Research Philosophy, Methodological Implications, and Research Design

Abstract

In this chapter, I explain the choices for the layers of the research approach applied in this book. Chiefly, I used a critical realist research stance and analyzed both qualitative case studies as well as survey data in a mixed methods approach. For the central qualitative research, I collected data through 81 problem-centered interviews, one focus group with 12 affected people, and discussions with over 60 experts. I analyzed the data through Qualitative Text Analysis to examine effects, mechanisms, social system dynamics, and structures. For the parallel quantitative study on the Coastal El Niño, I assessed extensive survey data through regression models. To evaluate differential displacement risk, I used a dataset collected by Peru’s National Institute of Statistics and Informatics directly after the disaster with close to 190,000 affected adults spread across all of Peru. Additionally, to identify the effects of displacement on well-being, I applied a customized, merged dataset of that survey and the National Census collected later in the same year. The chapter discusses the used data as well as the strengths and limitations of all chosen methods.

In this chapter, I explain the choices for the four different layers of the research approach applied in this dissertation, which are illustrated in Figure 3.1 (Saunders et al. 2011). I used a critical realist research stance, applied retroduction and abduction as modes of reasoning, and analyzed both qualitative case studies and survey data in a mixed methods approach.

Figure 3.1

Layers of research and application in this study. (Note: Blue boxes indicate the choices made in this study. Reproduced from Saunders et al. (2019: 130) and edited by the author.)

1 Critical Realist Stance and Implications

First, the philosophical (or research) stance as the outermost layer refers to assumptions about the nature of reality (ontology), valid knowledge and knowing (epistemology), and values and aims of research (axiology). Few migration studies explicate their philosophical stance (Castles 2012; Iosifides 2012). This dissertation is based on critical realism, which combines a realist ontology and relativist epistemology. This research philosophy has gained a standing in social sciences and serves this study for four reasons (Maxwell & Mittapalli 2010). First, its ontology allows for a complex analysis of why well-being changes in (im)mobilities occur, and what role structure and agency play therein. Second, its epistemology favors diversity in research perspectives, methods, and data, which is useful for examining alternative explanations of well-being impacts. Third, the stance makes it possible to incorporate the role of the human mind into research, for example, the insight that biases may lead migrants to misinterpret their well-being situations. Finally, critical realism has a strong value orientation, which is, in my view, essential when studying well-being in the context of climate injustices and in an unequal society such as Peru (Iosifides 2012; Maxwell 2012; World Bank 2021b).

Second, philosophic stances come with different approaches to theories or modes of reasoning. Theories are “the analysis and statement of how and why a set of facts relates to each other” (Kumar 2011: 21). Approaches to theories refer to different mental operations to construct order and logic in data and to connect data with theory. The critical realist goal is to develop hypothetical models for the mechanisms and structures behind empirically observed phenomena and build theories of them with multiple viewpoints (Lawson et al. 2009). Realism permits both using existing theories and provides guidance on how theory can be developed. Induction (going from data to broader theory) and deduction (testing theory-derived hypotheses with data) serve as the foundation for abduction and retroduction (Hartwig 2007).¹ In other words, I will “continue to ask the question why?” (Easton 2010: 124), use counterfactual thinking, study extreme or surprising cases, and compare cases to identify generative mechanisms (Danermark et al. 2002).

Third, methodology applies the research stance and modes of reasoning systematically to the research (Castles 2012). It discusses how scholars can retrieve and produce knowledge about the social world and why which type of methods can provide valid data (Teddlie & Tashakkori 2010). Given the critical realist premises, both quantitative and qualitative research methods can produce valid knowledge under certain conditions (Iosifides 2012; Maxwell & Mittapalli 2010).² Given these complementarities of qualitative and quantitative approaches, it is methodologically sensible to combine them (Sayer 1992; Seawright 2016). Qualitative methods can discern social action, intentions, and meanings around (im)mobilities and well-being. They can address context, complexity, and diversity, and shed light on generative mechanisms. Conversely, quantitative methods are valuable to systematically inquire diversity and regularities in well-being effects; to scale and compare scales; to measure the strength of influences; and to test and refine hypotheses about mechanisms. Various prior mixed methods studies have used critical realism (Shannon-Baker 2016) and several authors have called for mixed methods to study migration (Castles 2012; Iosifides 2012, 2017). I explain the research design in detail in section 3.2.

Before, I close with discussing how the critical realist value base shapes this research. Critical realist research aims at reducing domination and expanding freedom or flourishing (Maxwell 2012).³ It is based on a similar argument as critical social theories that theory should serve emancipation and not mere knowledge creation (Horkheimer 1982; Lawson et al. 2009). I agree that in a world where migration offers opportunities for few wealthier people while many marginalized groups confront restrictions and control, “realist explanatory critiques of social relations of injustice and of their effects and consequences are urgently needed” (Iosifides 2012: 47). All knowledge generation is a social practice with impacts; it should aim to inform those affected by domination and inequality to empower them in their struggles for self-determination. In this study, I attempt to do so by revealing structures of domination, control, oppression, and exclusion before and after people leave areas facing climate hazards; how these structures shape the uneven distribution of opportunities to migrate in the first place and under humane conditions; and how they shape chances to preserve well-being. I also attempt to expose mechanisms behind well-being changes of migrants and stayers, and how climate (im)mobilities modify, reduce, reproduce, or reinforce such structural inequalities. In doing so, I refrain from dominant discourses of managing and controlling migrants.

