Classifying climate vulnerability and inequalities in India, Mexico, and Nigeria: a latent class analysis approach

The climate crisis exacerbates social, economic, and health disparities. This study employs innovative methods to identify distinct groups affected by recent climate events. A mobile phone-based survey was conducted in April 2022 with individuals residing in multiple climate-affected states across three countries: India (n = 1020), Mexico (n = 1020), and Nigeria (n = 1021). Latent class analysis and classification and regression tree analysis were used to identify the groups most exposed to climate events, the effects and responses taken, and then to identify the characteristics associated with group membership. Effects included housing damage or lost work, while responses included actions such as borrowing money or dropping out of school. Findings revealed four distinct groups: Group 1 reported low exposure, no effects, or responses (49% of respondents in India, 43% in Mexico, and 27% in Nigeria); Group 2 experienced multiple hazards with moderate effects and some responses; Group 3 was characterized by drought exposure with more effects and responses taken; Group 4 was affected by heavy flooding and rainfall with varied effects. Notably, India had the largest proportion of respondents in Group 3 (17%), in Mexico over a quarter (29%) were in group 4, while over half of Nigerian respondents were in Group 2 (52%). Characteristics associated with membership in each group varied by country. Overall, men from rural areas with lower incomes and reliant on agriculture experienced the highest levels of exposure and vulnerability, while urban women from higher-income households were the least affected. This study underscores the importance of considering the intersectionality of risk and vulnerability when formulating policies and programs to address the impacts of climate change. Results emphasize the need for multi-sectoral policies that target the needs of different groups, to reduce inequalities and tailor to the context-specific needs of the most vulnerable people and households.

and political factors [3].Exposure to climate hazards can amplify existing vulnerabilities differentiated by demographic characteristics such as age, gender, urban residence, or type of employment.Effective climate adaptation programs and policies must account for this differential vulnerability and tailor responses to the needs of diverse contexts and populations.
When exposed to a climate hazard, individuals and households may experience various direct effects or harms, such as housing loss or damage, lost income, as well as harm to their health and wellbeing.They and their households must also make choices about how they respond.Adaptation refers to the various ways that people alter their behaviors and systems to adjust to climate change [4].This may include, for example, decisions to migrate to safer geographies, or to diversify livelihoods to protect household incomes.Ultimately, the goal is to build resilience and ensure that individuals, households, and their communities are positioned to prepare for, respond to, and recover from multi-hazard climate threats.Certain responses such as moving away from agricultural livelihoods or dropping out of school can have lifelong implications for earning potential, future health, and well-being.While adaptation responses are often considered as steps to build resilience, in some cases they can also undermine it [5].Some responses to climate change may be necessary in the short-term but may create long-term harms, potentially exacerbating existing inequalities for vulnerable groups, termed 'maladaptation ' [6, 7].Examples from around the world are increasingly highlighting the heterogeneity in exposure and vulnerability to climate change, with specific groups at higher risk.For example, those in rural households, particularly ones reliant on rainfed agriculture for their livelihood, including smallholder farmers [8].These households may be limited in the types of adaptive responses available to them.
A growing number of policies and programs are being developed to protect, support, and enable vulnerable communities to build resilience to climate change.Climate-smart agriculture, application of nature-based solutions and indigenous farming practices, policies around safe migration, and access to climate education all show promise [6].However, targeting these solutions to the most vulnerable people and communities, and adapting them to different contexts are critical to achieve an impact.Language around loss and damage is increasingly used, referring to the consequences of climate change that go beyond what people can adapt to, or when options exist but cannot be adopted because of a lack of resources or access [6,9].Strategies to identify the highest risk and most vulnerable individuals and communities are important to inform interventions and policies that are context specific, accessible, and impactful.However, identifying who is most vulnerable and effectively reaching them with messaging, programming and policies can present a major challenge.
This analysis uses novel statistical approaches to differentiate exposure, effects and responses to recent climate events reported in regions of India, Mexico, and Nigeria to identify vulnerable groups that require targeted intervention and support.The methodology is an innovative way to define categories or types of people within a dataset, and then discern characteristics associated with membership in each category.This approach may enhance results and better delineate characteristics of each group.This analysis seeks to identify the characteristics associated with disproportionate exposure, effects, and responses in each country to highlight the inequalities of climate change and its context-specific effects.By identifying categories and types of respondents, this analytic approach can inform the types of interventions or policies required to reach groups that may face different exposures and harms within and across contexts.
Data was collected in India, Mexico and Nigeria, as these are three of the world's fastest growing countries and economies [10] that each face significant climate related threats according to the Intergovernmental Panel on Climate Change or IPCC [11] and the Global Climate Risk Index [12].Though to varying degrees, all three countries experience a range of different climate hazards, including flooding, drought and heat waves, and all three have high economic and social inequality that may exacerbate vulnerability [13].This study aims to test statistical approaches to identify groups that are disproportionately exposed to, harmed by, and taking different responses in the face of climate events, a methodology that can be applied to better target programs and policies to the most vulnerable and can be replicated in other settings and expanded using additional data.

