What Influences the Demand for a Potential Flood Insurance Product in an Area with Low Previous Exposure to Insurance? – A Case Study in the West African Lower Mono River Basin (LMRB)

Wagner, Simon; Thiam, Sophie; Dossoumou, Nadège I. P.; Daou, David Article — Published Version What Influences the Demand for a Potential Flood Insurance Product in an Area with Low Previous Exposure to Insurance? – A Case Study in the West African Lower Mono River Basin (LMRB) Economics of Disasters and Climate Change Provided in Cooperation with: Springer Nature Suggested Citation: Wagner, Simon; Thiam, Sophie; Dossoumou, Nadège I. P.; Daou, David (2023) : What Influences the Demand for a Potential Flood Insurance Product in an Area with Low Previous Exposure to Insurance? – A Case Study in the West African Lower Mono River Basin (LMRB), Economics of Disasters and Climate Change, ISSN 2511-1299, Springer International Publishing, Cham, Vol. 8, Iss. 1, pp. 1-32, https://doi.org/10.1007/s41885-023-00138-w This Version is available at: https://hdl.handle.net/10419/318348 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/4.0/ Vol.:(0123456789) Economics of Disasters and Climate Change (2024) 8:1–32 https://doi.org/10.1007/s41885-023-00138-w 1 3 RESEARCH What Influences theDemand foraPotential Flood Insurance Product inanArea withLow Previous Exposure toInsurance? – ACase Study intheWest African Lower Mono River Basin (LMRB) SimonWagner1,2· SophieThiam3· NadègeI.P.Dossoumou4· DavidDaou2 Received: 12 August 2023 / Accepted: 19 November 2023 / Published online: 18 December 2023 © The Author(s) 2023 Abstract Floods portray a severe problem in the riverine areas of West Africa while more frequent and intense heavy precipitation events are projected under climatic change scenarios. Already, floods cause manifold impacts, leaving the population to cope with the financial impacts of floods through their own means. As formal risk transfer mechanisms (e.g., insurance) are not yet widely available to the population, efforts to increase their accessibility are being intensified. However, studies assessing flood insurance demand currently mostly focus on regions with more established markets. Also, they are majorly applying conventional statistical modeling approaches that consider only a small number of parameters. Contrarily, this study aims to provide an approach for assessing flood insurance in a context of low previous exposure to such products, to allow for a better consideration of the research context. Therefore, a parameter selection framework is provided and machine learning and deep learning models are applied to selected parameters from an existing household survey data set. In addition, the deep learning sequential neural networks outperformed all machine learning models achieving an accuracy between 93.5— 100% depending on the loss function and optimizer used. The risk to be covered, insurance perception, no access to any source, access to support from community solidarity funds, access to governmental support, or drawing upon own resources for financial coping, financial recovery time, lack of means and prioritizing more essential needs emerged as important model parameters in researching insurance demand. Future roll-out campaigns could consider the parameters pointed out by this study. Keywords Floods· Machine learning· Deep learning· Willingness to insure· Togo· Benin Introduction Over the past decades, there have been observations of an increasing trend of hydrological extremes (i.e. maximum peak discharge) in West Africa, leading to an increase of disastrous flood events in areas located in proximity to large rivers (Ranasinghe etal. 2021). Extended author information available on the last page of the article 2 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Moreover, while overall precipitation is projected to decrease in West Africa, heavy precipitation events are expected to occur more frequently and intensively according to scenarios considering medium to high emission levels, which leads to accumulated hydroclimatic stress through drought and flood events in the region (Trisos etal. 2022; Giorgi etal. 2019). Already, floods cause a wide variety of impacts in West Africa, such as damaged buildings, disruption of livelihoods, damaged goods, fatalities, displacement, sickness and spreading of diseases, damaged infrastructure and crop damage (Wagner etal. 2021; Afriyie etal. 2018; Brisibe and Pepple 2018; Addo and Danso 2017; Ahadzie etal. 2016; Enete etal. 2016; Adewole etal. 2015; Adelekan and Fregene 2015; Codjoe etal. 2014). With regards to the financial implications of flood impacts in the Lower Mono River Basin (LMRB) in particular, it was found that floods regularly affect households financially through agricultural (lost investments through loss and destruction of crops and plantations, loss of livestock), material (repair and replacement cost for damage or destruction of residential houses and personal material belongings), health (sickness and subsequent payment for medical care), and commercial/trade impacts (lost income from damaged stored products for sale, lack of market access, and affected marketplaces) (Wagner etal. 2022). While mutual support among affected households, especially in the phases of response and reconstruction (especially hosting flood victims and helping neighbors to rebuild) (Lamond etal. 2019; Amoako etal. 2019; Ahadzie etal. 2016; Codjoe and Issah 2016; Adelekan and Asiyanbi 2016), seems to be very prevalent in the West African region, there appears to be a lack of risk transfer instruments that are designed to address the financial consequences of floods (Wagner etal. 2021). Thus, people in the region frequently resort to informal mechanisms that are not originally designated for alleviating the diverse financial implications of flood impacts, which sets households back in their financial achievements (Wagner etal. 2022; Boubacar etal. 2017; Addo and Danso 2017). Moreover, the frequency and