Transcript No Slide Title

The good, the bad and the ugly
(evaluating empirical climate and health studies)
18 July 2006
Sari Kovats
Lecturer, Public and Environmental Health Research Unit,
LSHTM
Outline



Basic environmental epidemiology
 Study designs
 Data issues (exposure and outcome measures)
 Systematic reviews
Discuss abstracts
Climate and health studies
 Time series (again)
 Inter-annual variability
 Trends: early effects of climate change?
Environmental epidemiology

Disease driven approach


Identification of disease endpoints, followed by the
examination of potential hazards in effort to establish
causation
Exposure-driven approach

Identifying potential hazards and then examining their
effects on human health
Exposures and outcomes
In an epidemiological study there are:
(a) the outcome of interest
(b) the primary exposure (or risk factor)
of interest
(c) other exposures that may influence
the outcome (potential confounders)
EPIDEMIOLOGICAL STUDIES
OBSERVATIONAL (NON-EXPERIMENTAL)
We observe only
INTERVENTION (EXPERIMENTAL)
We allocate exposure
EPIDEMIOLOGICAL STUDIES
INTERVENTION (EXPERIMENTAL)
OBSERVATIONAL (NON-EXPERIMENTAL)
DATA FROM
GROUPS
DATA FROM
GROUPS
DATA FROM
INDIVIDUALS
COMMUNITY
TRIAL
(f)
DESCRIPTIVE
(a)
DATA FROM
INDIVIDUALS
ANALYTIC
DESCRIPTIVE
ECOLOGICAL
STUDY
CROSS-SECTIONAL
STUDY
(b)
(c)
CLINICAL TRIAL,
INDIVIDUAL FIELD
TRIAL (g)
ANALYTIC
COHORT
STUDY
(d)
CASE-CONTROL
STUDY
(e)
Ecological studies use..

Average exposure for a group


E.g. temperature, rainfall
A population measure of outcome –


Risk or Rate
Counts of events
Ecological studies

Strengths
 Quick and relatively inexpensive
 Simple to conduct
 Availability of data from surveillance programs and disease
registries

Limitations
 Difficulties in linking exposure with disease
 Limitations in controlling for potential confounding factors
 time series avoids some confounding issues….
 “Ecological fallacy” – making a causal inference about an
individual phenomenon or process on the basis of group
observations
Situations where group level variables may
be better


Exposures without much within group variability
(salt consumption in U.S.)
Exposures which can only be measured at
population level



Herd immunity in studying infectious disease
(vaccination levels may be more informative than
individual behavior)
Social capital
Climate
Cross-sectional studies


also called survey or prevalence study
measures exposure and outcome at the same point in
time




involves disease prevalence
usually involves random sampling and questionnaire
measurement
cannot distinguish whether hypothesized cause preceded the
outcome
Spatial/geographical studies: links environmental data with
survey data
Case control studies

Example. Chicago heat wave 1999 Naughton et al.




Cases: 63 deaths from heat stroke during heat wave
Control – 77 alive controls, matched on age and
neighbourhood. Cases Range of social, environmental risk factors for heat wave
deaths
“Working air conditioner at home” Odd Ratio 0.2 (95% CI
1.0, 0.7)


Must consider selection of controls
Cannot calculate rates or attributable risks
Bias

Selection bias



how were subjects selected for investigation
how representative were they of the target population with regard to
the study question?
Information bias (recall bias)



what was the response rate, and might responders and nonresponders have differed in important ways?
how accurately were exposure and outcome variables measured?
Random vs. systematic errors – have different implications for final
estimate
Chance

Hypothesis testing


p-value
Precision of estimate

Confidence intervals

Assumes estimates/data are unbiased
Beware of multiple testing!

