Research Policy 54 (2025) 105173

Contents lists available at ScienceDirect

Research Policy
journal homepage: www.elsevier.com/locate/respol

Digital technology and inclusive development during global crisis: Evidence

from a randomised experiment in Bangladesh
Xiaolan Fu a,* , Pervez Ghauri b, Jianan Lu c
a
Technology and Management Centre for Development (TMCD), Oxford Department of International Development (ODID), University of Oxford, UK
b
Department of Strategy and International Business, Birmingham University, University of Birmingham, UK
c
School of Accounting, Economics and Finance, University of Portsmouth, UK

A R T I C L E I N F O A B S T R A C T

Keywords: We investigate the role of digital technology in inclusive development by conducting a randomised control trial
Digital technology (RCT) in Bangladesh. We analyse the mechanisms through which access to content-based digital platforms
Digital skills training empower marginalised communities’ entrepreneurship and employment and hence contributes to their income
Randomised control trial
and resilience. Based on a sample of 2974 respondents, we found that the treatment group, who received our
Marginalised communities
Inclusive development
content-based digital platform usage training, manifested higher income resilience and better employment status
during the COVID-19 pandemic in 2020–2021. The results are underpinned by sub-group studies, real-usage
studies, mechanism validity checks and qualitative analyses. This study contributes to the literature by ana-
lysing the impact of digital entrepreneurial support programmes in developing economies and the transmission
mechanisms through which they include marginalised communities in economic activities. It also provides the
first RCT-based empirical evidence on the developmental impact of content-related digital platform technology
that requires low capital investment.

1. Introduction are seminal studies investigating the impact of ‘first generation’ mobile
technology-based inclusive innovation, such as mobile payments and
Digital technology-enabled innovations and their impact on inclu- their usage (Lashitew et al., 2019; Pal et al., 2021). In recent years, there
sive development have become central to contemporary debates are also some pioneering studies evaluating the developmental impact of
(Nambisan et al., 2019; Fu et al., 2022) and form a critical area in need digital platforms. Kelley et al. (2024) evaluated whether digital plat-
of policy and regulatory attention (Pansera and Owen, 2018; Eiteneyer forms improve employment outcomes among vocational training grad-
et al., 2019; Lashitew et al., 2019; Ghauri et al., 2021; Mansell, 2021). uates in India using a randomised control trial. The results were however
Research should answer questions about why and how new technologies not conclusive, as they found evidence of voluntary unemployment
come into being, with the aim of investigating their broader conse- among the graduates. Studies on the impact of emerging content-based
quences for society (Freeman, 1994, 2007; Chataway et al., 2014). The digital platform based inclusive innovation and related training and user
implications for marginalised communities during global crises demand experience are limited with a few exceptions (Helfat and Raubitschek,
particular attention (Dosi and Soete, 2022). Therefore, it is important 2018; Fu et al., 2022). Digital skills/digital literacy in employment and
not only to adjust to the digital realities (Mansell, 2021), it is also entrepreneurship has garnered increasing attention (Pallais and Sands,
important to understand the societal outcomes of novel digital in- 2016; Chan and Wang, 2018; Adermon and Hensvik, 2022), and dis-
novations and how these influence ecosystems and communities cussions on who benefits from entrepreneurship training have been
(Nambisan et al., 2019). raised (Lyons and Zhang, 2018; Anderson and McKenzie, 2022; Kotha
An important area awaiting further research is understanding the et al., 2023). However, our understanding of the impact of digital
various societal impacts of these digital technology-based inclusive in- platforms and related digital skills training on the overall capabilities
novations and related training and user experiences, especially the and human development in marginalised communities is still limited.
transmission mechanisms, and the relevant policy implications. There The new developments in digital technologies such as new content-

* Corresponding author at: Technology and Management Centre for Development, Department of International Development, University of Oxford, 3 Mansfield
Road, Oxford OX1 3TB, UK.
E-mail address: [email protected] (X. Fu).

https://doi.org/10.1016/j.respol.2025.105173
Received 17 December 2022; Received in revised form 16 October 2024; Accepted 4 January 2025
Available online 22 January 2025
0048-7333/© 2025 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
X. Fu et al. Research Policy 54 (2025) 105173

based digital platform-enabled inclusive innovation enable market COVID-19 crisis, resulting from better employment and entrepreneur-
participation and organisational change (the rise of micro digital en- ship capabilities. Second, this study examines the mechanisms through
trepreneurs) and may also lead to social and economic empowerment in which access to digital technology brings about treatment effects. We
the marginalised society. This presents a novel and interesting area for find that our fieldwork intervention helps marginalised communities to
research. There is still a dearth of literature empirically analysing the reduce subjective barriers of access to digital technology and reduce
mechanisms through which this new digital inclusive innovation and their information poverty. Meanwhile, the treatment group manifests
related skills training enable and empower marginalised communities. higher aspirations, social capitals, better learning and capability devel-
Although there are some qualitative discussions (Şimşek et al., 2022; Fu opment, more job opportunities, and better utilisations of digital tech-
et al., 2022; Ghauri et al., 2022), there remains a lack of empirical ev- nology during economic hardships.
idence from a large-scale dataset. There is in particular a lack of ex-ante We employed a battery of approaches to scrutinize the robustness of
and ex-post evidence on this topic, based on longitudinal data to provide our results and gain further insights. First, our baseline result is sub-
reliable arguments for effective policy interventions (Pal et al., 2021), stantiated by the Unobservable Selection Bias test, indicating that our
during global crises such as the COVID-19 pandemic (Dosi and Soete, regression specifications are unlikely to be affected by unobservable
2022; Canale et al., 2022). Our understanding of the impact of the factors. Second, sub-group studies with interaction terms were con-
related digital skills training is also limited. ducted, revealing that the subgroup differences in response to the
In this study, we take the view that poverty is not the lack of income treatment are not statistically significant. Our treatment effects are
or access to commodities, it equates with social exclusion, chronic in- evident in improving the lower-income residents. Third, our interven-
equalities and limited freedom to choose (Sen, 2000; De Haan, 2001; tion not only enhances the quantity but also the quality of residents’
Walton, 2011). We argue that along with financial constraints’ reduc- usage of digital platforms, contributing to their income resilience during
tion and technological de-skilling, content-based digital platforms pro- the global crisis. Fourth, we examined the validity of the mechanisms to
vide inclusiveness for marginalised people who are excluded from observe how they contribute to the resilience elements. Finally, as a
traditional economic and e-commerce activities. We examine the extent significant supplementary component, we conducted qualitative
of the effect and the transmission mechanisms through a fieldwork research by interviewing a group of Haate Haat App stakeholders,
experiment of a Randomised Control Trial (RCT) in Bangladesh which including users, fieldwork operators, and the App manager. Consistent
spanned the period of the COVID-19 crisis. with our RCT results, the qualitative study further reinforces the
We designed and rolled out a content-based digital platform (A mechanisms suggested in our theoretical discussions.
mobile App named ‘হােত হাত’ (Haate Haat) - in Bengali, which means This research contributes to the literature in several ways. Firstly,
‘hand-to-hand’) in Bangladesh. We developed this App based on another our study aligns with a growing body of research analysing entrepre-
successful App in China, Kuaishou (Kwai), which does not use algo- neurial support or digital training programmes in developing economies
rithms to promote more ‘engaging’ content or content from celebrities so (Armanios et al., 2017; González-Uribe and Leatherbee, 2018). While
that ‘normal people can be seen’ (Fu et al., 2022). The RCT intervention the literature has examined how to enhance entrepreneurship in
was applied during the App’s promotional campaign when local mar- developing countries and the effects of training programmes, the results
ginalised residents were randomly selected to receive our digital tech- have yielded mixed implications (De Mel et al., 2014; Dutt et al., 2016;
nology utilisation training programme. In the empirical analysis, we González-Uribe and Reyes, 2021; McKenzie, 2021). There remains an
compare the treatment group with the non-treatment group (control underdeveloped literature on human development using the capability
group) to examine the mechanisms’ validity and evaluate the in- approach proposed by Sen (1995, 2000), who asserts that poverty
tervention’s true impact on the marginalised community. As our field- reduction, income, resources and wealth are means to achieve human
work and intervention were conducted during the most severe period of development and not an end in themselves. Human development lies in
the COVID-19 crisis, this provides a unique and timely setting for us to the freedom to choose, other means such as income and resources, help
measure the resilience of the marginalised communities who were in removing the hindrances to freedom (Sen, 2000) and to achieve in-
empowered by digital technology in fighting the economic hardship of clusive development. We shed light on this important discussion by
the COVID-19 pandemic (Vargo et al., 2021; Canale et al., 2022). providing experimental evidence and qualitative support to argue the
The RCT method offers the advantages that it prevents the deliberate mechanisms for this progress.
manipulation of results, provides immediate comparative results, and Secondly, while the impact of digital platforms on employment and
reflects the true impact of the treatment (Banerjee et al., 2007; Duflo user income has been explored in the literature (Helfat and Raubitschek,
et al., 2011; Banerjee et al., 2015; World Bank, 2022). As suggested by 2018; Fu et al., 2021; Lee et al., 2022; Şimşek et al., 2022), much of the
Banerjee and Duflo (2011), radical thinking is needed to alleviate evidence is derived from case studies or statistical analyses of secondary
poverty; the RCT approach can assist this in policy design and imple- data. There remains a dearth of empirical evidence, such as the RCT
mentations by providing causative rather than summative evaluations. approach, to thoroughly analyse mechanisms and offer actionable policy
We also conducted a qualitative analysis to supplement the empirical implications (Pal et al., 2021; Bohnsack et al., 2021). Concurrently, the
evidence from RCT and strengthen our discussions with more detailed global economic crisis triggered by COVID-19 has exacerbated hardships
and vivid evidence (Bakhshi et al., 2015; Lashitew et al., 2019). among marginalised populations worldwide, underscoring once more
Therefore, we distinguish this study from classical RCT literature by the importance of inclusive development and the role of digital tech-
examining “how it works” beyond “what works”, on which traditional nology (Pansera and Owen, 2018; Dosi and Soete, 2022; Chakravorty
RCT studies largely focused (Reddy, 2012; Faulkner, 2014; Kelley et al., et al., 2023). Our study contributes empirically by applying the RCT
2024). We combine these methods to provide a rigorous and in-depth approach to investigate digital technology and inclusive development
study of the impact of content-based digital platforms and related (Duflo et al., 2011; Banerjee et al., 2015), utilising the unique context of
training on social and economic development in a marginalised society COVID-19 to examine the resilience and empowerment generated by
during a global pandemic. emerging technology for marginalised communities. Thirdly, this
The empirical results of our study provided several insights. First, we research examines the effect of a digital platform on inclusive devel-
find that the treatment group, the marginalised respondents who opment. The experimental evidence not only allows us to compare pre-
received our short video platform function and used the training, man- and post-intervention changes but also enables us to contrast these
ifested better status regarding inclusive development and resilience changes with those in a similar control group of participants, identifying
during the COVID-19 economic hardship. For example, when compared the true effect of digital technology on jobs and income in marginalised
with the control group, the treatment group showed a 9.6 % smaller communities. Results from empirical studies by Kelley et al. (2024) on
decrease in household income (greater income resilience) during the the impact of digital technology on employment are inconclusive. Kelley

2
X. Fu et al. Research Policy 54 (2025) 105173

et al. (2024)’s study is on digital platforms and their effect on employ- economically and socially marginalised communities from enrolling into
ment and job search in India. The conclusions of the study are mixed as value-creating activities and obtaining social wellbeing. Second, the
the good job offers fail to materialise on the platform, mainly because of traditional “bricks-and-mortar” retail sector and classical e-business
the doubts of the youth about the effectiveness of the matching facili- models require initial capital investment. This creates a financial entry
tated by digital platforms. barrier for BOPs (people at the “base of the pyramid” (Prahalad and
Moreover, our study engages with the literature on the “capability Hammond, 2002) to enrol in business activities and create value.
approach”, as articulated by Amartya Sen (Sen, 1990), and the debate on
technology and human development. Sen argues that development en- 2.2. Content-based digital platform for inclusive development
tails capability enhancement and the freedom to access opportunities,
ensuring that the voices of marginalised communities are included. Recently, content-based digital platforms have emerged as digital
Capability thus refers to a person’s freedom to choose one type of life technology that can be realised through mobile applications (Apps), that
from another (Sen, 1995). Human development also requires that people are reframing the business involvement among marginalised commu-
have the goods and services needed for living, as well as be empowered nities and re-evaluating the discussion of inclusive development (i.e., Fu
to enjoy self-esteem, the feeling of worthiness, respect, dignity and and Akter, 2016; Helfat and Raubitschek, 2018; Fu et al., 2022), and the
integrity. Our study investigates how this new content platform-based innovative channels it creates for BOPs (Prahalad and Hammond, 2002).
approach can help disadvantaged individuals who may be excluded All these mechanisms facilitate entrepreneurship and job creation in
from economic activities to be able to gain income, and also build ca- marginalised communities. Fig. 1 illustrates the analytical framework.
pabilities, have freedom of choices, and enhance self-esteem. Lowered
entry barriers, and better access to information, markets, and resources, 2.2.1. Financial constraint reduction and technological de-skilling
can directly enhance users’ skills, knowledge, and connections, and Content-based digital platforms enable a key innovative feature: the
open new windows of opportunities, which also manifest in a higher users of the platforms can easily become content providers, rather than
capability to cope with economic hardships. terminal customers. By sharing their daily life skills, experiences, and
The evidence from this research provides valuable implications for even their normal life activities through content-based digital platforms,
policymakers, aiding them in designing effective policies to help mar- culturally or socially marginalised communities can enrol in new busi-
ginalised communities utilise and benefit from the new opportunities ness activities (Sotiriadis, 2017). This digitalised platform lowers
brought by digital technology. financial entrance barriers (a mobile device for a family’s daily use) and
requires no initial capital investment (the contents are from daily lives
2. Conceptual framework and no stock is needed). These features enable marginalised commu-
nities to generate income and create social value (Helfat and Raubit-
2.1. Digital technology and inclusive development schek, 2018; Sen, 2000; Fu et al., 2022). Moreover, the content-based
digital platforms are providing innovative and interactive terminals
In recent years, “inclusive development” has received increasing for content to be transferred digitally rather than physically (Teece,
attention (Banerjee and Duflo, 2007; Ranieri and Almeida Ramos, 2013; 2018; Si et al., 2022), which reduces the literacy barrier (no need to
Pike et al., 2016) which calls for innovative policy to reduce poverty, read, write or type) and skill barriers (simple to learn).
inequality, and promote social inclusion (UNDP, 2022). Development By providing the features of financial constraint reduction and
can be described as a process of expanding the freedom of the members technological de-skilling, the content-based inclusive digital platform
of society (Sen, 1990), Development can thus be achieved by empow- has the potential to empower marginalised communities (Fu et al.,
ering marginalised communities and creating opportunities for them to 2022). However, there is a lack of evidence that provides measurable
be involved in economic activities. As a result, members of society find policy implications. We summarise the working mechanisms and
an expansion of freedom in making life choices and creating their values, construct corresponding measurements to provide a systemic examina-
which leads to reduced inequality, poverty alleviation and inclusive tion and discussion.
development (Robeyns, 2009; Anand et al., 2013; Miletzki and Broten,
2017). 2.2.2. Digital technology and entrepreneurship
The creation and diffusion of new technology is bringing new in- Based on the mechanisms discussed above, digital technology en-
sights and critical discussions regarding inclusive development (Paunov, ables an entrepreneurial pursuit for marginalised people. It is only
2013; Chataway et al., 2014; Fu, 2020), and hence it is important to recently that studies started to connect the fields of digital technology,
ensure that the direction of technical change serves to create inclusive entrepreneurship, and inclusive development (i.e., Fu et al., 2022;
and sustainable economic and social development (Fu and Shi, 2022). Ghauri et al., 2022). Content-based digital platforms provide an inter-
The development of digital technology facilitates the diffusion of in- active resource for BOPs to enrol into the digitalised business activities.
formation, which improves the efficiency of communication (Lal, 1999; By sharing their life experiences, such as farming and crafting skills,
Bardhan et al., 2006), stimulates resource allocation and distribution marginalised people create their values for society and obtain economic
(Chun et al., 2015; Leite, 2019) and shapes the culture and structure of benefits through the monetisation process.
social organisations (Bloomfield and Coombs, 1992; Saldanha et al., Besides providing an interactive platform, digital technology em-
2021). Digital technologies are found to offer an opportunity to poor powers marginalised people by raising grassroots entrepreneurs. As
communities by helping them to overcome the constraints in accessing illustrated in our conceptual framework (Fig. 1), by alleviating infor-
information on markets and resources (Orser et al., 2019; Rodima- mation poverty, promoting aspiration and role modelling for entrepre-
Taylor and Grimes, 2019; Fu et al., 2021; Ghauri et al., 2022). Digital neurs, enhancing social capitals and providing resources for learning
technology facilitates openness and transparency of the ecosystem that and capability development, access to content-based digital platforms
encourages collaborative entrepreneurship (Dammert et al., 2015; thus provides employment opportunities that contribute to inclusive
Nambisan et al., 2019). development and resilience to crisis. Adjusting to the digital age would
However, two notable barriers hinder marginalised communities have profound societal impacts (Nambisan et al., 2019; Mansell, 2021),
from accessing digital technology, namely digital skills and capital re- and the inclusiveness of this transmission merits attention and effort.
quirements. First, access to information requires matching digital liter- We agree that poverty is rooted in global and local power relations
acy and skills (Gallivan et al., 2005; Orser et al., 2019). The “information and cultural-political issues and is difficult to resolve with micro in-
poverty” (Norris, 2001; Britz, 2004) and “digital divide” (Lam and Lee, terventions (Arora and Romijn, 2011). However, considering our earlier
2006) have been widely discussed as the barriers preventing stance on poorer communities, we believe that through stimulating their