severity of flood impact levels in the LMRB require more concerted risk reduction activities before establishing risk transfer mechanisms, such as insurance, that enable spreading the risk of financial losses across a larger pool of beneficiaries (Wagner etal. 2022). Also, whether insurance is an appropriate risk management tool in developing economies or not remains a contested issue (Pill 2022; Mechler and Deubelli 2021; Dehm 2020; Linnerooth-Bayer etal. 2019; Schäfer etal. 2019; Gewirtzman etal. 2018). While there are increased efforts to raise insurance penetration and insurance coverage against climate-related extreme events in developing economies (InsuResilience Global Partnership 2021), insurance protection against flood impacts remains difficult to be established, even globally (Léger 2022; Flood Resilience Initiative 2020; Lloyd’s 2018). In addition, much of the research on the uptake of or willingness to pay for flood insurance focusses on the Asian, North American and European region, in which the establishment of flood insurance in the market and familiarity with such products are very different from the West African region. Aside from a few studies (Berg etal. 2022; Oduniyi etal. 2020; Navrud and Vondolia 2020; Adzawla etal. 2019), this topic has not been widely researched in the African context. Also, insurance penetration on the African continent in general is only half of the global average while also the average premiums per person are eleven times lower (Bagus etal. 2020). Thus, to better inform future roll-out campaigns of flood insurance products it is important to research the parameters that are associated with insurance take-up in settings where a large number of people at risk have not yet been insurance customers, such as the LMRB. Most studies researching the willingness to insure (WTI) against floods/willingness to pay (WTP) rely on parameter selection directly based on literature and subsequently apply regression methods (Netusil etal. 2021; Robinson and Botzen 2019; Reynaud etal. 2018; 3 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Fahad and Jing 2018; Turner etal. 2014; Botzen etal. 2013, Botzen and van den Bergh 2012), that usually only consider a low number of parameters. Contrarily, it presents a challenge to derive such parameters from a considerable body of studies for the West African region, due to the limited number of available publications from this area. Thus, established frameworks or reasons for parameter inclusion from other contexts might not be the best fitting for this research context. To address this gap, this study investigates the following central research question: Which parameters influence the decision-making process of households to take up a potential insurance product against flood damages in a setting with low previous exposure to such products, such as the LMRB? Constrained by the limited literature base for the West African region, this study initially reviews literature on WTI against floods/WTP for flood insurance on a global scale. Based on this body of literature, a framework is developed that summarizes six thematic areas of parameters (subjective perception of flood risk, objective flood risk, interactions with insurance institutions, Interaction with other institutions & social environment, attributes of HH/individuals, assets to be potentially insured) to guide which factors are influential on the demand for insurance in the research setting. To structure the parameter selection, feature columns for the entire data set were initially assessed for the entire data set. Then, the remaining parameters were categorized into the six thematic areas of the framework. Moreover, the grouped parameters were assessed through pairplots and a heatmap correlation matrix. As a final step of verification, crosstabs were used for assessing the correlation between the parameters and the output value. This data-driven parameter selection approach is deemed suitable for this study due to researching a context in which people at risk have not been widely exposed to insurance products. Subsequently, on the basis of the selected parameters, machine learning and deep learning models are trained that serve in explaining the observed demand for a potential flood insurance product in the research area. Background Insurance andRisk Transfer forFloods inTogo andBenin Currently, insurance products against the impacts of floods are not widely offered on a household level in Togo and Benin. The insurance industry is mostly centered around motorcycle/car insurance and less on natural hazards (Meton 2019). In addition, there are efforts in Benin to establish health insurance in pilot communities free of charge for its beneficiaries in the first three years (Government of the Republic of Benin 2021). With regards to floods, calls for a feasibility assessment of a flood insurance system through a national insurance fund are even dating back to at least 2011, as stated in a post-disaster needs assessment of the 2010 floods (Government of the Republic of Benin 2011). Also, the Togolese government expressed a strong interest in feasibility studies of an agricultural insurance system within its National Adaptation Plan (Government of the Republic of Togo 2017). In addition, in 2018 Togo was chosen by the pan-African risk pool mechanism African Risk Capacity (ARC) to serve as a pilot country for the implementation of a flood insurance scheme (Akoda 2018). However, no information on its current status could be found, and the most recent available report for the Togolese Republic only contains information for the event of drought (African Risk Capacity 2021b), similarly for Benin (African Risk Capacity 2021a). Moreover, the Beninese government also stated a practical 4 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 absence of an insurance system for climate-related impacts, such as floods, droughts, wind storms, or heat waves, despite their potentially high impact on the country’s gross