A systematic review on lagged associations in climate–health studies

Abstract

Background

Lagged associations in climate–health studies have already been ubiquitously acknowledged in recent years. Despite extensive time-series models having proposed accounting for lags, few studies have addressed the question of maximum-lag specification, which could induce considerable deviations of effect estimates.

Methods

We searched the PubMed and Scopus electronic databases for existing climate–health literature in the English language with a time-series or case-crossover study design published during 2000–2019 to summarize the statistical methodologies and reported lags of associations between climate variables and 14 common causes of morbidity and mortality. We also aggregated the results of the included studies by country and climate zone.

Results

The associations between infectious-disease outcomes and temperatures were found to be lagged for ∼1–2 weeks for influenza, 3–6 weeks for diarrhoea, 7–12 weeks for malaria and 6–16 weeks for dengue fever. Meanwhile, the associations between both cardiovascular and respiratory diseases and hot temperatures lasted for <5 days, whereas the associations between cardiovascular diseases and cold temperatures were observed to be 10–20 days. In addition, rainfall showed a 4- to 10-week lagged association with infectious diarrheal diseases, whereas the association could be further delayed to 8–12 weeks for vector-borne diseases.

Conclusion

Our findings indicated some general patterns for possible lagged associations between some common health outcomes and climatic exposures, and suggested a necessity for a biologically plausible and reasonable definition of the effect lag in the modelling practices for future environmental epidemiological studies.

Climate, health, lag, delay, infectious disease, non-communicable disease

Introduction

Time-series regression has become one of the most widely adopted statistical methods for exploring the associations between environmental exposures and various health outcomes.¹ In recent years, mounting attention from environmental epidemiologists has been paid to the climate-change-related impact on population health. Climatic conditions, such as ambient temperature, humidity, wind, sunshine and rainfall, have been found to be associated with not only mortality,² but also the morbidity of communicable and non-communicable diseases.^3–5 The associations between environmental predictors and health outcomes have been reported with an established lag by preceding studies employing various types of statistical models^4,6 and a certain time-series-model technique has been developed for lagged association estimation.¹ Theoretically, there is an inherent lag, usually variable and unpredictable, between any environmental stressor and the human response, which has been supported by previous epidemiological studies.^4–11 Significantly extended durations were observed particularly for the impacts of cold temperatures,^4,5,8 whereas the effects of hot temperatures were more acute and possibly associated with mortality/morbidity displacement.^9–11 For infectious diseases, the lagged associations could last for weeks or even months, since certain intermediary agents that are also potentially affected by weather could be involved in the transmission.¹² Furthermore, the impacts of temperatures and other climate variables on health outcomes could be significantly influenced by their seasonal fluctuations, which makes the bidimensional modelling of exposure–lag–response associations more intricate.

Key Messages

• Reasonable specification of the effect delay of climate variables in epidemiological studies is essential for avoiding considerable deviations in effect estimates.

• Multiple types of lagged associations with climate variables were found in the existing literature.

• General patterns were identified for the lagged associations between different types of health outcomes and climatic exposures.

• A better understanding of biological mechanisms of the association investigated can be of significance in both modelling practices and the estimation of lagged risk patterns.

Earlier studies often used unconstrained/polynomial distributed lag models or moving-average values of climate variables on a certain time scale to account for the potential effect lags of predictors.^13,14 Over the last few years, a newly developed technique has been commonly employed to model the delayed effects of environmental exposures on health outcomes. The distributed lag non-linear model (DLNM),^15,16 taking into consideration the various lag weights of predictors of the previous days/weeks/months, could be exploited as an alternative advanced method to model non-linear exposure–lag–response dependencies. However, constrained DLNM requires deliberate adjustment of the maximum-lag argument to obtain the optimal model fit. In most contexts, the specification of this argument is based on biological plausibility and previous modelling experience in related research areas. In practice, the estimated lag-dependent associations may be under- or overestimated with misspecification of the allowed maximum duration of effect delay in the ‘crossbasis’ of DLNMs, possibly due to model under- or overfitting or a harvesting effect from a longer lag structure. Crossbasis is a bidimensional function defined by the combination of two sets of basis functions measuring the exposure–response and lag–response relationships simultaneously.¹⁵ Some recent studies used 21 days as the maximum lag to extract the overall cumulative temperature–mortality risk.^2,17 However, the optimal choice on the lag specification for different causes of mortality or morbidity remains inconclusive. Likewise, other lag-assessment methods such as the moving average technique employing a start and end point of time to incorporate a lag in the model also require a plausible and justifiable maximum-duration definition to obtain a better model fit and optimized parameter estimation.