Finally, I also attempt to approach the subjects of inquiry (self-)critically. Producing knowledge is a social practice shaped by politics, power, and by researchers themselves. Evaluating knowledge requires awareness that it is produced by communication, which, in turn, typically occurs in unequal social settings that favor certain narratives. My own values, socialization, and biases can have influenced this research. As a relatively young, male, white academic from the Global North, my socialization is different from that of most interviewees. I interviewed people of all different ages, also much older ones; of different ethnicities and religions; as well as with a different upbringing and socioeconomic situation. While I have studied and lived in Latin America, speak Spanish, and prepared scientifically and culturally for the fieldwork, these differences have shaped interviews, the analysis, and interpretations. I am aware that relationships with the respondents were often unequal. Lastly, while I have tried to be as impartial as possible, I acknowledge that all collected data on empirical events linked to concepts like (im)mobilities and well-being are value-laden and do not represent one objective truth.

2 Mixed Methods Research Design

After having explained the three outer layers of this research approach, I turn to discuss the concrete choices for the research design and methods in the next section. Methods are the procedures and practices chosen to collect and analyze data, as justified by the methodology (Castles 2012).

To study the well-being impacts of climate (im)mobilities in Peru in this dissertation, I use an ex-post-facto, convergent parallel design with qualitative methods weighted more heavily than quantitative ones (Creswell & Clark 2017). An ex-post design is appropriate here given the absence of experimental options, which could have reduced the influence of unobserved third factors, such as self-selection of migrants (McKenzie et al. 2010; McKenzie & Yang 2010; Stillman et al. 2015). This choice was most realistic for the time and resource horizon of this study and is in line with prior guidance for studies in this research field (Banerjee et al. 2013; Melde et al. 2017; Milan & Gioli 2015). I partially address the lack of experimental setup through method and results triangulation.

The chosen critical realist stance favors mixed methods approaches, which bring several benefits for studying climate (im)mobilities and well-being. Foremost, social science studies apply mixed methods to use strengths of both qualitative and quantitative strands while reducing their individual limitations (Kelle 2014; Teddlie & Tashakkori 2010). Beyond, mixed methods allow testing whether both components produce convergent results (corroboration); shedding light on respective blind spots (completeness); and raising the integrity of findings (credibility) (Bryman 2006; Kelle 2014; Schoonenboom & Johnson 2017; Tashakkori & Teddlie 2010). These advantages lead eminent scholars like Stephen Castles to argue that “most forms of migration research are likely to require ‘mixed-methods approaches’” (2012: 21; Fauser 2018).

In mixed methods designs, qualitative and quantitative strands can be weighted differently and integrated at different points (Kelle 2014; Schoonenboom & Johnson 2017). In this study, I prioritized the qualitative component due to its unique adeptness to assess the meaning of people’s climate-related experiences in the social world (Nature Climate Change 2021). For comparability, I conducted the same qualitative methods in all three large zones of Peru (highlands, rainforest, and coast). Moreover, data and time constraints allowed for one additional quantitative analysis of the Coastal El Niño case. I performed both components concurrently but separately to preserve data independence and triangulation options, and integrated them later through meta-inferences (Tashakkori & Teddlie 2010). This approach is coined convergent parallel or parallel mixed design (Creswell & Clark 2017; Schoonenboom & Johnson 2017). Figure 3.2 provides an overview of the applied research design.

Figure 3.2

Overview of the applied mixed methods design. (Note: Created by the author)

In this paragraph, I briefly outline the applied methods before I explain them in detail below. I started the central qualitative research with a review of the evidence (see chapter 4). Afterwards, I collected data through 81 problem-centered interviews (Witzel & Reiter 2012), one focus group with 12 affected people (Morgan 1999b), and discussions with over 60 experts. Next, I analyzed the data through Qualitative Text Analysis to examine effects, mechanisms, social system dynamics, and structures (Kuckartz & Rädiker 2019). For the parallel quantitative study on the Coastal El Niño, I assessed extensive survey data through regression models. To evaluate differential displacement risk, I used a dataset collected by Peru’s National Institute of Statistics and Informatics (INEI)⁴ directly after the disaster with close to 190,000 affected adults spread across all of Peru. Additionally, INEI on request created a customized, merged dataset of that survey and the National Census collected later in the same year, which I analyzed to identify the effects of displacement on well-being.

2.1 Qualitative Methods

I mainly used qualitative methods to analyze affected people’s narratives on the experienced well-being changes and underlying mechanisms of action. I collected data during several weeks of research in the Peruvian communities of interest during three visits in 2018 and 2019. Another scheduled visit in 2020 to present results and liaise with stakeholders was held virtually due to COVID-19 restrictions. The collected data included (a) problem-centered interviews with 81 (36 m / 45f) migrants and family members to explore their perceptions on hazards and well-being impacts of (im)mobilities (Witzel & Reiter 2012); (b) one focus group with 12 (3 m / 9f) pupils in a sending community to cover an important group underrepresented in the interviews (Vogl 2014); and (c) more than 60 discussions with experts such as policy makers, researchers, and practitioners to gain background insights into structural conditions that shape well-being effects (Gläser & Laudel 2010; Helfferich 2014).

The qualitative strand is case-oriented and uses the comparative method for “rich descriptions of a few instances” of typical cases of villages of departure or immobility and areas of arrival, focusing on “context, complexity and difference” in the chosen cases (Della Porta 2008: 216, 221). The dense knowledge created in this small-N case comparison is useful for discovering well-being effects and mechanisms. While the three cases in Peru are distinctively configured in space and time, the knowledge gained in these in-depth studies can help to build more generalized concepts “that transcend the validity of individual cases” (Della Porta 2008: 206). I explain the site selection below.

2.1.1 Site Selection

I collected data from Peru’s three major regions to cover the following cases (Figure 3.3):

• (1) Long-distance rural-to-urban migration from two villages in the highlands of the Lima Region and immobility in these areas, influenced by gradual glacier recession and rainfall changes;

• (2) short-distance, attempted planned relocation (community-wide migration) of two villages in the rainforest Region of San Martín due to abrupt floods, resulting in entrapment and only one eventual relocation; and

• (3) short-distance displacement (acute, forced migration) from several villages in the coastal Region of Piura, forced mainly by abrupt flooding.

