Divided we call: disparities in access and use of mobile phones in Rwanda.
"Once the toys of rich yuppies, mobile phones have evolved in a few short years to become tools of economic empowerment for the world's poorest people. These phones compensate for inadequate infrastructure ... making markets more efficient and unleashing entrepreneurship."
--The Economist, September 2009
In the popular media, as well as in the development community, many are optimistic about the potential uses of the mobile phone in the developing world. Called a "lifeline for the world's poor" by the BBC, mobile phones are reaching the world's poor at an impressive rate (Anderson, 2007). Already, over two-thirds of the world's mobile phones are in developing countries, and Nokia estimates that, by 2012, more than 90% of sub-Saharan Africa will have mobile coverage (United Nations, 2009).
The potential impact of the mobile phone has not been lost on the research community. A wealth of recent ethnographic research has sought to characterize mobile phone use in the developing world, while a growing parallel body of quantitative work attempts to estimate the impacts of these technologies on local and national economies (Blumenstock, Eagle, & Fafchamps, 2011; Donner, 2008; Jensen, 2007). A separate strain of research seeks to leverage this knowledge by designing mobile-based technologies for deployment in developing countries (Brewer et al., 2005; Parikh, Javid, Ghosh, & Toyama, 2006).
Given this heightened interest in mobile phone use in developing countries, it is surprising how many basic gaps exist in our understanding of how phones are being used on a day-to-day basis by the average person. For instance, it is well known that many phones in East Africa are shared by multiple individuals, but there are few reliable estimates regarding the overall prevalence of phone sharing. For this and other phenomena, even less is known about the subtler dynamics within the population: Do women share phones more than men? Do they call a more diverse network of contacts? Do poor people use their phones differently from rich people?
This article seeks to fill a number of these gaps in our understanding through a detailed quantitative analysis of phone use in Rwanda. The analysis is divided into three sections. First, we compare the overall demographic composition of Rwanda with the demographic composition of a representative sample of mobile phone users, exposing systematic differences between those who own phones and those who do not. Second, we examine new survey data on phone use, paying particular attention to reported behaviors of phone ownership and sharing. Third, we analyze the call histories of our survey respondents, as recorded by the mobile operator, to better understand normal patterns of utilization. Some representative findings include the following insights:
* Section 4: Phone users are disproportionately male, better educated, and older, and they also come from larger households than normal Rwandans. Using an econometric model, we estimate the annual expenditures of phone users to be over twice that of ordinary citizens.
* Section 5: The vast majority of those surveyed report owning the phone they use, and roughly one-third say they share their mobile phone with friends and family. We note statistically significant differences between men and women in patterns of sharing and the types of calls made.
* Section 6: The length of the average call in Rwanda is extremely short, roughly 32 seconds. While men and women spend approximately the same amount of time per day on the phone, there are subtle differences in use by gender. We also observe vastly different patterns of use between the upper- and lower-income quartiles of phone owners.
While the primary focus of this article is to provide a quantitative perspective on mobile phone use in Rwanda, we also contribute to the literature by describing a methodological innovation that may be useful to other researchers interested in studying information and communication technologies (ICTs) in developing countries. This innovation is to combine data collected in structured phone interviews with the call detail records (CDR) that are logged by mobile phone operators. Thus, for a geographically stratified random sample of roughly 900 mobile phone users, we have obtained not only basic demographic and socioeconomic information, but also a detailed history of all phone calls made and received. Our analysis leverages this novel source of data, pointing to many possible extensions for future work.
The remainder of the article is organized as follows: Section 2 discusses related work, and Section 3 describes the principal datasets used in the analysis. Section 4 presents a quantitative comparison of the population of mobile phone users to the greater population of Rwandans. Sections 5 and 6 analyze reported and observed patterns of mobile phone use, first using data collected in phone interviews, and then incorporating the data obtained from the phone company. Section 7 concludes.
2. Related Work
To our knowledge, this is the first article to study phone use through a joint analysis of large-scale household surveys, follow-up phone surveys, and call detail records (CDRs) obtained from the phone company. However, in addition to the review articles mentioned in the introduction, we highlight the results of three separate strands of research that are directly relevant to the analysis that follows.
First, a small group of studies has previously attempted to quantify patterns of phone use in the developing world at a level of detail exceeding the cross-country statistics provided by organizations such as the International Telecommunication Union. In particular, Gillwald (2005) conducted household surveys in 10 African nations, in an effort to measure how individuals and households use different types of ICTs. Using data collected in 2004 and 2005, the author supplies reference statistics that provide a useful context for some of the numbers reported in this article. Separately, Scott, McKemyey, and Batchelor (2004) conducted 1,800 household interviews in Uganda, Botswana, and Ghana, focusing on gender-disaggregated access to ICTs. They found that men and women had remarkably similar patterns of use. By contrast, Huyer, Hafkin, Ertl, and Dryburgh (2005) combine aggregate statistics from various sources to characterize the "gender divide" in access to, and use of, ICTs, finding women at a disadvantage with respect to several metrics of phone access and use. Our findings are generally more consistent with those of Huyer et al. (2005); we are also able to highlight a number of gender-specific differences which, due to a lack of suitable data, were not tested in prior work.
Second, a nascent body of literature has begun to use CDRs to understand underlying dynamics of human behavior. For instance, Gonzalez, Hidalgo, and Barabasi (2008) use CDRs to analyze the trajectories of 100,000 people in a European country to study patterns of human mobility, and Eagle, Pentland, and Lazer (2009) examine the structure of friend networks using data from 100 specially programmed smartphones. There are only a few examples of this type of analysis in the context of the developing world (Blumenstock, 2012; Blumenstock, Gillick, & Eagle, 2010; Frias-Martinez, Frias-Martinez, & Oliver, 2010), but the number of studies is rapidly increasing as data become more readily available.