domestic product (Government of the Republic of Benin 2020). Regarding the LMRB in particular, a recent study points out a strong need for risk-reducing flood adaptation measures and that a conventional, market-based flood insurance approach could be impractical due to the high severity and frequency levels of reported flood impacts from a household perspective (Wagner etal. 2022). As a consequence, this study aims to show relevant insights into the potential flood insurance market, for the case that risk-reducing flood adaptation measures are successfully implemented in the LMRB. Moreover, the research provides insight for insurers to see if they could help to opening a market for themselves by contributing to investing into flood adaptation measures in the area. Finally, this research could benefit the previously mentioned endeavors of establishing flood insurance that are already taking place and support their potential rollout campaigns. Studies Researching theDemand forFlood Insurance Various studies on the demand for insurance and their influential factors have been published in the past years under the fields of willingness to pay (WTP) or willingness to insure (WTI). Whereas the former stride is mainly focusing on calculating a premium that potential insurance clients are willing to pay, the latter usually researches the general interest level among targeted groups. The latter aspect also portrays the main focus of this study. However, only a small number has researched the influential factors on demand for flood insurance in the African context (Berg etal. 2022; Oduniyi etal. 2020; Navrud and Vondolia 2020; Adzawla etal. 2019). The major share of studies from that stride of research focused on the Asian (Hossain etal. 2022, Senapati 2020a, b, Liu etal. 2019, Dewi etal. 2018, Reynaud etal. 2018, Sidi etal. 2018, Fahad and Jing 2018, Arshad etal. 2016, Ren and Wang 2016, Abbas etal. 2015, Aliagha etal. 2015, Aliagha etal. 2014, Turner et al. 2014, Hung 2009), North American (Darlington and Yiannakoulias 2022; Huang and Lubell 2022; Netusil etal. 2021; Thistlethwaite etal. 2020; Atreya etal. 2015; Oulahen 2015; Kousky 2011; Browne and Hoyt 2000) or European contexts (Osberghaus and Reif 2021; Robinson and Botzen 2020, 2019; Botzen etal. 2013; Seifert etal. 2013, Botzen and van den Bergh 2012) – areas in which flood insurance systems and insurance in general are more widely established. In studies from this stride of research, the influential factors mentioned have often been grouped into different categories to provide better orientation for researchers in the selection of relevant parameters (summarized in Table1). For example, Seifert etal. (2013) state the influence of perceptions of flood risks (subjective views), experiences with flood impacts (objective views) as well as factors relating to interactions with disaster assistance from institutions (humanitarian/public compensation). Similarly, Netusil etal. (2021) also point out the importance of factors expressing subjective and objective views on flood risk, while adding the characteristics of residential houses (assets) and demographic characteristics of the respondents (attributes of HH/individual). Aliagha etal. (2014) as well raise the influence of objective and subjective views on flood risk and socio-economic/demographic factors. To achieve its objective, this study compiles further influential factors from further WTP/WTI studies from a global scope/various geographical contexts and grouped them as well into distinct categories while drawing upon and complementing the suggested categories from the previously mentioned studies. In that way, a framework to support the selection of influential factors was created for this study (Fig.1). 5 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 1 Summary of parameters mentioned in WTP/WTI studies Category Thematic area Parameter References Comparable parameter in survey data set Flood risk “Subjective” perception of flood risk Flood risk perception (Hossain etal. 2022, Reynaud etal. 2018, Oulahen 2015, Seifert etal. 2013, Botzen and van den Bergh 2012, Hung 2009) Yes Recently) experienced flood events (Osberghaus and Reif 2021; Senapati 2020a; Liu etal. 2019; Adzawla etal. 2019; Fahad and Jing 2018; Ren and Wang 2016; Atreya etal. 2015; Aliagha etal. 2014; Turner etal. 2014; Hung 2009; Browne and Hoyt 2000) Yes Perception on climate change (Adzawla etal. 2019; Oulahen 2015, Botzen and van den Bergh 2012)Yes Awareness (Senapati 2020b)Yes Anticipated worry and regret about uninsured losses (Robinson and Botzen 2020, 2019)Yes The observation of other’s losses (Turner etal. 2014)Yes “Objective” Flood Risk (Externally defined) level of flood risk (Huang and Lubell 2022; Netusil etal. 2021; Kousky 2011)Yes Proximity to rivers (Sidi etal. 2018, Botzen and van den Bergh 2012, Kousky 2011) Indirectly contained in other parameter of flood risk Living in a low lying area (Botzen and van den Bergh 2012) Indirectly contained in other parameter of flood risk House elevation (Aliagha etal. 2015)Yes Experienced flood impacts (Hossain etal. 2022, Osberghaus and Reif 2021, Paopid etal. 2020, Senapati 2020a, Liu etal. 2019, Fahad and Jing 2018, Reynaud etal. 2018, Arshad etal. 2016, Oulahen 2015, Atreya etal. 2015, Turner etal. 2014, Seifert etal. 2013, Hung 2009, Browne and Hoyt 2000) Yes Flood depth and duration (Paopid etal. 2020, Aliagha etal. 2015)Yes Presence of other risk-reduction measures/levee protection (Hossain etal. 2022; Thistlethwaite etal. 2020; Kousky 2011)Yes 6 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 1 (continued) Category Thematic area Parameter References Comparable parameter in survey data set Interaction Interaction with insurance institutions Price of insurance (Navrud and Vondolia 2020; Reynaud etal. 2018; Browne and Hoyt 2000)No Multi-year insurance policies/billing frequency (Reynaud etal. 2018; Botzen etal. 2013)No The amount offered in the insurance contract (Senapati 2020a; Reynaud etal. 2018)No