Figure 3.3

The three Regions for the qualitative data collection in Peru. (Note: The map on the left displays Peru’s location in Latin America, the one on the right the Regions within Peru where qualitative data was collected. Created by the author using paintmaps.com © and mapchart.net © and edited subsequently)

Figure 3.4 below specifies the distribution of these villages across Peru’s three large natural zones.

Figure 3.4

Sites for qualitative data collection across Peru’s three large regions. (Note: To protect the respondents, the pins indicate approximate locations only. Created by the author, based on CIA (1970))

I selected the areas of origin of migrants (and the homes of stayers) with a view to match three criteria:

• (1) Rural villages with similar, locally typical subsistence livelihood systems that

• (2) have experienced impacts of water-related climate hazards typical for Peru’s three large topographical zones (highlands, rainforest, and coast), which

• (3) have influenced (im)mobilities in forms characteristic for these hazards, but varied across cases, resulting in diverse well-being conditions.

First, I selected areas with livelihoods—and by extension with (im)mobility patterns—susceptible to climate hazards. The chosen villages primarily use ecosystem-based livelihoods and are typically home to smallholder subsistence farmers with low levels of income, education, and health, who tend to be among the groups most vulnerable to climate impacts (Cohn et al. 2017; Donatti et al. 2019; Niles & Salerno 2018). Selecting villages with these similar livelihood features reduced the number of confounding variables and facilitated better insights into well-being mechanisms; nonetheless, even similar villages are never the same and keeping all contextual variables constant is impossible.

Second, I chose home villages of migrants and stayers affected by either gradual or abrupt water-related hazards, which were either directly related to climate change or provided temporal analogs. To begin with, I set the focus on water (and related hazards) because it is one of Peru’s adaptation priorities in its Nationally Determined Contributions (NDCs) and National Adaptation Plan (NAP) (GoP 2015; MINAM 2021), while global reviews highlight its role in climate (im)mobilities (Nagabhatla et al.; Wrathall et al. 2018). Next, the ex-post design required to select areas where people could notice physical (for example, glacier retreat) or temporal effects of hazards (for example, changes in rainfall timing) which influence (im)mobilities (Laczko & Aghazarm 2009). I selected three hazard dynamics that the systematic review for this study identified as the most typical influences on (im)mobility patterns in Peru’s three large topographical zones: glacier recession (alongside rainfall changes) in the highlands (Sierra); floods in the rainforest (Selva); and El Niño events in the coastal zone (Costa) (Bergmann et al. 2021a; see also reviews in results chapters 5–7). On the one hand, I selected Sierra villages harmed by gradual hazards directly attributable to climate change, namely glacier recession (Seehaus et al. 2019) and changes in the rainfall regime (Heidinger et al. 2018). Studies demonstrate that both such glacier retreat (e.g. Alata et al. 2018; Altamirano Rua 2021; Figueiredo et al. 2019; Heikkinen 2017; Wrathall et al. 2014) and rainfall changes (e.g. Hook & Snyder 2021; Lennox 2015; Milan 2016; Milan & Ho 2014) can alter migration in the Sierra. On the other hand, I chose villages affected by two types of abrupt hazards for which climate change attribution is not as clear, but which provide temporal analogs for future climate impacts.⁵ To begin with, I selected two Selva villages harmed by floods, which periodically affect (im)mobilities in this region (e.g. Hofmeijer et al. 2013; Langill 2018; List 2016; Sherman et al. 2016). When habitability is threatened, the state has occasionally attempted to relocate entire communities (Bernales 2019; Desmaison et al. 2018; Estrada et al. 2018; Lopez 2018; Pittaluga 2019). While extreme floods have already increased in the Selva (Barichivich et al. 2018; Gloor et al. 2013; Marengo & Espinoza 2016), it remains unclear how much more likely climate change made the specific floods analyzed in this study. Yet, given that extreme floods have increased in this region overall, and climate change is projected to raise them further (Duffy et al. 2015; Langerwisch et al. 2013; Zulkafli et al. 2016), the cases do provide valuable insights into a dynamic with increasing importance. Moreover, I selected sites on the Costa harmed by the 2017 Coastal El Niño (CEN) floods. Peru’s coast is periodically affected by severe flooding due to El Niño events (Sanabria et al. 2018), which are among the main drivers of acute migration in this zone (Bayer et al. 2014; Ferradas 2015; French & Mechler 2017; Venkateswaran et al. 2017). Climate change made the specific 2017 CEN analyzed here at least 1.5 times more likely (Christidis et al. 2019). Even independently of the exact climate attribution for this event, the analysis of the 2017 CEN sheds light on a type of phenomenon that Peru will face more often due to climate change (Cai et al. 2015; IPCC 2019a; Peng et al. 2019). (Lastly, choosing one case per zone also did justice to Peru’s diverse topography and made the findings relevant for national policymakers, who typically think in these boundaries.)

Third, I selected departure and arrival points of diverse spatial and temporal forms of migration to observe varied conditions for well-being changes. Migration was either propelled suddenly (coast and rainforest) or driven over longer time frames (highlands), as shaped by the abrupt and gradual hazards discussed above. Moreover, I sought to investigate various forms of (im)mobilities along the spectrum of more voluntary (some cases from the highlands) and forced instances (highlands, coast, and rainforest). I also chose (im)mobilities involving different numbers of people, from individuals to households (highlands and some coastal cases) and entire communities (coast and rainforest). These choices intended to satisfy quality criteria for case selections (Gerring & Cojocaru 2016; Seawright & Gerring 2008).⁶

The local partners facilitating the selection of cases included the Mountain Institute for the highlands; San Martín’s Regional Office of Security and National Defense and the Peruvian National Center for Disaster Risk Estimation, Prevention and Reduction (CENEPRED) for the rainforest cases; as well as Caritas and the student group CIMA at the University of Piura for the coast.⁷ Gaining access to the research sites and subjects is a key task of empirical research, and these partners allowed me to enter the villages together with local experts who had known the respondents for years. This approach is common in studies on hard-to-reach migrant populations (Bloch 2007; Ho & Milan 2012). Once the sites were determined, sampling and interviewing followed to gather the qualitative data.