To date, there has been no universally agreed standard for the predefined maximum-lag specification for climatic predictors in time-series and case-crossover studies. Given the complexity of the association modelling in terms of its non-linearity, effect longevity and season sensitivity, the present review focused on the lagged associations between climate parameters and health outcomes. To provide empirical evidence on the lag definition in relevant statistical models, we aimed to review all environmental epidemiological studies with a time-series or case-crossover design to find out all possible delays of associations between climate variables and some common communicable and non-communicable diseases.

Methods

This review was conducted following the guidelines of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)¹⁸ and the full protocol for this review was registered with PROSPERO (registration number: CRD42020165702).¹⁹

Search strategy

The specific exposures of our interest were climate variables and thus we restricted our review to some climate-sensitive public-health issues of significant concern and of great burden globally or regionally. We classified the health outcomes into five categories: air-borne diseases (influenza and pneumonia), water/food-borne diseases (all-cause diarrhoea, cholera and salmonellosis), vector-borne diseases (malaria and dengue fever), non-communicable respiratory diseases [all respiratory diseases, asthma and chronic obstructive pulmonary disease (COPD)] and cardiovascular diseases (all cardiovascular diseases, stroke, coronary heart disease and myocardial infarction). We also tried some other commonly seen diseases that in our prior studies were also identified with delayed associations with climate variables.^8,20 However, a fairly limited number of studies have addressed these issues, so we only included the 14 most frequently studied outcomes. Relevant studies were identified by searching the PubMed (https://pubmed.ncbi.nlm.nih.gov/) and Scopus (http://www.scopus.com/) electronic databases, and further screening the reference lists of reviewed articles. We included the following key designated terms in one search string to narrow down the above search outcomes:

• predefined climate terms in the titles (‘weather’ or ‘climate’ or ‘climatic’ or ‘meteorology’ or ‘meteorological’ or ‘temperature’ or ‘heat’ or ‘cold’ or ‘humidity’ or ‘rain’ or ‘rainfall’ or ‘precipitation’ or ‘wind’ or ‘sunlight’ or ‘flood’ or ‘drought’); and

• each health-outcome term in the titles or abstracts (‘influenza’, ‘pneumonia’, ‘diarrhea’ or ‘diarrhoea’ or ‘diarrheal disease’ or ‘diarrhoeal disease’, ‘cholera’, ‘salmonellosis’ or ‘salmonella’, ‘malaria’, ‘dengue’, ‘respiratory’, ‘asthma’, ‘chronic obstructive pulmonary disease’ or ‘COPD’, ‘cardiovascular’ or ‘circulatory’, ‘stroke’, ‘myocardial infarction’, ‘coronary heart disease’ or ‘coronary artery disease’ or ‘ischemic heart disease’); and

• predefined methodological terms in the titles or abstracts (‘time series’ or ‘case-crossover’ or ‘model’ or ‘regression’).

The search was conducted in January 2020. The detailed search strings are provided in Supplementary Appendix 1, available as Supplementary data at IJE online.

Eligibility criteria

The search was restricted to journal articles in the English language with available full texts published during 2000–2019. Time-series studies with any kind of statistical models as well as studies with a case-crossover design on the delayed associations between climate factors and certain health outcomes were included in the review, with an exception of correlation analyses without regression employment. Only human-involved epidemiological studies were reviewed, and laboratory studies and studies only on animals or vectors were excluded. We also excluded studies on the global ocean–atmosphere interaction such as El Niño and Indian Ocean Dipole. Studies on extreme events (i.e. heatwaves and cold spells) were not reviewed in our study. Signal analyses such as wavelet transformation principally coping with non-stationary time series were excluded as well. To generalize the results of this review to a wider population, we only reviewed some common climate variables. Temperature change between neighbouring days and temperature variability were excluded from this review.