3
X. Fu et al. Research Policy 54 (2025) 105173

Fig. 1. Conceptual framework.

confidence, freedom to choose and sense of belonging, it is possible to long-term implications for social inclusiveness and economic develop-
help them become value creating entrepreneurs (Sen, 1995; London and ment (Knack and Keefer, 1997; Iyer et al., 2005).
Hart, 2004; Simanis and Hart, 2009).
2.2.3.4. Learning & capability development. “Capability approach” (Sen,
2.2.3. Working mechanisms 1990; Robeyns, 2009) has long argued for the capability enhancement
and freedom for people. According to the “user empowerment mecha-
2.2.3.1. Information poverty reduction. “Information inclusiveness” nism in digital platforms” (Fu et al., 2022), access to content-based in-
(Teece, 2018) has been argued as a key element for inclusive develop- formation enables illiterate or low-educated people to obtain useful
ment, especially during the crisis of the COVID-19 pandemic (Dosi and skills and knowledge and to capture opportunities. This helps margin-
Soete, 2022). Access to digital technology, especially utilising content- alised people enhance their capabilities to cope with economic hard-
based digital platforms, would facilitate marginalised people to obtain ships and promote inclusiveness for the society. Meanwhile, the
information, exchange ideas and capture opportunities (Ansari and outbreak of COVID-19 further shed light on the importance of digital
Garud, 2009). A number of scholars have presented the concept of technology utilisation, such as the capacity for remote working and
digital inequality, as disadvantaged communities often suffer from a remote learning (Belzunegui-Eraso and Erro-Garcés, 2020).
lack of accessible information channels (Koo and Eesley, 2021; Lam-
brecht and Tucker, 2019). We refer to Pal et al. (2021) for the question 2.2.3.5. Opportunity creation/job creation. Effective opportunities for
design to evaluate the information access but we extend their mea- people to live with freedom according to their values and choices
surement from “receiving information” to include interactive means. (Robeyns, 2006) have been argued as another dimension of the “Capa-
bility approach”. The diffusion of digital technology enabled trans-
2.2.3.2. Role modelling/aspiration of grassroots entrepreneurship. Con- formation from traditional product exchanging practice to an interactive
tent-based digital technology fosters grassroots entrepreneurship and business engagement that eliminates the disadvantage of the remote
has been highlighted as a solution to poverty (Bruton et al., 2013). location and geographic dispersion of marginalised people (Tarafdar
Importantly, the content-based digital platform not only decreases the et al., 2013). Users are the content receivers, but easily become the
barriers to grassroots entrepreneurship, but also inspires marginalised content providers (Fu et al., 2022), which creates opportunities for
communities to enrol in economic and social activities by exposure to marginalised people to learn and nurture entrepreneurship and inno-
grassroot role models (Sen, 2000; Fu et al., 2022.) The exposure of vation from the BOP and for other socially marginalised groups (Hall
grassroots entrepreneurs from marginalised communities to digital et al., 2012; Mansell, 2021; Chakravorty et al., 2023). This mechanism
technology helps them to create value. It will also help us to understand creates inclusive opportunities and occupational positions for margin-
how to transform the benefits of technology into jobs and income for alised communities to enrol in economic activities.
marginalised communities (Orser et al., 2019; Ghauri et al., 2022).
2.2.3.6. Subjective barriers’ reduction. It has been argued that subjective
2.2.3.3. Social capital enhancement. Social capital plays an important barriers, such as lack of recognition, experience, self-confidence, and
role in inclusive growth (Shortall, 2008; Dinda, 2014) and an in- ignorance regarding digital technology, lack of freedom to choose, all
dividual’s career development and business performance (Dimitriadis create exclusions for some social groups such as elders, and people with
and Koning, 2022). Content-based digital platforms, as a form of “in- limited digital capabilities (Leten et al., 2016; Bohnsack et al., 2021;
formation communication technology” (ICT), help to bridge interper- Dosi and Soete, 2022). These subjective constraints are related to the
sonal communications and connections, which facilitates social capital “self-efficacy in digitalisation” (Gallivan et al., 2005; Lam and Lee,
(Molony, 2009). The literature has confirmed that digitalised exchanges 2006). Meanwhile, Pal et al. (2021) argued that these barriers are
and correspondence facilitate social capital (Mansell, 2021; Eiteneyer constraint-based negative factors when individuals assess digital tech-
et al., 2019) and digitalised communications supplement physical in- nology to bring value to their daily lives.
teractions rather than replace them (Palmié et al., 2022). Moreover, This reality highlights the importance of providing training to pro-
digital technology brings alternative communication means for mar- mote marginalised people’s utilisation of digital technology. However,
ginalised communities, that bridges their connections over physical previous research largely focused on training at the enterprise or
distance and thereby enhances social capital (Molony, 2009). This has country level (Gallivan et al., 2005; Marler et al., 2006; Saldanha et al.,

4
X. Fu et al. Research Policy 54 (2025) 105173

2021), and the rather subjective evidence pervasively supported by discussing “how it works” but assesses the interventions to identify
qualitative methods (Chan and Ngai, 2007; Pal et al., 2021). Micro-level “what works”. In our study, we learn from Banerjee and Duflo (2011)
measurable results of digital technology adoption and utilisation and a few recent seminal works (i.e., Eesley and Wu, 2020; Pal et al.,
training among marginalised communities in developing countries are 2021; Couture et al., 2021; Anderson and McKenzie, 2022) and provide
rare. Our study fills this gap in the current literature by providing a systematic discussion for the mechanisms associated with our inter-
experimental evidence through our RCT fieldwork. vention that led to the economic outcomes. Second, we conducted
qualitative research to supplement the empirical evidence to provide a
2.2.3.7. COVID-19 related utilisation. COVID-19 has been argued as a more detailed discussion of the mechanisms (Bakhshi et al., 2015;
“syndemic” (Dosi and Soete, 2022) crisis that disproportionately Pansera and Owen, 2018; Lashitew et al., 2019). These results are dis-
affected different groups, resulting in increasing structural inequalities cussed in Section 5. Third, soon after the RCT experiment, we rolled out
and having implications for lasting impacts on generations (Hanushek our intervention nation-wide through 64 Tech Hubs of our local net-
and Woessmann, 2020; Golinelli et al., 2020). This emphasises the need works. In this effort, we fully addressed the “programme entry effects”
for proper utilisation of digital technology during a crisis to promote (Burtless, 1995), “generalizability” (Faulkner, 2014) and especially the
inclusiveness. Digital technology plays a vital role in pandemic pre- ethical considerations (treated-untreated inequality) behind the RCT
paredness and response, and better applications of it facilitated resil- design (Reddy, 2012).
ience against hardship during the COVID-19 crisis (Whitelaw et al.,
2020). For example, it helped people to receive timely health care 3.1.2. Fieldwork context, sample randomisation and collection
(Greenhalgh et al., 2020), to maintain working capacity (Belzunegui- We chose Bangladesh as the fieldwork location to deliver the RCT
Eraso and Erro-Garcés, 2020) and sustain their mental health (Almourad due to a few considerations. First, according to the World Bank,1
et al., 2021). Our fieldwork study overlapped with the COVID-19 crisis Bangladesh is a representative low-income developing country whose
and the accompanying economic hardship (Oct 2019-Apr 2021), which GDP per capita was $1991.5 in 2018. It has a large population size
enabled us to access this mechanism and provide timely implications for (161.3 m, 2018, World Bank2) while many live in marginal circum-
focused policies. stances. For example, the smartphone penetration rate (mobile internet
We admit that social justice, freedom to choose and democracy are subscribers) was only 28 % in 2020 (Okeleke et al., 2021). This context
very important for poverty reduction, but poverty elimination needs offers a large under-served user community where an online community
multiple approaches (Sen, 1981). Although individual actions and can be built to benefit many people. Meanwhile, Bangladesh is also one
choices are rooted in a macro context that cannot be ignored, these of the world’s fastest-growing economies, and the central government
actions can empower these individuals and communities. Our research has launched the “Digital Bangladesh - Vision 2021” programme to
aims to help the poor by enabling them to create value and income by invigorate its digital technology ecosystem.3 The dense population in
themselves. This process would help to develop their capabilities, aspi- Bangladesh and the use of English (formerly a British colony) enables a
rations, identities, and most of all confidence. This is also a proactive large potential domestic market for the new digital business model.
approach to eliminate poverty in addition to enhancing social justice Moreover, we employ networks that can obtain the local government’s
and democracy. support for the project and facilitate RCT fieldwork when needed.
Considering local socioeconomic significance, we conducted RCT in
3. Empirical design the marginalised wards (city districts) in Dhaka (the capital city of
Bangladesh, representative urban area which includes 57 marginalised
3.1. Empirical approach and fieldwork wards) and Bogura (a smaller city, representative semi-urban area which
includes 21 marginalised wards). The sample stratification was imple-
3.1.1. Randomised control trial mented in two stages. First, in consultation with local authorities, we
Randomised Control Trial (RCT) is an experimental impact evalua- built stratified sampling frames of all eligible wards based on key so-
tion design which has been well-established in studies of development cioeconomic and demographic indicators to obtain two wards in each
economics and social policies (Burtless, 1995; Pitt and Khandker, 1998; city. Second, we randomly selected one ward as the treatment group and
Bratberg et al., 2002). In recent years, RCT has received increasing another ward as the control group in each city. The intervention was
attention, especially in development policy evaluations regarding then conducted with all households within the two treatment wards.
poverty reduction (Banerjee and Duflo, 2007; Banerjee et al., 2015; Lee This stratification design (same city, but different yet comparable
et al., 2021), industrial or regional development (Duflo et al., 2011; wards) helps minimise the initial discrepancies between the treatment
Bakhshi et al., 2015), and for the development of enterprises or of in- and control groups—such as lifestyle, culture, climate, and urban
dividuals (Eesley and Wang, 2017; Eesley and Wu, 2020), especially for infrastructure—while reducing the spillover effect and enhancing the
entrepreneurs (Chatterji et al., 2019; McKenzie and Puerto, 2021). consistency of the treatment.
As a strict experimental method, the pleasing feature of the RCT Detailed statistics of this stratification can be found in Appendix
approach is its randomised allocation of subjects to a treatment and a Tables B1 and B2 and the geographical distribution of the wards during
control group, which removes systematic correlation between treatment our RCT can be found in Appendix Fig. A1. As a result, respondents in
status, averts observable and unobservable selection biases, and thus Dhaka North ward 6 and Bogura ward 9 were randomly selected as our
provides a causative rather than summative evaluation (Bratberg et al., treatment groups. In contrast, respondents from Dhaka South ward 3
2002; Banerjee et al., 2007; Elbers and Gunning, 2014). This causative and Bogura ward 17 are the control groups. In the research design, we
discussion helps researchers evaluate the effects of economic stimuli selected wards for treatment and control that are far away from each
attributed to intervention; it provides reliable evidence to government other to minimise the possible spillover of treatment effects. After
and international organisations to support policy making (Bakhshi et al., cleaning all the invalid observations and matching respondents from
2015). Given these advantages, we adopt RCT as our empirical baseline survey to endline survey, we obtained a working sample of
approach. 1104 households, which includes 2974 adult respondents in our RCT.
However, our research is distinguished from traditional RCT litera-
ture in three veins. First, the primary feature is that we discuss the
treatment effect (economic resilience) as well as the mechanisms that 1
https://data.worldbank.org/indicator/NY.GDP.PCAP.CD?locations=BD
contribute to the real impacts. As argued by Reddy (2012), classical RCT 2
https://data.worldbank.org/indicator/SP.POP.TOTL?locations=BD
work, such as that of Banerjee and Duflo (2011), lacks effort on 3
See https://www.undp.org/bangladesh/blog/digital-bangladesh-inno
vative-bangladesh-road-2041