Methods
We conducted a mobile phone survey in regions of India, Mexico and Nigeria using random digit dialing.The survey collected information about household characteristics, demographics, and experience with recent climate events.Our statistical analysis is then comprised of a two-step process.The LCA was run and resulted in four distinct groups in our dataset.Each observation was assigned to one of these groups.In the second step, we used classification and regression tree (CART) to predict the likelihood of being in one of the groups based on the demographic characteristics used in the CART model.The LCA results were used as a dependent variable in the CART model.Each step is detailed below.

Survey sampling and methods
Mobile phones are nearly ubiquitous, with over 5 billion mobile users globally [14].Mobile phones can reach large numbers of people living in otherwise difficult to reach areas [15], and for example during the COVID-19 pandemic when face-to-face surveys were largely halted.For this study a mobile phone-based survey was conducted in June 2022.A mobile phone data collection firm with offices based in each country conducted training in each country at local call centers, so live calls were made using random digit dialing.Inclusion criteria were that the respondent must be above the age of 18 yr, speak one of the languages offered (English, and then Assamese, Telugu, and Hindi for India; Hausa for Nigeria, and Spanish in Mexico), living in target regions (Andhra Pradesh, Bihar, Chhattisgarh, Assam, and Jharkhand in India; Kano, Gombe and Niger in Nigeria; and Veracruz, Puebla, Yucatan, Oaxaca, Chiapas, Guanajato, Ciudad de Mexico, and Estado de Mexico in Mexico) (supplementary file 1).These regions were selected because they experience different climate change hazards and overall are at high risk [16][17][18][19].The survey had a soft quota of at least 40% female respondents and otherwise followed a random digit dialing approach in selected regions.
Researchers developed a short survey instrument comprised of questions regarding respondent demographic characteristics (age, gender, income, marital status, household size, urban or rural residence, educational attainment, employment) and exposure to recent climate events (drought, heavy rains, unseasonal rains, floods or flash floods, sea level rise, or heat waves).A series of questions asked what the effects of exposure to this climate event were (including options such as loss of employment, food shortage, health impacted, housing damage, and other harms).Another question asked how the respondent and their household responded to the climate event (including marrying earlier than planned, dropping out school, borrowing money, spending savings, moving to a new place, switching to a nonagricultural livelihood, and other responses).
The study received expedited Institutional Review Board (IRB) approval given that data was collected via mobile phone, with adults, and presented minimal risk [20].IRB provides ethical review and guidance for research studies.All respondents gave verbal consent after being read information regarding the study for informed consent.At the beginning of each call respondents were explained their rights to skip questions or drop out of the survey, the steps taken to ensure confidentiality, and were offered a small incentive for their time.Country specific local phone numbers were shared with each respondent to contact in case of any ethical concerns or complaints.All respondents received an ID number and their identifying information was removed to ensure confidentiality.