Study selection

Due to the massive number of articles identified, precise measures following a PRISMA flow diagram were undertaken for screening and eligibility assessment (Figure 1). After removing the duplicated records, all article titles were screened by two authors individually to determine whether they investigated the relationships between climate factors and the outcomes of interest. Then these authors checked the abstracts and methods of the selected articles to assess whether they were human-involved epidemiological studies and whether a time-series or case-crossover design was employed. Next, all included studies were reviewed to examine whether lagged associations or delayed effects were investigated. Studies with a predefined fixed length of lag (i.e. using a single lag to capture the overall cumulative effect) or an undifferentiated lag preset for all exposures in the model [e.g. employing models with three predefined lags (1, 2 and 3 weeks) with the same lag for temperature and humidity, and finally reporting a 2-week lag for both variables because this model yielded the lowest Akaike information criterion value] were excluded from this review.

Figure 1

PRISMA diagram flow of systematic review

Data extraction

We established certain schemes to investigate the lagged climate–health associations, including author and year of publication, study period, study location, Köppen climate classification, target population, environmental exposures that were reported with a lag, outcome type (e.g. mortality or morbidity), statistical model, time scale, lag design, predefined maximum lag, reported lag and lag selection criteria. We reported both predefined maximum lags and observed lags of the corresponding associations in the current review to cover a wider aspect of methodology used in previous literature and to provide an empirical statistical summary for future studies. We did not extract the specific association estimates, since they were not our major concerns. For mortality studies in which we were able to see lag–response association plots, the effect duration for mortality displacement was not taken into consideration, as our study was primarily designed to investigate the acute or prolonged association with non-optimal temperatures and other climate variables.

Results

Descriptive summary

We identified 6935 published articles by our initial search on PubMed and Scopus, of which 406 were selected for the review based on our inclusion criteria and eligibility assessment, including 26 for air-borne diseases, 39 for water/food-borne diseases, 103 for vector-borne diseases, 88 for non-communicable respiratory diseases and 150 for cardiovascular diseases (Figure 1). A detailed review table for all health outcomes is presented in Supplementary Appendix 2, available as Supplementary data at IJE online. A total of 362 studies investigating the association between ambient temperature and health outcomes were included in this review, followed by 105 on rainfall and 63 on humidity (Table 1).

Table 1

The number of studies with lags for different predictors and outcomes

	Temp	Humidity	Rainfall	Wind	Pressure	Sun	Flood
Influenza	10	10	4	1	2	0	1
Pneumonia	10	1	1	0	0	1	0
Diarrhoea	12	4	9	0	1	0	6
Cholera	4	0	7	0	0	0	0
Salmonellosis	11	2	6	0	0	0	0
Malaria	26	10	32	1	0	3	1
Dengue fever	56	27	44	4	0	0	0
Respiratory	59	3	0	0	1	0	0
Asthma	15	2	1	0	0	0	0
COPD	11	3	0	1	0	0	0
Cardiovascular	83	1	0	0	1	0	0
IHD	19	0	0	0	0	0	0
MI	20	0	1	1	2	1	0
Stroke	26	0	0	2	0	0	0
TOTAL	362	63	105	14	7	5	8

Respiratory, non-communicable respiratory disease; COPD, chronic obstructive pulmonary disease; Cardiovascular, cardiovascular disease; IHD, ischaemic heart disease; MI, myocardial infarction; Temp, temperature; Wind, wind speed; Pressure, air pressure; Sun, sunshine.

Table 2 summarizes some basic characteristics of the studies included in this review. Of all the 406 studies, 251 (61.8%) were conducted in Asia. Studies with lags in Africa mainly focused on malaria and dengue fever, whereas studies in Europe were mostly about non-communicable respiratory and cardiovascular diseases. A generalized linear model (GLM) with different types of distribution assumed accounting for overdispersion was the major statistical method adopted in all studies, followed by a generalized additive model (GAM), particularly among studies involving non-communicable diseases. An autoregressive integrated moving average (ARIMA) model and its variants (ARMA/SARIMA/ARIMAX) were largely employed for malaria and dengue studies. A daily scale was the most common type of time unit for studies on cardiovascular and respiratory diseases as well as pneumonia, whereas most studies on malaria and dengue fever used a monthly scale to model the association, followed by a weekly scale. Different types of lag design were adopted for infectious disease modelling, including cross-correlation analysis, a single lag model, a moving average lag model, an unconstrained or polynomial distributed lag model (UDLM/PDLM) and a DLNM, whereas the majority of studies on non-communicable diseases assessed the lagged associations with the latest DLNM technique.