5
X. Fu et al. Research Policy 54 (2025) 105173

Appendix Table A1 shows the socioeconomic statistics of our sample We did the baseline survey face to face in October 2019 to collect the
respondents. demographic and socio-economic information from our sample resi-
As the respondents’ socioeconomic status, demographical features, dents. After a few months’ data compiling and analysis following the
infrastructures, and local policies were identical between the two baseline survey, we designed and released the App nationwide in
groups, our intervention reveals the difference between control and Bangladesh in February 2020. Our fieldwork intervention (treatment),
treatment groups after the experiment. As shown in Appendix Table A2, the digital technology application (mobile App) usage training pro-
1,413 individuals (538 households) were in the control group, and 1561 gramme, was launched in June 2020. Finally, we conducted the endline
individuals (566 households) were in the treatment group. We found no survey through online interviews in April 2021 to assess the impact of
statistical difference between these two groups before our RCT the treatment. In the survey, we invited a maximum of four individuals
experiment.4 from each household to take the intervention and answer the
There are two notable matters we would like to highlight here. questionnaire.
Firstly, due to the nature of the training intervention, our experiment During the App promotional campaigns, our fieldwork team deliv-
was not able to include “non-compliant” individuals, such as those who ered a series of App-usage training sessions to improve the digital lit-
refused to take part in the survey or were unable to participate in digital eracy and digital self-efficacy of the households in the treatment group
training at the beginning. Hence, referring to comparable literature such while introducing the features and benefits of the Haate Haat App and
as Banerjee et al. (2007) and Bakhshi et al. (2015), we acknowledge a the content-based digital business model where they can utilise the App.
potential upward bias in estimating our treatment effect. Secondly, our Based on the results of the sample randomisation as discussed in the
treatment was administered at the household level, ensuring that all above section, our fieldwork team physically visited the randomly
eligible household members received it. To gather data, we interviewed selected households in the treatment wards (Dhaka North ward 6 and
a maximum of four individuals within each household, thereby con- Bogura ward 9) to promote the Haate Haat App and deliver the digital
trolling for potential biases stemming from household size variations. It literacy training. The residents from the treatment groups were
is important to note that this study encompasses both household and encouraged to trial the Haate Haat App and receive the digital literacy
individual levels of analysis. For instance, questions pertaining to indi- training either via the illustration by the fieldwork team or by visiting
vidual income and employment status are addressed at the individual the SBK foundation’s local Technology Hubs (Tech-hubs). After the
level and vary from person to person. Conversely, questions regarding breakout of the COVID-19 pandemic, our local fieldwork team utilised a
household income represent the collective income of all household virtual method to implement the intervention. In total, the treatment
members, including rents or operating income, if applicable (refer to group received an average of 1.5 h’ digital literacy training per house-
Fig. 1). hold, which was combined with physical and virtual delivery across our
experimental period.6 Notably, although we endeavoured to encourage
3.1.3. Intervention implementation the audience to download the Haate Haat App, registration and main-
To apply our RCT intervention (treatment), we designed a content- tenance were not compulsory in order to receive our training. The
based digital platform named ‘Haate Haat’ (‘হােত হাত’ in Bengali, audience are free to make their own decision to use Haate Haat or other
(‘hand-to-hand’ in English)). Our Haate Haat App aims to help mar- digital platforms. Hence, we define the “treatment group” as the in-
ginalised communities and grassroots entrepreneurs establish mean- dividuals and households for whom our digital literacy training was
ingful collaboration, exchange information, and create income through provided, rather than the Haate Haat App active users.
the short videos they produce themselves. The key feature of the Haate Our fieldwork team emphasised a few objectives when delivering the
Haat App is that it is designed to help marginalised communities and training sessions. First, our fieldwork instructors helped audiences to
grassroots entrepreneurs to establish important networks and collabo- improve their digital literacy. For example, they showed the audiences
rations and create business and income through the contents (i.e., how to create a short video featuring their specific values (i.e., cooking /
photos, short videos) or to sell the products that they produce.5 Also, cultivating/ crafting skills/characteristic livelihood /talent show) and
through this digital platform, they can establish networks, develop upload to the digital platforms. Second, our fieldwork instructors
micro/small business, share or learn skills, and hence develop social endeavoured to explain the new business model to the audiences and
capital and capabilities and gain economic benefits. This App was make them realise the value of creating content and to utilise digital
developed based on another successful App, Kuaishou (Kwai), which did technology such as our Haate Haat App to achieve the economic and/or
not use algorithms to promote more ‘engaging’ content or content pro- social benefits. Third, they helped to improve the audience’s digital self-
duced by famous people so that ‘normal people can be seen’, and hence efficacy by illustrating how to make a proper utilisation of digital
became popular among farmers, villagers, migrants, and youth. We technology to access information, obtain/exchange knowledge, improve
develop and manage the App with our Bangladesh local operating social connection and start up an entrepreneurship through digital
partner, SBK Foundation. There was a 9-member technical team and a 4- platforms (such as Haate Haat, but it could be other Apps, depending on
member researcher team who collaborated to build up the Haate Haat their preferences). Importantly, our fieldwork team helped the audience
App. There were 30 local people (SBK staff and voluntary students from to realise the low cost (financial constraint reduction) and easy-access
Sher-e-Bangla Agricultural University) who contributed to our fieldwork (technological de-skilling) and that the content-based digital business
team and 31 people helped to collect the data through two waves of model enables them to participate in the main-stream society, and
surveys. illustrated how they can create a value with it.
We would like to note two matters. First, soon after our RCT

4
The only exception is the number of household members, where the control
6
group has 0.12 more household members compared to the treatment group. The designed training plan comprised a 40-min initiation training session at
However, we believe this does not affect the validity of our RCT: first, the each household, a 45-min session conducted by SBK in their ‘Tech-hubs,’ and a
difference is not due to selection bias or adverse selection, given our two-stage 5-min follow-up session. A total of 566 households received the digital training
sample stratification; second, the economic size of this difference is consider- program. Due to the COVID-19 outbreak, the intervention could not be
ably small. implemented simultaneously and was mostly conducted in a hybrid manner.
5
This App is available on Google Play (as currently we cannot anticipate that Unfortunately, we do not have records of whether individuals participated in-
marginalised communities would have access to the Apple system). For infor- person or virtually. However, we adhered to the principle that the planned
mation and download please see: https://play.google.com/store/apps/details? training should be delivered in a manner appropriate to the situation and in
id=com.syntecglobal.hand_to_hand&hl=en compliance with government COVID policies.

6
X. Fu et al. Research Policy 54 (2025) 105173

experimental period finished in April 2021, the Haate Haat App pro- each comparing group. Our working sample includes over a thousand
moting campaigns along with the training sessions were expanded to all observations for each group, far beyond the estimated minimum sample
the 64 districts in Bangladesh as SBK foundation has Tech-hubs for each size to secure estimating power.
of the districts. Social media campaigns and posters/banners in public
places were also introduced across Bangladesh. This effort aimed to 3.2.4. Selective migration
deliver our Haate Haat App and the content-based inclusive digital Some experimental studies may be concerned with selective migra-
business model to the broadest audiences to promote its social impacts. tion issues, especially when the intervention is bonus- or credit-related
Second, the design of the Haate Haat App is an effort to exclude any (i.e., Banerjee et al., 2015; Nakano and Magezi, 2020). This means
third-party interests when promoting digital literacy training, with the that respondents may selectively migrate to regions where the inter-
hope of obtaining a set of first-hand usage data from participants during vention was implemented to receive the benefits, which would skew the
the RCT. However, during the training, we also recommended all other statistics. However, selective migration did not seem likely to happen in
mainstream digital platforms to participants and explained how they our case, given that our intervention was not monetary-related and was
could benefit from them. Hence, a limitation of this research is that we a double-blind process.
did not obtain enough observations to quantitatively compare Haate
Haat App users and non-users, except for carrying out qualitative 3.3. Variable construction
research. We utilise the qualitative approach to provide supplementary
evidence for this paper in Section 5. This section constructs the variables of each element following our
conceptual framework discussed earlier. For the index measurements of
3.2. Validation and identification for the RCT the mechanisms, we followed the approach of Partial Least Square (PLS)
(Staples et al., 1999; Tenenhaus et al., 2005). For the outcome variables
To ensure the validation of the RCT experiment, we highlighted a such as income, we utilised the actual value of the local currency
battery of approaches during the assignment procedure: (Bangladeshi Taka) and a normalised value to construct a comparable
index. A tabular summary of the measurement construction and the
3.2.1. Double-blind corresponding literature can be seen in Table 1.
First, the research team (who designed the questionnaire and ana- Two matters need to be mentioned here. First, our study is contingent
lysed the data) and the survey team (who operated the survey) were on residents’ utilisation of content-based digital technology (mobile
separated; second, the survey team and treatment team (who delivered apps), so the observations are the individuals who can access mobile
the intervention) were separated. The survey team was only responsible applications, and we excluded the non-applicable respondents (e.g.
for conducting the interviews for the baseline and endline surveys, while don’t have a smart phone). Second, as we examine the real-income
the treatment team was only responsible for the consistent assignment of impact of respondents before and after the intervention, we have
intervention. Both teams were monitored by the research team to ensure excluded retired respondents from the set of questions related to income
that they received identical within-team training, and to implement and employment. This exclusion has resulted in a reduction of 103
qualified fieldwork. In other words, to illustrate this double-blind, the observations.
people who received and analysed the data did not know the re-
spondents throughout the experiment; meanwhile, the people who 3.3.1. Measurement of inclusive development and resilience
collected the data did not know whether each individual respondent Although different studies take different perspectives to discuss
received the treatment. This procedure secured the exclusion of sub- “inclusive development”, most of them take income, employment, and
jective biases, selective biases and physical cues that could potentially capabilities as footholds (Ranieri and Almeida Ramos, 2013). Hence,
taint the quality of the experiment (Bratberg et al., 2002; Banerjee et al., following our analytical framework, we measured inclusive develop-
2007). ment as a synthesised term with components that included income
(household and individual), employment, and entrepreneurship.
3.2.2. Hawthorne effect and spill-over During crises such as the economic hardships caused by the COVID-
The ‘Hawthorne effect’ (also known as the ‘participation effect’) 19 pandemic, inclusive development may also manifest as a greater
argues that respondents would adjust their behaviours and answers if resilience in income, employment and in maintaining existing or
they were aware that they were participants in an experiment creating new businesses for the residents. Our baseline and endline
(McCambridge et al., 2014). To avoid this, we included the procedure of survey overlapped the period when this crisis was most severe (Oct
third blinding – the respondents were not informed about the contrasting 2019-Apr 2021). As shown in Table 2, we observed a dramatic decrease
experiment. This meant that neither the treatment group nor the control of all the inclusive development indicators. However, the treatment
group knew of the existence of this comparison or about each other. This group was comparatively less affected by the hardship and reflected a
effort also helped us to exclude spillovers and contaminations and higher resilience.
identify intervention outcomes (Glennerster and Takavarasha, 2013). We normalised the absolute value of household income decrease
Moreover, the wards for treatment and control were selected as they are (taking the whole household as an observed unit) and the individual
far away from each other in order to minimise the possible spillover of income decrease (taking each individual as an observed unit), as the
treatment effects. As it was impossible to completely eradicate the measurements of income resilience. The indicators of employment and
concern of potential spillover throughout this year-long fieldwork, we entrepreneurship were dummy variables with a positive value that
are cautious about claiming a “triple-blind” for our RCT. indicated that this respondent is currently employed or maintained/
created a business during the previous year. Based on these elements and
3.2.3. Attrition and sampling power applying the Partial Least Square (PLS) approach, we constructed the
The potential attrition may be a critical perspective for experimental Index of inclusive development & resilience for residents as the primary
studies (Bakhshi et al., 2015). To address this concern, we adopted the outcome to identify in this study.
power calculation (Batistatou et al., 2014) to estimate sample size to
secure the minimum detectable effect and to address potential attrition. 3.3.2. Measurement of working mechanisms
Following Dupont and Plummer Jr (1990), we set the power of 80 % Consistent with the literature discussed in Section 2.2, we first argue
with a significance level of 5 % while selecting a standard deviation of five mechanisms through which access to content-based digital platform
0.49 and an expected average treatment effect (ATE) of 0.1. The output technology impacts inclusive development. Meanwhile, we include two
suggested that the sample size should be at least 336 observations for other mechanisms that our RCT intervention arguably delivered. All the

7
X. Fu et al. Research Policy 54 (2025) 105173

Table 1
Variable construction and the corresponding literature.
Measurement Description Corresponding literature & theories for variable
construction

Information poverty reduction Composed by survey questions: “Does your usage of Apps benefit or “Information poverty and digital divide”: Norris (2001);
improve the following aspects: Access to information; Access to news; Britz (2004); Lam and Lee (2006).
Knowledge exchange. “Information inclusiveness during digitalisation and in
COVID-19”: Teece (2018); Dosi and Soete (2022)
“Valence framework model” of mobile App usage: Li et al.
(2018); Pal et al. (2021).
Role modelling /Aspiration Composed by survey questions: “Does your usage of Apps benefit or “Grassroot entrepreneurship and poverty reduction”:
improve the following aspects: Getting business ideas (role modelling); Bruton et al. (2013)
Getting government policy guidance; Getting entrepreneurship “Digi-tech enabled grassroot entrepreneurship and
examples innovation”: Tarafdar et al. (2013); Nambisan et al.
(2019); Si et al. (2022); Fu et al. (2022)
Social capital enhancement Composed by survey questions: “Does your usage of Apps benefit or “Digital technology and social capitals” Molony (2009);
improve the following aspects: Social networking; Communication; Eiteneyer et al. (2019); Palmié et al. (2022).
Trust in each other” “Social capital and inclusive development”: Iyer et al.
(2005); Shortall (2008); Dinda (2014).
Learning & capability development Composed by survey questions: “Does your usage of Apps benefit or “Capability approach”: Sen (1990); Robeyns (2009)
improve the following aspects: Working skill learning or receiving “User empowerment mechanism in digital platforms”:
training; Language learning; Entrepreneurship skills/capability Helfat and Raubitschek (2018); Fu et al. (2022).
development”. “Teleworking capacity and application in COVID-19”:
Verstegen et al. (2019); Belzunegui-Eraso and Erro-Garcés
(2020)
Opportunity creation/ job creation Composed by survey questions: “Does your usage of Apps benefit or “Capability approach”: Robeyns (2006)
improve the following aspects: Opportunity to collaborate with other Content-based digital platforms as “multi-sided and
users/sellers; Opportunity to find business partners; Opportunity to find interactive content terminals”: Helfat and Raubitschek
jobs (2018); Teece (2018); Fu et al. (2022).
Subjective barriers to digital technology Composed with survey questions: “Do the following serve as barriers to “Self-efficacy in digitalisation and subjective constraints”:
your usage of mobile Apps: Lack of inclination; Lack of digital skills”. Gallivan et al. (2005); Lam and Lee (2006).
“Constraint-based negative valences”: Pal et al. (2021)
COVID-19 related utilisations: COVID-19 Composed by survey questions: “What specific activities did you “Health support on digi-tech during COVID-19”:
information and health advice; mental perform through mobile apps due to pandemic situation: COVID-19 Greenhalgh et al. (2020); Golinelli et al. (2020); Vargo
support during quarantine information & guidelines; COVID-19 health advice; Other health advice et al. (2021)
& online treatment; Receiving and providing mental support from “Communication on digi-tech and mental health during
others during quarantine; Expressing feelings and life experience during COVID-19”: Whitelaw et al. (2020); Vargo et al. (2021);
quarantine”. Canale et al. (2022)
“Mental health and digital wellbeing”: Vanden Abeele
(2021)
Index of inclusive development & resilience Composed by the indexes: household income resilience; employment; “Dimensions of inclusive development”: Ranieri and
for residents entrepreneurship Almeida Ramos (2013); Pike et al. (2016)

Table 2
The effect of COVID-19 on residents’ social-economic status.
Overall sample Control group Treatment group

Obs. Mean SD Obs. Mean SD Obs. Mean SD

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Panel A: Baseline survey (Oct 2019)

Household income (BDT) last year 2974 410,835.0 447,802.1 1413 416,723.5 460,144.7 1561 405,504.8 436,408.2
Personal income (BDT) last year 2974 139,094.9 258,506.7 1413 143,019.5 257,392.6 1561 135,542.4 259,542.3
Currently have a job 2971 0.4860 0.4999 1410 0.5007 0.5002 1561 0.4728 0.4994

Panel B: Endline survey (Apr 2021)

Household income (BDT) last year 2974 254,783.4 220,870.1 1413 238,533.28 200,835.0 1561 269,492.8 236,657.7
Personal income (BDT) last year 2873 91,688.6 159,684.6 1377 87,489.48 148,466.2 1496 95,553.6 169,312.8
Currently have a job 2871 0.4006 0.4901 1377 0.3914 0.4882 1494 0.409 0.4918

Note: This table reports the social-economic status of the respondents in our sample. Our baseline survey was conducted in October 2019 and the endline survey was
conducted in April 2021, which covered the most depressed period of COVID-19.

indexes are composed following the Partial Least Square (PLS) approach exchange information and knowledge, indicate a reduction in “infor-
(Staples et al., 1999; Tenenhaus et al., 2005). The original data are from mation poverty” (Britz, 2004; Lam and Lee, 2006).
the survey questions we collected from the respondents through the For the index “role modelling/aspiration”, we draw from the survey
baseline and endline surveys during our RCT fieldwork. questions about whether the respondents utilise digital technology to
obtain business ideas or see role models, get government policy guid-
3.3.2.1. Access to content-based digital platform technology. “Information ance, and learn from entrepreneurship examples.
poverty reduction” is constructed based on a set of survey questions The index of “social capital enhancement” is composed of survey
asking whether the respondents utilise content-based digital platforms questions about whether the respondents utilise mobile digital appli-
to access information, news, and knowledge exchange. A higher access cations to improve their social networking and communications and
to information, and better utilisation of technology to obtain and help them to establish trust among users.