Trust in insurers (Sidi etal. 2018; Reynaud etal. 2018; Aliagha etal. 2014)Yes Types of risk covered (Reynaud etal. 2018)Yes Previous insurance purchase (Senapati 2020a)Yes Insurance provider (Reynaud etal. 2018)Yes Perception of effectiveness of insurance (Abbas etal. 2015)Yes Awareness of insurance (understanding) (Oduniyi etal. 2020; Senapati 2020b)Yes Interaction with other institutions & social environment Perceived responsibility for preventing damage (Oulahen 2015)Yes Humanitarian/public compensation (Seifert etal. 2013, Botzen and van den Bergh 2012)Yes Flood risk communication (Botzen etal. 2013)Yes Flood prediction (warning) (Sidi etal. 2018)Yes Access to information and extension services (Hossain etal. 2022; Adzawla etal. 2019)Yes Membership in farmer’s groups (Hossain etal. 2022; Adzawla etal. 2019)Yes Perception towards government effort in handling flood (Sidi etal. 2018)Yes Risk sharing between agents (Berg etal. 2022)Yes Social influence (Lo 2013)No 7 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 1 (continued) Category Thematic area Parameter References Comparable parameter in survey data set Attributes Attributes (of HH/individual) Income (Dewi etal. 2018, Sidi etal. 2018; Arshad etal. 2016; Ren and Wang 2016; Aliagha etal. 2015, 2014; Abbas etal. 2015; Kousky 2011; Hung 2009; Browne and Hoyt 2000) Yes Education (Oduniyi etal. 2020; Adzawla etal. 2019; Sidi etal. 2018; Atreya etal. 2015)Yes Age (Oduniyi etal. 2020; Atreya etal. 2015; Abbas etal. 2015)Yes Ethnicity (Atreya etal. 2015)Yes Attitudes towards risk taking (e.g., risk averse) (Hossain etal. 2022; Reynaud etal. 2018, Botzen and van den Bergh 2012)Yes Internal locus of control (Robinson and Botzen 2020)Yes Ability to pay (Fahad and Jing 2018; Arshad etal. 2016)Yes Alternative income sources (non-agricultural) (Hossain etal. 2022; Adzawla etal. 2019; Abbas etal. 2015)Yes Preference uncertainty (Hung 2009)Yes Conservatism (Hung 2009)No Farmer’s experience (Oduniyi etal. 2020)Yes Marital status (Oduniyi etal. 2020)Yes HH dependents (Oduniyi etal. 2020)Yes Remittances (Adzawla etal. 2019)Yes Having the location of the house in an affluent area (Adzawla etal. 2019)No 8 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 1 (continued) Category Thematic area Parameter References Comparable parameter in survey data set Potential assets to be insured House price/dwelling value (Darlington and Yiannakoulias 2022, Paopid etal. 2020, Kousky 2011)No Amount of land owned (Kousky 2011)Yes Land status (ownership) (Dewi etal. 2018, Abbas etal. 2015)Yes Farm typology (Fahad and Jing 2018; Arshad etal. 2016)Yes Cultivated land size (Senapati 2020a)No Farm size (Dewi etal. 2018)No Seed prices (Senapati 2020a)No Fertilizer prices (Senapati 2020a)No Expenditure of farmer (Dewi etal. 2018)No House conditions (Hung 2009)Yes Commercial production (Adzawla etal. 2019)No 15 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 2 (continued) Parameters Responses Frequency Percentage Additional sources of income (multiple responses possible) Raising cattle 213 28.6 Fishing 86 11.6 Hunting 7 0.9 Local industries 188 25.3 Manufacturing industries 14 1.9 Construction and public works 13 1.7 Commerce, catering and accomodation 182 24.5 Transport and communication 26 3.5 Banks and insurance 1 0.1 No response 91 12.2 Currently owning any form of insurance Yes 17 2.3 No 727 97.7 Total 744 100 Previously owned insurance but terminated the contract Yes 8 1.1 No 736 98.9 Total 744 100 16 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Data Preparation andVariable Selection Initially, data had to be separated into categorical and numerical parameters while cleaning the data and removing NaN (Not a Number) values. The latter was necessary since the presence of NaN values will stop the calculation of fitting the model if not removed, but will also generate NaN values after calculation. For the creation of the model, one-hot encoding was used for the categorical parameters (transformation into binary 0–1 parameters) and standard scaling for the numerical data (discarding mean and scaling according to variance of the unit) to be able to create a processor for the model. The process of parameter selection is illustrated in Fig. 4. In order to begin the initial selection of relevant parameters, feature columns were assessed based on the p-value and (Spearman) correlation value to uncover the relationships between parameters. This steps allowed for a reduction of the initially more than 400 parameters to around 100. The remaining parameters were then grouped by topic into the six areas of the framework presented in Fig.1. Then, pair plots (showcasing pairwise bivariate distributions) and a (Pearson) correlation heat map were generated to further facilitate the selection of influential parameters. Based on the heat map correlation matrix, it was decided to use the parameters with low correlation values while disregarding the others, as the high correlation parameters can be connected and related in two ways: if the values of correlation are higher than + 0.5, then these parameters are directly correlated and if less than -0.5 then they are inversely correlated, which means if one parameter tends to increase, then the connected one decrease for negative values while it increases for positive values. For additional verification, cross-tabulations that illustrate the correlations between the parameters and the output parameter were used before further steps were conducted in the analysis. Moreover, it allowed for deciding which parameters to retain or drop. Comparison ofMachine Learning Models Machine learning models were tested by using the Scikit-learn sklearn package. For all models, the data was split into training (67%) and test data (33%). The first model was the multinomial logistic regression model, and is considered a supervised learning technique. This technique serves to predict if an object belongs to a certain class by providing 111 285 50 33 17 11 96 34 90 17 0 50 100 