The Evidence Mounts: Poverty, Inflation and Rwanda

By Sam Desiere

In a recent blogpost an anonymous researcher on showed that poverty in Rwanda has increased from 2011 to 2014 by 5 percentage points. This contradicts the official poverty statistics and narrative, which claim that poverty decreased by 5.8 percentage points, namely from 44.9% in 2011 to 39.1% in 2014 (NISR, 2015). Importantly, the author published the Stata-files used to analyse the data of the EICV 3 and EICV 4 household surveys, enabling other researchers to verify his claims.

Recently, I also calculated trends in poverty using the same datasets. Although I used a slightly different (and, arguably, less sophisticated) methodology, the results confirm that poverty did not decrease. In addition, I show that the poverty trends are very sensitive to the inflation rate used. With an inflation of 16.7% (as reported by the National Institute of Statistics of Rwanda, NISR), poverty indeed decreased by at least 5 percentage points. With an inflation rate of 30% – which is in my view more in line with the ‘real’ inflation rate – my estimates show that poverty increased by 1.2 percentage points.

The fact that two researchers arrive – independently from each other – at the same conclusion, strengthens my belief that the EICV surveys show that poverty in Rwanda has increased. This has important implications for the current debate about (rural) policies in Rwanda, but I leave a discussion of these implications to researchers and policy makers more familiar with the reality on the ground and focus in this blogpost on the technical aspects of estimating poverty trends.

In this this post, I briefly describe my methodology and key findings and discuss (food) price inflation, which turns out to be a critical parameter. The Stata do-files required to replicate my findings can be found here.


Rwanda’s poverty estimates are based on the Integrated Household Living Conditions Survey (EICV by their French acronym), which are conducted every three years. I used data from EICV 3, conducted in 2010/11 and EICV 4, conducted in 2013/14, which are made publicly available by the NISR. More specifically, I used the modules on food consumption purchased on the market and food consumption from own production. In both waves, the questionnaire of both modules is nearly identical. Food consumption is reported for more than 100 food items.

Unlike the anonymous researcher, I did not use the modules on non-food expenditure. I did so for two reasons. First, the NISR reports that most households spend over 60% of their budget on food. Hence, food expenditure is a good proxy of poverty. Second, non-food expenditure would require some additional data cleaning, which requires additional assumptions. Hence, I simply calculated food expenditure in both waves.

The meta-data of EICV 4 (available on NISR’s website) clearly stipulates that each sampled household in Kigali was visited 11 times over a period of 33 days. The modules on food consumption were administered during every visit. Rural households were visited 8 times over a period of 16 days. The meta-data of EICV 3, however, does not provide information on the number of times a household was visited. I simply assumed that the same methodology, for both rural and urban households, was followed in wave 3 as in wave 4. If this assumption is wrong – something I could not check – the results presented below will be erroneous.

In both waves, households reported how much they had spent on food purchased on the market by food item since the previous visit of the enumerator. I simply added up expenditure on all food items. Households also reported how much they had consumed from own production. Converting the consumption from own production in monetary values was more challenging. Households typically reported consumption from own production in kg. Some households also reported in the same module how much they would have paid on the market for this food item. I used this information to calculate the median, national price for each food item and used this price to convert consumption from own production in its monetary value. Since relatively few households reported prices, I did not attempt to calculate region specific prices nor did I correct for price seasonality. On this point my methodology differs from the anonymous researcher, who calculated a Laspeyres price index to account for spatial and temporal price variation.

To verify my assumptions, I checked whether my estimates of food expenditure are correlated with the household poverty status as reported by the NISR and included as a separate variable in the datasets. In both waves, food expenditure was lower for households classified by the NISR as extremely poor compared to household classified as poor, and the expenditure of this group was in turn lower than the expenditure of non-poor households. These results, available upon request, confirm that my assumptions are at least partially similar to the assumptions of the NISR.

Food expenditure can only be compared between the waves if the food inflation rate between 2010/11 and 2013/14 is known. I used two different inflation rates. First, I used an inflation rate of 16.7%, which is reported by the NISR (NISR, 2016, p. 43). Second, I estimated inflation based on food prices reported by the respondents, which I also used to convert food consumption from own production in monetary values. Inflation is then defined as a weighted average of the price increase of nine important crops. I used the same weights as those used by NISR to construct the 2013/14 adjusted food poverty line (NISR, 2015, table B4, p. 38). These estimates of inflation will be discussed in greater detail below.

