The impact of COVID-19 fiscal spending on climate change adaptation and resilience – Nature.com

This paper used multiple methods to analyse the potential impacts of COVID-19 fiscal spending on climate change adaptation and resilience (A&R). First, we expanded an existing taxonomy of fiscal spending to incorporate A&R-relevant policy measures50. Second, we applied this taxonomy to a database of ~8,000 policies implemented by 88 countries during the pandemic and analysed the A&R characteristics of this spending. Third, we used techniques in NLP and DL to consider how A&R versus mitigation themes feature in broader policy planning in a subset of 11 countries. Figures of country spending were created using Tableau 2021.4 and tables were generated using Excel 2016.

To assess the potential impact of fiscal policy on climate change adaptation and resilience, we developed a policy taxonomy, organized around the likely impacts of fiscal archetypes on climate A&R. We developed the taxonomy by expanding the original archetype set established by the GRO50, conducting an extensive literature review, and drawing upon existing adaptation and resilience frameworks, as outlined below.

The GRO is a policy database developed by the Smith School of Enterprise and the Environment at the University of Oxford, which tracked COVID-19 fiscal policy in 88 countries from March 2020 to December 2021 (see Supplementary Table 1 for a list of all included countries)11. In the database, policies are categorized into exhaustive and mutually exclusive subarchetypes, which are each associated with one overarching archetype and designated as functioning either for rescue (that is, initial pandemic relief) or economic recovery. An additional indiscriminate archetype captures spending that does not clearly fit into specific archetypes, typically due to a lack of specificity of the policy descriptions provided by governments. Indiscriminate archetypes are classified as unclear, rather than being allocated to either the rescue or recovery phase. The original GRO included 40 archetypes and 158 subarchetypes.

Each subarchetype was assessed in ref. 11 for its potential impact on short- and long-term greenhouse gas emissions. Other potential environmental, social and economic impacts were defined for an adapted archetype set in ref. 50. Archetypes were also tagged by sector. The archetypes were developed from first principles and tested against a preliminary set of 2,000 observed policies. Archetypes were developed with a focus on fiscal policy in contractionary macroeconomic environments.

We took the GRO taxonomy and analysed the ~8,000 COVID-19 policies recorded in ref. 11 to identify gaps where policies might have A&R characteristics but were categorized to archetypes that would traditionally not be considered A&R-related. For example, while most tourism incentives (archetype S) might be considered poor for A&R, Spains US$1.62 billion initiative to improve sustainability in the tourism sector (special attention to Balearic and Canary Islands) could have positive A&R impacts, hence a new subarchetype, Incentives for tourism with A&R conditions, was introduced21.

Next, the augmented taxonomy was compared to existing adaptation and resilience frameworks, such as the IPCCs adaptation actions22 and the European Unions sustainable finance taxonomy23, to identify further gaps. Various classification approaches already exist in climate A&R scholarship; however, they operate at a coarser level. The European Unions taxonomy for sustainable economic activities uses a sectoral classification approach to account for A&R activities alongside other sustainability criteria23. The United Nations Office for Disaster Risk Reduction (UNDRR)s climate resilience classification framework follows a similar sectoral approach51. The CRAFT framework adds a level of granularity by categorizing policies according to the type of activity they represent (rather than the sector), which incorporates both cross-sectoral and sector-specific policies, resulting in 293 subarchetypes (42 archetypes), triple the 88 of the European Union taxonomy.

Existing approaches to categorizing A&R spending tend to emphasize physical adaptation actions, failing to consider the broader impacts of spending on climate resilience. The assessment framework developed by multilateral development banks for aligning activities with the Paris Agreement, for example, focuses specifically on policies that manage physical climate change risks52. More broadly, A&R frameworks tend to evaluate only actions that are explicitly oriented towards adaptation and resilience. For example, the World Bank53 proposes six priority adaptation policy actions, spanning: inclusive development; facilitating adaptation and protection against shocks for firms, peoples, land and public assets; managing economic and financial risk; and monitoring of interventions. Similarly, the UNDRRs Budget Tagging guide for Disaster Risk Reduction and Climate Change Adaptation focuses only on activities explicitly oriented towards these objectives. CRAFT, by contrast, includes policies that are explicitly targeted at climate change adaptation actions (which we classify as direct A&R), alongside policies that are not explicitly climate-oriented, but which may have positive implications for climate adaptation or resilience (identified as indirect A&R). Importantly, CRAFT adds depth to existing approaches by covering policies that improve, reduce or have no impact on A&R, rather than focusing solely on actions that explicitly aim to enhance A&R, allowing us to provide a more holistic picture of the proportion of fiscal spending with potential positive, neutral and negative impacts on climate A&R.