Finally, there exists a handful of studies that provide excellent descriptions of different patterns of mobile phone use in specific communities throughout the developing world (see Burrell, 2010; Horst & Miller, 2006), with a few focused specifically on Rwanda (Donner, 2007; Futch & McIntosh, 2009). We draw on these insights in interpreting our quantitative results in the following sections. In particular, in the discussion and conclusion, we try to situate the quantitative results of this article within the qualitative findings of researchers who have worked on similar questions.
3. Data and Survey Methodology
The analysis relies on three sources of data: a phone survey of a representative sample of Rwandan mobile phone users, a detailed log of all phone activity by those individuals in the period of January 2005-December 2008, and a household-level demographic survey conducted by the Rwandan government. Further details on each dataset are provided in the following subsections.
3.1 Phone Survey
In the summer of 2009, employing a trained group of enumerators from the Kigali Institute of Science and Technology, we administered a short, structured interview to a geographically stratified group of mobile phone users. The survey instrument contained roughly 80 questions and took between 10 and 20 minutes to administer. We queried basic demographic and socioeconomic information, but we did not collect identifying information, such as the respondent's name, address, or identification numbers. The anonymized phone numbers were obtained from Rwanda's primary mobile phone operator, which had over 90% market share at the time of the survey.
The survey population was intended to be a representative sample of all active phone users. Thus, from the full database of 1.5 million registered phone numbers, we eliminated numbers that had not been used at least once in each of the three most recent months for which we had data (October-December 2008). Then, each one of the remaining 800,000 numbers was assigned to a geographic district based on the location of the phone for the majority of calls made. From each of the 30 districts, 300 numbers were selected randomly, creating a base survey population of 9,000 candidate respondents, where sampling weights for each district were determined based on the distribution of districts in the set of 800,000 active numbers. Given available resources, the team of surveyors was able to call 1,529 unique respondents who had been selected randomly from the pool of 9,000.
3.2 Phone Company Records
For each of the users whom we attempted to contact in the phone survey, we obtained from the phone company an exhaustive log of all phone-based activity that had occurred from the beginning of 2005 through the end of 2008. Thus, for every phone call made or received by one of the survey respondents, we had data on the time and date of the call, as well as the proximate location (based on the cell towers through which the call was routed) of both the caller and the receiver. This allowed us to compute several metrics of phone use and social network structure for each of the 1,529 users whom we attempted to contact. While most of the metrics we used are simple to understand and compute, a few require explanation:
* Activation date: The date on which the phone first appears in the transaction logs.
* Days of activity: The number of days on which the phone was used.
* Net calls: The number of outgoing calls minus the number of incoming calls.
* Degree: The number of unique contacts with whom the person communicated (called or received a call).
* Daily degree: The average number of unique people contacted on any given day that the phone was used.
* Recharge: Monetary value deposited on SIM card.
3.3 Rwanda Demographic and Health Survey
The final dataset we used is a large, representative household survey conducted by the Rwandan government in 2005. In the Demographic and Health Survey (DHS) of 10,272 households, detailed data was collected on demographic composition, asset and durable ownership, and a wealth of other socioeconomic indicators (Institut National de la Statistique du Rwanda [INSR] & ORC Macro, 2006). We used this data to compare the general Rwandan population to the population of phone users contacted in our phone survey.
3.4 Notes on the Data
Of the 1,529 numbers our surveyors attempted to dial, 588 (38%) never picked up the phone. The large number of unanswered calls is striking, but not surprising. As has been noted by other researchers (James & Versteeg, 2007), a large number of people own a SIM card (which costs roughly US$1) without actually owning a mobile phone (which costs closer to US$30). Moreover, SIM cards are commonly lost or stolen, and many people habitually leave their phones off due to the lack of reliable power in the country.
To the extent that these nonresponders are systematically different from responders, the external validity of our results could be limited to the population of individuals likely to answer the phone, rather than the broader population of individuals who have ever used a phone. For instance, if nonresponders tend to be poorer than responders, we might overestimate the wealth of the average phone owner if we base our estimates solely on information provided by respondents.
However, the quantitative evidence at our disposal suggests that these biases are likely to be small. For, although we were only able to collect demographic information for 901 respondents, we still have complete call usage information for the full sample of the 1,529 individuals whom we attempted to contact, and we are therefore able to compare the usage pattern of respondents to that of nonrespondents. We report these results in Table 1, where average values are computed separately for the set of numbers dialed (column 1), for survey respondents (column 2), for nonrespondents (column 3), and by response to this question: "Does anyone else use this phone regularly?" (columns 4 and 5). The final two columns present the p-values obtained by running two-sample t-tests comparing all respondents with all nonrespondents (column 6), and by comparing respondents who share their phone with those who do not share their phone (column 7).
In general, we observe only modest differences between the group of individuals who participated in the phone survey and those who did not. In particular, on the days the phone is used (i.e., activity "per day" in Table 1), behavior of nonrespondents is not statistically different from that of respondents. This is important, as we later assume that the sample of survey respondents is representative of the larger population of mobile phone users in Rwanda. However, the two groups are not identical. Namely, there is a significant difference in the number of days the phone is used. Based on the sum total of evidence presented in Table 1, we believe that the most likely explanation for nonresponse is that those individuals have discontinued use of their SIM card, either due to loss, theft, or replacement. An alternative explanation consistent with the data is that nonrespondents are, on any given day, less likely to use their phone, either because it is off, or because it is unavailable. However, the fact that respondents and nonrespondents act similarly when the phone is on (and in particular, that they make the same number of calls and consume the same amount of airtime) provides some reassurance that the two groups are likely to be comparable along dimensions that, for practical reasons, we are unable to directly observe.