Latent class analysis (LCA)
LCA was used on this cross-sectional survey data to examine intersectional exposures and harms related to climate change, separately within each country.LCA is a statistical tool that can be used to identify unobserved profiles or types of people within data based on multiple observed variables [21].LCA was run to create a class or group profile of respondents in each country that had experienced exposure to each type of climate hazard, the effects, and the adaptive responses reported.The included variables were: (1) exposure to climate events in the previous five years (respondents selected one, if any, from drought, heat waves, heavy rains, unseasonal rains, flooding, sea level rise); (2) experience of climate induced effects or harms (reduced agricultural production, family or themselves experiencing adverse health effects, housing damaged, or lost employment) and adaptive responses (skipping health services, borrowing money, spending savings, married earlier, dropping out of school earlier, women working more, skipping meals, or migrating).Including these measures of the effects and the adaptive response taken are important to highlight the multi-dimensional aspects of vulnerability in creating the groups.We used a threshold of 0.5 to describe the significant characteristics of each group.This cut off is commonly used to suggest above this point the likelihood of that outcome increases [22][23][24][25].
The analysis followed an iterative process entailing constructing a series of models (starting with a two-class model and expanding until statistically distinct groups formed), and refining the variables included.Each model yielded probabilities for membership in each class, and, for each class, item response probabilities for each indicator.Final model selection was based on model fit statistics and interpretability [21].This included assessing Akaike's information criterion and Bayesian information criterion to measure goodness of fit and parsimony.The Lo-Mendell-Rubin likelihood ratio test was used to select the optimum number of classes.The classification quality of the model was evaluated according to the entropy criterion, in which the values range from zero to one, where values close to one indicate good classification.LCA was conducted using Mplus software (v6.12).The model fit criteria (AIC, BIC, Entropy, LRT) for each country are presented in appendix, resulting in the selection of four classes for each country.

CART analysis
Once the final LCA models were selected, we assigned each respondent to their class based on their highest posterior latent class probability.We then assessed which respondents were more likely to be in each group based on their socio-demographic characteristics.CART analysis identified the factors influencing membership in each class using a non-parametric recursive partitioning technique [26,27].These were fit separately for each country.At the first split, the CART model evaluates the predictors and splits the model on the predictor variable to maximize intranode homogeneity and inter-node heterogeneity [26].The model continues splitting until maximization is achieved.CART analysis explored demographic characteristics associated with membership in the four classes defined by the LCA, including by gender, age, educational attainment, employment in agriculture, number of working household members/household size, urban or rural, and income level, each summarized in a binary variable.The CART analysis was conducted using SPSS version 28 [28] and a cross-validated model was used to interpret the CART results.

