International Journal of Public Health

, Volume 55, Issue 2, pp 85–96

Climate change epidemiology: methodological challenges

  • Wei W. Xun
  • Aneire E. Khan
  • Edwin Michael
  • Paolo Vineis

DOI: 10.1007/s00038-009-0091-1

Cite this article as:
Xun, W.W., Khan, A.E., Michael, E. et al. Int J Public Health (2010) 55: 85. doi:10.1007/s00038-009-0091-1


Climate change is now thought to be unequivocal, while its potential effects on global and public health cannot be ignored. However, the complexities of the causal webs, the dynamics of the interactions and unpredictability mean that climate change presents new challenges to epidemiology and magnifies existing methodological problems. This article reviews a number of such challenges, including topics such as exposure assessment, bias, confounding, causal complexities and uncertainties, with examples and recommendations provided where appropriate. Hence, epidemiology must continue to adapt by developing new approaches and the integration of other disciplines such as geography and climatology, with an emphasis on informing policy-making and disseminating knowledge beyond the field.


Climate change Epidemiology Methodology Bias Confounding 


Today, population health scientists are faced with challenges on an unprecedented scale, in the form of global climate change. The increases in average air and sea temperatures observed in the last several decades have been confirmed as “unequivocal”, and attributable to human activity by the Intergovernmental Panel for Climate Change (IPCC). Temperatures are predicted to further increase significantly over the next century (IPCC 2001, 2007; WHO 2008). Direct impacts on human health include increased frequency and severity of extreme weather events such as heat waves, droughts, storms, and floods, while indirect impacts can be mediated through the various biological systems that sustain life on Earth. Significant threats to human health highlighted by the IPCC and WHO include: temperature extremes and other extreme weather events, air pollution (ground ozone), water- and food-borne diseases, rodent- and vector-borne diseases, and disruptions of food and water supplies, along with effects of social, cultural, and political changes such as migration and conflicts (IPCC 2001, 2007; Martens and McMichael 2002; McMichael et al. 2003).

These expected outcomes imply that to assess the predicted impacts of climate change on human health, we must consider both the temporal and spatial scales over which risks and effects are likely to occur, the multifactorial sources and pathways that contribute to health change, and the dynamic and complex interactions between them. This clearly presents new challenges not only to current epidemiological methods regarding disease causation, but also to the type of approaches required to reliably investigate climate–health relationships (Baker and Nieuwenhuijsen 2008; March and Susser 2006; McMichael 1999; Pekkanen and Pearce 2001). Here, our aim is to review some key challenges that climate change presents to epidemiology.

Key challenges and their methodological implications

Scale of study

The scales involved in investigating climate–health relationships, in terms of time-span and geographical spread, are largely unfamiliar to population health scientists. Climate impacts on populations rather than individuals, therefore within a given area, identifying “unexposed” control group is often difficult as risks will be characterized by small yet significant differences in exposure among individuals. Assessment of likely risk is also complicated by the long-time scales involved in both risk evolution and disease causation as a result of climate change, necessitating long-term study designs (Baker and Nieuwenhuijsen 2008; March and Susser 2006). Hence, not only is there a need to carefully define the appropriate spatial–temporal scales for investigating climate–health relationships, but also the development of methods and frameworks to conduct research and combine evidence from various scales, from both the macro-population level to the micro-individual and even molecular levels (Pekkanen and Pearce 2001). Other related issues include to enable assessments of health over the life span perhaps using life-course approaches and historical trend analysis, and the use of multilevel modelling approaches to better relate macro causes (e.g. environmental and societal changes) and micro causes (e.g. individual vulnerability) of health change.

Exposure assessment

Although the recent unusual climatic trends have been confirmed by various international panels and organisations, translating climate and its alterations into meaningful exposure assessments presents challenges to epidemiologists. Climate is naturally variable, sometimes considerably so. It is defined using a set of weather “norms”: the expected averages values for a specific time-period and location. As weather and climate can vary over spatial (local, national, global) and temporal (day, season, year) scales, it is crucial that climate variability and climate change are clearly distinguished, as the effects of the latter are gradual and occur on decadal time-scales. Extensive and standardised data collection for extended periods of observation is required in order to establish baseline values, against which future climate changes can be measured (IPCC 2001, 2007; Martens and McMichael 2002).