2.1.2 Data Collection

The analytical units were individual migrants and members of migrant households who either accompanied these migrants or stayed at home (stayers). I targeted the heads of migrant households, and occasionally additional household members like spouses, to gain insights into their experiences related to hazards, (im)mobilities, and well-being. For families of migrant members who had moved away, I attempted to interview the new head of household in the village of origin.

I used non-probabilistic, iterative sampling orientated at contrasts, which some authors coin as theoretical sampling. I selected this strategy to systematically contrast cases and reveal themes, connections, and divergences; to compare the mechanisms which express themselves in the different cases; and to illustrate the diversity of well-being constellations, similar as in grounded theory (Corbin & Strauss 2014; Przyborski & Wolhrab-Sahr 2014; Strübing 2014).⁸ After interviews, I iteratively read through notes to find incipient patterns and themes around well-being effects and mechanisms, which guided the selection of new interviewees until returns of further interviews diminished and saturation was reached, which was the case after 81 interviews.⁹ The sampling differed slightly in the three cases. Migrants from the villages in the Selva and Costa moved in large clusters and over short distances, so that they could be readily tracked in destinations. Accompanied by local partners, I spent several days in these sites and went from home to home to select and interview migrants until saturation was reached. By contrast, sampling longer-distance migrants from the Sierra required two steps. I started by interviewing households in the Andean home villages affected by hazards, and then used snowball (or chain referral) sampling to trace migrants in urban areas.¹⁰ Regarding destinations, I focused on Junín’s Regional capital Huancayo and the national capital Lima for two reasons. First, interviewees in the villages observed that these were the main destinations. Second, both cities featured migrant hometown associations from the Province of origin, which organized events that offered chances to meet migrants. I conducted all interviews in Spanish without interpreters. As all inhabitants in the study areas spoke Spanish, no exclusions due to language had to be made.

For conducting the interviews, techniques with varying premises exist (Hopf 2015; Lamnek & Krell 2016). Broadly speaking, they are either like structured mining for information or narrative travelling (Kvale & Brinkmann 2009: 48–50).¹¹ I decided that combining structured and narrative interviewing served the research interest here best for two reasons. First, it puts researchers in an active position so they can use scientific research knowledge to structure key topics in the interview. Yet, second, it does not limit the proper local perspectives of respondents or impede the chance of discovering novel aspects. To this end, I used elements of the problem-centered method (Kurz et al. 2000; Witzel & Reiter 2012),¹² which brings together the knowledge of the researcher and respondents in a dialogue. Interviewees are competent (but partially biased) insider experts of their lives. Researchers enter as well-informed travelers with scientific knowledge to openly learn, and at once, to assist in reconstructing the meaning of the insider knowledge regarding the research interests.

Accordingly, a prerequisite for this research was compiling information on the interviewee’s living conditions. I had gathered this knowledge in a preliminary sensitizing framework that defined the direction of interest and initial priorities. Later, during the interviews, I assessed and situated new empirical observations by continuously mentally referring to this knowledge. Based on the framework, I developed a topical guide with a road map of key interview topics (Figure 3.5 and Electronic Supplementary Material). The guide provided structure and enabled me to re-center on the research interest during interviews, although the relevance and sequence of topics depended on respondents’ accounts and the guide was adjusted to new data received. In this way, the guide also ensured comparability across interviews by establishing similar topical complexes in each dialogue.

Figure 3.5

Topical guide and topical complexes. (Note: Created by the author)

Conducting the interview proceeded in several stages (Witzel & Reiter 2012). Bearing in mind that the questions were personal and partially sensitive, I left it to the respondents to decide on a setting in which they felt most comfortable to speak (and which still permitted decent recording). Often, we spoke at their homes but when outside, I asked to talk at a small distance from other people (Figure 3.6). Afterwards, a warming up phase with informal conversations with respondents followed to build a relationship. Then, I briefly explained the research project and answered initial questions. Afterwards, I provided an introductory explanation for the interview, including ethical and data protection information as well as a request for permission to record (see Electronic Supplementary Material). Opening questions followed to facilitate narrative accounts by the respondents; they prompted interviewees to tell me the story of how their lives and well-being had changed since they had migrated or stayed. These narrative accounts provided cues for the follow-up conversation on well-being effects and their causes. Next, I asked follow-up questions to encourage additional narrative accounts and to stimulate self-reflection, sporadically providing imaginative prompts or pre-interpretations. I also used strategies to improve understanding where suitable. When topics from the topical guide were omitted, I asked ad-hoc questions on them, usually toward the end. Closing the interview involved various steps. First, I collected data on age, gender, livelihoods, occupation, and other factors to compare profiles. The recordings stopped here. Second, I debriefed respondents and thanked them for the insights shared. I invited final questions or thoughts and provided information on how to contact me. Third, after leaving the interview site, I wrote postscripts that captured key information for self-debriefing, as sketches of the interviews with first interpretations and cues that would later support the analysis of the data.