Of those who answered their phones, only 16 (2%) refused to participate in the survey. We believe this very high response rate was due to several factors: First, incoming calls cost nothing to receive, and respondents were paid US$1 in airtime as compensation, a significant amount, given that GDP per capita is roughly US$1,000. Second, most Rwandans are unaccustomed to receiving a call lasting up to 20 minutes (40 times the length of the average phone call), and many seemed flattered to receive the extended attention of university researchers. Finally, respondents were generally more receptive than would be expected in most developed countries, where privacy concerns are rife. (1)
After discarding a handful of surveys that had imperfect data, we were left with a total of 901 valid surveys. The full breakdown of survey responses is given in Figure 1.
Finally, it is also worthwhile to note that aggregate usage on shared phones does not appear to be significantly different from aggregate usage on unshared phones (Table 1, column 7). This is useful, as it allows us to increase our statistical power by including shared phones in most of the later analysis. More generally, however, the result is surprising, as our expectation was that shared phones would show both a higher level of use, and a wider network of contacts. The fact that shared phones appear so similar to unshared phones could be due to a variety of factors: Non-owners might be using their own SIM cards; the owner of the shared phone might be the dominant user; or non-owners may use the phone in exactly the same way as the owner. These and other dynamics of phone sharing are discussed further in Section 5.1.
4. Comparison of Phone Users to At-Large Population
Though mobile phone penetration has risen rapidly in Rwanda over the past decade, only roughly one-third of the population currently owns a mobile phone (International Telecommunication Union, 2012). While it is generally assumed that these phone owners are not representative of the population at large, the nature and extent of these differences is not well understood. Here, we present a quantitative comparison of the representative population of mobile phone owners, as captured in the phone survey, with the representative sample of the at-large population, as recorded in the 2005 household survey. For both samples, reported statistics are weighted by sampling strata.
4.1 Demographic Composition
We begin by analyzing the demographic composition of the two populations. The most striking demographic difference is in gender composition. While 47% of Rwandans are male, males account for 67% of phone owners (see Table 2, panel A). Beyond gender, there are also significant differences in age, household size, and educational attainment. As is evident in Figure 2, the differences between phone users and the at-large population are systematic and occur throughout the demographic distribution. (2)
4.2 Socioeconomic Status
The demographic evidence seems to indicate that phones in Rwanda are owned primarily by the economically privileged. We now test this hypothesis directly. This test is not entirely straightforward, since, in practice, it is difficult to measure the socioeconomic status of a respondent, particularly in a short telephone interview. This difficulty arises because most Rwandans do not earn a fixed wage, and a large percentage of "income" is derived from home-produced goods or other informal channels. Thus, we employ two separate means of measuring socioeconomic status: asset ownership and predicted expenditures.
Asset ownership: In the Demographic and Health Survey (DHS), the Rwandan government collected data on a large number of indicators of wealth, such as housing characteristics and ownership of assets and durables. We obtained the data and questionnaires used in the DHS, and we asked the respondents in our phone survey a subset of these questions verbatim. Panel B of Table 2 reports the average levels of asset ownership among phone survey respondents (column 1) and Rwandan households measured in the DHS (column 2). The differences in asset ownership are stark, with phone users possessing a disproportionately large number of expensive assets. For instance, while only 2.4% of Rwandan households possess a TV, nearly 40% of phone users report TV ownership.
Predicted expenditures: The difference in asset ownership provides compelling evidence that phone users are better off than the general population. However, the underlying differences in wealth and well-being are still murky. For instance, it is hard to say whether a person with a TV and a bicycle is better off than someone with a radio and a refrigerator. Thus, we derive a second measure of socioeconomic status, predicted expenditures, that allows for a more direct comparison of well-being along a single dimension of wealth. While the precise method for computing predicted expenditures is described in a separate paper (Blumenstock, Shen, & Eagle, 2010), the basic idea is as follows: First, actual expenditures are captured in the DHS through an exhaustive series of questions about household consumption. For the DHS sample, we can therefore compute total expenditures by aggregating expenditures across these subcategories in a manner following Deaton and Zaidi (2002). We then fit a model to the DHS data that relates total expenditures to asset ownership. The estimated coefficients of three models are presented in Table 3. (3) We observe a strong relationship between asset ownership and total expenditures; using information on only eight attributes, the best model explains almost 60% of the variation in household expenditures. Finally, since each of these assets was also measured in the phone survey, we can then predict the level of expenditures that would be expected for each of the phone survey respondents, based on the assets already owned by the respondent.
In Table 2, panel C, we report the predicted annual expenditures for both populations, estimated with the above technique. Using the asset-based formula, we find that phone users have over twice the predicted expenditures of ordinary Rwandans. As before, this difference is not idiosyncratic at the mean. As can be seen in Figure 3, the entire expenditure distribution is shifted to the right. (4)
The aggregate socioeconomic differences between the two populations are notable, but they should be taken in the context of the limitations of the data. While Blumenstock, Shen, and Eagle (2010) provide a more complete discussion of these limitations, we briefly note three sources of concern. First, our measure of predicted expenditures is crude and requires many problematic assumptions, particularly about the relationship between assets and expenditures (see Filmer & Pritchett, 2001), and it glosses over distinctions among expenditures, consumption, and permanent income (see Deaton & Muellbauer, 1980). Second, there was a three-year interval between the times when the government data was collected and the phone survey was conducted, during which most Rwandans experienced substantial improvements in socioeconomic status. (5) Third, the data sets for the two populations were collected with different methodologies, and the self-reporting bias in asset ownership could conceivably be exaggerated in the phone survey. Whereas the government data was collected by enumerators at the place of residence and could be verified visually, the data collected over the phone could not be similarly confirmed. Despite these weaknesses, we believe the metric does provide a noisy indicator of socioeconomic status. In future work, we hope to do in-person follow-up interviews with a small subset of respondents to gauge the magnitude of potential biases.