Results
Table 1 highlights demographic variation in the sample by country.Across all three countries, respondents were about half female, over half were married, and between a quarter and half urban (24% in Nigeria, 30% in India, and 45% in Mexico) (table 1).Most participants were 25-34 yr of age and had completed secondary school, potentially reflecting the demographics related to mobile phone ownership being a requirement for study enrollment.Globally, mobile phone users tend to be younger with higher education and income [14].Over a third worked in agriculture in India and Nigeria, with only 8% in Mexico.In India and Mexico most households had one working household member (56% and 58%, respectively), whereas in Nigeria more than half of respondents had three or more working household members (57%).Our sample is based on random digit dialing in selected states and is not nationally representative.
Table 2 highlights the results from the LCA including the proportion of individuals in each class overall and the probability of reporting each climate related hazard, effect, and response.The table shows the probability that each characteristic is represented in the Group, using the cut off of 0.5.Group 1 reports almost no exposure to any climate event (except for heat waves in Mexico), and almost no effects or adaptive responses.Group 2 is multi-hazard, including heavy rains, and some heat waves, but overall has a less clear pattern suggesting some exposure to climate hazards and some effects and adaptive responses taken.This is a more moderately impacted group.Group 3 is multi-hazard but in all three countries includes drought.This is the group where drought exposure is concentrated, and respondents reported high impact including several effects and responses taken.Group 4 reports heavy rains and flooding with many effects and responses.Within each country, there are several specific effects and adaptive responses reported and some variation.
In India, about half of respondents fell into Group 1 (49%) (table 2).They do not report any climate hazards, effects on their household, and no responses taken.Group 2 reported experience of heavy rains (68%) and reduced agricultural production in the area as the main effect.But no adaptive responses were significantly reported.Group 3 was comprised of 17% of respondents in India; this group experienced drought, heat waves and heavy rains was also reported.These climate hazards resulted in reduced agricultural production in the area (62%), and several responses including women working more due to climate (93%), skipping meals (74%), marrying earlier than planned (65%), and dropping out of school (58%).Group 4 was comprised of only 11% of respondents in India.Group 4 reported high exposure to heavy rains and flooding that resulted in reduced agricultural production (79%) and housing damage (66%).This resulted in several responses, including borrowing money, dropping out of school, skipping meals, and women working more.Although not above the 0.5 threshold, migration was reported (24%) in Group 4, the highest in this group compared to any other.
In Mexico, most respondents were in Group 1 (43% of respondents).This group reported exposure to heat waves, but no significant effects and the only response reported was that women were working more due to climate change.Group 2 (12% of respondents) was also only exposed to heat waves (58%) and though they did not report many effects (losing work was reported by 41%), they did report many adaptive responses including that women were working more due to climate (85%), skipping meals (67%), marrying earlier than planned (59%), dropping out of school (57%) and migrating out (57%).Group 3 (15% of respondents) reported exposure to drought, as well as heavy rains and heat waves.This group did not report many effects but did report some economic responses including borrowing money (60%) and using savings (71%) to adapt.Overall few harms in this group overlapped with those reported in India.Group 4 (29% of respondents) reported exposure to flooding.The main effect was losing work (82%), and several responses were reported including marrying earlier (82%), women working more (81%), and dropping out of school (51%).
In Nigeria, only 27% of respondents were in Group 1, reporting no significant climate hazards, effects or responses (table 2).Most respondents in Nigeria were in Group 2 (54%) the moderately impacted group.They reported exposure to heavy rains, and housing damage (51%) was the main effect.Loss of work approached significance (44%).Several responses were reported mainly use of savings (56%), skipping meals more (59%), and women were working more due to climate (54%).Group 3 (11% of respondents) reported droughts (57%) and also exposure to heat waves and heavy rains.In Nigeria, this group reported many effects including reduced agricultural production, health impacts, and housing damage.They also reported many responses including using savings (98%), women working more (87%), migrating out (86%), and skipping healthcare (64%).Skipping meals approached significance with 48% reporting this response.Group 4 was only 8% of respondents in Nigeria, they reported exposure to flooding events that did not cause many effects, but respondents reported skipping meals (100%), women working more (67%) and borrowing money (59%).
The CART figures highlight the most likely pathways to membership in each of the four latent class groups.In India, the proportion of respondents in Group 3 was as high as 28.6% (compared to 16.6% overall) if the respondents were from rural areas, reliant on agricultural livelihoods, male, and with lower incomes (less than 10 000 INR) (figure 1).The proportion of respondents in Group 4 was as high as 30.9% (compared to 11.4% overall) if the respondents were also from rural areas, male, from low earning households but relying on non-agricultural livelihoods.The proportion of respondents from Group 1 was as high as 87.2% (compared to 49.4% overall) if the respondent was from an urban area, female, and from a high-income household.
In Mexico, the proportion of respondents in Group 3 was as high as 35.8% (compared to 15.3% overall) if the respondent had lower educational attainment (secondary school or less), from larger households with two or more working members, lower income (less than 8000 pesos), older age (35 and above), and slightly higher for those from urban areas (35.8%) (figure 2).The proportion of respondents in Group 4 was as high as 44.8% (compared to 29.2% overall) if the respondent had lower educational attainment, lived in smaller households with one or fewer working members, and in low-income households.The proportion of respondents in Group 1 was as high as 79.3% (vs 43.3% overall) if the respondent had higher educational attainment (graduate or post graduate degree), lived in households with one or fewer working members, and from urban areas.
In Nigeria, the proportion of respondents in Group 3 was as high as 19% (vs 11.5% overall) if the respondent was employed in agriculture, had lower income (less than 210 050 NGN), and lower educational attainment (up to secondary school) (figure 3).The proportion of respondents in Group 4 was as high as 67% (compared to 53.8% overall) if the respondent lived in rural areas, had lower income, was not employed in agriculture, and was from a larger home with two or more working members.They were slightly more likely to be in this group if they had a higher level of education.The proportion of respondents in Group 1 was as high as 45.5% (compared to 26.8% overall) if the respondents were in urban areas, with higher educational attainment (graduate or post-graduate degrees) and not reliant on agriculture for their livelihood.