Through an extensive literature review, each existing and new subarchetype was assessed for its potential impact on climate A&R. Archetypes were scored using a 3-point Likert scale (negative, neutral, positive) for two dimensions: direct and indirect climate A&R. We defined direct A&R as explicit efforts to adapt to current or expected climate effects, that is, policies that aim to implement direct adaptation actions. Examples of policies with potential positive impacts on direct climate A&R include the construction of seawalls or efforts to secure coastal ecosystems by planting mangroves54,55.

We defined indirect A&R as efforts that increase resilience or reduce vulnerability to climate change effects, regardless of whether the intention was to directly address climate risks. For example, policies that build capacity for local utilities were identified as having a potential positive impact on indirect climate A&R, because utilities provide services (water supply, waste and sanitation, energy distribution) that are crucial to the functioning and adaptive capacity of individuals, communities and systems24. Similarly, spending on education, even that which is not climate-specific, has been found to increase adaptive capacity and thus was scored positively for indirect climate A&R56,57. Healthcare systems are also crucial to ensuring the ability of populations to adapt and be resilient in the face of climate change58. Other policies that have expected positive impacts for indirect climate A&R include capacity building for subnational public entities, supply chain resilience measures, increasing social and political inclusion, enhancing managerial capacity, and providing access to institutions and information59.

All subarchetypes with a positive direct impact also have a positive indirect impact. This is because specific adaptation actions have broader impacts for climate change adaptation and resilience. For example, the construction of a seawall is also expected to enhance the economic resilience of coastal communities, hence this subarchetype, which was scored positively for direct A&R, was also scored positively for indirect A&R. By contrast, not all policies that were scored positively for indirect A&R were scored positively for direct A&R. For example, education investment that did not specify adaptation or resilience measures was not scored positively for direct A&R, even though it has a positive impact on indirect A&R by enhancing adaptive capacity more broadly. By scoring policies for both direct and indirect A&R impacts, we recognize that climate A&R extends beyond physical adaptation actions and intersects with social, political, economic and environmental resilience60,61. Supplementary Table 9 outlines all policy archetypes with a positive score for either direct or indirect climate A&R, while a literature review and justification for each score is provided in ref. 20.

Policy archetypes that are not expected to have a positive impact on direct or indirect A&R were treated in two ways. Policies that have little relevance to climate A&R were scored as neutral (0). For example, general tax cuts and interest rate reductions do not contribute to direct climate adaptation, and their short-term nature means that any savings they create for individuals or businesses do not contribute to climate resilience by building adaptive capacity. Some policies were scored as neutral for climate A&R because their impacts are limited to the COVID-19 pandemic. For instance, the provision of basic needs (shelter, food, social services), if secured beyond the pandemic, would contribute to adaptive capacity. However, short-term provision of basic needs, delimited to the pandemic, were scored as having a neutral climate A&R impact.

By contrast, policies that entail lock in of non-resilient infrastructure or promote maladaptation were scored as having a negative climate A&R impact. For example, spending on general transportation, energy and urban development infrastructure without regard to resilience is likely to result in lock in, whereby assets with long lifespans are maladapted to changing and uncertain local climate conditions26. There are a few exceptions, whereby infrastructural policies that are non-resilient are counterbalanced by the positive adaptation and resilience impacts of that archetype. For example, education and healthcare infrastructure constructed without regard for resilience may have lock-in potential; however, these impacts are counterbalanced by the adaptive capacity benefits of strengthening education and healthcare facilities, resulting in a neutral score. We did not score any liquidity policies negatively, as we do not expect this short-term funding to result in long-term infrastructure investments with lock-in potential.

Policies that are positive from a mitigation standpoint, such as the construction of renewable energy infrastructure, without consideration of infrastructure resilience, are not always positive for climate A&R. In terms of indirect A&R, clean energy infrastructure provides sustainable jobs and enhances access to energy, both of which are crucial to adaptive capacity24. However, these positive impacts are outweighed by the vulnerability of these facilities to future climate impacts if the new infrastructure is constructed without resilience in mind62,63. On balance, these policies are thus expected to have a potential negative impact on indirect climate A&R, despite their positive impact on mitigation.