Even in the case of extreme weather events, standardised definitions can be difficult as each phenomenon has features unique to its locality and time-span. Consider the example of a heat waves (Appendix 1): the absolute temperatures, duration, and even impact could be used to describe such an event, and all of the aforementioned defining factors can vary between geographical locations and time. Hence, a truly useful definition needs to be locality specific and take into account other relevant indicators like physiological, behavioural and other characteristics including adaptations of its population. Appendix 1 reviews of some current national heat-wave response plans with a focus on how they are defined.

The definition of exposure must also vary according to the type of health outcome to be studied. For example, many communities, especially in developing countries, are particularly vulnerable to rapid-onset flash floods, as by nature they provide little warning (Few et al. 2004) and therefore making preventative measures very difficult. The majority of flash-flood-related deaths are from drowning and injuries. This differs from slow-onset floods that are less likely to cause immediate mortality and morbidity in the onset phase (Few et al. 2004). In this case, a likely health outcome is the subsequent outbreak of an infectious disease. It is important to distinguish between the types of flood exposures for directing response measures and response planning following flood events.

Ecological and atomistic fallacies

Pure ecological approaches that lack the incorporation of any individual-level information tend to be prone to a methodological shortfall: the ecological bias (sometimes referred to as “ecological fallacy” or “cross-level bias”). This occurs when aggregate or group level assessments are used to infer associations at the individual level. Its implications for epidemiological research have been discussed in depth elsewhere (Greenland 2001; Portnov et al. 2007; Wakefield 2008). An approach to minimising such bias is to increase inter-group and reduce intra-group variance in exposure (i.e. more “centres” of internally equally exposed individuals). However, ecological bias could still arise from the loss of information and the reduction in exposure variance resulting from aggregation of data, a common practice due to practicality, confidentiality, or contextual reasons which directly define variables at the group level (i.e. the presence of a health-related resource at neighbourhood-level). This can distort the magnitude and even the direction of the ecological-level association observed in contrast to individual-level association, depending on the rate and distribution of these factors in the control population. Modern ecological studies routinely incorporate individual-level data to minimise ecological bias, and these are often referred to as “semi-individual” studies. However, this does not eliminate the problem of bias altogether. Abrahamowicz et al. (2004) simulated a semi-individual study with cluster level exposure and individual outcome measure and performed Cox regression analysis, a routine approach when investigating causal relationships in environmental epidemiology. They found that aggregation of individual data resulted in systematic underestimation of the causal relationship between exposure and outcome, which differed from confounding.

However, the preceding does not imply that aggregate-level data should never be used in epidemiological studies; sometimes it is the only form of information available. On the other hand, it is also important to consider populations not only as a collection of individuals, but also as entities with their own unique culture, history, and attributes (Pearce 1996). Population characteristics can have effects independent of the individual. This is demonstrated by McMichael (1999) using the example of herd immunity in the case of infectious diseases suggesting that inferring causation at the population level from individual level comparisons can lead to what has been termed as “atomistic fallacy” by Susser (1973). Any practical and cost-effective health intervention will no doubt also need to address the population as well as individuals. This has resulted in proposals for epidemiology to lean towards ecological or population-based study designs, while also incorporating multi-level information and interdisciplinary approaches, in order to rise to the challenges climate change presents (McMichael et al. 2003; McMichael 1999; Rothman 1993). In a two-pronged approach, one would thus use individual data as much as possible for initial aetiological research to identify any important associations (including vulnerability), while group-level data would be used for studies that ultimately aim to inform policy makers and adaptation, set at the level of populations.

Confounding and effect modification

Climate change research can be especially prone to certain causes of artefactual association: confounding and effect modification. A confounder is an independent risk factor for the outcome that is also associated with exposure, while an effect modifier is a third factor acting within the casual pathway that affects the outcome.

Relationships between climate and health are often mediated via complex webs of diffused and indirect pathways and cascades (i.e. through ecosystems), frequently with prolonged time-span for the associated impacts to become evident. Attribution to a particular exposure would thus be difficult as there would be many competing explanations for observed associations, which cannot be clearly and easily specified. Other complexities such as simultaneous exposures make attribution to each individual hazard using existing quantitative methods difficult.