Table 2

Summary of study characteristics

	Number of studies (N = 406)
	Air	Water/food	Vector	Resp	CVD
Region of study
Africa	2	5	21	0	0
Asia	13	22	65	52	99
Europe	2	3	0	14	23
North America	2	2	7	9	12
Oceania	4	6	2	8	8
South America	2	1	8	3	4
Across-region	1	0	0	2	4
Statistical models
GLM/GLMM	11	29	50	43	79
GAM/GAMM	6	4	10	28	41
GEE	0	0	9	1	1
ARMA/ARIMA/SARIMA	5	1	21	1	0
Case-crossover design	3	1	0	14	19
Mixed/others/unspecified	1	4	13	1	10
Timescale
Daily	11	10	7	83	139
Weekly/biweekly	13	16	31	3	5
Monthly	1	12	65	2	5
Mixed/others	1	1	0	0	1
Outcome
Mortality	2	0	2	42	89
Morbidity	10	4	10	43	41
Surveillance/registry	14	31	87	0	3
Years of life lost	0	0	0	3	9
Mixed/others	0	4	4	0	8
Lag assessment
Single lag model	2	10	15	6	7
Moving average lag model	0	2	2	8	10
Unconstrained distributed lag model	4	9	16	6	12
Polynomial distributed lag model	1	0	5	6	10
Distributed lag non-linear model	8	8	21	46	87
Cross-correlation/correlation	3	6	25	1	2
Mixed/others/unspecified	8	4	19	15	22

Air, air-borne disease; Water/food, water/food-borne disease; Vector, vector-borne disease; Resp, non-communicable respiratory disease; CVD, cardiovascular disease; GL(M)M, generalized linear (mixed) model; GA(M)M, generalized additive (mixed) model; GEE, generalized estimating equation; (S)AR(I)MA, (seasonal) autoregressive (integrated) moving average.

Lags for infectious diseases

Ambient temperature

For influenza, the associations with mean temperature and humidity were generally found to be lagged for 1–2 and 2–4 weeks, respectively. Out of 11 pneumonia studies included, 8 reported separate lagged effects for hot and cold temperatures or for different seasons. The heat effects lasted for <5 days, whereas the durations of cold effects remained inconclusive (see Supplementary Appendix 2, available as Supplementary data at IJE online, for details). After aggregation by country, ambient temperatures (all types of temperatures combined except for diurnal temperature range) were found to be associated with diarrhoea (all-cause, cholera and salmonellosis combined) for 3–6 weeks in most included countries (Figure 2). Among these studies, the mean temperature was examined in Asia, Europe and Canada, whereas the maximum or minimum temperatures were alternatively used in Oceania and Africa (see Supplementary Appendix 3, available as Supplementary data at IJE online, for the country-averaged results for the three types of temperatures). We also aggregated the associations by different Köppen climate classifications (Table 3; see Supplementary Appendix 4, available as Supplementary data at IJE online, for details). Except for a slightly longer lag in the tropical area, no significant distinctions of lags for diarrhoea between different climate zones were observed.

Figure 2

World map of country-averaged lag weeks of associations between climate variables and diarrhoea (all-cause and cause-specific combined) (one study in Europe included 10 countries)

Table 3

Summary of climate-zone-averaged lag weeks of associations between climate variables and selected infectious diseases

	Temperature	Hot effect	Cold effect	Humidity	Rainfall
Influenza/pneumonia
Tropical	–	1	8	2.2	4.8
Dry	4	–	–	4	–
Temperate	1.7	1.5	3.7	1.5	–
Continental	1.5	0.5	–	0	1
Diarrhoea
Tropical	3	5.5	4	–	–
Dry	–	5	2	–	–
Temperate	2.9	2.3	3.2	–	–
Continental	–	2	–	–	–
Malaria
Tropical	8.6	–	5.3	10.5	–
Dry	4.5	–	–	8	8.7
Temperate	7.5	–	–	6.4	10
Continental	3	–	–	7	7
Dengue fever
Tropical	10.3	–	–	8.9	9.8
Dry	11.7	–	–	–	10.5
Temperate	7.5	–	–	6.3	8.7
Continental	–	–	–	–	–