We recognize the distinction in the literature among the three dimensions of resilience: absorptive capacity, adaptive capacity and transformative capacity64. However, we did not score policy archetypes for their distinct impacts on each of these dimensions. As ref. 64 highlights, specific interventions are likely to have impacts on multiple dimensions, depending on the intensity of the disturbance and the time of exposure. Policies tend to vary widely on these dimensions for any given subarchetype, such that it would appear misleading to score a policy archetype for a specific dimension.

The CRAFT framework includes 42 archetypes and 293 subarchetypes. This represents a step forward in granularity of taxonomies for assessment of adaptation and resilience impacts of policy interventions. The European Unions taxonomy for sustainable economic activities incorporates only 88 policy types, while the UNDRRs Budget Tagging guide for Disaster Risk Reduction and Climate Change Adaptation classifies activities into 20 broad areas, further broken down into 77 action areas. CRAFT therefore offers more specificity in its assessments than existing approaches. Nonetheless, a taxonomic approach can never replace the specificity of individual policy-level impact assessments; necessarily, there will be variation in the types of interventions assigned to specific categories. For example, even within the subarchetype of agricultural investment with A&R conditions, there will probably be variation in the extent of impact of individual policies that cannot be captured through our Likert-scale assessments of positive, neutral and negative direct and indirect A&R impacts. Only impact evaluations at the policy level can truly capture the potential impact of specific policies on A&R; however, this is not always feasible for policymakers or researchers. A taxonomic approach thus enables an approximate assessment that is scalable, feasible and replicable. While there is likely to be some variation in policy impacts within subarchetypes, CRAFT offers a higher level of granularity than existing assessment approaches, thus offering useful insights for policymakers and researchers alike.

The new policy taxonomy developed for assessing potential A&R impacts was applied to the GRO database11. The GRO database records all fiscal policies implemented by 88 countries over the period of March 2020 to December 2021. Each policy is assigned to a subarchetype and thus takes on the direct and indirect A&R scores, which are implemented at the subarchetype level. Policy names, descriptions, local currency amount (and US$ equivalent) and several other fields are captured for each policy, enabling aggregations at the country and archetype level.

To test the validity of the taxonomy for the GRO, we conducted a robustness check for a subset of policies per archetype. We manually reviewed 4,459 policies out of a total set of 8,037 policies to ensure a 95% confidence interval at 5% margin of error for every archetype. The sample was randomly selected per archetype, with a minimum of 10 policies selected per subarchetype (unless the subarchetype contained less than 10 policies, in which case we reviewed every policy) to ensure coverage of all subarchetypes. For each selected policy, we evaluated whether the direct and indirect A&R scores assigned at the subarchetype level fit the policy description, examining the source documents where clarification was required. We assigned a confidence rating of High where the percentage of inconsistencies in the random sample was between 010%, Medium for 1020% and Low for 20100%. We found that 97% of total spending and 96% of recovery spending were associated with archetypes with fewer than 20% scoring inconsistencies (medium to high confidence) (Extended Data Table 2). We also found that 93% of all archetypes and 94% of recovery archetypes were identified as having a medium to high confidence rating. We report our results as a range, with the lower bound referring to high and medium confidence subarchetypes only, and the upper bound including subarchetypes of all confidence levels, except where all policies are high and medium confidence, in which case only one figure is reported.

In analysing the GRO data, we also evaluated correlations between A&R spending, country income levels and the vulnerability indicators developed by the ND-GAIN32. ND-GAIN defines vulnerability as the propensity or predisposition of human societies to be negatively impacted by climate hazards32. This vulnerability index is a compound measure of exposure, sensitivity and adaptive capacity. Exposure is defined as the physical factors external to the system that contribute to vulnerability. Sensitivity is the extent to which a country is dependent upon a sector negatively affected by climate hazard, or the proportion of the population particularly susceptible to a climate change hazard. Adaptive capacity indicates the availability of social resources for sector-specific adaptation, which can include sustainable adaptation solutions32. We extracted the indicators for our 88 studied countries for the year 2020. We also extracted World Bank31 data for country income levels (GDP per capita and GNI per capita) for 2020.