8
X. Fu et al. Research Policy 54 (2025) 105173

The “learning & capability development” is constructed to measure randomly allocated the digital training programme to treatment and
marginalised persons’ utilisation of content-based digital technology control groups, where no significant socioeconomic discrepancies were
platforms from aspects of learning/acquiring working skills, language captured before the intervention among these two groups, the differ-
and entrepreneurial capabilities. ences in the outcomes between the two groups contributed to the effect
The “opportunity creation/job creation” index is constructed from the of the RCT intervention (Banerjee and Duflo, 2007; Bakhshi et al., 2015).
survey questions about whether the respondents utilise digital technol- Yi,h represents the index of “inclusive development and resilience for
ogy to create an opportunity to collaborate with their business partners residents” to each of its elements: household income, individual income
or explore occupational opportunities. of each household member, employment status of each individual
A higher value of these indices indicates positive treatment impacts member, and entrepreneurship status of the household.
from our RCT intervention. Fʹi,h is a vector of individual-level or household-level control vari-
ables: age, age-squared, gender dummy, gender of household-head
3.3.2.2. Subjective barriers and COVID-19 related utilisations. Our RCT dummy, household-head age, and household member number. Zʹ is a
fieldwork provides training in digital technology utilisation, which may set of fixed effects: education level, household-head marriage status,
reduce the subjective barriers for residents who received the treatment. household-head religion, number of household members, and number of
Meanwhile, the utilisation of digital technology plays a positive role in minors (under 18 years old). The robust standard errors are clustered at
assisting marginalised communities in coping with a crisis and providing the household level in the regressions.
resilience for them during economic hardships. In the survey, we asked As discussed in Section 2, the treatment of our RCT intervention does
whether the respondent found a lack of inclination, lack of literacy skills, not directly generate income for respondents; it works through the
or lack of digital skills, as a barrier to accessing digital technology ap- mechanisms that help respondents access the short video platform
plications (Mobile Apps). By composing these three dummy variables technology-based novel business model. To examine this, we apply the
and applying the Partial Least Square (PLS) approach, we constructed Two-Stage Least Square (2SLS) regression approach (i.e., Blinder, 1973;
the measurement of “subjective barriers to digital technology”. Banerjee et al., 2007; Lee et al., 2021) to study the mechanisms through
Applying the same approach (PLS), we constructed the measure- which the treatment contributes to the socioeconomic impacts as argued
ments of “COVID-19 related utilisations” by taking the sum of the in Eq. (1).
following range of perspectives: COVID-19 information and guidelines; ( )
health advice; other health advice and online treatment; the receipt or Yi,h = α + β2 Mechanismsi,h = β1 Treatment i,h + Fʹi,h Γ + Zʹ + εi,h (2)
provision of mental support from others during quarantine; expressing Eq. (2) shows the first stage, which examines the impact of treatment
feelings and life experiences during quarantine. on the indicators of mechanisms. Then, the second stage shows the
impact of treatment on the outcome which was initiated by the treat-
ment. Here Mechanisms represents the seven mechanisms discussed in
3.4. Model specification and validity checks Section 3.3.2. The term Yi,h represents the index of “inclusive develop-
ment and resilience for residents” as the outcome variable. β1 and β2
3.4.1. Model specification represent the coefficients from the first and second stage, separately. The
Illuminated by the literature that conducted comparable research (i. treatment effect would be interpreted as β1 *β2 . According to Blinder
e., Duflo et al., 2011; Banerjee et al., 2015; Bakhshi et al., 2015; Lee (1973), we expect β1 *β2 to consist of β for the sign and significance, but a
et al., 2021), we constructed empirical models to evaluate the treatment larger magnitude as it includes more information through the structure
effects of our RCT intervention. This design is illustrated in Fig. 2. modelling.
Yi,h = α + βTreatment i,h + Fʹi,h Γ + Zʹ + εi,h (1)
3.4.2. Pre-experimental balance and attrition check
The Eq. (1) demonstrates the baseline model, where the “Treatment” An essential condition to secure the validation of RCT is that there
represents our key explanatory variable, which equals one if the are no significant pre-experiment statistical differences between the
respondent is from the treatment group and equals zero if not. As we

Fig. 2. Empirical results (2SLS) brief illustration.

9
X. Fu et al. Research Policy 54 (2025) 105173

treatment and control groups (i.e., Bakhshi et al., 2015). We conducted 4. Empirical results
balance checks in the baseline survey for all the concerned dimensions
related to respondents’ access to digital technology. As shown in Ap- 4.1. Baseline results
pendix Tables A1 and A2, we find no statistical difference between the
treatment and control groups, including demographic and socioeco- Table 3 shows the treatment effects of our RCT on the index of “in-
nomic status dimensions such as age, gender, household income or pre- clusive development and resilience for residents” and on each of its el-
experiment digital technology access and cognitions. ements (discussed in Section 3.3 and shown in Table 1). Consistent with
Meanwhile, we check the response rate (attrition) to see whether our Eq. (1), the treatment represents the marginalised people who randomly
data are exposed to survey response bias. Our endline survey covers received our training programme (treatment group), and its coefficient
93.3 % of the households and 91.7 % of the respondent individuals, indicates the statistical difference between the treatment group and
compared with the baseline survey. These attrition rates (6.7 % by people who did not receive the treatment (control group). Column (1) of
households and 8.7 % by individuals) do not show a significant differ- Table 3 shows that the treatment group significantly manifests a higher
ence between the control group (9.5 %) and treatment group (7.2 %). value of 0.13 for the index, which is 13 % of its standard deviation.
Compared with similar research (i.e., Banerjee et al., 2007; Duflo et al., From Columns (2) to (7) of Table 3, we take each element as an
2011; Eesley and Wu, 2020) that show above 15 % attritions, our outcome variable separately and find that the treatment plays a positive
attrition is arguably low and balanced between treatment and control role in each component. For the income elements, we compare the
groups and is not likely to trigger survey response bias. Please see Ap- indexed measurement and the actual number of incomes and see
pendix Table A3 for detailed illustration. consistent results. As demonstrated in Columns (2) and (3), the treat-
ment group exhibits a higher income resilience index of 0.095. Corre-
3.4.3. Randomisation checks - model specification and sample selection spondingly, the treatment group experienced less economic hardship, i.
To check the validity of our model specification, we conduct multi- e. a 39,562 Bangladeshi Taka (BDT) smaller decrease in household in-
variate regression analysis with our key explanatory variable treatment come during the COVID-19 pandemic. This treatment effect is equiva-
dummy as dependent variable and use the covariates and outcome var- lent to 9.6 % of the mean value of the household income in our overall
iables in the baseline regression as independent variables. An ideal sample’s baseline survey (410,835 BDT, see Appendix Table A1).
model specification is expected to demonstrate insignificant statistics. Meanwhile, Columns (5)–(6) report the marginal effect of Probit
Also, to check the multicollinearity we adopt O’Brien (2007)’s approach regression when observing the treatment effect on employment and
to conduct a variance inflation factors (VIF) test. As shown in Appendix entrepreneurship, for both variables have a binary outcome. We find
Table A3, the statistics indicate acceptance of our regression specifica- that respondents from the treatment group statistically have a 2.7 %
tion’s reliance. higher likelihood of maintaining an occupation and a 3.1 % higher
Besides the model specification, the selection of participants is also a likelihood of maintaining their business or creating a new business
matter for RCT design (Allcott, 2015). The target population of our during economic hardship. This evidence indicates the treatment group
experiment is the marginalised households with at least 3G internet has a higher resilience during the crisis and provides insight into their
coverage who are located either in a marginalised urban area or a semi- income performance compared with the control group.
urban area of Bangladesh. These eligibility criteria are based on the fact We would like to highlight the cost-effectiveness of our fieldwork
that our research requires essential internet connections and smart de- programme, as evidenced by the results presented here. The expenditure
vices such as a smartphone. for our intervention was approximately £50 k, covering 566 households
in the treatment group. When we convert the treatment effect of 39,562
BDT (the saved household income per household on average) into GBP

Table 3
Baseline results: Treatment effect on inclusive development & resilience for residents.
Index of inclusive development & Each element for the index
resilience for residents
Household income Individual income Employment Entrepreneurship

Resilience Real income Resilience Real income

index decrease (BDT) index decrease (BDT)

(1) (2) (3) (4) (5) (6) (7)

Treatment: 0.130*** 0.095** − 39,562.469** 0.073* − 15,853.90* 0.027** 0.031**

[0.001] [0.011] [0.011] [0.067] [0.067] [0.041] [0.021]
Individual control set Yes Yes Yes Yes Yes Yes Yes
Household control set Yes Yes Yes Yes Yes Yes Yes
Education fixed effect Yes Yes Yes Yes Yes Yes Yes
Household fixed effects Yes Yes Yes Yes Yes Yes Yes
Observations 2870 1075 1075 2872 2872 2870 2931
R2/Pseudo R2 0.105 0.150 0.150 0.084 0.084 0.4039 0.029
Specification OLS OLS OLS OLS OLS Probit Probit

Note: Resilience index is the normalised absolute value of income decreased during COVID-19 (Apr 2020 – Apr 2021). Real income decrease measures by how much
Bangladeshi Taka (BDT) the household’s real income decreased during COVID-19, where a lower value indicates a higher resilience during economic hardship and the
coefficient indicates the amount of income. Employment indicates whether the respondent has a job at the time of our endline survey (Apr 2021). Entrepreneurship
indicates whether the respondent’s household maintained their business or created a new business during Apr 2020-Apr 2021. For control settings: Individual control set
and Household control set include respondents’ age, age-squared, gender, gender of household-head, household-head age, and the number of household members.
Education fixed effect and Household fixed effects include the fixed effects of the respondents’ education level (12 categories), household-head marriage status,
household-head religion, number of household members, and number of household minors (under 18 years old). Each column shows the results from the corresponding
specification. Parentheses contain p-values where.
*
p < 0.1.
**
p < 0.05.
***
p < 0.01.

10
X. Fu et al. Research Policy 54 (2025) 105173

(approximately £265) and multiply it by the 566 treatment households, digital technology,’ the treatment predicts an increase in the index value
the total amount is about £150 k. In comparison to the £50 k cost of 0.131 (0.103*1.272) and 0.130 (− 0.122*-1.067), respectively.
incurred for the treatment, this represents a significant economic impact Corresponding to Fig. 2, we provide detailed statistics of the 2SLS
with notable policy implications. regression results in Table 4.1. This table also includes the results
Fig. 2 and Table 4.1 illustrate the results of Two-Stage Least Squares examining the validity of the 2SLS regressions, such as the Anderson LM
(2SLS) regressions for all the regression specifications in Eq. 2. This work under-identification test and the Sargan-Hansen over-identification test.
examines the mechanisms through which our RCT intervention has so- The statistics universally reject the null hypotheses, confirming the
cioeconomic impacts, where the coefficient results from the first stage validity of the 2SLS regression structure and the reliability of the results.
(treatment to mechanisms) and the second stage (mechanisms to out- In Table 4.2, we replicate the 2SLS regressions separately for each of
comes) are listed separately. Notably, as the variables utilised here are the four elements of the resilience index to observe each element’s
constructed in index values (discussed in Section 2), the coefficient re- contribution to the overall treatment effect. Furthermore, we test the
sults do not have specific economic interpretations. impact on income resilience individually for each of the four elements
Fig. 2 presents a visual summary of our 2SLS results. The middle by instrumenting each element with all six mechanism indicators. This
arrow represents the results in Column (1) of Table 2. The left and right approach allows us to discern how and which mechanisms contribute to
sets of arrows refer to the first and second stage results, respectively. We each element of resilience. Notably, consistent with column (4) of
find that the treatment significantly predicts better access to all five Table 3, we find that our treatment effects largely manifest at the
mechanisms of accessing digital technology, a lower subjective barrier household level, while the individual income resilience show a limited
to digital technology, and improved utilisation of digital technology significance of 10 %. This suggests that our program has a more sig-
during the COVID-19 pandemic. Similar patterns are evident in the nificant impact on improving employment and entrepreneurial out-
second stage results, as shown on the right side of the figure. The higher comes, particularly at the household level, rather than individual level.
values of the mechanism indicators positively predict a higher index of One explanation for these differences in effect sizes is that the treatment
inclusive development and resilience for residents, except for the ‘sub- was conducted at the household level, which is typically the unit
jective barriers to digital technology,’ which exhibits a negative rela- accessing digital platform content, especially given that our respondents
tionship, as expected. When moving from the first stage to the second are predominantly from marginalised communities. This highlights the
stage, the treatment significantly predicts a higher index of inclusive importance of designing policies that consider households as the pri-
development and resilience for residents. For example, from the mech- mary unit when promoting digital training programs.
anisms of ‘information poverty reduction’ and ‘subjective barriers to Overall, our empirical results correspond to our discussions in

Table 4.1
Two-Stage Least Square (2SLS) regression results.
Panel A: First stage - The impact of treatment on various mechanisms.

Dependent variables are the mechanisms in each column

Information Role modelling Social capital Learning & capacity Opportunity Subjective barriers to COVID-19 related
poverty reduction /Aspiration enhancement development creation/job digital technology utilisations
creation

(1) (2) (3) (4) (5) (6) (7)

Treatment 0.103*** 0.052* 0.096*** 0.079*** 0.061** − 0.122*** 0.118***

[0.000] [0.059] [0.000] [0.004] [0.027] [0.000] [0.000]
All control Yes Yes Yes Yes Yes Yes Yes
variables
Observations 2870 2870 2870 2870 2870 2870 2870

Panel B: Second stage - The impact of mechanisms on inclusive development & resilience.

Information Role Social capital Learning & Opportunity Subjective COVID-19

poverty modelling / enhancement capacity creation/ job barriers to digital related
reduction Aspiration development creation technology utilisations

(1) (2) (3) (4) (5) (6) (7)

Dependent Index of inclusive 1.272*** 2.527* 1.354** 1.648** 2.128* − 1.067*** 1.101***
development &
variable resilience for [0.009] [0.086] [0.012] [0.026] [0.058] [0.005] [0.005]
residents
Anderson LM 0.000 0.000 0.000 0.000 0.000 0.000 0.000
under-
identification
C-D-W weak- 0.000 0.000 0.000 0.000 0.000 0.000 0.000
identification
Sargan-Hansen 0.000 0.000 0.000 0.000 0.000 0.000 0.000
over-identification
All control Yes Yes Yes Yes Yes Yes Yes
variables
Observations 2870 2870 2870 2870 2870 2870 2870

Note: All control variables include individual control set, household control set, education fixed effect and household fixed effects, which is consistent with Table 2. The
Anderson LM under-identification represents the p-value of the Anderson L.M. under-identification test, which has a null hypothesis that the equation is under-identified.
The C-D-W weak-id statistics represents the p-value of the Cragg-Donald-Wald weak-identification test, which tests whether the exogenous regressors are only weakly
correlated with the endogenous regressors. The Sargan-Hansen over-id test represents the p-value of the Sargan-Hansen over-identification test, which tests whether the
regression is over-identified when introducing the exogenous regressor.