150 200 250 300 Very likely Likely Indifferent Unlikely Very unlikely Togo (n=496)Benin (n=248) Fig. 3 Distribution of responses within outcome variable (likelihood of purchase of a potential flood insurance product) 17 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 a probability on a range between 0 and 1 (James etal. 2021). Furthermore, the Histogrambased Gradient Boosting classifier model was applied, which considers gradient values obtained by prior update steps from moving into the steepest direction of descent (Feng et al. 2018). Also hyperparameter tuning and gridsearch were applied to this classifier, which however did not lead to a satisfactory improvement of the model accuracy. Finally, additional machine learning tests were applied by using decision trees, a method drawing upon the Gini-Index (James etal. 2021). In addition, bagging was applied to the decision trees to lower the variance in the prediction function, as well a random forest model, drawing upon an assembly of various decision trees (Hastie etal. 2009). Deep Learning Model (Sequential Neural Network) In order to attempt achieving better results than the ones obtained from more conventional machine learning approaches (see 3.2.2), this study added a deep learning (DL) model (sequential neural network model) to the analysis using both the TensorFlow and Keras packages. Sequential models are part of artificial neural networks, which usually consist of several layers (input layer, hidden layers, and output layer) that each are equipped with several nodes/neurons, containing activation functions, that are connected through weighted connections between the layers (Jung 2022; James etal. 2021). In general, a sequential model processes the inputted data in a one-directional, linear sequence from the input layer, passing through the hidden layers, and arriving at the output layer (Chollet 2021). Usually, DL approaches are chosen in cases where extremely large data sets are processed and when the possibility to interpret the model does not play and important role (James etal. 2021). Still, this study applied this approach to clarify if a DL model would improve the accuracy of prediction. With regards to the large amount of categorical data, that were encoded, it also helped to consider a larger Fig. 4 Selection process of the final set of model parameters 18 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 amount of available data. To analyze numerical and categorical features in a combined manner in this DL model, feature columns were defined by using a Dense Features layer and using it as an input into the Keras model. The sequential model built for this study uses the Relu (Rectified Linear Unit) activation function for the input layer, not allowing activation of the neuron if input values are below 0 (James etal. 2021), and a Softmax function for the output layer, which is best suited if a categorical output is desired (Klimo etal. 2021). Each neuron of the input layer receives a variable of the dataset and passes that information to another neuron, which leads to a higher number of neurons with a higher number of variables. This model contains 256 neurons. Besides, the Softmax layer must have the same number of nodes as the output layer, which is five in the case of this model (Fig.5). The activation layer is actually the nonlinear function and it transforms the values of the first hidden layer into weighted sums to the next layer. In addition, the Adam as optimizer with a cross entropy and 200 epochs was applied for fitting the model. To compare this model, a second DL model was generated containing 50 neurons, the he_uniform function as kernel initializer, drawing samples from a truncated normal distribution centred on 0 and the stochastic gradient descent (SGD) optimizer. Sequential models bear the disadvantage that they only allow to provide input into the model only once at the beginning, in contrast to functional models in which layers can be connected to one another in a multi-directional way, allowing for feed-back loops (Chollet 2021). Yet, sequential models still better allow for a consideration of a large number of input parameters in comparison to a conventional regression model approaches, as currently widely used in the field of WTP/WTI. In addition, in comparison to conventional ML approaches a neural network can learn from the data in a better and more complex way and even work with unstructured data (Janiesch et al. 2021) and thus better reflect the research context. This consideration was of high importance to this research project to not directly infer findings and assumptions from studies in regions with more established insurance markets. Instead this study wants to consider a wider range of parameters to better represent the interest levels of a population that has not been widely exposed to the usage of such products before. Fig. 5 Application of Softmax on the DL model output layer 19 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Results Selected Relevant Parameters According toPairplots, Correlation Matrix andCross Tabs For parameter selection, feature columns for the entire data set were initially assessed for the entire data set. Then, the remaining parameters were categorized into the six thematic areas of the framework (Fig.1). Moreover, the grouped parameters were assessed through pairplots and a heatmap correlation matrix. As a final step of verification, crosstabs were used for assessing the correlation between the parameters and the output value. The relevant parameters reflected all six thematic areas of the presented framework on influential factors on insurance demand. As visualized in Table1, parameters on potential assets to be covered were only sparsely represented in this data set, which can be seen as the reason for them only appearing once in the final selected set of parameters. Finally, 38 parameters (including one output parameter) make up the final set of selected parameters (Table3). The selected