Studies reporting lagged associations between temperatures and malaria were mostly conducted in Asia and Africa (Figure 3), whereas those reporting delayed effects of temperatures on dengue fever were predominantly performed in the Asia-Pacific area and Latin America (Figure 4). Contrary to those for diarrhoea, the associations between temperatures and vector-borne diseases persisted for a significantly longer period. A 7- to 12-week lag of association between malaria and temperatures was identified in most countries, whereas a 2- to 5-week averaged lag was reported in South Korea, Burundi and South Africa (Figure 3). The associations lasted significantly longer in areas with a tropical or temperate climate than in those with a dry or continental climate (Table 3). Further delayed associations were observed for dengue fever, ranging from 6 to 16 weeks (Figure 4). A shortened lag of 3–4 weeks in Nepal and Brazil and an extended lag of 25 weeks in Mexico were also discovered (Figure 4). In addition, the delayed effects reported in temperate regions were shorter than those observed in dry and tropical areas (Table 3).

Figure 3

World map of country-averaged lag weeks of associations between climate variables and malaria

Figure 4

World map of country-averaged lag weeks of associations between climate variables and dengue fever

Rainfall and humidity

The lagged effects of rainfall—another significant climate variable associated with infectious diseases—have also been extensively investigated. Its associations with diarrhoea (all-cause, cholera and salmonellosis combined) were found to last for 4–10 weeks except for shorter durations reported in Ecuador, Haiti, Singapore and Zambia (Figure 2). Most countries studying malaria and dengue fever discovered a similar lag period of 8–12 weeks (Figures 3 and 4). Several exceptions were reported as well, including a 4- to 6-week lag in Ecuador, Thailand and Niger, and a significantly longer lag of 20–28 weeks in Curaçao and Panama (Figures 3 and 4). Humidity has been less commonly studied in terms of its delayed effects on infectious diseases. The current aggregated results showed a lag of 2–6 weeks for diarrhoea, 4–8 weeks for malaria and 5–14 weeks for dengue fever (Figures 2–4). Similarly to temperature, humidity showed longer associations with vector-borne diseases in tropical regions (Table 3).

Lags for non-communicable diseases

Regarding non-communicable respiratory and cardiovascular diseases, the associations with temperatures were massively studied. Commonly, heat and cold effects (mostly 1st/3rd/5th and 99th/97th/95th percentiles of a relevant temperature variable) were separately assessed, regardless of the variations in temperature ranges across different study locations. We summarized all studies reporting separate associations between high and/or low temperatures and respiratory and cardiovascular outcomes (all-cause and cause-specific combined) (Figure 5). Both cardiovascular and respiratory studies found that high temperatures were acutely associated with the outcomes, mostly with a 5 days’ lag or shorter. To demonstrate a clearer pattern of the associations between all-cause respiratory diseases and low temperatures, we removed one study in Palermo, Italy in which an extra-long lagged cold effect was observed (55 days; see Supplementary Appendix 2, available as Supplementary data at IJE online, for details). The delayed associations of cold temperatures were more variable for both diseases. For cardiovascular diseases, the majority of studies reported a 10–20 days’ lag forming an approximately normal distribution, whereas the patterns for respiratory diseases remained inconclusive (Figure 5; see Supplementary Appendix 2, available as Supplementary data at IJE online, for details).

Figure 5

Histograms of lag days of associations between temperatures and cardiovascular and non-communicable respiratory diseases (all-cause and cause-specific combined)

After aggregation by country, both the heat- and cold-delayed effects showed significant variations (Figures 6 and 7). A long lag of ≥10 days for hot temperatures was observed in Italy and Chile for respiratory diseases and in Taiwan, the UK and a study including 12 European countries for cardiovascular outcomes (Figure 6). For associations with cold temperatures, an unexpected short lag of <10 days was found in Russia for respiratory diseases and in South Korea, Thailand and Germany for cardiovascular diseases (Figure 7). The aggregated results by Köppen climate classification showed a consistent pattern of the heat effect with a lag of ≤5 days in all climate zones (Figure 8). In addition, cold temperatures manifested similar lagged associations with cardiovascular outcomes for ∼15 days in all regions. However, the associations between cold temperatures and respiratory diseases displayed significant variations, with the longest duration of 30 days in the tropical zone and a lag of <15 days in areas with a dry or continental climate.