[FIGURE 2 OMITTED]
5. Reported Patterns of Phone Use
The previous section highlights the demographic and socioeconomic differences between average Rwandans and Rwandans with mobile phones. For the remainder of the article, we restrict our attention to the population of mobile phone users, and focus on analyzing reported and observed patterns of mobile phone use. Reported behaviors are based on data gathered through phone interviews; observed patterns are computed from the CDRs obtained from the phone company.
5.1 Ownership and Sharing
While the vast majority of mobile phones in industrialized countries are owned and used by individuals, the situation in developing countries is different Steenson & Donner, 2009). In East Africa, phone sharing is common. In Uganda, for instance, ethnographers have noted intricate social norms of sharing that systematically exclude women and other subpopulations Burrell, 2010). Using the data from the phone survey, we can provide a quantitative perspective on these dynamics.
Extrapolating from the representative survey to the larger population, we estimate that 30% of Rwandans share their phone, where sharing is defined as an affirmative response to the question, "Does anyone else use this phone regularly?" Of those who reported letting others use the phone, 42% reported that someone else had used their phone in the last day, and 78% reported that someone had used the phone in the last week. These and other statistics are presented in Table 4, panel A, column 1). Also worth noting is the fact that nearly 98% of those surveyed reported that they owned the phone they were using. Taken in the context of the statistics on phone sharing, this leads us to believe that, regardless of whether or not other people have access to a phone, it is the owner of the phone who typically answers incoming calls from unknown callers.
Do these numbers match the observations of other researchers in similar contexts? The only other statistics we have seen on phone sharing in Rwanda estimate that between 2% and 70% of people share their phones, but such a range is so large as to permit only minimal comparison Nsengiyumva & Stork, 2005). In other African nations, estimates of phone sharing tend to be higher, typically in the range of 30% to 70% (Gillwald, 2005). However, given the large differences in mobile access and ownership between nations, the numbers are hard to compare. Moreover, the data in Gillwald (2005) was gathered in 2004, when fees were higher and mobile penetration was lower.
[FIGURE 3 OMITTED]
Columns (2), (3), and (4) of Table 4 highlight differences between genders with respect to phone sharing. In our representative sample, female respondents disproportionately reported that the phone was shared. However, this difference is only marginally significant, statistically. Also noteworthy is the fact that men and women report that a comparable number of different people have used their phones in the past 24 hours or 7 days. This is likely due to the fact that both genders report that their spouse is the main other person to use the phone (38% for women, 43% for men). Finally, we observe modest differences in the gender composition of owners (22% female) vs. non-owners (37% female), but due to the small sample size of non-owners (19 of 901 respondents), the difference is not statistically significant. (6) We discuss the implications of this gender divide in Section 7.
More generally, we checked a variety of other socioeconomic and demographic factors to see whether any particular subpopulation was unusually likely to report using a shared phone. However, phone sharing appeared to be evenly distributed across the population. For instance, we observed only modest differences by geographic location. Similarly, a probit regression of phone sharing on our measure of predicted expenditures yielded a statistically insignificant coefficient. Finally, there was no clear relationship between years of schooling and phone sharing, nor between household size and phone sharing. (7)
5.2 Mobile Relationships
Finally, we asked all survey respondents about the people with whom they talk on the phone regularly. Respondents were asked to estimate how many times in the past week they had spoken to contacts in the following three categories: friends, family, and business. If the respondent was unable to provide an estimate, the surveyor asked about the past 24-hour period, and multiplied the response accordingly. Thus, the estimates are noisy because of measurement error and reporting bias, and also because many respondents did not draw clear distinctions among the different types of contacts. For instance, while the "family" category was relatively unambiguous, some respondents found our distinction between "friends" and "business contacts" to be somewhat contrived.
With these caveats in mind, we do note significant differences in the reported behavior of men and women. As can be seen in Table 4, panel B, men report a larger number of total calls, as well as more frequent contact with friends and business contacts. Women, on the other hand, report more frequent contact with family, though this last difference is not statistically significant. These trends are generally consistent with qualitative observations of gender dynamics surrounding mobile phone use in developing countries. (8) However, in other dimensions of phone use, the behavior of men and women appears similar (see Table 4, panel C). Unfortunately, our current analysis is limited by the coarseness of the survey questions. In future work, we hope to further probe gender differences in reported phone usage.
6. Observed Patterns of Phone Use
Until now, we have focused on the reported use of mobile phones, as described by the respondents during phone interviews. As has been noted previously, however, such data are likely to be noisy and biased. Fortunately, we have a more reliable measure of actual use: The call detail records (CDRs) obtained from the mobile operator provide an itemized list of all network activity for each of our respondents. In Table 5, we summarize this usage using the same metrics as in Table 1. In addition, we compute the following:
* In/out-degree: Number of different people to/ from whom calls were made/received
* Clustering: Percentage of first-degree contacts who have contacted each other
* Betweenness: Average shortest path between user and 50 randomly sampled numbers
* Me2U transfers: Interpersonal transfers of airtime made over the network
* Districts: Number of political districts in which the phone was used (Rwanda has 30 districts)
Aggregate statistics on phone use are presented in Table 5, column (1). The average Rwandan completes 190 calls per month, each of which lasts an average of 32 seconds. It is difficult to find recent, comparable figures from other countries, but both numbers are lower than the corresponding figures are likely to be in most industrialized nations. For instance, estimates of use in the United States are closer to 204 calls per month, lasting roughly three minutes each; in India, the industry average is 377 minutes of use per month (ZDNet Research, 2005). These differences are most likely due to the per-second fee structure and the high cost of a phone relative to daily income. To provide some context, a three-minute call in Rwanda costs roughly US$0.60, which amounts to 0.06% of the average GDP per capita (GDPpc). The corresponding igure in the United States is US$0.60 for a three-minute call (0.001% of GDPpc); in India, a three-minute call costs only US$0.04 (roughly 0.003% of GDPpc). (9)
6.1 Differences by Gender
Within the sample of phone users, there are large differences in phone use across demographic groups. In column (6) of Table 5, we highlight the differences between men and women. To summarize the results: Between genders, there are significant differences in the length of calls made (women talk longer), in the direction of the calls (women receive more calls than they make; men are the opposite), in international calling (both men and women receive more than they make, but women receive even more than men), and in airtime gifts using the Me2U service (women receive more airtime). More broadly, men and women have comparably sized networks of contacts, but the networks of men tend to be more tightly clustered than those of women. Finally, we note that, contrary to the large and significant differences in total calls reported by male and female respondents (discussed in the previous section), the actual difference is small and statistically insignificant.