Figure 3.6

Photo of an interview with an affected farmer. (Note: Photo taken by colleagues from the Mountain Institute)

Besides individual interviews, I convened one focus group with adolescents, as they were previously underrepresented in the data (Figure 3.7). This method brings together people from a target group to engage in a moderated discussion and interaction, which provides different types of insights than individual interviews (Krueger & King 1999; Morgan 1999b). Twelve pupils (3 m / 9f) aged 14 to 16 years old participated. The sampling was purposive: through local partners in the school, pupils in the final classes before graduating from school—and thus facing the decision whether to stay or migrate—were invited. Questions followed the topical guide for the interviews in a discreetly structured approach. I allowed participants to open their own directions but also applied moderation tools to refocus group dynamics on the research interests. To this end, I used a funnel approach, moving from initially broader, open-ended questions encouraging narration to the central topics, and finally, to specific questions on the research interests (Krueger 1999; Morgan 1999a).

Figure 3.7

Photo of the focus group with pupils in a study site in the highlands. (Note: Photo by the author)

The charts in Figure 3.8 below summarize key data of the 93 affected people. The tables in the respective results chapters 5–7 provide information disaggregated by regions. They illustrate that while most interviewees were at working age, I also covered younger and older groups. Women are slightly overrepresented in the data. While most respondents were mestizo, I was able to sample one indigenous village. Primarily, most interviewees worked in agriculture, and almost all households were agricultural. Finally, across Peru’s three large zones, I interviewed similar shares of migrants, displaced persons, relocatees and those trapped but aspiring to relocate, as well as other stayers.

Figure 3.8

Qualitative data profiles of 93 affected people. (Note: The graphs illustrate the profiles of 81 interviewees and 12 focus group participants. Created by the author)

I also conducted discussions with experts for background context on the larger structural factors and processes behind the well-being effects of (im)mobilities in Peru. I identified the experts through desk research and referral from authorities, civil society, and international organizations working on related topics. They included experts at higher state levels, such as staff in national ministries, and at the local level, such as village heads. In total, I discussed with more than 60 policy makers, officials, practitioners, academics, and activists working in diverse entities (Figure 3.9).

Figure 3.9

Experts consulted across administrative scales and fields of expertise. (Note: Boxes colored beige indicate discussions with experts from the Costa, gray from the Sierra, and green from the Selva. V1 and V2 = village 1 and village 2 in the Sierra; H and L = Huancayo and Lima; V3 and V4 = village 3 and 4 in the Selva; LP and UP = Lower and Upper Piura on the Costa. Created by the author)

Discussions with experts are not a method as such; rather, they are defined by the target group of respondents, namely experts (or key informant), and their special knowledge, position, and access to information about climate change, (im)mobilities, and well-being (Witzel & Reiter 2012). While the interviews with affected people aimed at distilling their subjectivity, discussions with experts intended to find more neutral views on the effects of (im)mobilities held by people who are not research objects themselves (Bogner & Menz 2009; Gläser & Laudel 2010; Helfferich 2014).¹³ To this end, I used elements of the problem-centered method (Witzel & Reiter 2012).¹⁴ These discussions fed into the analysis via field notes taken and were not recorded or transcribed.

2.1.3 Transcription and Text Analysis

The next step for analyzing the information contained in the recorded interviews with affected people was transcribing them into text. Transcription is an integral part of qualitative analysis processes because it requires selective decisions that imply a first sampling and analysis of the oral material, and results in interpretive constructions (Davidson 2009; Kvale 2007; Sandelowski 1994; Wellard & McKenna 2001). To guarantee careful transcription, the EPICC project at PIK hired a Peruvian student assistant who typed the Spanish transcriptions manually. I provided the assistant with detailed notation, confidentiality, and data protection instructions as well as information on the study purpose, as recommended by the literature (Stuckey 2014; Wellard & McKenna 2001). The transcriptions are based on intelligent verbatim guidelines, with cues of some nonverbal behavior, an approach which can increase reliability, dependability, and trustworthiness of the results (Easton et al. 2000; Stuckey 2014). In this way, the assistant only discreetly adjusted information for readability, without changing the core of what was said. Finally, the assistant proofread all transcripts and I checked and listened to some of the transcribed tapes for quality control (MacLean et al. 2004). After transcription, I deleted any data that could identify the interviewee, such as names, workplaces, and specific positions (Stuckey 2014). The transcription guidelines are in the Electronic Supplementary Material.

Amon the many approaches used for analyzing qualitative data (Flick 2009; Gläser & Laudel 2010; Mayring 2014), I selected thematic and evaluative Qualitative Text Analysis (QTA) after Kuckartz (2010, 2014b; Kuckartz & Rädiker 2019) as the central method for analyzing the transcribed interviews in this dissertation. Thematic (or content-related) analysis enables “identifying, systematizing, and analyzing topics and subtopics and how they are related”, while evaluative analysis is about “assessing, classifying, and evaluating content” (Kuckartz 2014b: 68). I used this combination to understand factual changes in well-being as well as underlying processes.¹⁵ The QTA followed a five-step approach with reference to the research questions (Kuckartz 2014b) (Figure 3.10).

Figure 3.10

The analytical process of Qualitative Text Analysis. (Note: Reproduced and edited by the author, based on Kuckartz (2014b: 40))

Using MaxQDA software, first I added several variables to the cases for comparative analysis later (age; gender; interview site; occupations; and (im)mobility status). Then, I systematically read entire interviews with a view to understanding their meanings for the research questions (Kuckartz 2014b).