Figure 6

World map of country-averaged lag days of associations between hot temperatures and cardiovascular and non-communicable respiratory diseases (all-cause and cause-specific combined) (one study in Europe included 12 countries)

Figure 7

World map of country-averaged lag days of associations between cold temperatures and cardiovascular and non-communicable respiratory diseases (all-cause and cause-specific combined) (one study in Europe included 12 countries)

Figure 8

Averaged lag days of associations between temperatures and cardiovascular and non-communicable respiratory diseases by Köppen climate classifications

Discussion

The present review was designed to provide epidemiological evidence and recommendations for further modelling of associations between climate parameters and common causes of mortality and morbidity. We identified 406 relevant articles for 14 health outcomes with some lag patterns disclosed. Ambient temperature, the most widely studied climate variable, showed clear and distinct delayed effects on different health outcomes. For infectious-disease outcomes, a lag of 1–2 weeks was found for influenza, 3–6 weeks for diarrhoea, 7–12 weeks for malaria and 6–16 weeks for dengue fever. The associations between both cardiovascular and non-communicable respiratory diseases and hot temperatures lasted for <5 days, whereas those between cardiovascular diseases and cold temperatures were observed to be 10–20 days. In addition, rainfall, a potential risk factor for infectious diseases, showed a 4- to 10-week lagged association with diarrhoea, whereas the effects were found to be further delayed to 8–12 weeks for vector-borne diseases.

Lag investigation has become of increasing interest for environmental epidemiologists in recent years. It not only enables researchers to capture the overall or maximum effects of non-optimal weather conditions, but also provides evidence for authorities to establish early-warning systems for the minimization of relevant adverse health impacts. Different biological, physiological and behavioural mechanisms of distinct lagged associations have been considered. Nonetheless, the estimated lags varied across studies, sometimes significantly, even with the same targeted health outcome. One possible reason is that the pre-existing quantitative evidence required to establish the framework in assessing the lagged effect was complicated and indecisive for most health outcomes of interest.¹²

For all kinds of infectious diseases, the incubation period from infection to manifestation is a critical factor causing delays. Due to the mildness when symptoms start to present, there could be further delays before medical consultation is sought. In addition to these common causes, there have been some specific explanations indicated for certain communicable diseases. For instance, pre-existing infection in animals before slaughter could induce a lag for food-borne diseases because of the time interval for production and distribution between slaughter and consumption, especially for ready-to-eat or frozen food.²¹ Moreover, apart from the proliferation and survival of pathogens in the environment that can be directly affected by ambient temperature, eating habits (more barbecues and picnics in summer) and the consumption of raw or cross-contaminated food (prone to the influence of temperature during its production, transport and storage) are both risk factors for contracting food-borne diseases.⁵ For waterborne diseases, such as cholera in our case, a relatively longer effect duration of rainfall was observed. Rainfall-caused nutrient runoff and seepage first drive a plankton bloom and then there would be a greater chance of pathogen cholerae in estuarine or marine environments to facilitate disease transmission through plankton blooms.²² In addition, developing countries lacking efficient sanitary infrastructures would also suffer from the contamination of surface water and groundwater used for drinking and bathing by sewage overflow due to heavy rainfall. Therefore, the course of association with rainfall is likely to persist for weeks. The lag was found to be even more extended for vector-borne diseases. Except for the incubation period in the human body, there are two more components contributing to vector-borne disease transmission compared with those for other infectious diseases: the growth of female mosquitoes from eggs to adults for parasite transmission and the development of parasites from gametocytes to sporozoites that can infect humans.²³ This could be a possible explanation of the wider variation in the aggregated lagged associations that we found between temperature and vector-borne diseases (mostly 7–12 weeks for malaria and 6–16 weeks for dengue fever) compared with other infectious outcomes. And a longer lag for rainfall was expected because of the cumulative time for runoff and seepage to collect in low-lying breeding sites.²⁴