Given the impersonal nature of our metrics, it is not simple to interpret these statistics. Evidence from the United States and Norway suggests that gender differences in phone use are not unique to developing countries (Cotten, Anderson, & Tufekci, 2009; Ling, 2001). Whether the differences seen in Rwanda reflect benign cultural differences or more insidious dynamics of power and patriarchy is a deeper question that we touch on in the conclusion.
6.2 Differences by Socioeconomic Status
While the differences by gender are somewhat ambiguous, the differences between socioeconomic groups are striking. To analyze phone use by socioeconomic strata, we ranked each of the respondents by predicted expenditures--a measure based on known asset ownership, as discussed in section 4.2--and then we separately computed averages for the upper and lower quartiles. These statistics are presented in columns (4) and (5) of Table 5; the test for a difference between the two populations appears in column (7).
Above and beyond the differences between phone owners and non-owners (Figure 3), we note large and consistent differences in usage within the population of owners, and in particular, between the richest 25% and the poorest 25% of phone users. Across nearly every measure, the richer people use their phones more: in number of calls, length of calls, number of days on which the phone is used, size and structure of the social network, etc. (10) While some of these differences are not statistically significant, the overall relationship between use and socioeconomic status remains strong. (11)
7. Discussion and Conclusion
The preceding analysis provides a quantitative perspective on the demographic and socioeconomic structure of mobile phone use in Rwanda. Though the analytic results are diverse, a relatively consistent picture begins to emerge: Mobile phone use in Rwanda is far from uniform. There are significant and systematic differences not only in who owns the phone (see Section 4), but also in how different types of owners use the phone (see Sections 5 and 6). Specifically, phone owners are much more likely to be male, they are better educated, they come from larger households, and they are substantially wealthier than those without mobile phones. Within the population of phone owners, there are differences in usage between men and women, particularly in reported phone sharing and the types of calls that are made. Most notable, however, is the vast difference in use between poorer and richer phone owners, such that the highest income quartile uses their phones 30%-100% more than lowest income quartile, depending on the measure of use.
Taken together, the evidence in this article indicates that it is the privileged, male members of Rwandan society who disproportionately own and use mobile phones. Unfortunately, this pattern does not seem to be unique to Rwanda; similar patterns have been observed in East Africa (Burrell, 2010) and other countries around the world (Huyer et al., 2005). Moreover, the same trends can be seen with other technologies in other contexts. For instance, Toyama et al. (2005) and Kiri and Menon (2006) observe that use of telecenters is dominated by younger, more educated men.
We believe the preceding analysis to be useful for a few distinct reasons. First, we believe there is intrinsic value in developing insight into the daily patterns of use of such a massively popular technology, in part to help scholars and practitioners better understand how phone-based technologies are likely to be received and used. As we have seen, traditional Western models of phone use--and the potential design assumptions they impose--do not necessarily apply to the Rwandan context. Second, we hope our methods and analysis can inspire and be improved on by other researchers. In particular, the method of coupling anonymous call detail records with structured phone interviews should provide fertile ground for future work. Finally, by providing more reliable estimates of the distribution of phone access and use, we seek to inform policy makers about the potential distributional impacts of phone use in countries such as Rwanda. Given the considerable attention and investment devoted to mobile telephony in developing countries, it is important to better understand who is--and who isn't--reaping the benefits of the new technology.
The authors would like to thank Jenna Burrell, Tapan Parikh, Kentaro Toyama, and the three anonymous referees for their thoughtful comments and suggestions. We are grateful to Santhi Kumaran and the faculty and students at the Kigali Institute of Science and Techhnology for helping with the phone survey, and to Nadia Musaninkindi and Jean Marie Nyabyenda for their assistance with the DHS data. Ye (Jack) Shen and Paige Dunn-Rankin provided excellent research assistance.
Anderson, T. (2007, February 19). Mobile phone lifeline for world's poor. BBC News. Retrieved from http://news.bbc.co.uk/
Average American spends 619 minutes a month on the cell phone. (2005, January 25). ZDNet Research. Retrieved from http://www.zdnet.com/
Blumenstock, J. (2012). Inferring patterns of internal migration from mobile phone call records: Evidence from Rwanda. Information Technology for Development. 18(2), 107-125.
Blumenstock, J. E., Eagle, N., & Fafchamps, M. (2011). Charity and reciprocity in mobile phone-based giving: Evidence from Rwanda. Centre for the Study of African Economies Working Paper #WPS/2011-19. Oxford, UK: University of Oxford.
Blumenstock, J. E., Gillick, D., & Eagle, N. (2010, March). Who's calling? Demographics of mobile phone use in Rwanda. AAAI Symposium for Artificial Intelligence and Development (pp. 116-117). Palo Alto, CA: Stanford University.