Afterwards, I created a combination of thematic and evaluative categories in a mixed, concept- and data-driven approach.¹⁶ In a first step, I derived concept-driven, thematic and evaluative categories and sub-categories from the research questions, central concepts, theories, and topical guide in this study. For example, for categories on objective well-being, I adjusted and extended previous findings from ressearch with deprived groups in Peru (Copestake 2008c) (see section 2.3). Initial categories also evolved from the topical interview guide, for example, on migration capabilities, aspirations, and drivers. The coding started with these categories. Second, while coding the first 30% of all interviews, I added new, data-driven categories using a subsumption strategy (Kuckartz 2014b; Mayring 2010): I probed all text step by step to find new topics around the research questions. Then I subsumed aspects already covered by existing categories under those. Finally, I created new (sub-)categories for new aspects. For evaluative categories (such as well-being changes in health), I defined three ordinal levels: positive/improving, neutral, or negative/deteriorating. While coding the first 30% of the material, I also adjusted the concept-driven categories as needed. Third, I compiled all text segments for each category, developed category definitions and anchor examples (and differentiation from other codes, where needed), and fixed the category system. Finally, I used this system to code the whole material. The category system is detailed in the Electronic Supplementary Material.

Subsequently, I used three tools to analyze the data based on these categories (Kuckartz 2014b; Kuckartz & Rädiker 2019). First, I focused on topics and sub-topics, analyzing each main category regarding what was discussed and what was omitted or evaded, as well as what tendencies and singularities emerged across cases. I thereby aimed to account for the criticism that QTA tends to overstress frequently mentioned topics, reproduce mainstream and dominant narratives, and suppress or deny other contents and their absence (George 1959). Second, I examined relationships between main categories and their sub-categories. For example, I analyzed how well-being components within the category development from a secure base (livelihoods, education, health and food security) related to each other, and also how this main category related to the other three main categories. Third, I examined trends across groups, for example by comparing views of people engaging in varied types of (im)mobilities, driven by either abrupt or gradual hazards. Building on these tools, I drew conclusions on the research questions and identified new questions arising from the analysis.

2.1.4 Ethical Considerations and Data Protection

Research with human subjects must address ethical challenges (Friedrichs 2014), especially when asking migrants or stayers, some of whom in vulnerable situations, about sensitive topics (van Iiempt & Bilger 2012). Ethics require taking responsibility for the researchers’ actions as well as providing accountability and redress options (Dench et al. 2004). The principle of “do no harm” is key for qualitative studies, which imply personal and little standardized interactions. Guidelines and regulations commonly highlight the Belmont principles.¹⁷ The German Professional Association of Sociologists and the American Sociological Association share similar criteria (Friedrichs 2014).

To comply with these standards, I asked respondents for their written informed consent to participate in interviews (see Electronic Supplementary Material).¹⁸ Further, I explained which information would be collected and how it would be used. I also detailed the research procedures and products as well as related potential benefits and risks. As migration research often influences real policies, scholars need to be aware of possible impacts on their respondents and reflect on which data truly needs to be collected (van Iiempt & Bilger 2012).¹⁹ Next, prior to the interviews, I explained that confidential information would be treated as such, and that the data would not be used in ways that could compromise respondents. I stressed that I would never reveal people’s clear names or the names of their hometowns, and since I interviewed respondents from small settlements, I carefully assessed if they could be identified despite the deletion of these names. In the analysis, I use a numbering system (for example, V1–4 for respondent 4 from village 1) and broad categories (such as age group) to refer to interviewees. Then, I asked for written permission to record, transcribe, and use the information academically. Finally, I restricted access to recordings and transcripts to myself and the student hired for transcription, under strict data protection policies. Focus groups require equal attention to ethical principles (Morgan 1999b), especially as the one conducted here was with adolescents.²⁰ One overarching ethical challenge in the qualitative part was dealing with inequalities in the relationship with the interviewees (Lammers 2007). I, as a foreign, privileged researcher, met people in often-vulnerable situations in which power relations, hierarchies, and strong socio-economic differences were salient. I attempted to be aware of these factors to avoid that people participated against their will, for example, due to social pressure or fear of negative consequences, and bearing in mind that there might be personal reasons to participate (Glazer 1982). I emphasized that the interviews had academic character and would not entail financial compensation, which was key as many deprived respondents hoped for support.²¹ Ethical considerations also applied to the time after collecting the qualitative data. (Most of these considerations also applied to the quantitative strand discussed further below).²²

2.1.5 Limitations

The research design offered various strengths—which are discussed in the conclusions (chapter 9)—but also implied limitations. First, the site selection was strongly shaped by what local partners suggested as accessible locations. Although I chose sites representing diverse conditions, partners did not propose areas that would be too dangerous for an outsider. Thus, the study might not cover well-being processes of people in insecure vicinities. I attempted to compensate for this possible limitation through discussions with experts and the quantitative strand, which provides data for all settings.

Second, not all migrants of interest could be sampled and interviewed. For example, men, adolescents, and older adults are underrepresented in the data, and I did not interview children due to ethical concerns. In particular, the snowballing technique applied for tracing migrants from the Sierra might have created biases and blind spots (Jacobsen & Landau 2003). People without close contacts in their villages of origin were possibly not reached and respondents’ personal situations might have further shaped the reach. For example, some migrants may have declined interviews as they were either ashamed of their situations or doing so well that they did not care to spend time with an outsider. In addition, not all migrants came to hometown association meetings where most interviews took place, some possibly because they lacked money for the necessary travel or time due to their hard work. Nevertheless, snowballing was the most robust option available for the set-up of this study and built upon prior studies in this field (Koubi et al. 2016; Laczko & Aghazarm 2009). In research with hard-to-reach populations, accurate sampling frames tend to be unavailable or too expensive to create, as was the case here (Bloch 2007). In such cases, chain referral through intermediaries, service providers, and local organizations—such as the migrant hometown associations here—is common. In addition, as the new respondents often have friendly and trusted ties with the chain referrers, such sampling can build more motivation and higher response rates among otherwise hard-to-reach groups than other methods (Bloch 2007; Faugier & Sargeant 1997). Building such access and personal relationships is key for interviewing people who may be otherwise reluctant to participate and allows for an efficient use of time and resources (Atkinson & Flint 2001; Heckathorn 2002; Rodgers 2004).