We reported a pooled result of lagged associations of low and high temperatures with both cardiovascular and non-communicable respiratory diseases separately. It has been suggested that the use of identical lag structures for heat and cold effects when assessing temperature–mortality/morbidity association was not appropriate.²⁵ Our findings, together with previous epidemiological evidence,²⁶ indicated that cold affects human health in a more indirect fashion than heat, causing unbalanced association delays. Early experimental studies showed that both heat and cold increase blood viscosity and plasma-cholesterol concentration.^27–29 Heat further induces significant sweating, elevates the heart rate and decreases arterial pressure, whereas cold causes raised arterial pressure and plasma-fibrinogen concentration.^27–29 Physiologically, the first human response to non-optimal heat is to increase the surface blood circulation from vital organs and sweat to help cool down. However, the body’s mechanism of thermoregulation can fail, with excess blood diverted underneath the skin, causing overloaded stress on the heart and lungs.³⁰ Under the situation of increased cardiac workload, together with dehydration and salt depletion, heart-failure-induced death may occur. Further, the increased blood viscosity and plasma-cholesterol levels may introduce a higher risk of thrombosis that could trigger myocardial ischemia and stroke. Additionally, under heat stress, interleukins regulating local and systemic acute inflammatory responses could be released, which is another risk factor for heart failure by inducing damage to heart tissue and inflammation.^31,32 These acute cardiovascular responses to the failure of body thermoregulation explain the short duration of heat effects. On the contrary, in addition to the same mechanism of thrombosis formation as mentioned above, cold may lead to the activation of a sympathetic nervous system and increases catecholamine secretion, which could accordingly elevate the heart rate, peripheral vascular resistance and then blood pressure.^33,34 These changes could indirectly decrease the ratio of myocardial oxygen supply to demand, due to which myocardial ischemia or myocardial infarction may eventually occur.³⁴ Although a sudden exposure to cold may also be associated with an acute cardiac response, the longer course of impact of cold on the human cardiovascular system than that of heat supported the existing evidence of varying lagged associations with hot and cold temperatures.

The physiological mechanism for the associations between heat or cold and chronic respiratory diseases has been less discussed. Breathing both cold dry and hot humid air have been found to be able to trigger bronchoconstriction, which could exacerbate asthma or COPD.^35–37 Excessive heat exposure not only attracts a defensive thermoregulatory response increasing cardiac output and skin blood flow, but also incites the release of systemic inflammatory factors and enhanced pulmonary ventilation, which would possibly induce thermal hyperpnea.^31,38,39 In addition, a laboratory study on the biological mechanism of acute lung injury demonstrated that heat exposure could impair lung tissues within a short time by directly inducing IL-1β and TNF-α, two pro-inflammatory cytokines.⁴⁰ Thus, COPD characterized by ventilatory impairment and persistent pulmonary inflammation could be further exacerbated by heat.^41,42 Anderson hypothesized an acute respiratory effect of breathing hot air and did not find significant effect modification by local climate, suggesting that very brief exposure to heat could trigger direct human airway responses in a short period.⁴¹ This rapid respiratory response has also been indicated in asthma patients.^37,43 Relatively, the development of bronchoconstriction caused by the hyperventilation of cold dry air was slower.⁴³ The cold-air hyperpnea-induced bronchoconstriction along with some frequent nasal symptoms are rapidly developing short-term responses provoked by sudden cooling of the airways, whereas the long-term responses caused by repeated and long-standing cooling and drying of the airways would damage the airway epithelium and lead to changes in the wall structure and function of the airway.⁴⁴ Both responses could induce a chronic inflammatory disorder, which is the major mechanism of asthma and COPD exacerbation.^45,46 Apart from respiratory smooth muscles, both the pulmonary and the tracheobronchial vasculatures could also be major effectors in response to cold exposure.⁴⁷ Nevertheless, the above-mentioned explanations on the thermoregulatory challenge remain speculative and the physiopathological mechanisms of the differentiated lasting course of temperature effects on both cardiovascular and respiratory disorders require further investigation.

