of an appropriate output metric, commentary on potential mechanisms, use of controls or counterfactuals, and clarity of findings, will be considered and reported in the descriptive analysis. Potential sources of bias will be scored using a traffic-lights system (indicating high risk, medium risk and low risk of bias, as appropriate). The following categories of bias will be assessed for experiments and quasi-experimental studies (Waddington et al., 2012): confounding and sample selection bias; performance bias; motivation bias; reporting biases; and other sources of bias (see Annex 3). Sources of bias for statistical modelling studies will include factors relating to model specification (e.g. source of model coefficients) and methods of inference (e.g. use of systematic sensitivity analysis ⁵ ).

These scorings will not be used as weights in the analysis. We will firstly report direct and indirect outcomes separately along the causal chain, and secondly explore sensitivity using risk of bias categories.

Where information regarding the reasons for heterogeneity is presented in the studies, it will be recorded, distinguishing whether this information is based on empirical evidence or author's hypotheses.

Moderator variables

This review will assess findings from a wide range of studies conducted under widely differing circumstances. Data on relevant contextual factors (as indicated in Figure 1) will be collected and taken into account as ‘effect modifiers’ in analysis and interpretation of the results of the systematic review. Any effect modifiers identified a posteriori will be indicated clearly as such in the report. The following moderators have been identified (Annex 3).

Subsidy design and implementation: inputs subsidised, targeting and access, beneficiary household characteristics, complementary inputs and investments including credit, subsidy scale and rationing (for example, value per beneficiary) and technical fit (for example, marginal productivity).

Output market characteristics: tradable/non-tradable, staple/non staple foods and cash crops, input and product demand elasticities (allowing for close substitutes).

Livelihood and economy characteristics: wealth, economic activity and activities such as different farm and non-farm activities, diversity within and between livelihoods, production systems, asset holdings and attributes (e.g. land, machinery and education).

Infrastructure: roads (for example, all weather or seasonal roads), information communications technology (ICT), access to international markets and irrigation.

Other interventions, institutions and policies: land tenure and property rights, access to social protection, produce pricing, farmer credit access, other forms of subsidy and taxes affecting agriculture.

Agro-ecological, climate and weather conditions: spatial and temporal rainfall distribution, droughts, soils, temperatures, pests and diseases.

Measures of the treatment effect

The statistical evidence in the papers will be extracted with the intention of comparing effects of input subsidies on outcomes. A sample of data extraction will be done by two team members to ensure consistency. We will examine changes in comparable criteria at fixed periods after subsidies have been introduced, or have been withdrawn. We will collect data at multiple follow-up periods and synthesise these appropriately across studies.

We will extract data to compute standardised mean difference effect sizes for continuous outcomes, and odds ratios for dichotomous outcomes, using methods outlined in Lipsey and Wilson (2001). Methods for computing effect sizes for treatment effects estimated using regression adjustment are presented in Annex 4. We will ensure all effect sizes are calculated consistently, so that the direction of change reflects a uniform increase or decrease in the outcome variable across studies (e.g. where studies estimate effects of introducing or removing subsidies).

Criteria for determining independent findings

We will only include independent effect sizes in any single meta-analysis. This means that if we are pooling across studies, we will only take a single effect estimate for that study. It also means that when we have multiple studies on the same programme, we will seek an appropriate method of generating a single study effect. We will use the following decision criteria to determine independent findings: 1) where multiple specifications are presented for a single study, we will choose the method with the lowest risk of bias (usually the least parsimonious in terms of covariates for quasi-experiments); 2) as used in Baird et al. (2013), where we have multiple independent estimates for sub-populations we will calculate a ‘summary effect size’ using inverse-variance weighted random effects meta-analysis; 3) where we have multiple dependent estimates we will calculate a ‘synthetic effect size’ using the approach given in Borenstein et al. (2009; Chapter 24).

Unit of analysis errors

We will use the appropriate unit of analysis for clustered studies when calculating standard errors of the effect. For clustered studies, if the authors do not state they have done so, we will adjust the standard error upwards using the standard formula in the Cochrane Handbook (Higgins & Green, 2011).

Dealing with missing data

To calculate standardised mean differences data on the standard deviation of the outcome variable are needed. Where this is not reported, we will contact primary study authors to obtain this, where possible, or use an effect size measure which does not require outcome standard deviation such as the response ratio (see Annex 4). Where data are not reported on confidence intervals in simulation studies (e.g. due to lack of sensitivity analysis), we will report effect sizes only.

Treatment of qualitative research

We do not plan to include qualitative data in this review.

Methods of synthesis

We will present effect sizes and 95 per cent confidence intervals using forest plots. Where constructs are considered sufficiently similar, we will also estimate pooled effect sizes across studies using inverse-variance weighted random effects meta-analysis using Stata software (Stata Corp, TX, USA). We will present results from studies measuring direct and indirect outcomes, and studies using experimental, quasi-experimental and simulation models separately. We will examine sensitivity of findings to study design and risk of bias assessment, and other factors such as length of follow-up period. We will conduct moderator analysis according to the moderators defined above, and if sufficient studies permit we will conduct analysis of publication bias using conventional methods such as funnel graphs. If sufficient studies are available, we also plan to use meta-regression to examine sensitivity to study design parameters, moderator effects and other effects such as small sample sources of publication bias (Egger's test (Egger et al., 1997)).