Third, in some cases it was not possible to follow through with the interview techniques suggested by the problem-centered method. Respondents were often on the move or occupied, so that conversational instead of overly formalized approaches were required. Moreover, many respondents did not provide long narrative accounts in response to opening questions or further prompts, which led to some situations where question-response schemes prevailed. In addition, as interviews mostly took place in places familiar for respondents, occasionally, more people joined in and created small group discussions. These additional accounts often opened new views, but occasionally, they also changed the conversation dynamics. In such situations, social desirability, hierarchies, and fear of over-disclosure may have shaped the main respondents’ answers (Reczek 2014).

Third, I initially had envisaged more focus groups, yet time, resource, and later COVID-19 constraints impeded this goal. The focus group in the Sierra provided valuable insights and might have been usefully replicated in other settings to explore narratives of other specific groups. For example, distilling female group views would have been interesting to contrast male narratives, since gender aspects are often salient in rural areas in Peru (Milan 2016). However, with around 60% of the interviewees being women, female views are still duly accounted for. Valuable insights could also have been gained through additional focus groups with members of receiving communities or with groups divided between migrants faring better and those faring worse in destinations. I accounted for this change in plans by considering results across varied sub-groups of respondents in the analysis.

Fourth, Qualitative Text Analysis also implied certain limitations. To start with, additional coders or reviewers could have increased the reliability and quality of the category system (Kuckartz & Rädiker 2019) but were not available due to resource constraints. Beyond, QTA alone may not pierce through the surface of all interview content (Rosenthal 2018), and as a code-based analysis, it risks detaching text from the original context (Hitzler & Honer 1997). I countered this constraint by accounting for the sequential structure and Gestalt of key cases, which raised the understanding of the meaning of the texts and their contexts (Hopf 1995; Hopf & Hopf 1997; Hopf & Schmidt 1993). Finally, because I met several experts in the context of work trips for the EPICC project at PIK, some of the discussions were infused with discussions around project needs and results, which occasionally conflicted with a structured interview approach. For this reason and due to time and resource constraints, these discussions with experts were not recorded or transcribed; rather, I used notes taken from the conversations with experts mostly as contextual information for the analysis.

Lastly, while several features of the study design raised the validity of results—including in-depth interviews with affected people and triangulation with experts—findings should still be read with two limitations in mind. First, the cross-sectional data may mask longer-term changes in OWB and SWB or lagged interactions. Intergenerational and life-course views would provide additional value for time-dependent effects (Dustmann & Glitz 2011; Singh et al. 2019) and longitudinal data could provide supplementary insights (KNOMAD 2015). For example, the lack of long-term data impeded an evaluation of possible long-term, positive side-effects of the 2017 CEN on the Costa (such as more pasture, planting areas, and forests, which were witnessed in prior events (Sperling et al. 2008)), which could influence people’s well-being. Finally, for the Selva and Costa cases, limits of temporal analogs must be kept in mind, so that the results of this study may be transferable to a large degree to future El Niño events or rainforest floods, but not fully (Berrang-Ford et al. 2011; Ford et al. 2010). As just one example, governance strongly shapes the emergence of disasters (e.g. Ahrens & Rudolph 2006; UNDRR 2020) and strongly affected the well-being effects for displaced persons and relocatees in this study; however, it remains unclear how Peruvian institutions, policies, and governance may change in the future, and how these changes would affect well-being outcomes in turn.

2.2 Quantitative Methods

This section explains the data and methods used in the statistical analyses to study differential displacement risk and the effects of displacement on people’s well-being after the Coastal El Niño (CEN) floods in March 2017. I give additional details in the full empirical case study in chapter 7.

2.2.1 Data

The quantitative analyses make use of two datasets compiled by INEI. First, INEI collected data from households and public buildings in areas affected by the CEN through a survey conducted between mid-April and end of April 2017. Through this “CEN Survey”²³, it aimed to improve the understanding of damages and the characteristics of affected people, their dwellings, and public infrastructure. To gather the data, INEI asked local authorities in the 892 districts declared in a state of emergency due to the CEN (Table 3.1) to identify all affected rural villages as well as the affected blocks in urban areas, in which enumerators then recorded data from all heads of households and information about all public buildings (INEI 2017a, 2017c).²⁴ Altogether, the CEN Survey registered 398,148 persons in 199,938 dwellings and 2,615 public buildings. This analysis focuses on the 186,437 adult respondents whose homes where directly affected and experienced at least minor damages.²⁵ The extensive CEN Survey provides a first valuable data point about the most affected areas in Peru shortly after the main floods had affected Peru in March 2017.

Table 3.1 Areas for data collection in the CEN Survey

The second dataset is the Peruvian National Census 2017 (INEI 2018c),²⁶ which was by chance enumerated seven months after the CEN disaster and thus six months after the CEN Survey. To support this research, INEI searched for the 398,148 respondents of the CEN Survey of April 2017 among the 29.4 million entries of the National Census collected on the 22^nd October 2017 (INEI 2018c).²⁷ INEI found 342,009 CEN respondents in the Census (87.2%), whereas 49,933 persons (12.7%) could not be cross identified. The well-being analysis here focuses on the 186,437 adult CEN Survey respondents with affected homes, of whom 164,084 (88%) could and 22,353 (12%) could not be cross-identified in the National Census data. This attrition could be due to various reasons. For example, persons surveyed in the CEN could have passed away, moved abroad, lived in areas that could not be surveyed, or refused to cooperate in the enumeration. However, because the differences between the identified and non-identified groups are not large, they should not lead to a strong systematic attrition bias in the analyses. The summary statistics for the CEN Survey respondents with homes affected by the disaster demonstrate that the respondents who could not be identified in the National Census did not differ substantially from the cross-identified population regarding key social factors (Table 3.2). The two groups had almost identical rates of secondary education, civil status, and disabilities. In the group cross-identified in the Census, approximately five percentage points less respondents lived in small rural villages and around five percentage points more were unemployed or female compared to the non-matched group.