To assess how A&R themes feature in broader policy planning, we first identified a set of 78 core policy papers (Supplementary Table 8) that were framed as plans for economic recovery, covering 11 of the G20 countries (that is, all those with policy documents published in English). The policy paper corpus was selected from source documents provided by the GRO, supplemented by key-term database searches to add missing budget documents. We analysed the English corpus for text related to A&R and climate mitigation using a climate dictionary expanded from previous papers with techniques in deep learning. The corpus was limited to English documents as vocabularies differ considerably across languages; we leave this exercise to be repeated by future works in other languages, but do not expect the direction of the results to change.

Creating bespoke dictionaries for NLP analysis is notoriously difficult45,65. The objective is to identify a complete set of terms that are broad enough to capture all mentions of a particular theme but precise enough to exclude irrelevant themes. One method for dictionary creation involves surveying subject matter experts, but experts are prone to missing important terms45,65,66. Supervised and active unsupervised methods both offer useful advances44,67, but previous applications struggle to fully address the limits of setting an appropriate starting dictionary. Reference 68 build on ref. 66 to demonstrate a classification approach that iteratively identifies keywords relevant to the emergent themes of a prescribed document set. In our case, where A&R is often a very minor theme in the policy documents and the corpus is small, the classification approach is unlikely to generate substantial additional terms. Instead of a classifier approach, we adopted the method of ref. 45 to iteratively expand a starting dictionary on the basis of embedding models of the target corpus itself; this is a similar method to the later work of ref. 69. The dictionary expansion process began with the full set of policy papers. From these papers, terms (words, bigrams and trigrams) were embedded using the word2vec neural network model, resulting in three separate 100-dimensional embedding spaces. In each space, every term (1:n) was combined with every other term, resulting in a total of n factorial possible term pairs. The vectorized positions of each pair were then averaged and the term closest to this average position was added to the dictionary. An initial set of vectorized A&R terms served as a starting point for this search. Once added, these new terms were manually reviewed. This procedure was iteratively performed until no additional terms emerged. Similar to the classifier approach, the method is somewhat limited by the low overall prevalence of A&R discussion in the policy paper corpus (topic model available in Supplementary Information; also see Extended Data Figs. 2 and 3).

As an alternative approach to dictionary building, we experimented with deep learning BERT70, expanding and categorizing lists of climate A&R and mitigation terms sourced from ref. 45. The transformer architecture is unique in its use of self-attention to differentially weight the significance of input data, processing an entire input simultaneously rather than sequentially. This allows for better contextual learning than previous approaches such as Recurrent Neural Networks71. BERT is the preeminent transformer model, pretrained on a corpus of 3.3 billion words to understand how English words relate to each other. It learns a target words meaning on the basis of the full sentence in which it occurs, whereas a popular alternative, Generative Pretrained Transformers, learns meaning from words that occur only earlier in a sentence than the target word. BERT has been applied to countless topics, including climate mitigation issues (see ClimateBert in the study of refs. 72,73 on electric vehicles), but not, so far as we know, to questions of climate A&R.

We used BERT to identify terms relevant to concepts of climate A&R and mitigation, supplementing language from ref. 45 and that provided by experts. To do so, we fine-tuned a BERT model using policy names and descriptions provided by the GRO dataset of ~8,000 COVID-19 fiscal policies. The model was subsequently trained to identify fiscal provisions that supported direct and indirect adaptation on the basis of policy titles and descriptions. The trained model was then applied to the policy document corpus to identify language consistent with strong climate A&R and mitigation. A subset of selected clauses was manually reviewed to identify new terms for the base dictionary. Dictionary terms were categorized into those that support adaptation, mitigation and both/unclear using the climate A&R taxonomy impact assessment matrix, CRAFT20. The full dictionary and categorizations are included in Supplementary Table 10, and Extended Data Figs. 2 and 3.

Applying the BERT-supplemented dictionaries, we used basic NLP techniques and manual sorting to categorize all 124,593 clauses in the 78 policy papers into those that pertained to topics of general climate, climate mitigation, climate A&R, other forms of A&R and other. Table 2 provides a preliminary statistical account of term usage across the policy papers. These results are helpful for direct comparisons within a country or proportional comparisons between countries. They are unsuitable for direct comparisons between countries as the typology of policy documents vary considerably. Recorded mentions of non-climate A&R are likely to underestimate usage as the dictionaries were developed to target climate topics.

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

See the article here:

The impact of COVID-19 fiscal spending on climate change adaptation and resilience - Nature.com