Figure 1 illustrates the complexity of performing studies in real-life climate change scenario of saltwater intrusion in Bangladesh. For decades, salinity levels in surface and groundwater in coastal Bangladesh have been rising at unprecedented rates (Mirza 2004; Tanner et al. 2007). Higher sea levels are likely to increase salinization (Mondal et al. 2001; Salim et al. 2007). Saltwater from the Bay of Bengal is reported to have penetrated over 100 km along tributary channels currently affecting 20 million people and 830,000 hectare of arable land by varying degree of salinity in Bangladesh (Muir and Allison 2006). This has raised serious public health concerns as various salt-related diseases have been reported in those areas, specifically hypertension and eclampsia in pregnancy (Khan et al. 2008), and cholera outbreaks as a consequence of changes in water quality and temperature that facilitate the proliferation of Vibrio cholerae (Khan et al. 2008).
Fig. 1

Confounder and effect modifier in epidemiology. a Salinity in surface water and its related health impacts are determined by several factors including (1) rising sea levels (2) shrimp and prawn cultivation in the coast, and (3) reduced freshwater flow from upstream, partly controlled by the Farakka barrage. The causal pathway between sea-level rise (SLR) and salinity-related health effects is complex; health effects differ for varying levels of other factors. b To investigate the association between climate change/SLR and water salinity related diseases, banned antibiotics used in prawn farming (climate change and SLR can lead to lower productivity of rice growing farms (i.e. through droughts and salinization of soil) so that rice fields are converted for shrimp and prawn farms, which in turn, lead to salinization) would be considered as (a) a ‘confounder’, while the ‘freshwater flow upstream’ would be considered as (b) an effect modifier. Asterisks climate change and SLR can lead to lower productivity of rice growing farms (i.e. through droughts and salinization of soil) so that rice fields are converted for shrimp and prawn farms, which in turn, lead to salinization

Other factors also contribute to the effects of climate change. In particular, the shrimp farming business, which requires high levels of salt in pond water for cultivation, has risen in the same region and has become a major export industry (Mondal et al. 2001), thus further worsening the ecological situation. In addition, farming of the freshwater prawn (Macrobrachium rosenbergii) has spread in the region, and the latter group of farmers have started using banned antibiotics like nitrofurans, due to temperature-related bacterial proliferation (Ahmed et al. 2008). The diseased prawns that are rejected from international markets (sometimes found to be contaminated with bacteria like V. cholerae) are distributed and consumed by the local communities, resulting in outbreaks of cholera, diarrhoea, dysentery, and skin diseases (Ahmed et al. 2008; USAID 2006). The causal pathway between SLR and salinity-related health effects is therefore confounded by this complex scenario, since cholera outbreaks can either be due to climate change (via modifications in water salinity, pH, and temperature) or to the consumption of infected shrimps, and can even be exacerbated by exposure to antibiotics. The association becomes even more complex when we address additional factors that modify the salinity and health relationship, as illustrated in Fig. 1.

The case-crossover study design maybe used to overcome some of the challenges related to confounding by combining elements from the case–control and cross-over designs. Pioneered by Maclure (1991); here cases serve as their own controls as they pass between windows of exposure and non-exposure. Exposure is then related to the subsequent occurrence of the event of interest. The strengths of this study design include overcoming the difficulty of finding appropriate controls and eliminating within-individual confounding by personal characteristics that remain constant in the short term; also its bi-directional selection of control periods allows adjustment for seasonal and secular trends (Jaakkola 2003). However, it is best applied to relatively rapid-onset, transient, and/or reversible health impacts such as cholera outbreaks; it is unsuitable for many diseases that are chronic in nature (with long lag-times) and have complex aetiology involving many causal factors, such as hypertension.