11
X. Fu et al. Research Policy 54 (2025) 105173

Table 4.2
2SLS results when unbundling the resilience index into each element (Second stage).
Information Role modelling Social capital Learning & capacity Opportunity Subjective barriers COVID-19
poverty reduction /Aspiration enhancement development creation/job to digital technology related
creation utilisations

(1) (2) (3) (4) (5) (6) (7)

Dependent variables
Household 0.698*** 2.092* 0.767*** 1.247** 1.046** − 0.639*** 0.804***
income
resilience index [0.004] [0.088] [0.006] [0.019] [0.020] [0.003] [0.004]
All control Yes Yes Yes Yes Yes Yes Yes
variables
Observations 1075 1075 1075 1075 1075 1075 1075
Individual income 0.716* 1.423 0.762* 0.928 1.198 − 0.601* 0.620*
resilience index [0.077] [0.160] [0.082] [0.102] [0.138] [0.067] [0.065]
All control Yes Yes Yes Yes Yes Yes Yes
variables
Observations 2870 2870 2870 2870 2870 2870 2870
Employment 0.910*** 1.220*** 0.942*** 1.030*** 1.193*** − 0.792*** 0.865***
[0.003] [0.000] [0.002] [0.000] [0.000] [0.006] [0.007]
All control Yes Yes Yes Yes Yes Yes Yes
variables
Observations 2870 2870 2870 2870 2870 2870 2870
Entrepreneurship 1.261*** 0.995*** 1.059*** 1.168*** − 0.849*** 0.868*** 1.261***
[0.000] [0.000] [0.000] [0.000] [0.002] [0.002] [0.000]
All control Yes Yes Yes Yes Yes Yes Yes
variables
Observations 2870 2870 2870 2870 2870 2870 2870

Note: In the first section, where the dependent variable is the household income resilience index, the regression is conducted at the household level, with a total of 1104
observations. In the following three sections, individual-level regressions are performed, with a total of 2974 observations. Parentheses contain p-values where * p <
0.1, ** p < 0.05, *** p < 0.01. All control variables include individual control set, household control set, education fixed effect and household fixed effects, which is
consistent with Table 2.

Section 2. Access to digital technology empowers marginalised com-

Table 5
munities through the five mechanisms, facilitates their capacity for
Robustness test: unobservable selection bias.
employment and entrepreneurship, results in income growth, and ach-
ieves social inclusiveness. Meanwhile, through reduced subjective bar- Independent variable: Treatment group
riers and better usage during economic hardship, residents who have the Oster’s (2019) Oster’s (2019) β value
advantage of recognising and utilising digital technology manifest a δ value when:
higher resilience during crises such as COVID-19. A comprehensive Setting δ Setting δ
discussion, including the policy implications and potential threats to our =0 =1
results, is provided in Section 6. (1) (2) (3)

Index of inclusive development & 38.077 0.132 0.130

4.2. Robustness: unobservable selection bias test resilience for residents
Information poverty reduction 13.157 0.196 0.190
Role modelling /aspiration 60.824 0.110 0.111
A frequent concern for social science study is the unobservable fac-
Social capital enhancement 13.529 0.185 0.180
tors that may drive the empirical results. We apply the approach Learning & capability development 14.075 0.152 0.148
developed by Oster (2019) that considers the coefficient movement and Opportunity creation/ job creation 11.283 0.119 0.115
the corresponding R-squared movement when evaluating the omitted Subjective barriers to digital 19.493 − 0.242 − 0.240
variables bias. This approach includes two optional indices: δ, which technology
COVID-19 related utilisations 106.504 0.240 0.245
indicates the weight for the unobservable that is needed to make an All control variables Yes Yes Yes
equivalent impact to the observables; β, this index reflects the treatment
effect under the assumption of equal selection (δ = 1) and no informa- Note: Each row shows the results from the corresponding specification. Each cell
shows the index value when applying Oster’s (2019) approach.
tive relationship (δ = 0) between unobservable and observable. We
adopt both indices in our tests. Each row in Table 5 represents the result
for one regression specification corresponding to our mechanism dis- among the regression specifications.
cussion in Fig. 1.
Our results universally reject the hypothesis of unobservable selec- 4.3. Implications in income and age groups during “syndemic”
tion bias. For example, Column (1) of Table 5 reports the estimated
value of Oster’s δ for each corresponding specification. In Row (1) of Literature has long called for attention to digital inclusiveness
Column (1), the value of 38.077 indicates that the selection of unob- regarding different age-groups, especially the older people (i.e., Lam and
servable should weigh >38 times relative to observable to make the Lee, 2006). Under the circumstance of global crisis, appropriate inter-
treatment effect equal to zero, which is way higher than the benchmark vention on specific social groups would reduce the digital divide which
(1, which value indicates equipotent selection) as proposed by Oster has been exacerbated by existing inequalities among different social
(2019). For the β index, we see that none of the coefficients in Columns groups during the “syndemic” such as COVID-19 (Dosi and Soete, 2022).
(2) and (3) change their sign under the assumption of δ = 0 (no infor- To analyse these implications among our sample, we sub-classify the
mative relationship between observable and unobservable) and δ = 1 baseline working sample according to respondents’ income and age
(equal selection between observable and unobservable). This evidence groups. We trisect the sample according to their income into lower in-
argues that our baseline results are not driven by unobservable factors come, middle income and higher income levels in our observed sample.

12
X. Fu et al. Research Policy 54 (2025) 105173

Similarly, the age groups are: 18–30 as the younger age group; 30–50 as abject poverty. In Panel B, we find that although older people naturally
the middle age group; 50 and above as the older age group. We apply the have a disadvantage regarding resilience, our intervention does not
model specifications 3.1 and 3.2 when adding the interaction terms of show better significance among younger people. This evidence might be
different age groups and income groups to the main effects, which en- explained by the accumulated wealth differences corresponding to the
ables us to observe the differential effects across different social groups. residents’ age, even though younger residents have an advantage in
assimilating digital technology (Couture et al., 2021).
Yi,h = α + β1 Treatment + β2 Middle incomei,h + β3 High incomei,h
+ β4 Treatment × Middle incomei,h + β5 Treatment × High incomei,h
4.4. More and better usage of mobile apps?
+ Fʹi,h Γ + Zʹ + εi,h
(3.1) Our preceding empirical results support the treatment effect result-
ing from our fieldwork intervention and validate the underlying mech-
Yi,h = α + β1 Treatment + β2 Middle Agei,h + β3 Old Agei,h anisms. As an extensive and supplementary piece of evidence, we
+ β4 Treatment × Middle Agei,h + β5 Treatment × Old Agei,h (3.2) analyse whether our training treatment has an impact on residents’
+ Fʹi,h Γ + Zʹ + εi,h actual utilisation of mobile apps and whether it influences their income
resilience.
The results are shown in Panel A and Panel B of Table 6. Notably, the As depicted in Table 7, we apply the baseline 2SLS regression spec-
results from the lower-income group and younger age group are natu- ifications (corresponding to Table 3) to address extensive questions and
rally absorbed. Meanwhile, we portray the results from this section in find that: First, our fieldwork intervention, the digital platform uti-
Fig. 3 to provide supplementary information. We find that the subgroup lisation training programme, increased the treatment group’s usage of
differences in response to the treatment are not universally statistically mobile apps. Second, residents in the treatment group exhibit enhanced
significant. In Panel A, we find that the treatment shows the most sig- usage of mobile apps, as observed in obtaining market information,
nificant impact when interacting with the middle-income group. acquiring skills or undergoing online training, and engaging in social
Considering that our sample residents are already from marginalised networking or communications. Third, the evidence can elucidate the
regions, this evidence supports the validity and effectiveness of our income resilience observed among residents in the treatment group. In
intervention when delivered to low- and mid-income people. As dis- summary, we argue that our intervention enhances both the quantity
cussed in Section 3.4.3, our research is not focused on a population in and quality of residents’ usage of digital platforms, thereby contributing

Table 6
Implications in age groups and income groups.
Panel A: Interaction of income groups

Index of inclusive Information Role modelling Social Learning & Opportunity Subjective COVID-19
development & poverty /Aspiration capital capacity creation/job barriers to digital related
resilience for reduction enhance- development creation technology utilisations
residents ment

(1) (2) (3) (4) (5) (6) (7) (8)

Treatment 0.1831*** 0.1191*** 0.0318 0.0492 0.0190 − 0.0620 − 0.1166*** 0.1959***

Middle income 0.3921*** − 0.0035 − 0.1882*** − 0.0830 − 0.0819** − 0.1011** 0.1262** 0.0770**
group
Higher income 0.4962*** 0.1423*** 0.1581*** 0.1129** 0.1003** − 0.1521*** − 0.0630 0.1577***
group
Treatment× 0.2191*** 0.0330 0.1177** 0.1375** 0.1413** 0.1237** − 0.0594 − 0.1091*
Middle
income
Treatment× 0.0321 − 0.0953 − 0.0659 − 0.0093 0.0240 0.2492*** 0.0410 − 0.1368**
Higher
income
Full control Yes Yes Yes Yes Yes Yes Yes Yes
variables
Observations 2870 2870 2870 2870 2870 2870 2870 2870

Panel B: Interaction of age groups

Index of inclusive Information Role Social capital Learning & Opportunity Subjective COVID-19
development & poverty modelling enhance- capacity creation/job barriers to digital related
resilience for reduction /Aspiration ment development creation technology utilisations
residents

(1) (2) (3) (4) (5) (6) (7) (8)

Treatment 0.0881* 0.1339** 0.1087** 0.1343*** 0.1733*** 0.0481 − 0.1753*** 0.1563***

Middle age − 0.1145** 0.0269 0.1968*** − 0.0228 0.0272 − 0.0388 − 0.0402 − 0.0086
group
Older age − 0.2577** 0.0162 0.2102*** − 0.0209 − 0.0238 − 0.0337 − 0.0405 0.0385
group
Treatment × 0.0478 − 0.0291 − 0.0864 − 0.0296 − 0.1406** 0.0162 0.0834 − 0.0345
Middle age
Treatment × 0.1151 − 0.1042 − 0.0880 − 0.1422* − 0.1665** 0.0307 0.0846 − 0.1259*
Older age
Full control Yes Yes Yes Yes Yes Yes Yes Yes
variables
Observations 2871 2871 2871 2871 2871 2871 2871 2871

13
X. Fu et al. Research Policy 54 (2025) 105173

Fig. 3. Treatment effects across age and income groups.

Table 7
More and better usage of mobile Apps? 2SLS estimation results.
Panel A: First stage

The number of mobile Apps you are currently Whether the utilisation of mobile apps helps your life from the following dimensions:
using
Getting market Learning skills / receiving online Social networking /
information training communication

(1) (2) (3) (4)

Treatment 0.132** 0.148*** 0.021** 0.060***

[0.029] [0.008] [0.038] [0.000]
All control Yes Yes Yes Yes
variables
Observations 2870 2870 2867 2867

Panel B: Second stage

The number of mobile Apps you are Whether the utilisation of mobile apps helps your life from the following dimensions:
currently using

Getting market Learning skills / receiving online Social networking /

information training communication

(1) (2) (3) (4)

Index of inclusive 0.986* 0.878** 6.068* 2.169**

development &
resilience for residents [0.061] [0.031] [0.075] [0.010]
Anderson LM under- 0.000 0.000 0.000 0.000
identification
C-D-W weak-identification 0.000 0.000 0.000 0.000
Sargan-Hansen over- 0.000 0.000 0.000 0.000
identification
All control variables Yes Yes Yes Yes
Observations 2870 2870 2867 2867

14
X. Fu et al. Research Policy 54 (2025) 105173

to their income resilience during the global crisis. methodology and empirical sections. We find supportive evidence that
corresponds to the mechanisms which have been empirically discussed.
5. Qualitative analysis For example, the Haate Haat App provides a platform for marginalised
people to access information, get connections and solve their business or
To supplement the empirical evidence of the RCT approach, we daily life problems:
conducted a follow-up qualitive research by interviewing a group of I was doing vegetable cultivation in front of my home and the vegetables
stakeholders including six Haate Haat App users, three of our fieldwork were attacked by some insects. I uploaded the picture in the App and one
operators, and one Haate Haat App manager. These users were randomly person from another district suggested that I can use these sorts of pesticides
selected from those who participated in the endline survey; the field- […] these insects be removed. (Haate Haat user 1, countrywoman).
work operators were randomly selected from those SBK nation-wide Also, marginalised communities lack financial resources to access the
Tech Hub staff who were involved in the fieldwork. This effort is in formal market and conduct commercial activities, they even lack access
line with a stream of recent research that uses qualitative evidence to to traditional E-commerce as it requires financial investment and storage
support the empirical discussions and to overcome the limitations of the capacity. In contrast, the Haate Haat App creates profit opportunities
RCT approach (i.e., Bakhshi et al., 2015; Lashitew et al., 2019). Here we with only minimal capital investment, suggesting a financial con-
use in-depth semi-structured interviews to qualitatively examine the straints’ reduction function of the Haate Haat App’s inclusive business
mechanisms and supplement our theoretical arguments and RCT find- model (As discussed in Section 2.2.1). For example, a craftsman user
ings with detailed and vivid cases. The average length for each interview said:
was about 40 min. The Haate Haat App was designed in Bengali (Ban- Before it normally took a month to sell my handicrafts. But now when I
gladesh’s local language) and users largely use Bengali as their daily use the App, I can sell within a week as more people can find me. Now I can
language. There were English-Bangladeshi translators who attended the sell four products in a month. (Haate Haat user 2, craftsman).
interviews when needed. The respondents were generally asked what We find rural people are more interested in our App. In the rural area
they think about the Haate Haat App, give an example of their utilisation there are only two bazaars in a week, normally only Monday and Thursday.
of the Haate Haat App, how they find out about the digital literacy With the App they can sell their products other than in a fair and don’t need a
training, and the role of digital technology/digital skills for them during budget to attend the bazaar. The App also allows rural people to earn income
COVID-19 and how the RCT training sessions helped them. purely by sharing their skills. (Haate Haat App manager).
Table 8 shows the qualitative results: the representative quotes are After the Haate Haat related digital trainings, some of the girls I worked
listed in the third column with a note on their origin; the second column with started their own cooking channel. They feel it is so easy and no cost to
shows the second-order brief summary for the quote; the first column start these things, like some new sector, a kind of freelancing. (Haate Haat
shows the corresponding mechanisms as we discussed in the user 4, social worker).