parameters of the model covered the following categories of parameters from the framework: Perception on climate change; Flood risk perception; Experienced flood impacts; (Externally defined) level of flood risk; Awareness of insurance (understanding); Trust in insurers; Perception of effectiveness of insurance; Previous insurance purchase; Insurance provider; Types of risk covered; Perceived responsibility for preventing damage; Humanitarian/public compensation; Membership in farmer’s groups; Risk sharing between agents; Income; Marital status; Ability to pay; Preference uncertainty; Land status (ownership). Model Accuracies All models were applied to three separate data sets each, namely one overall data set containing submissions from both Togo and Benin (n = 744) as well as two subsets from Togo (n = 496) and Benin (n = 248) exclusively. Initially, six machine learning models were run on the data sets and compared by their model accuracy. The applied model types for the classification are logistic regression, a histogram-based gradient boosting classifier, an optimized histogram-based gradient boosting classifier, decision trees, a bagging trees classifier, and a random forest classifier. Moreover, a sequential neural network was applied to the data sets to compare if a DL model would yield higher accuracies than the conventional ML models. As illustrated in Table4, almost all models (except for the optimized histogram-based gradient boosting classifier) returned the highest accuracies for the Togo subset. The logistic regression classifier returned an accuracy of 54.0% (stdv = 0.029) for the combined data set, 48.0% (stdv = 0.0042) for the Benin subset, and 61.7% (stdv = 0.049) for the Togo subset. Overall, this classifier therefore ranked among the ones with the weakest performances of the conventional ML models. The histogram-based gradient boosting classifier achieved 64.0% (stdv = 0.00) for the combined data set, 55.5% (stdv = 0.00) for the Benin subset, and 65.3% (stdv = 0.00) for the Togo subset. Thus, it ranked among the better performing conventional ML models, especially for the combined data set and the Benin subset. The model was even improved further through hyperparameter tuning and applying grid search. The model then achieved 67.0% (stdv = 0.00) accuracy for the combined data set, 58% percent (stdv = 0.00) for the Benin subset, which were the highest for all conventional 20 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 3 Summary of included model parameters for assessing the demand for flood insurance Category Thematic area Associated category of parameters from framework Description of selected parameters from the survey data set Flood risk “Subjective” perception of flood risk Perception on climate change Interviewee heard of climate change before Flood risk perception Perceived likelihood of future flooding “Objective” Flood Risk Experienced flood impacts Financial recovery time from commercial impacts Frequency of commercial impacts (past 20years) Intensity of commercial impacts (past 20years) Financial recovery time from all four impact types combined Frequency of all four impact types combined (past 20years) Severity of all four impact types combined (past 20years) (Externally defined) level of flood risk Flood risk zone based on distance to the river, elevation, and reports of flood affectedness Interaction Interaction with insurance institutions Awareness of insurance (understanding) Understanding of how insurance works No previous insurance purchase due to lack of information Trust in insurers Level of trust that insurance companies will deliver payout as promised No previous insurance purchase due to general lack of trust in companies Perception of effectiveness of insurance Insurance as an instrument only suited for the needs of wealthy people No previous insurance purchase due to too much paperwork Previous insurance purchase Household has access to insurance in case of experiencing flood impacts Insurance provider No insurance provider/products present in the area Types of risk covered Desired risk to be covered in potential flood insurance product Interaction with other institutions & social environment Perceived responsibility for preventing damage Desiring to have access to remittances to deal with flood impacts Humanitarian/public compensation Household has access to governmental support in case of experiencing flood impacts Household has access to NGO support in case of experiencing flood impacts Membership in farmer’s groups Household has access to support from cooperatives in case of experiencing flood impacts Risk sharing between agents Household is member of a savings group Household has access to credits from banks in case of experiencing flood impacts Household draws upon their own resources in case of experiencing flood impacts Household has access to support from community solidarity funds in case of experiencing flood impacts Household has access to credits from savings groups in case of experiencing flood impacts Household has access to credits from a private lender in case of experiencing flood impacts Household has no access to any previously mentioned source in case of experiencing flood impacts Household has not bought any insurance previously because they had access to other mechanisms of coverage 21 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 3 (continued) Category Thematic area Associated category of parameters from framework Description of selected parameters from the survey data set Attributes Attributes (of HH/individual) Income Household income per year Marital status Household is female-headed Ability to pay Fear that insurance purchase will affect more essential needs of the household to be covered Household has not bought any insurance before due to lack of means Preference uncertainty Household has not bought any insurance before due to not being interested in the topic Uncertainty on the reason why no insurance has been purchased before Assets