Bias in climate change

In addition to the issue of confounding as suggested previously, climate change research is affected by potential bias in still unexplored ways. A common classification of bias includes selection bias (in study conception, design and realization), and information bias. Selection bias occurs when the groups that are compared (e.g. exposed and unexposed) are recruited in a way that leads to an artificially unbalanced representation of the outcome variable. According to Rothman (1993), selection bias is “a distortion of the effect measured, resulting from procedures used to select subjects that lead to an effect estimate among subjects included in the study different from the estimate obtainable from the entire population theoretically targeted for the study”. For example, in a cohort study for mortality, if a national death registry is not available, then it is necessary to attempt to contact each subject or his next-of-kin to verify vital status (i.e. whether the subject is still alive). Bias could occur if the response rate is higher in the exposed persons with the disease (i.e. participation depends on exposure and disease status).

In the case of climate, in an extreme example subjects exposed to high climate variability can be selected from an area with endemic malaria while subjects “unexposed” to weather variability are selected from a non-malarial area. This will of course lead to a gross overestimation of the association between malaria and climate changes.

Formally speaking, the probability of being “represented” in the sample should be the same for all the subjects who are eligible (e.g. as controls in a case–control study), i.e. the probability should be equal to the sampling fraction. Sampling should not be influenced by disease status in a cohort or by exposure status in a case–control study. Information bias arises when information, e.g. on the outcome variable, is influenced by knowledge of exposure status. For example, clinical data on hypertension can be collected with great accuracy from an area affected by water salinization, while incomplete hospital records will be used to identify those with hypertension in an unexposed area. Such extreme cases of bias rarely occur, while more commonly selection bias is related to low response rates in interview-based epidemiological studies. Table 1 shows how response rates in epidemiological research are associated with social class (deprivation index), thus creating opportunities for both selection and information bias (Goodman and Gatward 2008).
Table 1

Non-participation rate in children by deprivation index IMD decile. (Goodman and Gatward 2008)


Non-participation rate (%)

1 least deprived


















10 most deprived


IMD index of multiple deprivation

The increasing use of objective information such as GIS-based maps and routinely collected data makes both selection and information bias less likely, but there are exceptions. Appendix 2 shows two examples of potentially biased research on the field of climate change.

Complexities: nonlinearity and vulnerability

Many exposure–response relationships in epidemiology often do not follow linear trends, but rather have more complicated dynamics; well known examples include correlations between body mass index and all cause mortality (Engeland et al. 2003), alcohol intake and coronary heart disease (Bagnardi et al. 2008), and air pollution (particulate matter) and daily mortality (Daniels et al. 2000), to name a few. To put this in the context of climate change, this means that it is not sufficient for epidemiological studies to identify associations but to characterise observed relationships in enough detail in order to be informative for public health policy-making. Appendix 3 illustrates this in more detail using the example of heat-waves Heat-Health Watch Warning Systems example continued from Appendix 1.

Also, the relevance of purely aetiological research on health effects of climate change is dwindling, given that intertwining social, cultural, and other determinants of health within populations are also central to their susceptibility. Population vulnerability is an important characteristic which is difficult to assess using traditional quantitative methods and whose definition varies across different disciplines (Martens and McMichael 2002). It is a product of the extent a group of individuals will be exposed to the change in climate, as well as the sensitivity to the potential impacts of such changes, with adaptive capacity, the ability to cope and minimise impacts (IPCC 2001; WHO 2000). In terms of health research, vulnerability is shaped by the risk related to exposures (e.g. geographical location), current status of health, economic growth, healthcare infrastructure, availability and security of resources, along with various political, social, and cultural characteristics that determine the distribution of the above factors within a population (Fig. 2).
Fig. 2

Population vulnerability interaction diagram. Also see Fig. 3 for a specific example in the case of cholera