Table 8
Qualitative research: quotes from interviews and corresponding mechanisms.
Corresponding mechanism 2nd order construct Representative quotes for underlying mechanisms

Information poverty Find a solution for cultivation I was doing vegetable cultivation in front of my home and the vegetables were attacked by some insects. I
reduction uploaded the picture in the App and one person from another district suggested that I can use these sorts of
pesticides […] these insects be removed. (Haate Haat user 1, countrywoman)
Get news and information Through the App I see all the information or all the news. (Haate Haat user 4, social worker)
Role modelling/aspiration Start their own cooking channels After the Haate Haat-related digital training, some of the girls I worked with started their own cooking
channel. They feel it is so easy and at no cost to start these things, like some new sector, a kind of freelancing.
(Haate Haat user 4, social worker)
Get aspiration of doing business (Before the Haate Haat-related training) I never, ever thought that I could be an entrepreneur or realise how
to start a business. (Haate Haat user 6, Female entrepreneur)
Entrepreneur training and skills Empowerment means independence and freedom to me. The Haate Haat App campaigns give out sort of
transfer entrepreneur training and skills transfer and other information through our App as an online platform.
(Haate Haat fieldwork operator 1)
Social capital enhancement Promote connections Haate Haat App helps to connect the dots and they have many opportunities by this. (Haate Haat user 4,
social worker)
Improves self-confidence Better digitalisation training with confidence improves people’s confidence. (Haate Haat fieldwork operator
3)
Learning & capability Realise the economic usefulness of After the training, I realised that social media is not just for fun. It can also be very useful or helpful for your
development social media economy’s growth. (Haate Haat user 6, Female entrepreneur)
Digital literacy for employment We are teaching everyone about technology. And nowadays we need to have basic digital literacy when we
go for a job. (Haate Haat fieldwork operator 2)
Improves independence The digital trainings make me feel independent. […] (Haate Haat user 5, ICT teacher)
Opportunity creation/job Create opportunities from off-bazaars We find rural people are more interested in our App. In rural areas there are only two bazaars in a week,
creation and lower the cost normally only Monday and Thursday. With the App, they can sell their products other than at a fair and don’t
need a budget to attend the bazaar. (Haate Haat App manager)
Sell handcraft quickly Before it normally took a month to sell my handcrafts. But now when I use the app, I can sell within a week as
more people can find me. Now I can sell four products in a month. (Haate Haat user 2, craftsman)
Opportunity for poor people Without the Haate Haat App some poor women that I am working with would actually be unable to start.
(Haate Haat user 4, social worker)
Subjective barriers to digital Realise the helpfulness and reduce It is indeed harmful if they are addicted to the time spent on mobile phones. We give them various kinds of
technology the harmfulness training like values, basic training, ethics […] now the users or the kids’ guardians realise that it will
improve their child and it will be helpful for them. (Haate Haat fieldwork operator 3)
COVID-19 related Better utilisation of digital After the training, we met Covid19. This is when I realised that I could work from home with the help of
utilisations technology during Covid-19 digital technology. (Haate Haat user 6, Female entrepreneur)
Digitalisation is our weapon during Covid19, it enables us to continue what we do during the pandemic and
quarantine. (Haate Haat fieldwork operator 3)
Be more prepared for the post-covid The training we delivered during Haate Haat App campaigns made them more prepared for the digital world
digital world. – the post-COVID digital world. (Haate Haat App manager)

Note: quotes may implicate multiple mechanisms, here we only choose the most representative link between the quotes and mechanisms.

15
X. Fu et al. Research Policy 54 (2025) 105173

This girls’ cooking channel story not only shows the financial con- contributes to inclusiveness for marginalised communities. Finally, our
straints’ reduction benefit of this content-based digital business model, fieldwork study overlaps the period of the COVID-19 pandemic. This
but also the benefit of lowering the skill and organisational barriers for unique setting enables us to examine and argue the resilience by which
marginalised people to access the mainstream economy and create their digital technology empowers marginal communities during a crisis and
own values by making contents. economic hardship.
Findings from our qualitative interviews also show how digital This study has important policy implications. First, our study pro-
technology and our digital literacy training played a role during a global vides RCT-based rigorous findings suggesting the social and economic
crisis such as the COIVD-19 pandemic. A female entrepreneur said: benefits from access to content-based digital platform combined with
After the training we met Covid-19. This is when I realised that I can work training on a mobile phone-based App. Therefore, policies should be
from home with the help of digital technology. (Haate Haat user 6, Female introduced to encourage and support the development and adoption of
entrepreneur). content-based digital platforms, considering the inclusiveness and eco-
In summary, the qualitative results are consistent with and supple- nomic benefits they bring to marginalised communities. Providing tar-
ment our RCT results. The research examines the mechanisms suggested geted support for marginalised communities and enabling them to
in our theoretical analysis and confirms the impacts of the intervention participate in the content-based income creation activities will include
on inclusive development and resilience in the marginalised commu- some of the otherwise left-out communities being included in the digital
nities during a global crisis. economy.
Our study also presents a case of cost-effective fieldwork imple-
6. Conclusions and policy implications mentation. As outlined in the baseline results, an initial investment of
£50 k during our intervention yielded an overall treatment effect of
In this paper, we analyse the role of emerging digital technology £150 k. We anticipate that the experience of our project will encourage
(content-based digital platforms) in inclusive development through the policymakers to explore the cost-effectiveness of implementing similar
lens of inclusive innovation. Through an RCT in Bangladesh, we provide digital training programmes on a larger scale.
measurable evidence to examine the working mechanisms and the real Secondly, policies shall be introduced to establish a “content-based
income impact for marginal communities when promoting the uti- digital platform usage training” scheme to raise awareness of the ben-
lisation of content-based digital platforms to marginalised communities. efits of new forms of digital technology, and help people living in remote
We argue that access to digital technology, especially content-based communities make greater use of them, identify job opportunities,
digital platforms, empowers marginalised communities through the develop capabilities, create new income sources, and receive
mechanisms of information poverty reduction; role modelling/aspira- information.
tion; social capital enhancement; learning, freedom to choose and thus Thirdly, governments shall design bespoke training programmes for
capability development; and opportunity creation/ job creation. Mean- young people and women a priority, especially for communities in
while, exposure to relevant digital skills’ training programmes would remote regions. These measures would lead to increased entrepreneur-
significantly help marginalised communities reduce their subjective ship and better employment for marginalised people, and subsequently
barriers to digital technology and improve utilisations during economic lead to income growth and greater resilience in marginalised commu-
hardships. These factors significantly enabled the marginalised com- nities, both in normal times and during a global crisis such as a
munity to use and benefit from digital technology and participate in pandemic.
economic activities through a digital platform technology-based new Finally, policies shall be introduced to incentivise the platform op-
content-based business model. As an outcome, proper intervention fos- erators and ecosystem players to enhance the job and knowledge in-
ters resilience in marginalised communities during a crisis such as the formation provision on the platform and nurturing positive role models
COVID-19 pandemic and alleviates their economic hardship. Mean- on the platform as these are also found to be effective transmission
while, our program has a more pronounced effect on enhancing mechanisms that lead to inclusive development and resilience.
employment and entrepreneurial outcomes, particularly at the house- There are several potential confounding effects or mechanisms that
hold level, which is typically the unit accessing digital platform content. may influence or affect our results, warranting a detailed discussion.
This evidence suggests the validity and necessity of targeted policy First is the competition or network effect. Treated individuals may gain
intervention to provide inclusive digital technology trainings. an advantage in competition over non-treated individuals at a certain
Our findings contribute to multiple strands of literature. Firstly, we stage, potentially introducing an upward bias to the treatment effect.
provide evidence and representative fieldwork intervention on how to This effect has been observed due to partial equilibrium during the RCT
materialise the economic benefits of emerging digital technology to design (e.g., Burtless, 1995). Second is the peer learning effect within
marginalised communities. This effort corresponds to the SDGs (UN, households, as members in larger families are more likely to learn from
2015) and to the study of inclusive development (UNDP, 2022) from the each other, suggesting that individuals are not independent during
perspective of solving “information poverty” (Britz, 2004; Lam and Lee, observation. However, our concern on this point is likely reduced, given
2006). Secondly, we provide timely empirical evidence that examines that the difference in household size between the control and experi-
the mechanisms through which content-based digital platform mental groups is only 0.12 and is statistically insignificant. Second is the
technology-based inclusive innovation promotes inclusive development peer learning effect within households. In larger families, members are
(Helfat and Raubitschek, 2018; Fu et al., 2022). Our findings have broad more likely to learn from each other, suggesting that individuals are not
implications of innovation policy for grassroot entrepreneurship and independent during observation. This also raises concerns about the
social empowerment for marginalised communities. Thirdly, this study statistically different number of household members between the
is in line with the literature of “capability approach” (Sen, 1990; treatment and control groups, albeit with a difference of only 0.12
Robeyns, 2009), that highlights the importance of making development persons. Third, the COVID-19 pandemic has created a unique context
as the capability and freedom to choose for people to create their own where residents increasingly rely on digital means. This may amplify
value. We argue that promoting inclusive access to digital technology observed treatment effects and make it difficult to achieve similar
and popularising the content-based business model would shed light on magnitudes of outcomes in other settings. Finally, due to the nature of
the practice of these theories, as an inclusive innovation to “leave no one our experiment, it was challenging to fully dissect the mechanism
behind” (George et al., 2012; Hall et al., 2012; Verstegen et al., 2019; Fu, behind income changes in residences, especially since we were unable to
2020). Meanwhile, this experimental study provides quantitative evi- conduct a social desirability bias check or involve third parties for more
dence which empirically argues the mechanisms through which the representative verification. Considering the above concerns, we are
content-based business model, enabled by digital technology, cautious in asserting that our experimental results can be expected to be

16
X. Fu et al. Research Policy 54 (2025) 105173

observed with comparable magnitudes when involving a broader focused on a purposely marginal society-oriented digital platform, for
segment of the population or in similar experiments conducted in which we designed the “Haate Haat” APP based on a successful App
different settings. “Kwai” and managed the App through a local non-profit NGO. This may
In addition to the above potential limitations of this study, there are hamper our implications for commercial practices. Moreover, an in-
also potential selection bias. As a “strategy field experiment”, there is a depth case study should be carried out to fully understand the impact
potential non-compliance to contend with (Chatterji et al., 2016) during of digital platform participation through role models and aspirations, on
our RCT experiment. Not all who were randomised to the treatment women and youths. Finally, it is worth noting the limitation in the
group probably attended or used the app and some who were assigned to research design which makes it difficult to disentangle the impact of the
the control group probably did actually use the app platform so there are availability of the app platform itself from the digital skills provided in
takers and “never takers”. Hence, we are analysing a local average the training sessions. Future research should revisit the project by car-
treatment effect (LATE), rather than a population average treatment rying out another survey of the treatment and control groups after a few
effect (PATE). In other words, there could be a self-selection bias which years of active use of the App by the treatment group users to examine
results in an upward estimation of treatment effect. This is particularly the impact of active use of the App in different depths and modes.
the case in randomised controlled trials with imperfect compliance, if
treatment assignment differs from treatment delivered, the effects of CRediT authorship contribution statement
random assignment are called intention-to-treat (ITT) effects (Angrist
and Pischke, 2014). This may also explain why age effects were not Xiaolan Fu: Writing – review & editing, Writing – original draft,
statistically significant in our study. Although our endline survey covers Supervision, Project administration, Methodology, Investigation,
93.3 % households and 91.7 % respondent individuals compared with Funding acquisition, Formal analysis, Data curation, Conceptualization.
the baseline survey, and their attrition rates are small and similar be- Pervez Ghauri: Writing – review & editing, Writing – original draft,
tween control and treatment groups (2–3 % difference), given the Supervision, Investigation, Funding acquisition, Formal analysis,
possible ITT effects, our conclusions should be drawn with some caution. Conceptualization. Jianan Lu: Writing – review & editing, Writing –
Finally, there are also challenges of generalizing findings to other original draft, Visualization, Methodology, Formal analysis, Data
contexts, especially considering the unique circumstances of the COVID- curation.
19 pandemic, which we should treat with caution. The COVID-19
pandemic and the mobility restrictions introduced as a public health Declaration of competing interest
policy by the government made people’s reliance on digital platform
greater than in normal circumstances. Hence, the role of digital tech- The authors declare that they have no known competing financial
nology and skills training may be bigger than normal situation. There- interests or personal relationships that could have appeared to influence
fore, while the pandemic provided a unique context for us to understand the work reported in this paper.
the role of digital technology and related training during a global crisis,
on the other hand, we need to be cautious in drawing generalisable Acknowledgements
conclusions, and future research shall study what transmission mecha-
nisms are still significant and what mechanisms’ impact increase or The authors are grateful to Torbjorn Fredriksson, Sonia Kabir, Elvis
decrease in normal times comparing with during a global pandemic. Avenyo; Shaheen Akter; Md. Sadique Rahman, and participants at the
Moreover, the pandemic caused uncertainties for the fieldwork, and IDMODEL project conferences at the University of Oxford, 6th Sino-UK
lockdown policies varied over time across different cities. We admit that Innovation and Development Forum, 3rd Sustainable Development
these difficulties during the pandemic resulted in our planned return Conference (Michigan), 2024 America Economic Association Annual
visits not being systematic. Conference, Royal Economic Society 2024 Annual Conference, and SBS-
The future research should examine how regulations should be TMCD joint workshop participants at the University of Oxford for
introduced to protect consumers’ rights and data privacy while not helpful comments, to SBK foundation to support the fieldwork inter-
hindering digital innovation; and how the perceived barriers can be vention and data collection, and to the Economic and Social Research
overcome by training and information provision. Also, our research Council (ESRC) for financial support (no. ES/S001336/1).

Appendix A

Appendix Table A1
Statistical summary.

Variable N Mean SD Min Max

Respondent’s age - baseline 2974 35.84 13.46 18 100

Respondent’s gender 2974 0.476 0.500 0 1
Respondent’s education level 2972 9.432 4.337 0 16
Household head: Female 2973 0.070 0.256 0 1
Household head: Marital status 2974 1.182 0.645 1 4
Household head: Religion 2974 1.042 0.205 1 3
Household members: Total 2974 4.112 1.303 1 10
Household members: Minors 2974 1.038 0.946 0 5
Household income (BDT) - baseline 2974 410,835 447,802 0 5.11e+06
Household income (BDT) - endline 2974 254,783 220,870 0 2.50e+06
Personal income (BDT) - baseline 2974 139,095 258,507 0 4.80e+06
Personal income (BDT) - endline 2873 91,689 159,685 0 2.50e+06
Currently employed - baseline 2971 0.486 0.500 0 1
Currently employed - endline 2871 0.401 0.490 0 1
(continued on next page)

17
X. Fu et al. Research Policy 54 (2025) 105173

Appendix Table A1 (continued )

Variable N Mean SD Min Max

Maintained business operation or created new business last year 2974 0.151 0.358 0 1
Household income resilience during COVID-19: Resilience index 2974 0 1 − 10.50 3.269
Individual income resilience during COVID-19: Resilience index 2873 0 1 − 19.15 5.749
Household income decrease during COVID− 19 2974 156,052 4.15e+05 − 1.20e+06 4.51e+06
Individual income decrease during COVID-19 2873 46,994 2.17e+05 -1.20e+06 4.20e+06
Index of inclusive development & resilience for residents 2871 0 1 − 15.23 5.453
Information poverty reduction 2871 0 0.720 − 1.220 1.160
Role modelling /aspiration 2871 0 0.720 − 1.427 2.609
Social capital enhancement 2871 0 0.720 − 3.151 0.668
Learning & capability development 2871 0 0.720 − 0.507 3.014
Opportunity creation/ job creation 2871 0 0.720 − 0.150 9.207
Subjective barriers to digital technology 2871 0 0.720 − 0.956 3.229
COVID-19 related utilisations 2871 0 0.720 − 0.832 3.415

Appendix Table A2
Balance check for pre-RCT difference in the baseline survey.