Compared with the traditional moving-average approach, DLNM has been found by a simulation study to be robustly reliable in time-series analysis, particularly for long lagged associations and in the presence of strong seasonal trends.⁴⁸ However, in real-life modelling practices, it requires considered inputs for a few model parameters that may significantly affect the results. Among them, the maximum lag of the crossbasis function is influential for both the shape and the magnitude of the associations of interest. During our review process, it was not uncommon to see studies reporting a lagged association over the entire predefined lag, and the actual association could still be observed beyond that lag according to the predictor-specific plot in DLNM, particularly when a cold effect was investigated. There have also been studies using a prolonged lag (up to months) with the intention to capture the persisting associations as well as the mortality displacement during the entire course of that lag period. It was worthwhile noting that an inadequate or overlong lag would possibly lead to underestimated or spurious associations due to model underfitting or overfitting. In our own practices, an extended maximum-lag setting might also result in extended observed lag–response associations, particularly when the actual length of the lag is close to the maximum lag predefined. Further experimental studies are therefore warranted for the theoretical validation of model fitting when the length of the allowed maximum lag is misspecified.

Among all the included studies, different lag measures for climatic predictors have been used. Studies employing distributed lag models, including UDLM, PDLM and DLNM, usually reported the entire duration of the association (i.e. from the effect beginning until its attenuation to null). Yet there were some studies suggesting a lag when the association/risk was maximized. We extracted the whole course including the non-significant risks from the effect plots to discover the possible maximum cumulative associations. There were two major types of moving-average lag definitions in the reviewed literature: average values from the current day to a certain lagged day (e.g. average for day 0–1, 0–2, 0–3, …) and average values in multiple strata (e.g. average for day 0–1, 2–6, 7–14, …). Most studies using single-lag models (i.e. one-unit lag in one model) reported an individual risk per lag/model, whereas studies using the ARIMA model and its variants mostly reported a lag when the cross-correlation was the strongest. As mentioned previously, DLNM has strong strength in assessing the sliced lag-specific and predictor-specific associations, and thus studies using this technique to test the effect delay for a specific value in the climate variable (e.g. 1st/99th percentile) would have greater power in our review.

Timescale was another factor influencing the lag assessment considerably. While non-communicable-disease studies primarily used daily-count outcomes, included studies on infectious diseases usually conducted analyses on a weekly or monthly scale, particularly for vector-borne diseases. However, the use of a longer time unit may underestimate the overall risk when the actual lag of the exposure is shorter (e.g. a monthly analysis was conducted whereas the incubation period of an infectious disease is only a few days), as also indicated by another methodological review on infectious diseases.¹² Therefore, appropriate study timescale should be selected based on both data availability and biological plausibility in further research.

Despite being the first comprehensive systematic review on the delayed effects of climate predictors on health to date, our study has a few limitations. First, the extent of association delays could potentially be affected by multiple options of statistical methodology and data availability. For instance, it is highly likely that the incorporation of a moving average lag and distributed lag could yield different assessment results, and the same rule applies to different timescales (e.g. daily/weekly/monthly), types of temperatures (e.g. maximum/minimum/mean temperature) and types of health outcomes (e.g. mortality/hospital admission/outpatient visit/emergency-room visit). Therefore, our aggregation of results from different study settings might deviate the true lags to a certain degree, even though the included separate studies were conducted at the same location in the same country. We strongly recommend readers to refer to the specific studies included in this review before applying to their future research. Second, we did not include the direction and magnitude of the association reported with the purpose of an explicit focus on the lags. However, it is also possible that significant associations could last for a relatively long period but with opposite effects (e.g. mortality displacement or protective effect). Hence, a clearer picture could be shown if the overall association was reported along with its duration. Third, we summarized the results by different countries as well as climate zones to account for possible lag differentiation across geographic locations and climate conditions. However, elaborative thinking is still necessary wherever generalization is needed to investigate the associations on a local basis, considering the wide variations in climate zones within the same country, and variations in socio-economic and cultural settings and health-preventive measures in the same climate zone.

In conclusion, our study synthesized prior epidemiological knowledge to form a strategic explorative modelling framework for future studies in a relevant area. With an increasing interest in investigating the lagged associations between environmental stressors and health outcomes, a better understanding of biological mechanisms can be of significance in both statistically robust modelling practice and the estimation of lagged risk patterns.

Supplementary data

Supplementary data are available at IJE online.

Funding

None.

Acknowledgements

This systematic review does not need institutional ethics approval since we did not collect personal or confidential information from participants and all evidence collected are publicly accessible. No new data were generated or analysed in support of this research.

Conflict of interest

None declared.