Blumenstock, J., Shen, Y., & Eagle, N. (2010, December). A method for estimating the relationship between phone use and wealth. QualMeetsQuant Workshop at the 4th International IEEE/ACM Conference on Information and Communication Technologies and Development. London.
Brewer, E., Demmer, M., Du, B., Ho, M., Kam, M., Nedevschi, S., et al. (2005). The case for technology in developing regions. Computer, 38(6), 25-38.
Burrell, J. (2010). Evaluating shared access: Social equality and the circulation of mobile phones in rural Uganda. Journal of Computer-Mediated Communication, 15(2), 230-250.
Cotten, S. R., Anderson, W. A., & Tufekci, Z. (2009). Old wine in a new technology, or a different type of digital divide? New Media Society, 11 (7), 1163-1186.
Deaton, A., & Muellbauer, J. (1980). Economics and consumer behavior. Cambridge, UK: Cambridge University Press.
Deaton, A., & Zaidi, S. (2002). Guidelines for constructing consumption aggregates for welfare analysis. Washington, DC: World Bank Publications.
Donner, J. (2007). The use of mobile phones by microentrepreneurs in Kigali, Rwanda: Changes to social and business networks. Information Technologies & International Development, 3(2), 3-19.
Donner, J. (2008). Research approaches to mobile use in the developing world: A review of the literature. The Information Society, 24(3), 140-159.
Eagle, N., Pentland, A., & Lazer, D. (2009). Inferring friendship network structure by using mobile phone data. Proceedings of the National Academy of Sciences, 106(36), 15274-15278.
Filmer, D., & Pritchett, L. H. (2001). Estimating wealth effects without expenditure data--or tears: An application to educational enrollments in states of India. Demography, 38(1), 115-132.
Frias-Martinez, V., Frias-Martinez, E., & Oliver, N. (2010). A gender-centric analysis of calling behavior in a developing economy. AAAI Symposium for Artificial Intelligence and Development, in process.
Futch, M., & McIntosh, C. (2009). Tracking the introduction of the village phone product in Rwanda. Information Technologies & International Development, 5(3), 54-81.
Gillwald, A. (2005). Towards an African e-index: Household and individual ICT access and usage across 10 African countries. Wits, South Africa: LINK Centre, Wits University, School of Public and Development Management.
Gonzalez, M. C., Hidalgo, C. A., & Barabasi, A.-L. (2008). Understanding individual human mobility patterns. Nature, 453(7196), 779-782.
Horst, H., & Miller, D. (2006). The cell phone: An anthropology of communication. London: Berg Publishers, Incorporated.
Huyer, S., Hafkin, N., Ertl, H., & Dryburgh, H. (2005). Women in the information society. In G. Sciadas (Ed.), From the digital divide to digital opportunities: Measuring infostates for development (pp. 135-196). Montreal: Orbicom.
Institut National de la Statistique du Rwanda (INSR) & ORC Macro. (2006). Rwanda Demographic and Health Survey2005. Calverton, MD: INSR and ORC Macro.
International Telecommunication Union. (2012). Mobile-cellular telephone subscriptions [data file]. Retrieved from http://www.itu.int/ITU-D/icteye/
James, J., & Versteeg, M. (2007). Mobile phones in Africa: How much do we really know? Social Indicators Research, 84(1), 117-126.
Jensen, R. (2007). The digital provide: Information (technology), market performance, and welfare in the South Indian fisheries sector. The Quarterly Journal of Economics, 122(3), 879-924.
Kiri, K., & Menon, D. (2006). For-profit rural kiosks in India: Achievements and challenges. Information for Development, 4, 14-28.
Ling, R. (2001). We release them little by little: Maturation and gender identity as seen in the use of mobile telephony. Personal Ubiquitous Computing, 5(2), 123-136.
Nsengiyumva, A., & Stork, C. (2005). Rwanda. Towards an African e-index: Household and individual ICT access and usage across 10 African countries (pp. 120-129). Wits, South Africa: LINK Centre, Wits University, School of Public and Development Management.
Parikh, T. S., Javid, P. K. S., Ghosh, K., & Toyama, K. (2006, April). Mobile phones and paper documents: Evaluating a new approach for capturing microfinance data in rural India. Proceedings of the SIGCHI conference on Human Factors in Computing Systems (pp. 551-560). Montreal: ACM.
Scott, N., McKemyey, K., & Batchelor, S. J. (2004). The use of telephones amongst the poor in Africa: Some gender implications. Gender Technology and Development, 8(2), 185-207.
Steenson, M., & Donner, J. (2009). Beyond the personal and private: Modes of mobile phone sharing in urban India. In S. W. Campbell & R. Ling (Eds.), The reconstruction of space and time: Mobile communication practices (pp. 231-250). Piscataway, NJ: Transaction Publishers.
The power of mobile money. (2009, September 24). The Economist. Retrieved from http://www.economist.com/
Toyama, K., Kiri, K., Menon, D., Pal, J., Sethi, S., & Srinivasan, J. (2005, March). PC kiosk trends in rural India. Proceedings of Policy Options and Models for Bridging Digital Divides. Tampere, Finland: University of Tampere.
United Nations. (2009). Compendium of innovative e-government practices. New York: United Nations Publication.
World Bank. (2010). Rwanda GDP per capita [data file]. Retrieved from http://www.tradingeconomics.com/
Joshua Evan Blumenstock
University of Washington
School of Information
Mary Gates Hall, Suite 370
Seattle, WA 98195
Massachusetts Institute of Technology
Cambridge, MA 02139
(1.) We had multiple respondents call us back at the call center at their own expense to thank us for taking an interest in their affairs, and to request that we call their friends and family as well. Such requests were politely declined.