Overall sample Control group Treatment group Control-Treatment

Difference

Obs. Obs. Mean Obs. Mean Coeff. p-value

(1) (2) (3) (4) (5) (6) (7)

Age 2974 1413 35.903 1561 35.785 0.1169 0.8132

Gender 2974 1413 0.4848 1561 0.4689 0.0159 0.3875
Head age 2974 1413 26.507 1561 27.204 − 0.6976 0.1270
Household members 2974 1413 4.1776 1561 4.0532 0.1244 0.0093
Household incomes last year 2974 1413 416,723.5 1561 405,504.8 11,218.7 0.4952
Electricity access 2963 1411 0.9901 1552 0.9942 − 0.0041 0.2017
Uploaded videos before 2969 1408 0.0952 1561 0.1121 − 0.0169 0.1314
Know how to uploads videos 2969 1408 0.6443 1561 0.6522 − 0.0079 0.7539
Usage of Apps: Social networking/communication 2970 1409 0.4222 1561 0.4798 − 0.0007 0.9559
Usage of Apps: Increase income 2970 1409 0.0220 1561 0.0212 0.0008 0.6008
Barriers to access digital technology: Lack digital skills 2970 1409 0.2703 1561 0.2792 − 0.0088 0.9994
Barriers to access digital technology: language literacy 2970 1409 0.1008 1561 0.1006 0.0002 0.8753
Barriers to access digital technology: general skills 2970 1409 0.1228 1561 0.1441 − 0.0213 0.9560
Barriers to access digital technology: infrastructures 2970 1409 0.1008 1561 0.1006 − 0.0228 0.9402

Appendix Table A3
Survey response bias check.

Baseline survey (Oct 2019) Endline survey (Apr 2021) Attrition rate Attritions (completion) in comparable research

Observed households 1183 1104 6.7 % Banerjee et al. (2007): 17 % and 18 %

Observed respondents 3244 2974 8.7 % Duflo et al. (2011): 24.9 %
Control group respondents 1562 1413 9.5 % Banerjee et al. (2015): 25 %
Treatment group respondents 1682 1561 7.2 % De Mel et al. (2014): 584 out of 624
Eiteneyer et al. (2019): 942 out of 15,414

Appendix Table A4
Multivariate regression analysis: Treatment as dependent variable.

Treatment (Dependent Var) Variance Inflation Factors 1/ VIF

Respondent’s age - baseline − 0.001 1.32 0.756399

[0.570]
Respondent’s gender − 0.008 1.81 0.553564
[0.822]
Respondent’s education level − 0.000 1.16 0.865415
[0.964]
Household head: Female 0.058 1.04 0.961125
[0.237]
Number of the household members − 0.023** 1.19 0.843060
[0.036]
Created new business last year 0.070* 1.04 0.963884
[0.056]
Household income last year − 0.000 2.15 0.464410
[0.947]
(continued on next page)

18
X. Fu et al. Research Policy 54 (2025) 105173

Appendix Table A4 (continued )

Treatment (Dependent Var) Variance Inflation Factors 1/ VIF

Individual income last year − 0.000 2.73 0.366301

[0.265]
Currently employed - baseline − 0.051 2.36 0.423258
[0.197]
Currently employed - endline 0.036 2.74 0.364956
[0.393]
Constant 0.654***
[0.000]
Observations 2869 Mean VIF 1.84
Adjusted R2 0.037
Note: VIF indicates Variance Inflation Factors according to O’Brien (2007), which method argues that a value of 3 or lower indicates the reliance of
regression specification.

Appendix Table A5
Survey items and corresponding measurements.

Household Member ID 1 2 3 4 Corresponding variable in regression models

A. Does your usage of mobile applications (Apps) over the last 12 months benefit or improve the following aspects of your life? Answer: (1, not useful 2, little useful 3, somewhat useful,
4 useful, 5 very useful)
A.1.1 Access to information Information poverty reduction
A.1.2 Knowledge sharing
A.1.3 Access to news
A.2.1 Getting business ideas (role modelling) Role modelling /Aspiration
A.2.2 Receiving government information
A.2.3 Getting Entrepreneurship skills
A.3.1 Social networking /communication Social capital enhancement
A.3.2 Trust each other
A.4.1 Learning new skills for work Learning & capability development
A.4.2 Language learning
A.4.3 Capability development
A.5.1 Opportunity to collaborate with others users/ Opportunity creation/ job creation
sellers
A.5.2 Opportunity to collaborate with potential
distributor?
A.5.3 Opportunity to change a job/ career

B. Do the following serve as barriers to your usage of mobile Apps? Answer: (1, not useful 2, little useful 3, somewhat useful, 4 useful, 5 very useful)
B.1 Lack of motivation/need Subjective barriers to digital technology
B.2 Lack of digital skills

C. What specific activities did you or will you perform through mobile apps due to the pandemic situation? Answer: (1, not useful 2, little useful 3, somewhat useful, 4 useful, 5 very
useful)
C.1 COVID-19 information & guidelines COVID-19 related utilisations: COVID-19 information and health advice; mental support during
C.2 COVID-19 Health advice quarantine
C.3 Other health advice & online treatment
C.4 Express feelings and life
C.5 Mental support
Note: to observe the real-income impact of respondents through intervention, we exclude retired respondents for these sets of questions.

Appendix Table B1
Stratified and randomly selected wards.

City Total marginalised wards Ward number in our RCT

Dhaka South 34 3
Dhaka North 23 6
Bogura 21 9 and 17
Total 78 4

19
X. Fu et al. Research Policy 54 (2025) 105173

Appendix Table B2
List of marginalised wards across the three selected cities.

City Zone Ward number City Zone Ward number

Dhaka North 2 2 Dhaka South 3 28

Dhaka North 2 3 Dhaka South 3 29
Dhaka North 2 4 Dhaka South 4 32
Dhaka North 2 5 Dhaka South 4 33
Dhaka North 2 6 Dhaka South 4 34
Dhaka North 2 7 Dhaka South 4 35
Dhaka North 2 8 Dhaka South 4 36
Dhaka North 2 15 Dhaka South 4 37
Dhaka North 3 21 Dhaka South 4 38
Dhaka North 3 22 Dhaka South 4 42
Dhaka North 4 9 Dhaka South 4 43
Dhaka North 4 10 Dhaka South 5 39
Dhaka North 4 11 Dhaka South 5 40
Dhaka North 4 12 Dhaka South 5 41
Dhaka North 4 13 Dhaka South 5 44
Dhaka North 4 14 Dhaka South 5 45
Dhaka North 4 16 Dhaka South 5 46
Dhaka North 5 28 Dhaka South 5 47
Dhaka North 5 29 Bogura – 1
Dhaka North 5 30 Bogura – 2
Dhaka North 5 31 Bogura – 3
Dhaka North 5 32 Bogura – 4
Dhaka North 5 33 Bogura – 5
Dhaka South 2 1 Bogura – 6
Dhaka South 2 2 Bogura – 7
Dhaka South 2 3 Bogura – 8
Dhaka South 2 4 Bogura – 9
Dhaka South 2 5 Bogura – 10
Dhaka South 2 6 Bogura – 11
Dhaka South 2 7 Bogura – 12
Dhaka South 2 8 Bogura – 13
Dhaka South 2 11 Bogura – 14
Dhaka South 3 14 Bogura – 15
Dhaka South 3 22 Bogura – 16
Dhaka South 3 23 Bogura – 17
Dhaka South 3 24 Bogura – 18
Dhaka South 3 25 Bogura – 19
Dhaka South 3 26 Bogura – 20
Dhaka South 3 27 Bogura – 21

20
X. Fu et al. Research Policy 54 (2025) 105173

Appendix Fig. A1. Maps of wards across fieldwork areas.

Data availability Anderson, S.J., McKenzie, D., 2022. Improving business practices and the boundary of
the entrepreneur: a randomized experiment comparing training, consulting,
insourcing, and outsourcing. J. Polit. Econ. 130 (1), 157–209.
Data will be made available on request. Angrist, J.D., Pischke, J.S., 2014. Mastering Metrics: The Path from Cause to Effect.
Princeton University Press.
Ansari, S., Garud, R., 2009. Inter-generational transitions in socio-technical systems: the
References
case of mobile communications. Res. Policy 38, 382–392.
Armanios, D.E., Eesley, C.E., Li, J., Eisenhardt, K.M., 2017. How entrepreneurs leverage
Adermon, A., Hensvik, L., 2022. Gig-jobs: stepping stones or dead ends? Labour Econ. 76, institutional intermediaries in emerging economies to acquire public resources.
102171. Strateg. Manag. J. 38 (7), 1373–1390.
Allcott, H., 2015. Site selection bias in programme evaluation. Q. J. Econ. 130 (3), Arora, S., Romijn, H., 2011. The empty rhetoric of poverty reduction at the base of the
1117–1165. pyramid. Organization 19 (4), 481–505.
Almourad, M.B., Alrobai, A., Skinner, T., Hussain, M., Ali, R., 2021. Digital wellbeing Bakhshi, H., Edwards, J.S., Roper, S., Scully, J., Shaw, D., Morley, L., Rathbone, N., 2015.
tools through users lens. Technol. Soc. 67, 101778. Assessing an experimental approach to industrial policy evaluation: applying RCT+
Anand, R., Mishra, M.S., Peiris, M.S.J., 2013. Inclusive Growth: Measurement and to the case of creative credits. Res. Policy 44 (8), 1462–1472.
Determinants. International Monetary Fund. Banerjee, A., Duflo, E., 2011. Poor Economics: A Radical Rethinking of the Way to Fight
Global Poverty. Public Affairs, New York.