Discussion
The primary contribution of this paper is identifying climate vulnerable groups in different contexts using an innovative statistical approach.Our approach identified types of people in our sample that fell into one of four categories based on exposure to climate events, effects experienced, and responses taken.Although one group (Group 4) clearly faced flooding exposure and one group (Group 3) faced drought exposure, overall respondents experienced multi-hazard contexts with exposure to several types  of climate events.Different demographic characteristics were associated with group membership, with implications for resilience and adaptive capacity, with some variation by country.For example, men from low-income households and reliant on agriculture were more likely to be in Group 3 (drought exposed) in India and Nigeria, but in Mexico this was more related to low educational attainment, larger households, older age, and slightly more likely to have an urban residence.This may be because in Mexico, over 80% of the population is classified as urban, and few exclusively work in agriculture (though our older age range may suggest some income from agricultural work, as older adults are more likely to work in this sector in Mexico [29]).This information can be used to inform and guide more targeted programs and policies that account for local demographic characteristics, context, and climate hazards.
Our approach highlights the demographic characteristics most likely to result in membership in each group identified, with men in rural settings and reliant on agriculture particularly vulnerable.Our multi-country approach highlights that this method can be replicated across settings, despite variation in the demographic characteristics of each country.For example, Mexico is one of the most urban countries in the world (80% of its population) [30]; compared to India where the majority of the population is rural and almost three-quarters of India's rural households rely on agriculture for their livelihoods [31], while Nigeria is one of the most rapidly urbanizing countries in the world [32] that also has a large agriculture sector.Some of our results in part highlight differences in underlying demographics and contexts.
Despite demographic, economic, and geographic variation, a clear pattern emerged across all three countries for Group 1 that was least exposed to and least impacted by climate hazards.Those in urban areas, with higher income, and higher educational attainment were most likely to be in this group.In India, this was strongly tied to gender, with urban women reporting the least exposure and least harm.High income and higher education are generally protective, as those with greater access to resources live in housing with better infrastructure and are better able to adapt [3,33].Those in urban areas may have improved access to amenities, information, and services, as well as more diversified incomes less impacted by climate change, whereas in rural areas people engaged in livelihoods highly dependent on natural resources and weather [3].Though our respondents did not report this, literature does increasingly show urban areas are not immune to the harms of climate change, and in fact may be at high risk, particularly the poorest and most marginalized urban residents [34,35].Potentially, living in an urban area creates more distance between the person and perceived climate problems in rural areas (though evidence remains mixed) [36].Research on differential vulnerability within urban areas is critical, as our survey may not have reached the most marginalized groups, or their experiences may not have been sufficiently reflected given the limited sample size.
Rural areas reported experience of multi-hazard climate risks and significant social and economic harms.In Group 3, where drought exposure was concentrated, common characteristics across the three countries were reliance on agriculture for livelihoods and lower income households.In Mexico, older adults were more likely to report drought exposure and harms (older adults are more likely to rely on agriculture for income in Mexico), and in India this group was also mostly men.Responses for the drought group varied, with some more financial responses (borrowing money and using savings were reported in Mexico and Nigeria), while some were more social responses, particularly in India where respondents dropped out of school, skipped meals, and said women were working more.Marrying earlier than planned was reported in India and approached significance in Nigeria (44%).Droughts have been linked to early marriage in certain settings, as families offset financial harms from crop loss by securing dowries [37][38][39] or reducing the number of mouths to feed.Droughts have also been linked with food insecurity [40][41][42], children joining families in the brick-kiln industry [43] and dropping out of school [44,45].Rural agricultural households may be more likely to be harmed by droughts, resulting in economic and food insecurities.Compounding and multi-dimensional harms create cycles of poverty and exacerbate inequalities.