Since I did not calculate total expenditure, but only food expenditure, I could not use the poverty lines proposed by the NISR. I therefore followed the ‘inverse’ methodology. First, I assumed that the NISR correctly estimated poverty in 2010/11 (44.9%) and used this information to determine the food expenditure threshold in 2010/11 prices that corresponds with this poverty rate. Second, I deflated food expenditure in 2013/14 using two different inflation rates, namely 16.7% and 30%. The first inflation rate corresponds with the inflation rate used by the NISR and thus allows me to replicate the findings of the NISR. The second inflation rate corresponds with my own estimate of inflation using the price data from EICV 3 and EICV 4. Third, I used the food expenditure threshold as an alternative to a poverty line to estimate the poverty rate in 2013/14. This approach is valid because I am not interested in ‘absolute’ poverty figures, but only in poverty trends.

In all analyses, I used the population weights to make the results nationally representative.


Poverty trends

Using the EICV 3 and EICV 4 datasets, I calculated food expenditure per adult equivalent, respectively in 2010/11 prices and 2013/2014 prices. In order to estimate poverty trends, food expenditure in 2013/14 has to be deflated to express it 2010/11 prices. Poverty trends are very sensitive to the inflation rate used to deflate food expenditure. Results are presented for two inflation rates: (1) an inflation rate of 16.7% as reported by NISR and (2) an inflation rate of 30%, which is at the lower end of my inflation estimates based on ESOKO price data or EICV price data (see below for a discussion of inflation).

Figure 1 shows cumulative frequency distributions of food expenditure for these two situations, while table 1 summarizes poverty trends

With an inflation rate of 16.7% (left panel, figure 1), real food expenditure per adult equivalent increased for all households from 2010/11 to 2013/14 and, as a result, poverty decreased. Assuming a poverty rate of 44.9% in 2010/11 (which corresponds to a food poverty line of 100,232 RWF per adult equivalent), poverty decreased by 7.9 percentage points. This poverty reduction is even more pronounced than reported by official statistics, which states than poverty decreased by 5.8 percentage points.

With an inflation rate of 30% (right panel, figure 1), food expenditure does no longer increase between 2011 and 2014 for all households. Again assuming that poverty is 44.9% in 2010/11, poverty even increased by 1.2 percentage points.

Figure 1: Cumulative distribution of food expenditure per adult equivalent for EICV 3 and EICV 4 for an inflation rate of 16.7% and 30%

Table 1: Poverty trends in function of the inflation rate

  Inflation: 16.7% Inflation: 30%
Poor HH EICV 3 (official statistics) 44.9% 44.9%
Poor HH EICV 4 (own estimates) 37.4% 46.1%
Trends in poverty (percentage points) -7.5 +1.2

In sum, the poverty trends are very sensitive to the inflation rate. With an inflation of 16.7% from 2011-2014, poverty decreased by at least 5 percentage points, which is in line with the official reports. With an inflation rate of 30%, poverty does not decrease. The question thus boils down to an accurate estimation of the inflation rate between 2011 and 2014.

Inflation rate

The EICV survey is not an ideal dataset to estimate inflation, because it does not contain much information on food prices. As explained earlier, some households report prices for those food items consumed from own production. This does not only mean that the number of observations is relatively limited, but also that households report prices of those items they did not buy on the market. I nevertheless used this information to calculate mean and median average prices by food item. I calculated national averages without taking into account price seasonality or regional price differences. In order to estimate ‘average’ inflation, a weighted average is taken over nine crops. The weights are proportional to the weights used for the construction of the 2013/14 adjusted food poverty line (NISR, 2015, table B4, p. 38). These nine crops account for 86% of the total calorific intake of the food basket. Two crops dominate this index: cassava (fermented) (weight: 38%) and dry beans (weight: 25%).

Figure 2 shows the increase in mean and median prices between 2010/11 and 2013/14 for nine crops, while the horizontal lines indicate the weighted average. The increase in median prices ranges from 10% for sorghum to 50% for cassava (both flour and roots). Median and mean inflation are 33% and 42%, respectively. This corresponds to an annual inflation of 9.5% and 12.5%, respectively.

Figure 2: Price increase for nine crops from 2010/11 to 3013/14 (mean and median prices)


These inflation estimates are substantially higher than the ones reported by NISR, which states that food prices increased by 16.7% between Jan 2011 and Jan 2014 (NISR, 2016, p. 43). Moreover, the estimates based on the EICV surveys are remarkably similar to the estimates based on detailed ESOKO price data, where I estimated inflation at 30.5% over the 2011-2014 period (details not reported here).