Table 3.2 Summary statistics on CEN Survey respondents cross-identified in the National Census and those not re-identified

2.2.2 Regression Models

These datasets were then used to analyze the research questions explained above through several regression models. The first analysis estimated how different environmental, socioeconomic, and demographic factors influenced the displacement risk of the households. Because the outcome is binary coded, the estimation was completed with logistic regression models. Model 1 considered only the influence of exogenous environmental factors, such as topographical and rainfall data.²⁸ This baseline model was then gradually extended by including further information on household composition and demographic characteristics (model 2) as well as on livelihood factors and wealth (model 3). I detail the model parameters in the empirical section 7.3.

The second analysis centered on how displacement affected people’s well-being. It started by comparing the well-being of the displaced households to those whose houses were affected but who could remain at home directly after the disaster, based on summary statistics of the CEN Survey. Because this exceptional sample covers close to the full affected population, summary statistics render robust results on people’s well-being outcomes. Then, five linear regression models were specified to explore the impact of the displacement on well-being seven months after the CEN under control of a broad set of environmental, demographic, and socioeconomic variables. The sample in this part of the analysis were the affected adult CEN Survey respondents who could be tracked in the National Census. To understand the displacement effects, a well-being index based on indicators available in the data was built, mainly using items for a space to live better and, to some degree, items for development from a secure base (see section 7.3 for details). The impacts on well-being were then analyzed through various models. Baseline model 1 comprised displacement as the only parameter. The next models added gradually more control variables for environmental factors (model 2), household composition and demographics (model 3), livelihood characteristics and wealth (model 4), and individual characteristics (model 5). The models thereby control for the potential non-randomness of the displacement risk. People do not randomly migrate or flee but factors such as age, sex, and well-being can systematically shape the probability of movement (Aksoy & Poutvaara 2021; Borjas et al. 1992; Kaestner & Malamud 2014). The controls are needed since the observed well-being outcomes might therefore not be due to the displacement itself, but due to pre-movement factors that made displacement more likely in the first place.

2.2.3 Limitations

The quantitative work allowed for a novel analysis of differential displacement risk and well-being impacts in an extensive sample of affected people from all of Peru. Thereby, the work complemented the in-depth qualitative analysis of well-being effects and mechanisms in coastal Piura usefully. Despite generating this added value, the results should be read with the following limitations in mind.

First, because data was not available for all parameters, the analyses operated with a subsample of the respondents (see section 7.3). Data was not consistently enlisted for those households whose homes had remained unaffected by the disaster. Therefore, the analyses focused on the respondents who had indicated that the disaster had affected their houses negatively, and for whom data was available. The differences in displacement risk and well-being might be even larger if compared to the unaffected. In addition, as the study excludes respondents below 18 years to avoid double counting and due to missing data, it allows insights into children’s situation by extension only.

Second, because the CEN Survey did not contain an explicit question on displacement status, the analysis is based on proxies that may be noisy. The assumption that uninhabitable homes equaled displacement (see section 7.3) is a plausible basis for the analysis. However, people with intact homes could still have fled, for example, because they were afraid of the disaster, had lost their livelihoods or health, or complied with the issued early warnings. Conversely, respondents whose homes were destroyed could still have decided to remain in place. Additionally, while the data on habitability allowed to infer that people were displaced one month after the CEN, information was missing if they had returned or remained in displacement due to the event seven months later.

Third, a possible attrition bias must be discussed for the well-being analysis because 12% of the subsample of interest was lost when merging the surveys. The remaining sample is still large, but if the attrition is not random, then the differences between the dropped-out and the remaining respondents could introduce a bias into the results and decrease the internal validity of the study (the identified relationships between variables). Yet, the summary statistics document that the differences between the remaining and the dropped-out respondents are marginal (Table 3.2). Additionally, the attrition affects the external validity less (the generalizability to the original population) as the sample still includes almost the entire possible population of the Peruvian households affected by the CEN.

Fourth, the surveys collected by INEI could not reflect the full range of well-being indicators of interest in this dissertation (see framework developed in section 2.3). Primarily, the data did not contain indicators on social relatedness and subjective well-being. While more data was available for the components development from a secure base and a space to live better, information was missing for several key subitems of these components, such as education or physical security. Therefore, the quantitative well-being analysis offers a robust indication of the life situations of a large group of affected people, but the scope of well-being which could be analyzed was limited. The qualitative analysis was a critical complement to understand the broader range of well-being changes of interest.

Fifth, there might be additional, district- or community-level factors that this analysis could not control for, but which could have influenced the well-being results. Examples include the quality of community networks and support, social participation, neighborhood infrastructure, local leadership and governance, and resource equity (Berkes & Ross 2013; Koliou et al. 2018). While the statistical analysis could not rule out indirect effects through these factors, the qualitative analysis partially compensates for this lack of data and offers insights into some of the possible influences.

Finally, the survey data offered two data points for up to seven months after the CEN, but neither allowed for insights into people’s gradual development of well-being nor into the outcomes over the long term. Given that many persons displaced by the CEN have remained in prolonged displacement (AFP 2021; IOM 2017c, 2018), it would have been interesting to see how different groups have recuperated over time, and which factors have aided or impeded recovery. The qualitative data collected one year after the Census helped to discern some of these longer-term phenomena.

Despite these limitations, the analyses of the secondary quantitative data provide extensive information on the differential displacement risk and well-being of a large group of affected people across the entire country, which usefully complements the analysis of the primary qualitative data.