(2.) In Figure 2, we exclude persons under 15 years of age to highlight the fact that the difference in mean is not caused solely by the fact that children do not own phones.
(3.) Predicted expenditures are estimated with a flexible polynomial regression of the logarithm of total expenditures on a variety of assets and durables. Column (1) reports estimates from ordinary least squares; column (2) adds district- level fixed effects, and column (3) additionally controls for livestock possession.
(4.) With this measure of predicted expenditures, it is possible to further characterize economic stratification within the population of mobile phone users. For instance, we estimate that, while 77.8% of phone owners live on less than US$2 per day and 51.8% live on less than US$1 per day, only 6.9% live below the national poverty line of US$0.43 per day. In the at-large population, we compute the corresponding rates to be much higher, at 94% (less than US$2/day), 82.5% (less than US$1/day), and 48.2% (below the poverty line).
(5.) The World Bank estimates an increase in per-capita GDP of roughly 26%, from $255 in 2005 to $321 in 2009 (World Bank, 2010)
(6.) It is possible that the observed differences in ownership are driven by a disinclination among women to answer a call from an unknown caller, but we have no evidence to support this conclusion.
(7.) There was, however, a statistically significant correlation between the number of adults in the household and the likelihood of the phone being shared, presumably due to the increased demand for the phone by individuals proximate to the phone owner.
(8.) In Ghana, for instance, Scott, McKemyey, and Batchelor observed,
[M]en are more likely to use the phone to communicate with friends, to make business and work-related calls, and to make calls relating to religious affairs, although this is still only a relatively minor use. On the other hand, a greater proportion of women make family calls. (2004, p. 200)
(9.) Numbers based on author's calculations from http://www.mtn.co.rw/, http://www.airtel.in/, and http://www. boostmobile.com.
(10.) The only exception is in the case of Me2U--the system for interpersonal airtime transfers--of which the poorer phone users appear to be significantly more active in the number of transfers. We believe this is because Me2U serves as a substitute for traditional financial services, and we intend to investigate this phenomenon further in future work.
(11.) To provide some indication of the joint effect of all of these factors, we note that a regression of predicted expenditures on the variables listed in Table 5 yields an [R.sup.2] of 0.15; a more flexible specification has a corresponding [R.sup.2] of 0.31.
Table 1. Summary Statistics: Survey Respondents, Nonrespondents, and Shared Phones. (1) (2) (3) Dialed Respondents No Answer Activation date 2/9/08 1/12/08 4/5/08 Days of activity 672.2 770.3 540.3 Avg. call length 32.3 31.7 33 Calls per day 6.24 6.25 6.23 Net calls per day (out-in) 0.4 0.087 0.82 Degree 797.8 734.0 883.6 Daily degree 3.81 3.78 3.86 Int'l calls per day 0.09 0.084 0.099 Credit used per day 184.6 163.5 212.9 Max. recharge value 3,391.6 2,756.3 4,246.4 Calls per day (out) 3.32 3.17 3.52 Calls per day (in) 2.92 3.08 2.71 N 1,529 901 628 (4) (5) (6) (7) Shared Unshared RvN SvNS Activation date 1/2/08 1/12/08 -- -- Days of activity 702.3 799 0.0002 0.31 Avg. call length 31.5 31.8 0.49 0.9 Calls per day 6.32 6.22 0.98 0.94 Net calls per day (out-in) 0.54 -0.1 0.19 0.46 Degree 882.9 671.3 0.67 0.55 Daily degree 3.98 3.7 0.91 0.72 Int'l calls per day 0.083 0.084 0.53 0.97 Credit used per day 151 168.8 0.3 0.62 Max. recharge value 2,609.8 2,818.300 0.28 0.62 Calls per day (out) 3.43 3.06 0.63 0.69 Calls per day (in) 2.89 3.16 0.28 0.58 N 239 661 -- -- Notes: Mean values, weighted by sampling strata, are reported for all statistics except activation date, where the median is reported. Columns (6) and (7) report p-values from adjusted Wald test for difference in means between columns (2) and (3), and (4) and (5), respectively. Table 2. Phone Users vs. General Populace. (1) (2) (3) Phone Users All Rwandans T-stat Panel A: Demographic indicators Age 32.03 21.37 32.03 Household size 5.87 4.98 11.56 Percent male 66.6% 47.4% 15.76 Completed sec. school 35.71% 1.60% Panel B: Socioeconomic status Owns a car 19.1% 0.1% 6.35 Owns a bicycle 38.6% 12.9% 19.51 Owns a refrigerator 16.7% 1.2% 4.33 Owns a landline 2.8% 6.2% -17.33 Owns a radio 94.3% 52.9% 82.78 Owns a TV 39.4% 2.4% 12.53 Panel C: Expenditures Predicted expenditures $1,725 $753 24.05 Notes: Mean values reported, weighted by sampling strata. Column (3) reports t-statistics testing for a difference in means between columns (1) and (2). All differences are significant with at least 99.99% confidence. Predicted expenditures are computed using a conversion rate of RWF550 = US$1. Table 3. Regression of Expenditures on Assets. (1) (2) (3) Assets + District FE + Livestock HH size 0.115 0.123 0.110 (31.20) (35.24) (26.94) Car/truck 0.650 0.661 0.545 (8.12) (8.76) (4.81) Bicycle 0.329 0.350 0.327 (12.64) (13.65) (12.06) Refrigerator 0.404 0.293 0.351 (5.70) (4.40) (3.61) Landline 1.06 0.800 0.779 (28.97) (22.41) (15.66) Goats 0.024 (6.42) Pigs 0.027 (2.66) Rabbits 0.005 (0.99) District FE No Yes Yes [R.sup.2] 0.520 0.577 0.487 N 6,900 6,900 4,739 Notes: Outcome is log of total household expenditures. T- statistics reported in parentheses. Regressions also include motorcycle, TV, radio, cattle, sheep, and chickens. Table 4. Reported Phone Use. (1) (2) All Men Panel A: Phone ownership and sharing Do you own this phone? 