In sum, I believe that the ‘real’ food inflation rate is substantially higher than the one used by NISR to estimate poverty trends. This probably explains why I find that poverty increased, while the NISR reported that poverty decreased. These findings raise concerns, not only for Rwanda’s (rural) policies, but also for international donors that have presented Rwanda as a model for development because of the supposedly strong poverty reductions.

Sam Desiere is currently a senior researcher at HIVA, the research institute for work and society of the University of Leuven, Belgium. In 2015 he obtained a PhD in agricultural economics from Ghent University, Belgium, which focused on data quality of household surveys in developing countries.

Featured Photograph: As part of the DFID funded Vision 2020 Umurenge Programme (VUP), Rwanda’s flagship Social Protection Programme, women and men in northern Rwanda work on a public works site in 2012, building terraces to prevent soil erosion 


  1. After reviewing the do-files carefully, we suggest two small corrections that would bring the author’s estimates more in line with the official calculations and with the previous blogpost. First, on line 128 of the EICV4 do-file (and EICV3-different line) we suggest to use actual or at least regional prices in the first instance, and only use national averages where price information is missing. This is because regional price variations can be significant in Rwanda, thus leading to an overestimation of the value of consumption in poor regions. Secondly, on line 151, and 225, we suggest using information about the number of months in which the item was consumed (s8bq2 and s8cq2) in order to scale up to yearly consumption values. This is because poor households may only consume the crop at harvest season, whereas richer households may be able to consume the crop all year round. So we would underestimate the gap between poor and rich households’ consumption without this adjustment.
    If these two changes are made, the increase in poverty at 30% inflation goes from +1.2% to +5.2%, which is very close to earlier obtained results. The rest of the difference can be accounted for by the fact that this study does not consider non-food consumption, and other methodological differences between the two papers.
    With this small amendment, the most important contribution of this paper is probably that it has confirmed the increase in poverty using what appears to be a third source for the price data, namely the ESOKO price dataset. So it seems that we now have up to three different sources indicating that poverty increased significantly between 2010 and 2014 in Rwanda: (1) The “Institute of Statistics of Rwanda/ Price Statistics Division” cited in the original NISR study (NISR 2015, p.16); (2) The EICV price data used in this study and the previous ROAPE blogpost; (3) The ESOKO price data. This suggests that the findings are very robust and puts the onus on NISR to explain why the CPI data used in the official report yields such different results.
    As the author rightly points out, understanding what happened to inflation is key to understanding poverty dynamics in Rwanda, due to the type of transmission mechanisms at work in the agricultural sector. There are at least two possible reasons why CPI data may not give an adequate picture of poverty dynamics in Rwanda (assuming that the CPI data are not just wrong or doctored). First, CPI data are not modelled on the consumption patterns of the poorest households. In Rwanda, consumption patterns of poor rural susbsistence farmers can be quite different from those of the average consumer, due to high inequality and a strong urban/rural divide. Second, CPI data focuses exclusively on consumer end-prices and therefore does not capture the complex mechanisms through which impoverishment occurs in Rwanda: Forced monoculture, coupled with poorly functioning markets and infrastructure, means that over-production of mandated crops often translates into localised price collapse at harvest season. The surplus production is purchased at low prices by local authorities and stored in warehouses to be re-sold when prices have recovered. This means that farmers sometimes have to buy back their own production at a loss, leading to further impoverishment of already vulnerable farmers. Those familiar with Rwandan history will recognise worrying patterns from the country’s previous disastrous experimentation with agricultural reform and mono-cropping in the early 1990s. Understanding exactly how these mechanisms work is key to preventing policy failure that are having devastating consequences on Rwandan farmers, and could have wide-ranging implications for national and regional stability. Unfortunately, the total impossibility to acknowledge, let alone discuss, these issues in Rwanda, means that corrective measures cannot be taken.

  2. Many thanks for your valid comments and for carefully examining my do-files!

    As you discuss, our findings are qualitatively similar in the sense that we both agree that poverty has increased in recent years. According to the calculations of the anonymous researcher(s), poverty increased by 5 percentage points, while my estimates show that poverty increased by 1 point. It is unavoidable that the estimates differ because we use different methodologies. You suggest that those differences are mainly driven by two assumption, which I discuss in turn:

    Regional prices: I agree that there is quite some variation in prices between regions and that it is better to use regional prices when converting own food production in its monetary value. The reason why I calculated national prices is that there are relatively few observations (i.e. often less than 200 for the main crops per province). I am not sure if we can accurately estimate regional prices given the limited number of observations.