Heavy rains and flooding were reporting in all three countries, across Groups 2, 3, and 4. Floods can result in losses of property, infrastructure, businesses, and increase the risk of diseases and other harms [46,47].In all three countries, Group 4 reported floods, with varied effects.In India and Nigeria this resulted in reduced agricultural production, lost work in Mexico, and housing damage in India.Effects included borrowing money reported in India and Nigeria, dropping out of school in India and Mexico, and skipping meals in India and Nigeria.In all three countries, respondents reported women having to work more, potentially that women are taking on paid roles to offset household economic losses.Respondents in Group 4 were more likely to be rural, but not rely on agricultural livelihoods.Floods may result in worse outcomes for more vulnerable households, including for employment, food security, and social harms [45][46][47][48][49]. Heat waves were reported in Mexico in Groups 1, 2, and 3, but only Group 3 in India and Nigeria.
Our study has several limitations.First, using random digit dialing means potential for selection bias as we could only include respondents with a mobile phone, potentially missing the most vulnerable people in each setting who may not have access or be fully representative [50][51][52].We used random digit dialing with soft quotas for urban or rural location and by gender, to try to get representative respondents.We also employed computerassisted telephone interviewing led by a real person, which tends to have better response rates than other methods [53].Second, our survey was short, to ensure a high completion rate since the responses were given over the phone.Third, given the mobile phone-based nature of the survey, we were not able to collect contextual or location information to verify climate hazards occurred.Therefore, this information is based on perceptions and experiences, and selfreported responses that may have recall bias.Lastly, our approach integrating LCA and CART methods has not been used often to date, but our results suggest it is a promising statistical technique to differentiate groups and better delineate (versus just control for) other characteristics.
Our study utilized a novel approach to understand groups within our surveyed respondents and the characteristics associated with belonging to each sub-group.The groups are differentiated by type of climate hazard exposed to in the last five years, and the various economic, social and health related harms experienced, and adaptation responses taken as a result.Across three diverse countries, we found several similarities and differences in these groups, highlighting the need for targeted interventions and policies that address the needs of vulnerable households and individuals.Our results also suggest that climate related harms are multi-dimensional and compounding, with some households experiencing multiple exposures and harms.This highlights the unequal effects of climate change, and the disproportionate impact concentrated on certain groups and even specific households or people.These compounding harms lead to reinforcing cycles of poverty and exclusion that will only exacerbate inequalities between communities as climate change continues [54,55].Our study approach highlights a methodology that can be implemented to identify high risk households and communities, characteristics of households most likely to be these groups, to identify their needs and experiences and better direct resources, programs, and policies to the most vulnerable.

Figure 1 .
Figure 1.Classification and regression tree (CART) results for demographic characteristics and group membership in India.

Figure 2 .
Figure 2. Classification and regression tree (CART) results for demographic characteristics and group membership in Mexico.

Figure 3 .
Figure 3. and regression tree (CART) results for demographic characteristics and group membership in Nigeria.

Table 1 .
Demographic characteristics of participants in each country.

Table 2 .
Latent class membership and probabilities for each country.