to be covered Land status (ownership) Household is owner of the house they are living in 22 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Table 4 Model accuracies of ML/DL models applied to the selected parameters Conventional Machine learning Deep learning Logistic Regression Histogram-based Gradient Boosting Classifier Optimized Histogram-based Gradient Boosting Classifier Decision Trees Bagging trees classifier Random Forest Classifier Sequential Neural Network First model Sequential Neural Network second model Accuracy both countries (n = 744) 0.540 ± 0.029 0.64 ± 0.000 0.67 ± 0.000 0.437 ± 0.034 0.612 ± 0.045 0.636 ± 0.035 1 ± 5.67 × 10–5 0.9350 ± 0.2329 Accuracy Benin subset (n = 248) 0.480 ± 0.0042 0.550 ± 0.000 0.58 ± 0.000 0.476 ± 0.051 0.552 ± 0.035 0.585 ± 0.048 1 ± 0.0013 0.9756 ± 0.1614 Accuracy Togo subset (n = 496) 0.617 ± 0.049 0.653 ± 0.000 0.69 ± 0.000 0.534 ± 0.049 0.704 ± 0.041 0.716 ± 0.051 1 ± 8.17 × 10–5 0.9512 ± 0.1291 23 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 ML models, and 69% (stdv = 0.00) for the Togo subset. Moreover, a decision tree classifier was applied, which merely reached 43.7% (p = 0.034) for the combined data set, 47.6% (stdv = 0.051) for the Benin subset, and 53.4% (stdv = 0.049) for the Togo subset. As a consequence, this classifier achieved the lowest accuracies among all conventional ML models. However, it was improved by applying bagging to then reach 61.2% (stdv = 0.043) for the combined data set, 55.2% (stdv = 0.035), and even 70.4% (stdv = 0.041) for the Togo subset. Finally, as the last conventional ML model, a random forest classifier was applied achieving 63.6% (stdv = 0.035) for the combined data set, 58.5% (stdv = 0.048) for the Benin subset, and even 71.6% (stdv = 0.051) for the Togo subset. These results clearly show that the datasets of Togo rendered the highest accuracies. The latter is due to the fact that there is higher correlation in the answers provided by respondents in Togo. Since the accuracies of the conventional ML models did not yield higher accuracies (over 75–80%), two sequential neural networks from the realm of DL were applied as a comparison. The first sequential neural network model returned 100.0% of accuracy for the combined data set, as well as for the Benin and Togo subsets. As a consequence, it yielded the best performance by far in comparison to the applied conventional ML models. This finding emerged somewhat surprising, since deep learning is rather recommended for data sets that are much larger than the survey data set. The second model however exhibited a slightly lower accuracy with 93.5% for the combined data set, 97.6% for the Benin subset and 95.12% for the Togo subset. A more detailed overview on the loss, precision, F1 score and recall are provided in Annex 1 as well as a confusion matrix in Annex 2 in the supplementary information to this article. Contribution ofParameters toPredicting Likelihoods ofInsurance Purchase intheDeep Learning Model For the sequential neural network model an overview of the most important parameters based on the feature importance value was generated (Fig.6). The feature importance value expresses the level of influence of a parameter on the output variable of the model (likelihood of insurance purchase). When identifying the most important features, a subset of relevant features can be selected for use in building a model. Therefore, the dimensionality is reduced as well as noise in the data. Moreover, the model interpretability is improved in that way. The selection of feature importance furthermore assists in reducing the number of parameters, therefore reducing the data and decreasing the time needed to obtain the results. The feature importance values were generated for the combined data set of both countries, as well as for the Togo and Benin subsets. In general, it can be observed that the feature importance varies in parts to a large extent across the parameters for the individual data sets. With regards to the parameter categories outlined by the framework presented in the study, interaction-related parameters were the most important category of parameters by far. Important parameters related to the thematic area of interaction with insurance institutions were the desired risk (agricultural, material, health, or commercial impacts) to be covered in potential flood insurance product (Togo). Also, the degree to which insurance was perceived as an instrument only suited for the needs of wealthy people (all) exhibited a high feature importance. In addition, parameters relating the interaction with other institutions and the social environment emerged as the thematic area with the most numerous important values. Feature importance was high when a household had no access to any source mentioned in the questionnaire for financial coping in case of experiencing flood 24 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 070605040302010 Likelihood of future flooding (perception) Climate change knowledge Financial recovery time (all) Frequency of impacts (all) Severity of impacts (all) Financial recovery time (commercial) Frequency of impacts (commercial) Intensity of impacts (commercial) Flood risk level Level of comprehension Lack of information Trust in service delivery Lack of trust in companies Perception of insurance Expected paperwork Access to insurance Lack of presence Risk type covered by insurance Remittances (desired for coping) Access to governmental support Access to NGO support Access to cooperatives Household savings group member Access to credits (bank) Coping with own means Access to community solidarity funds Access to credits (savings groups) Access to credits (private lender) No access to any source Access to other sources Income per year Female-headed household Prioritizing more essential needs Lack of means Lack of interest Uncertainty House ownership “Subje ctive” percept ion of flood risk “Objective” Flood Risk Interaction with insurance institutions Interaction with other institutions & social environment Attributes (of HH/individual) Ass ets to be cov ere d Flood riskInteractionAttributes Feature importance all Feature importance Togo Feature importance Benin Fig. 6 Feature importance of parameters in the sequential neural network model 31 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Oulahen G (2015) Flood insurance in Canada: implications for flood management and residential vulnerability to flood hazards. Environ Manage 55:603–615. https:// doi. org/ 10. 