    Number of months the item was consumed. This is in my view a trickier point and I do not fully agree with your suggestion to use this variable to scale up to yearly consumption values.

    The main reason is that it’s unclear to me how I should do this. Simply multiplying food consumtion and the number of months in which it was consumed does not seem correct. Assume for instance that the respondent did not consume the food item prior to the interview, but reports having consumed the item 11 out of 12 months (i.e. the only month it was not consume was the month of the interview). In this case, I cannot scale up because I have no idea how much the household consumed in every month. Similarly, even if the household consumed the item prior to the interview, I don’t like to assume that the household consumed exactly the same amount in all the other months in which the item was consumed, and that’s precisely what we are doing when using this variable to scale up to yearly consumption.

    I nevertheless tried to follow your suggestion (but still using national prices), but could not replicate your finding. I even found that poverty did not longer increase (with an inflation of 30%), but remained constant. It is, of course, possible that I did not program it correctly or did not correctly interpret your suggestion.

    I did not discuss it in the blogpost, but as a robustness check, I also verified whether the number of food items consumed and the number of food items times the number of months in which it was consumed increased between EICV 3 and EICV 4. The assumption is that dietary diversity should increase if poverty decreases. Using these rough measures of dietary diversity, I observed that it remained constant between the waves.

  3. Thanks for these clarifications. In the interest of clarity, we copy below the exact syntax required for the suggested modifications. In the syntax files of the original (anonymous) blog post, you will find the full set of assumptions and syntax codes used in the original study, which we believe address several of the concerns raised in Sam’s response. Until NISR publishes its own syntax files, however, we cannot know which set of assumptions most closely correspond to those used in the official calculations.
    As Sam rightly points out, the two studies use very different methodologies and are therefore not directly comparable, so we should not expect to find identical results. The fact that we still find similar trends (although at different levels), should, however, give us a great deal of confidence in validity of the results. Furthermore, Sam has added a number of important robustness checks, which further confirm the findings.
    With all this body of publicly verifiable evidence, it is very surprising that neither NISR nor the researchers involved in the official study, have come out to defend their results and to subject them to public scrutiny, as we have done here. It is also surprising that none of the donors who funded the EICV study, nor those who fund Rwanda’s poverty reduction strategy, have come out to request a clarification and verification of these issues. In fact, the World Bank and UNDP still quote the official EICV4 poverty figures on their country websites as proof of the success of their support to the Rwandan government. Their misguided songs of praise echo their own writings on Rwanda from before 1994 (



    *Consumption in monetary value
    gen consumption_item_price=consumption_item*s8cq14
    replace consumption_item_price=consumption_item*price_av if (s8cq14==0 | s8cq14==.)

    *Drop missing variables
    *drop if s8cq2==.

    *Count if items is consumed
    gen consumed=(s8cq2>0)

    *Count number of months
    gen consumed_months=s8cq2 //if (s8cq2>0)

    egen medmth = mean(s8cq2)
    replace s8cq2 = medmth if (s8cq2==0 | s8cq2==.) & (consumption_item>0 & consumption_item0 & s8cq0>11 & s8cq00 & s8cq0>11 & s8cq0<21)
    recode consumed_meat_months .=0

    collapse poverty Consumption province (sum) s8cq2 consumption_item_price consumed consumed_meat_months consumed_months (max) consumed_meat,by(hhid)
    gen consumption_ownFood=consumption_item_price*s8cq2 if province==1
    replace consumption_ownFood=consumption_item_price*2*s8cq2 if province!=1
    drop consumption_item_price

  4. […] Pour Sam Desiere, spécialiste de la question des statistiques dans les pays en voie de de développement, le taux de pauvreté au Rwanda n’aurait pas baissé de six points, mais augmenté d’un peu plus d’un point. Sur la même période, le taux d’inflation ne serait pas de 17%, mais de plus de 30%, ce qui entraînerait une plus grande vulnérabilité, notamment dans les couches les plus pauvres de la population rwanda…. […]

  5. […] Pour Sam Desiere, spécialiste de la question des statistiques dans les pays en voie de de développement, le taux de pauvreté au Rwanda n’aurait pas baissé de six points, mais augmenté d’un peu plus d’un point. Sur la même période, le taux d’inflation ne serait pas de 17%, mais de plus de 30%, ce qui entraînerait une plus grande vulnérabilité, notamment dans les couches les plus pauvres de la population rwanda…. […]


Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.