For instance, poverty is widely considered to effectuate inequalities for a large number of health determinants including: access to healthcare, poor occupational health, low social capital, low education attainment and health literacy, slow uptake of beneficial interventions and lifestyle, all of which can hinder effective adaptation (O’Neill et al. 2007). Many impoverished populations already shoulder a large share of harmful exposures, such as: environmental and occupational pollution, experiences of food- and water-stresses, and settlements on hazardous sites (e.g. low-lying lands and hillsides) which are more prone to natural disasters. They also tend to rely heavily on climate-dependent industries such as agriculture for livelihood, which is predicted to be most severely affected by the alterations in precipitation and temperature in long-term climate change (Stern 2006). Gender is another known factor that can introduce inequalities in population health and shares many of its mechanistic pathways with poverty. Women have been found to be over-represented in agricultural and other informal labour forces compared to men (Costello et al. 2009; Stern 2006). In disaster events, women have shown increased likelihood of experiencing domestic abuse, post-traumatic stress disorders, and displacement (Dankelman et al. 2008). However, they are also central to local post-event-management and adaptation, while social networks maintained by women can be greatly beneficial, as seen in recent cyclones and floods in India and Pakistan, respectively (Briceño 2002; IPCC 2007).

Use of epidemiology in health impact assessment

Risk assessment aims to comparably identify, characterise, and control environmental risks to health, and it is consequently heavily reliant on epidemiology (Nurminen et al. 1999). By the addition of a comparative element, the comparative risk assessment (CRA) enables quantification of the contributions to overall disease burden from a number of risk factors (Murray et al. 2003). Alternatively, the health impact assessment (HIA) framework, which is defined by a consensus as “a combination of procedures, methods and tools by which a policy, programme or project may be judged as to its potential effects on the health of a population, and the distribution of those effects within the population” (WHO 1999). It is traditionally a way to systematically assess impacts of a policy or intervention on health in a given community, and its multidisciplinary approach makes it appropriate to investigate the real-life complexity associated with climate change and other environmental issues since it evaluates the contribution of a wide range of health determinants, including economical, social, psychological, environmental, and organisational factors (Scott-Samuel et al. 2006). In the case of climate change, these may be adaptive and mitigative strategies. Yet, there are inherent drawbacks to these approaches, for example, the convention for HIA is to assess impacts of individual factors rather than complex causal webs, and it has also been criticised to lack quantitative rigour (Veerman et al. 2005).

A methodology for an integrated conceptual and analytic framework to assess health impacts of certain environmental exposures has been proposed recently. Briggs (2008) describes an integrated environmental HIA as “a measure of assessing health-related problems deriving from the environment, in ways that take account of the complexities, interdependencies and uncertainties of the real world”. It combines a number of existing methodologies to provide a more comprehensive assessment framework that encompasses comparative risk- and health impact-assessments, adaptive responses and policy-making, which is specifically focused for environmental health risks.

An alternative conceptual approach to deal with complexities in causal relationships of interest is to take a less reductionism view of causality. One can imagine instances in climate change research, where interactions between risk factors for the same and even different endpoints, can be just as interesting in regards of their potential as targets for prevention or adaptation measures. For example, a flood can directly cause immediate death, physical injury, and disruption to infrastructure, which can then lead to communicable diseases, poisoning in the short term, and also post-traumatic stress disorders in the long term. In addition, the attributable fraction of each factor may add up to more than 100%, and results will also vary with time and place of study, making interpretation difficult.

Uncertainties and unpredictable outcomes

Accounting for the inherent uncertainty in models of ecological and physical processes related to global climate change is necessary but difficult in practice. Uncertainties exist regarding climate–health relationships, as well as around the extent that socioeconomic adaptation, technological advances, and human behaviour will influence such relationships in the future (Haines et al. 2006). Recent work in the area of climate change and health has raised the importance of dealing with ‘biocomplexity’, defined as “properties of a system emerging from the interplay of behavioural, biological, chemical, physical, and social interactions that affect, sustain, or are modified by living organisms” (Martens and Huynen 2008; McMichael 2001; Michener et al. 2001). This increasing focus on understanding the biocomplexities of the climate–health relationship first came about from the growing awareness that traditional study methods principally based on empirical observations, which by definition, cannot be readily used to predict future events and their impacts, as a result of system uncertainties, interactions spanning multiple hierarchical levels as well as temporal and spatial scales, and complex dynamics leading to multiple emergent states (Michener et al. 2001). Some uncertainties generally occur from deficient understanding of the actual processes, such as ascertaining key parameter values in models, reasonably foreseeing the future (as population vulnerability may change over time), and sampling variation. Highly non-linear interactions and feedback loops between various components and pathways linked to the health impact of climate change would lead to the occurrence of chaotic and even unpredictable behaviours and thresholds for health change between populations. This system dynamics perspective suggests that ultimately the study of climate–health relationship will need to move beyond simply using empirical studies to the use of a systems approach based on well-validated mathematical modelling (Costello et al. 2009). Indeed, the challenge here is how best to use empirical studies to inform, validate, and test a mathematical modelling approach to enhance our understanding of climate epidemiology.