1007/ s002670140416-6 Paopid S, Tang J, Leelawat N (eds) (2020) Willingness to pay for flood insurance: a case study in Phang Khon, Sakon Nakhon Province, Thailand. IOP Conference Series: Earth and Environmental Science Parkoo EN, Thiam S, Adjonou K, Kokou K, Verleysdonk S, Adounkpe JG, Villamor GB (2022) Comparing expert and local community perspectives on flood management in the lower Mono River Catchment, Togo and Benin. Water 14:1536. https:// doi. org/ 10. 3390/ w1410 1536 Pill M (2022) Towards a funding mechanism for loss and damage from climate change impacts. Clim Risk Manag 35:100391. https:// doi. org/ 10. 1016/j. crm. 2021. 100391 Ranasinghe R, Ruane AC, Vautard R, Arnell N, Coppola E, Cruz FA, Dessai S, Islam AS, Rahimi M, Ruiz D (2021) Climate Change Information for Regional Impact and for Risk Assessment. In: Masson-Delmotte V, Zhai P, Pirani A, Connors SL, Péan C, Berger S, Caud N, Chen Y, Goldfarb L, Gomis MI, Huang M, Leitzell K, Lonnoy E, Matthews J, Maycock TK, Waterfield T, Yelekçi O, Yu R, Zhou B (eds) Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge Ren J, Wang HH (2016) Rural homeowners’ willingness to buy flood insurance. Emerg Mark Financ Trade 52:1156–1166. https:// doi. org/ 10. 1080/ 15404 96X. 2015. 11348 67 Reynaud A, Nguyen M-H, Aubert C (2018) Is there a demand for flood insurance in Vietnam? Results from a choice experiment. Environ Econ Policy Stud 20:593–617. https:// doi. org/ 10. 1007/ s100180170207-4 Robinson PJ, Botzen WJW (2019) Determinants of probability neglect and risk attitudes for disaster risk: an online experimental study of flood insurance demand among homeowners. Risk Anal 39:2514–2527. https:// doi. org/ 10. 1111/ risa. 13361 Robinson PJ, Botzen W (2020) Flood insurance demand and probability weighting: the influences of regret, worry, locus of control and the threshold of concern heuristic. Water Resour Econ 30:100144. https:// doi. org/ 10. 1016/j. wre. 2019. 100144 Schäfer L, Warner K, Kreft S (2019) Exploring and managing adaptation frontiers with climate risk insurance. In: Mechler R, Bouwer LM, Schinko T, Surminski S, Linnerooth-Bayer J (eds) Loss and damage from climate change. Springer International Publishing, Cham, pp 317–341 Seifert I, Botzen WJW, Kreibich H, Aerts JCJH (2013) Influence of flood risk characteristics on flood insurance demand: a comparison between Germany and the Netherlands. Nat Hazards Earth Syst Sci 13:1691–1705. https:// doi. org/ 10. 5194/ nhess1316912013 Senapati AK (2020b) Insuring against climatic shocks: Evidence on farm households’ willingness to pay for rainfall insurance product in rural India. Int J Disaster Risk Reduct 42:101351. https:// doi. org/ 10. 1016/j. ijdrr. 2019. 101351 Senapati AK (2020a) Do farmers value insurance against extreme droughts and floods? Evidence from Odisha, India. Glob Bus Rev: 097215092095761. https:// doi. org/ 10. 1177/ 09721 50920 957616 Sidi P, Mamat MB, Sukono, Supian S, Putra AS (2018) Demand analysis of flood insurance by using logistic regression model and genetic algorithm. IOP Conf Ser Mater Sci Eng 332:12053. https:// doi. org/ 10. 1088/ 1757899X/ 332/1/ 012053 Thistlethwaite J, Henstra D, Brown C, Scott D (2020) Barriers to insurance as a flood risk management tool: evidence from a survey of property owners. Int J Disaster Risk Sci 11:263–273. https:// doi. org/ 10. 1007/ s1375302000272-z Trisos CH, Adelekan IO, Totin E, Ayanlade A, Efitre J, Gemeda A, Kalaba K, Lennard C, Masao C, Mgaya Y (2022) Africa. In: Pörtner H-O, Roberts DC, Tignor M, Poloczanska ES, Mintenbeck K, Alegría A, Craig M, Langsdorf S, Löschke S, Möller V, Okem A, Rama B (eds) Climate Change 2022: Impacts, Adaptation and Vulnerability. Contribution of Working Group II to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, pp 1285–1455 Turner G, Said F, Afzal U (2014) Microinsurance demand after a rare flood event: evidence from a field experiment in Pakistan. Geneva Pap Risk Insur Issues Pract 39:201–223. https:// doi. org/ 10. 1057/ gpp. 2014.8 Wagner S, Souvignet M, Walz Y, Balogun K, Komi K, Kreft S, Rhyner J (2021) When does risk become residual? A systematic review of research on flood risk management in West Africa. Reg Environ Change 21:84. https:// doi. org/ 10. 1007/ s1011302101826-7 Wagner S, Thiam S, Dossoumou NIP, Hagenlocher M, Souvignet M, Rhyner J (2022) Recovering from financial implications of flood impacts—the role of risk transfer in the West African context. Sustainability 14:8433. https:// doi. org/ 10. 3390/ su141 48433 Wanyan R, Yang L, Pu M, Zhao T, Zeng L (2022) The nexus between air pollution and life insurance demand in china: evidence from deep machine learning. In: Sun X, Zhang X, Xia Z, Bertino E (eds) 32 Economics of Disasters and Climate Change (2024) 8:1–32 1 3 Advances in artificial intelligence and security, vol 1586. Springer International Publishing, Cham, pp 524–539 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Authors and Affiliations SimonWagner1,2· SophieThiam3· NadègeI.P.Dossoumou4· DavidDaou2 * Simon Wagner s6siw[email protected] 1 Agricultural Faculty, University ofBonn, Meckenheimer Allee 174, 53115Bonn, Germany 2 United Nations University – Institute forEnvironment andHuman Security (UNU-EHS), UN Campus Platz der Vereinten Nationen 1, D-53113Bonn, Germany 3 Center forDevelopment Research (ZEF) - Zentrum für Entwicklungsforschung, Genscherallee 3, 53113Bonn, Germany 4 West African Science Service Center onClimate Change andAdapted Land Use (WASCAL) B.P., 1515Lomé, Togo