21
X. Fu et al. Research Policy 54 (2025) 105173

Banerjee, A., Duflo, E., Glennerster, R., Kinnan, C., 2015. The miracle of microfinance? Freeman, C., 1994. ’The diffusion of information and communication Technology in the
Evidence from a randomised evaluation. Am. Econ. J. Appl. Econ. 7 (1), 22–53. World Economy in the 1990s. In: Mansell, R. (Ed.), The Management of Information
Banerjee, A.V., Cole, S., Duflo, E., Linden, L., 2007. Remedying education: evidence from and Communication Technologies: Emerging Patterns of Control. Aslib - The
two randomised experiments in India. Q. J. Econ. 122 (3), 1235–1264. Association for Information Management, London, pp. 8–41.
Banerjee, A.V., Duflo, E., 2007. The economic lives of the poor. J. Econ. Perspect. 21 (1), Freeman, C., 2007. ’The ICT paradigm’. In: Mansell, R., Avgerou, C., Quah, D.,
141–168. Silverstone, R. (Eds.), The Oxford Handbook of Information and Communication
Bardhan, I., Whitaker, J., Mithas, S., 2006. Information technology, production process Technologies. Oxford University Press, Oxford, pp. 34–54.
outsourcing, and manufacturing plant performance. J. Manag. Inf. Syst. 23 (2), Fu, X., 2020. Innovation under the Radar: The Nature and Sources of Innovation in
13–40. Africa. Cambridge University Press, Cambridge.
Batistatou, E., Roberts, C., Roberts, S., 2014. Sample size and power calculations for trials Fu, X., Akter, S., 2016. The impact of mobile phone technology on agricultural extension
and quasi-experimental studies with clustering. Stata J. 14 (1), 159–175. services delivery: evidence from India. J. Dev. Stud. 52 (11), 1561–1576.
Belzunegui-Eraso, A., Erro-Garcés, A., 2020. Teleworking in the context of the Covid-19 Fu, X., Avenyo, E., Ghauri, P., 2021. Digital platform and development: a survey of
crisis. Sustainability 12 (9), 3662. literature. Innov. Dev. 11 (2–3), 303–321.
Blinder, A.S., 1973. Wage discrimination: reduced form and structural estimates. J. Hum. Fu, X., Ghauri, P., Ogbonna, N., Xing, X., 2022. Platform-based business model and
Resour. 436–455. entrepreneurs from base of the pyramid. Technovation 119, 102451.
Bloomfield, B.P., Coombs, R., 1992. Information technology, control and power: the Fu, X., Shi, L., 2022. The direction of innovation in developing countries. WIPO working
centralisation and decentralisation debate revisited. J. Manag. Stud. 29 (4), paper no. 69. Available at: https://www.wipo.int/publications/en/details.jsp?
459–484. id=4596&plang=EN.
Bohnsack, R., Kurtz, H., Hanelt, A., 2021. Re-examining path dependence in the digital Gallivan, M.J., Spitler, V.K., Koufaris, M., 2005. Does information technology training
age: the evolution of connected car business models. Res. Policy 50 (9), 104328. really matter? A social information processing analysis of coworkers’ influence on IT
Bratberg, E., Grasdal, A., Risa, A.E., 2002. Evaluating social policy by experimental and usage in the workplace. J. Manag. Inf. Syst. 22 (1), 153–192.
nonexperimental methods. Scand. J. Econ. 104 (1), 147–171. George, G., McGahan, A.M., Prabhu, J., 2012. Innovation for inclusive growth: towards a
Britz, J.J., 2004. To know or not to know a moral reflection on information poverty. theoretical framework and a research agenda. J. Manag. Stud. 49 (4), 661–683.
J. Inf. Sci. 30 (3), 192–204. Ghauri, P., Fu, X., Minayora, A., 2022. Digital technology-based entrepreneurial pursuit
Bruton, G.D., Ketchen Jr., D.J., Ireland, R.D., 2013. Entrepreneurship as a solution to of the marginalised communities. J. Int. Manag. 28 (2), 100948.
poverty. J. Bus. Ventur. 28 (6), 683–689. Ghauri, P., Strange, R., Cooke, F.L., 2021. Research on international business: the new
Burtless, G., 1995. The case for randomised field trials in economic and policy research. realities. Int. Bus. Rev. 30 (2), 101794.
J. Econ. Perspect. 9 (2), 63–84. Glennerster, R., Takavarasha, K., 2013. Running Randomised Evaluations. Princeton
Canale, N., Marino, C., Lenzi, M., Vieno, A., Griffiths, M.D., Gaboardi, M., Giraldo, M., University Press, In Running Randomised Evaluations.
Cervone, C., Massimo, S., 2022. How communication technology fosters individual Golinelli, D., Boetto, E., Carullo, G., Nuzzolese, A., Landini, M., Fantini, M., 2020.
and social wellbeing during the COVID-19 pandemic: preliminary support for a Adoption of digital technologies in health care during the COVID-19 pandemic:
digital interaction model. J. Happiness Stud. 23 (2), 727–745. systematic review of early scientific literature. J. Med. Internet Res. 22 (11), e22280.
Chakravorty, B., Bhatiya, A.Y., Imbert, C., Lohnert, M., Panda, P., Rathelot, R., 2023. González-Uribe, J., Leatherbee, M., 2018. The effects of business accelerators on venture
Impact of the COVID-19 crisis on India’s rural youth: evidence from a panel survey performance: evidence from start-up Chile. Rev. Financ. Stud. 31 (4), 1566–1603.
and an experiment. World Dev. 168, 106242. González-Uribe, J., Reyes, S., 2021. Identifying and boosting “gazelles”: evidence from
Chan, J., Wang, J., 2018. Hiring preferences in online labor markets: evidence of a business accelerators. J. Financ. Econ. 139 (1), 260–287.
female hiring bias. Manag. Sci. 64 (7), 2973–2994. Greenhalgh, T., Wherton, J., Shaw, S., Morrison, C., 2020. Video consultations for Covid-
Chan, S.C., Ngai, E.W., 2007. A qualitative study of information technology adoption: 19. Br. Med. J. 368, m998.
how ten organisations adopted web-based training. Inf. Syst. J. 17 (3), 289–315. Hall, J., Matos, S., Sheehan, L., Silvestre, B., 2012. Entrepreneurship and innovation at
Chataway, J., Hanlin, R., Kaplinsky, R., 2014. Inclusive innovation: an architecture for the base of the pyramid: a recipe for inclusive growth or social exclusion? J. Manag.
policy development. Innov. Dev. 4, 33–54. Stud. 49 (4), 785–812.
Chatterji, A., Delecourt, S., Hasan, S., Koning, R., 2019. When does advice impact startup Hanushek, E. and L. Woessmann. 2020. The economic impacts of learning losses. OECD
performance? Strateg. Manag. J. 40 (3), 331–356. education working papers, no. 225, OECD publishing, Paris, https://doi.org/10.1
Chatterji, A.K., Findley, M., Jensen, N.M., Meier, S., Nielson, D., 2016. Field experiments 787/21908d74-en.
in strategy research. Strateg. Manag. J. 37 (1), 116–132. Helfat, C.E., Raubitschek, R.S., 2018. Dynamic and integrative capabilities for profiting
Chun, H., Kim, J.W., Lee, J., 2015. How does information technology improve aggregate from innovation in digital platform-based ecosystems. Res. Policy 47 (8),
productivity? A new channel of productivity dispersion and reallocation. Res. Policy 1391–1399.
44 (5), 999–1016. Iyer, S., Kitson, M., Toh, B., 2005. Social capital, economic growth and regional
Couture, V., Faber, B., Gu, Y., Liu, L., 2021. Connecting the countryside via e-commerce: development. Reg. Stud. 39 (8), 1015–1040.
evidence from China. American Economic Review: Insights 3 (1), 35–50. Kelley, E.M., Ksoll, C., Magruder, J., 2024. How do digital platforms affect employment
Dammert, A.C., Galdo, J., Galdo, V., 2015. Integrating mobile phone technologies into and job search? Evidence from Inda. J. Dev. Econ. 166, 103176.
labor-market intermediation: a multi-treatment experimental design. IZA J. Labor Knack, S., Keefer, P., 1997. Does social capital have an economic payoff? A cross-country
Dev. 4 (1), 11. investigation. Q. J. Econ. 112 (4), 1251–1288.
De Haan, A., 2001. Social Exclusion: Enriching the Understanding of Deprivation. Paper Koo, W.W., Eesley, C.E., 2021. Platform governance and the rural–urban divide: Sellers’
prepared for the World Bank Report Forum on Inclusion, Justice and Poverty responses to design change. Strateg. Manag. J. 42 (5), 941–967.
Reduction. Kotha, R., Vissa, B., Lin, Y., Corboz, A.V., 2023. Do ambitious entrepreneurs benefit more
De Mel, S., McKenzie, D., Woodruff, C., 2014. Business training and female enterprise from training? Strateg. Manag. J. 44 (2), 549–575.
start-up, growth, and dynamics: experimental evidence from Sri Lanka. J. Dev. Econ. Lal, K., 1999. Determinants of the adoption of information technology: a case study of
106, 199–210. electrical and electronic goods manufacturing firms in India. Res. Policy 28 (7),
Dimitriadis, S., Koning, R., 2022. Social skills improve business performance: evidence 667–680.
from a randomized control trial with entrepreneurs in Togo. Manag. Sci. 68 (12), Lam, J.C. and Lee, M.K. 2006. Digital inclusiveness–Longitudinal study of Internet
8635–8657. adoption by older adults. J. Manag. Inf. Syst., 22(4), pp. 177–206.
Dinda, S., 2014. Inclusive growth through creation of human and social capital. Int. J. Lambrecht, A., Tucker, C., 2019. Algorithmic bias? An empirical study of apparent
Soc. Econ. 41 (10), 878–895. gender-based discrimination in the display of STEM career ads. Manag. Sci. 65 (7),
Dosi, G., Soete, L., 2022. On the syndemic nature of crises: a Freeman perspective. Res. 2966–2981.
Policy 51 (1), 104393. Lashitew, A.A., van Tulder, R., Liasse, Y., 2019. Mobile phones for financial inclusion:
Duflo, E., Kremer, M., Robinson, J., 2011. Nudging farmers to use fertiliser: theory and what explains the diffusion of mobile money innovations? Res. Policy 48 (5),
experimental evidence from Kenya. Am. Econ. Rev. 101 (6), 2350–2390. 1201–1215.
Dupont, W.D., Plummer Jr., W.D., 1990. Power and sample size calculations: a review Lee, J.N., Morduch, J., Ravindran, S., Shonchoy, A., Zaman, H., 2021. Poverty and
and computer programme. Control. Clin. Trials 11 (2), 116–128. migration in the digital age: experimental evidence on mobile banking in
Dutt, N., Hawn, O., Vidal, E., Chatterji, A., McGahan, A., Mitchell, W., 2016. How open Bangladesh. Am. Econ. J. Appl. Econ. 13 (1), 38–71.
system intermediaries address institutional failures: the case of business incubators Lee, J.Y., Yang, Y.S., Ghauri, P., Park, B.I., 2022. The impact of social media and digital
in emerging-market countries. Acad. Manag. J. 59 (3), 818–840. platforms experience on SME international orientation: the moderating role of
Eesley, C., Wang, Y., 2017. Social influence in career choice: evidence from a randomized COVID-19 pandemic. J. Int. Manag. 28 (4), 100950.
field experiment on entrepreneurial mentorship. Res. Policy 46 (3), 636–650. Leite, A.C.C., 2019. China and global value chains: globalisation and the information and
Eesley, C., Wu, L., 2020. For startups, adaptability and Mentor network diversity can be communications technology sector. Reg. Stud. 53 (1), 156–157.
pivotal: evidence from a randomized experiment on a MOOC platform. MIS Q. 44 Leten, B., Belderbos, R., Looy, B.V., 2016. Entry and technological performance in new
(2). technology domains: technological opportunities, technology competition and
Eiteneyer, N., Bendig, D., Brettel, M., 2019. Social capital and the digital crowd: technological relatedness. J. Manag. Stud. 53 (8), 1257–1291.
involving backers to promote new product innovativeness. Res. Policy 48 (8), Li, L., Su, F., Zhang, W., Mao, J.Y., 2018. Digital transformation by SME entrepreneurs: a
103744. capability perspective. Inf. Syst. J. 28 (6), 1129–1157.
Elbers, C., Gunning, J.W., 2014. Evaluation of development programmes: randomised London, T., Hart, S.L., 2004. Reinventing strategies for emerging markets: beyond the
controlled trials or regressions? World Bank Econ. Rev. 28 (3), 432–445. transnational model. J. Int. Bus. Stud. 35, 350–370.
Faulkner, W.N., 2014. A critical analysis of a randomized controlled trial evaluation in Lyons, E., Zhang, L., 2018. Who does (not) benefit from entrepreneurship programmes?
Mexico: norm, mistake or exemplar? Evaluation 20 (2), 230–243. Strateg. Manag. J. 39 (1), 85–112.

22
X. Fu et al. Research Policy 54 (2025) 105173

Mansell, R., 2021. Adjusting to the digital: societal outcomes and consequences. Res. Robeyns, I., 2006. The capability approach in practice. J Polit Philos 14 (3), 351–376.
Policy 50 (9), 104296. Robeyns, I., 2009. Capability Approach. Edward Elgar Publishing, In Handbook of
Marler, J.H., Liang, X., Dulebohn, J.H., 2006. Training and effective employee economics and ethics.
information technology use. J. Manag. 32 (5), 721–743. Rodima-Taylor, D., Grimes, W.W., 2019. International remittance rails as infrastructures:
McCambridge, J., Witton, J., Elbourne, D.R., 2014. Systematic review of the Hawthorne embeddedness, innovation and financial access in developing economies. Rev. Int.
effect: new concepts are needed to study research participation effects. J. Clin. Polit. Econ. 26 (5), 839–862.
Epidemiol. 67 (3), 267–277. Saldanha, T.J., John-Mariadoss, B., Wu, M.X., Mithas, S., 2021. How information and
McKenzie, D., 2021. Small business training to improve management practices in communication technology shapes the influence of culture on innovation: a country-
developing countries: re-assessing the evidence for ‘training doesn’t work’. Oxf. Rev. level analysis. J. Manag. Inf. Syst. 38 (1), 108–139.
Econ. Policy 37 (2), 276–301. Sen, A., 1981. Ingredients of famine analysis: availability and entitlements. Q. J. Econ. 96
McKenzie, D., Puerto, S., 2021. Growing markets through business training for female (3), 433–464.
entrepreneurs: a market-level randomized experiment in Kenya. Am. Econ. J. Appl. Sen, A., 1990. Development as capability expansion. The Community Development
Econ. 13 (2), 297–332. Reader 41, 58.
Miletzki, J., Broten, N., 2017. An Analysis of Amartya Sen’s: Development as Freedom. Sen, A., 1995. Inequality Reexamined. Harvard university press.
Macat Library. Sen, A., 2000. Development as Freedom. Oxford University Press, New Delhi.
Molony, T., 2009. Carving a niche: ICT, social capital, and trust in the shift from personal Shortall, S., 2008. Are rural development programmes socially inclusive? Social
to impersonal trading in Tanzania. Inf. Technol. Dev. 15 (4), 283–301. inclusion, civic engagement, participation, and social capital: exploring the
Nakano, Y., Magezi, E.F., 2020. The impact of microcredit on agricultural technology differences. J. Rural. Stud. 24 (4), 450–457.
adoption and productivity: evidence from randomised control trial in Tanzania. Si, S., Hall, J., Suddaby, R., Ahlstrom, D., Wei, J., 2022. Technology, entrepreneurship,
World Dev. 133, 104997. innovation and social change in digital economics. Technovation 119, 102484.
Nambisan, S., Wright, M., Feldman, M., 2019. The digital transformation of innovation Simanis, E., Hart, S., 2009. Innovation from the inside out. MIT Sloan Manag. Rev. 56
and entrepreneurship: Progress, challenges and key themes. Res. Policy 48 (8), (4), 77–86.
103773. Şimşek, T., Öner, M.A., Kunday, Ö., Olcay, G.A., 2022. A journey towards a digital
Norris, P., 2001. Digital Divide: Civic Engagement, Information Poverty, and the Internet platform business model: a case study in a global tech-company. Technol. Forecast.
Worldwide. Cambridge University Press. Soc. Chang. 175, 121372.
O’Brien, R.M., 2007. A caution regarding rules of thumb for variance inflation factors. Sotiriadis, M.D., 2017. Sharing tourism experiences in social media: a literature review
Qual. Quant. 41, 673–690. and a set of suggested business strategies. Int. J. Contemp. Hosp. Manag. 29 (1),
Okeleke, K., Shah, R., Nesbitt-Ahmed, L. and Cruz, G. 2021. Achieving Mobile-Enabled 179–225.
Digital Inclusion in Bangladesh. GSMA National Dialogues. Available at: Staples, D.S., Hulland, J.S., Higgins, C.A., 1999. A self-efficacy theory explanation for the
https://www.gsma.com/mobilefordevelopment/resources/achieving-mobile-enable management of remote workers in virtual organisations. Organ. Sci. 10 (6),
d-digital-inclusion-in-bangladesh/. 758–776.
Orser, B., Riding, A., Li, Y., 2019. ’Technology adoption and gender-inclusive Tarafdar, M., Singh, R., Anekal, P., 2013. Impact of ICT-enabled product and process
entrepreneurship education and training’. Int. J. Gend. Entrep. 11 (3), 273–298. innovations at the bottom of the pyramid: a market separations perspective. J. Inf.
Oster, E., 2019. Unobservable selection and coefficient stability: theory and evidence. Technol. 28 (4), 279–295.
J. Bus. Econ. Stat. 37 (2), 187–204. Teece, D.J., 2018. Profiting from innovation in the digital economy: enabling
Pal, A., Herath, T., Rao, H.R., 2021. Why do people use mobile payment technologies and technologies, standards, and licensing models in the wireless world. Res. Policy 47
why would they continue? An examination and implications from India. Res. Policy (8), 1367–1387.
50 (6), 104228. Tenenhaus, M., Vinzi, V., Chatelin, Y.M., Lauro, C., 2005. Partial least squares.
Pallais, A., Sands, E.G., 2016. Why the referential treatment? Evidence from field Computational Statistics & Data Analysis 48 (1), 159–205.
experiments on referrals. J. Polit. Econ. 124 (6), 1793–1828. UNDP, 2022. Our focus: Inclusive growth. Available at: https://www.undp.org/eurasi
Palmié, M., Miehé, L., Oghazi, P., Parida, V., Wincent, J., 2022. The evolution of the a/our-focus/inclusive-growth (Accessed: 23 Oct 2022).
digital service ecosystem and digital business model innovation in retail: the United Nations, 2015. The 17 Goals. Available at: https://sdgs.un.org/goals.
emergence of meta-ecosystems and the value of physical interactions. Technol. Vanden Abeele, M.M., 2021. Digital wellbeing as a dynamic construct. Commun. Theory
Forecast. Soc. Chang. 177, 121496. 31 (4), 932–955.
Pansera, M., Owen, R., 2018. Framing inclusive innovation within the discourse of Vargo, D., Zhu, L., Benwell, B., Yan, Z., 2021. Digital technology use during COVID-19
development: insights from case studies in India. Res. Policy 47 (1), 23–34. pandemic: a rapid review. Human Behavior and Emerging Technologies 3 (1),
Paunov, C. 2013. Innovation and Inclusive Development: A Discussion of the Main Policy 13–24.
Issues. OECD Science, Technology and Industry Working Papers, No. 2013/01, Verstegen, L., Houkes, W., Reymen, I., 2019. Configuring collective digital-technology
OECD publishing, Paris, Available at: https://doi.org/10.1787/5k4dd1rvsnjj-en. usage in dynamic and complex design practices. Res. Policy 48 (8), 103696.
Pike, A., Rodríguez-Pose, A., Tomaney, J., 2016. Local and Regional Development. Walton, M., 2011. Culture matters for poverty, but not because of a culture of poverty:
Routledge. Notes on analytics and policy. In: Platteau, J.P., Peccoud, R. (Eds.), Culture,
Pitt, M.M., Khandker, S.R., 1998. The impact of group-based credit programmes on poor Institutions and Development. Routledge, London, pp. 195–233.
households in Bangladesh: does the gender of participants matter? J. Polit. Econ. 106 Whitelaw, S., Mamas, M.A., Topol, E., Van Spall, H.G., 2020. Applications of digital
(5), 958–996. technology in COVID-19 pandemic planning and response. The Lancet Digital Health
Prahalad, C.K., Hammond, A., 2002. Serving the world’s poor, profitably. Harv. Bus. Rev. 2 (8), 435–e440.
80 (9), 48–59. World Bank, 2022. The State of Economics: The Influence of Randomized Controlled
Ranieri, R., Almeida Ramos, R., 2013. Inclusive Growth: Building up a Concept, vol. No. Trials on Development Economics Research and on Development Policy. Available
104. Working paper. at: https://olc.worldbank.org/content/state-economics-influence-randomized-con
Reddy, S.G., 2012. Randomise this! On poor economics. Review of Agrarian Studies 2 trolled-trials-development-economics-research-and (Accessed: 23 Oct 2022).
(2369–2021–102).

23