The recent studies of cholera outbreaks provide a good example of how our understanding of an infectious disease has evolved from a linear empirical model based on oral-faecal transmission of a water-borne bacterium to a human host, to a more complex ecological model based on the climatic, chemical, biological and social factors that underpin transmission. Figure 3 adapted from Wilcox and Colwell (2005), illustrates specificities of the multi-factorial cross-scale environmental transmission of cholera and the hierarchical multi-level analysis framework required to investigate such diseases).
Fig. 3

Biocomplexity and cholera. This figure illustrates the inter-woven climatic, biological and social influences of cholera transmission, and the hierarchical multilevel framework approach needed for investigating climate–cholera relationships successfully. Adapted from Wilcox and Colwell (2005) [34]

Knowledge dissemination

The prompt dissemination of knowledge and experience gained through climate change research to governments and health organisations around the world is imperative. In addition to the climate change-health working groups already established such as the IPCC and at WHO, a number of international databases have been set up to promote the sharing of knowledge such as the UK Climate Impacts Programme and the Copenhagen UN Framework Convention on Climate Change (UNFCCC) (Costello et al. 2009).

In order for the significance of epidemiological findings to translate to a wider audience (i.e. other than fellow scientists), the traditional effect-indicators used for reporting findings, such as the odds ratio and risk ratio may need to be extended to measures of health impacts such as attributable risk fraction in terms of disease burden, or even quality-adjusted-life-years (QALY’s) and disability-adjusted-life-years (DALY’s), where possible and appropriate. Causal diagrams can also be valuable tools for describing complex causal systems and for controlling confounding (Joffe 2006), by incorporate changing scenarios: (1) spontaneous natural changes, which could include weather fluctuations, climate change, and population growth, (2) policy-induced changes, and (3) scenarios for use in modelling (Joffe 2006). The use of causal diagrams is considered to be advantageous because they are tractable and can readily connect with policies and other initiatives that could affect health.


Gaining a comprehensive a priori knowledge of well-validated relationships between climate-related factors and health, even in the absence of climate change, should be a high priority for health researchers. To ensure the best chances of success in detecting early climate-related changes, the choice of health outcomes needs to be based, at least in part, on a priori knowledge of the mechanisms of pathogenesis: i.e. diseases that have shown climate- or weather-sensitivity either directly (e.g. temperature extremes-related mortality) or have such a component within their aetiological pathways (e.g. cholera outbreaks). The International Institute for Environment and Development (IIED), IPCC and WHO have recommended a number of priority disease categories: heat- and cold-stresses; allergic disorders from natural and anthropogenic sources; natural disasters (sea-level rises, wind, storms, droughts, floods); food security and safety; water- and vector-borne diseases and ultraviolet radiation.

Perhaps the most unfamiliar challenge for epidemiologists is the need for integration into the wider context of public health and aim to inform policy, through effective dissemination of epidemiological knowledge beyond the specialist field.

In summary, climate change requires epidemiology to endeavour to identify research needs and gaps in empirical knowledge and prioritise according to need, to develop methodologies that address the “real-world” complexities such as multi-factorial causal webs and non-linear dose–response relationships, and to make communication and dissemination of results more accessible to a wider audience.


This research has been made possible by a contribution of the Grantham Institute for Climate Change to Aneire Khan.

Conflict of interest statement

The authors declare no conflicts of interest.

Copyright information

© Birkhäuser Verlag, Basel/Switzerland 2009

Authors and Affiliations

  • Wei W. Xun
    • 1
  • Aneire E. Khan
    • 1
  • Edwin Michael
    • 1
  • Paolo Vineis
    • 1
  1. 1.Department of Epidemiology and Public HealthImperial College London, St Mary’s CampusLondonUK

Personalised recommendations