97.9% 97.4% Do you own another SIM card? 34.7% 35.4% Does anyone else use this phone regularly? 29.9% 25.2% ... How many different people used it in the last 0.7 0.7 24 hours? ... How many different people in the last 7 days? 2.2 2.4 Panel B: Regular contacts Roughly how many times per week do you talk to ... ... Friends (boy/girlfriend included) 20.9 25.4 ... Family (spouse included) 11.0 10.0 ... Business contacts 23.5 29.6 Total calls per day (computed from above) 8.1 9.4 Panel C: Types of calls made Have you ever used your phone to ... ... Seek help in an emergency? 26.8% 28.2% ... Find a doctor? 31.1% 29.3% ... Find a job? 45.2% 49.3% ... Get advice on farming? 25.0% 27.2% N 901 645 (3) (4) Women p-value Panel A: Phone ownership and sharing Do you own this phone? 98.9% 0.411 Do you own another SIM card? 33.3% 0.806 Does anyone else use this phone regularly? 38.6% 0.105 ... How many different people used it in the last 0.7 0.925 24 hours? ... How many different people in the last 7 days? 1.9 0.362 Panel B: Regular contacts Roughly how many times per week do you talk to ... ... Friends (boy/girlfriend included) 11.8 0.002 ... Family (spouse included) 13.1 0.323 ... Business contacts 11.4 0.027 Total calls per day (computed from above) 5.3 0.014 Panel C: Types of calls made Have you ever used your phone to ... ... Seek help in an emergency? 24.1% 0.578 ... Find a doctor? 34.6% 0.524 ... Find a job? 36.8% 0.147 ... Get advice on farming? 20.7% 0.308 N 256 -- Notes: Percentages correspond to the proportion of affirmative responses (Panels A and C) or mean values (Panel B). All values weighted by sampling strata to produce averages representative of entire phone population. Sharing within last 24 hours and 7 days is conditional on the phone being shared at all. Table 5. Actual Phone Use Computed from Transaction Logs. (1) (2) (3) All Men Women Panel A: Domestic and international calls Activation date 1/12/08 1/29/08 12/26/07 Days of activity 770.3 743.4 823.8 Avg. call length 31.7 29.7 35.7 Calls per day 6.25 6.32 6.09 Net calls per day (out-in) 0.087 0.31 -0.37 Int'l calls per day 0.084 0.071 0.11 Net int'l calls (out-in) -0.014 -0.0018 -0.038 Panel B: Social network structure Degree 734 772.6 657.2 In-degree 488.2 488.5 487.6 Out-degree 433 475.9 347.7 Daily degree 3.78 3.87 3.61 Net daily degree (out-in) 0.00027 -0.17 0.34 Clustering 0.063 0.065 0.058 Betweenness 2.72 2.74 2.69 Panel C: Other behaviors Credit used per day 163.5 176.2 138.2 Max. recharge value 2,756.3 2,775.1 2,718.9 Avg. districts per day 1.36 1.37 1.34 Avg. districts contacted 1.21 1.2 1.22 Me2U transfers per day 0.044 0.041 0.05 Net Me2U transfers per day 0.00038 0.0066 -0.012 N 901 645 256 (4) (5) (6) (7) "Rich" "Poor" MvW RvP Panel A: Domestic and international calls Activation date 07/08/06 02/05/08 -- -- Days of activity 994.6 548.1 0.38 0.0001 Avg. call length 39.8 28.4 0.014 0.0001 Calls per day 8.42 6.47 0.82 0.26 Net calls per day (out-in) 0.76 -0.31 0.02 0.29 Int'l calls per day 0.13 0.066 0.11 0.065 Net int'l calls (out-in) -0.031 -0.028 0.031 0.89 Panel B: Social network structure Degree 1,240.7 498.8 0.56 0.037 In-degree 721.5 369.1 0.99 0.02 Out-degree 798.1 280.8 0.43 0.1 Daily degree 5.08 3.77 0.63 0.17 Net daily degree (out-in) -0.47 0.41 0.15 0.19 Clustering 0.056 0.057 0.067 0.88 Betweenness 2.61 2.77 0.27 0.0033 Panel C: Other behaviors Credit used per day 246.9 138.9 0.17 0.025 Max. recharge value 3,816.1 2,228.5 0.89 0.013 Avg. districts per day 1.51 1.47 0.8 0.81 Avg. districts contacted 1.4 1.28 0.81 0.48 Me2U transfers per day 0.037 0.083 0.43 0.012 Net Me2U transfers per day 0.0082 -0.012 0.011 0.14 N 180 180 -- -- Notes: Mean values reported, weighted by sampling strata to produce averages representative of entire phone population. "Rich" and "poor" are defined as those respondents in the top and bottom 20% of the predicted expenditure distribution, respectively. Columns (6) and (7) report p-values from adjusted Wald test for difference in means between columns (2) and (3), and (4) and (5), respectively. Figure 1. Survey Population. Busy 2% midway 1% Completed survey 60% Never answered 36% Rejected 1% Note: Table made from pie chart.
|Printer friendly Cite/link Email Feedback|
|Title Annotation:||Research Article|
|Author:||Blumenstock, Joshua Evan; Eagle, Nathan|
|Publication:||Information Technologies & International Development|
|Date:||Jun 22, 2012|
|Previous Article:||From the guest editor.|
|Next Article:||Understanding the links between ICT skills training and employability: an analytical framework.|