Confounding
Question: Is alcohol consumption during pregnancy
associated with increased risk of low birthweight
Low birth weight
outcome
Alcohol during
pregnancy
exposure
Smoking during pregnancy
potential confounding factor
Time series- consider time varying confounders
High temperature
Daily mortality
exposure
outcome
Air pollution
potential confounding factor
Epidemiological data

Routine sources of health data





Vital Registration (births, deaths)
Hospital statistics (admissions, clinic attendance)
Primary care
Laboratory data (notifiable diseases)
Health Surveys


Epidemiological Studies (cohort or longitudinal studies,
cross-sectional surveys)
Demographic and Health Surveys (low and middle
income countries)
Applications of different observational and analytical study designs
Ecological
Crosssectional
Casecontrol
Cohort
++++
-
+++++
-
Investigation of rare exposures
++
-
-
+++++
Examining multiple outcomes
+
++
-
+++++
Studying multiple exposures
++
++
++++
+++
Measurement of time
relationship between exposure
and outcome
+
-
+
+++++
Direct measurement of
incidence
-
-
+
Investigation of long latent
periods
-
-
+++
Investigation of rare disease
1
+++++
+++
2
1 Unless the sampling fraction is known for both cases and controls; i.e. unless the proportion of
cases and proportion of controls sampled from the population is known.
Strengths and weaknesses of different observational
analytic study designs
Ecological
Cross sectional
Case control
Cohort
Probability of:

selection bias
NA
medium
high
low

information bias
NA
high
high
low

loss to follow-up
NA
NA
NA
high

confounding
high
medium
medium
low
Time required
low
medium
medium
high
Cost
low
medium
medium
high
1. But high if you are not aware of, or do not measure, confounding factors
Reviewing the literature


Develop a clear written Search strategy
Clear research question

Inclusion/exclusion criteria

Search >1 database, plus hand searching, snowballing..

Some assessment of quality of studies


Limit to peer review published articles only.
Beware publication bias


Language bias
Climate change bias! – editors like novel or hot topics
Reviews- you need a “search strategy”
Ahern et al. 2005
Quality control: flooding and health studies






Clearly stated hypothesis
Individuals included in the study and how they were selected (i.e. using
some form of randomisation or probability sampling procedure)
Sample to include those who were affected by the flood event, and those
who were not. The latter are often referred to as the ‘control’ or
‘comparison’ group
Data collection in both the pre- and post-flood period. Prospective data
collection is given higher weighting than retrospective data collection, as
the latter is particularly susceptible to recall bias
Results should include p-values or confidence intervals, and limitations of
the study should also be highlighted
Clinical (e.g. mental health outcomes) or laboratory (e.g. leptospirosis)
diagnosis is given greater credence than self-reported diagnosis.
Ahern et al. 2006 Flood Hazards and Health. EarthScan Book.
Abstracts

Identify





Exposure measure
Outcome measure
Study design
Measure of uncertainty?
Confounders?
Climate and health studies
Three research tasks
Empirical studies
[epidemiology]
Risk Assessment
Scenario
Early effects?
detection
attribution
1970s=?
Sensitivity
Mechanisms
Responses
Causality?
present
future
IPCC: different types of evidence for health
effects




Health impacts of individual extreme events (heat waves,
floods, storms, droughts);
Spatial studies, where climate is an explanatory variable in
the distribution of the disease or the disease vector
Temporal studies (time series),
 inter-annual climate variability,
 short term (daily, weekly) changes (weather)
 longer term (decadal) changes in the context of detecting
early effects of climate change.
Experimental laboratory and field studies of vector,
pathogen, or plant (allergenic) biology.
Exposures: climate/weather parameterization

Long-term changes in mean temperatures, and other climate "norms"
o

Interannual climate variability
o

including indicators of recurring climate phenomena – [El Niño years or SOI]
Short term variability [weather]
o

climate change requires changes over decades or longer.
including monthly, weekly or daily meteorological variables.
Isolated extreme events
o
o
simple extremes, e.g. of temperature/precipitation extremes.
complex events such as tropical cyclones, floods or droughts.
Time series analysis: weekly Salmonellosis and Temp
Threshold
oC
%
95% CI
0
5 00
1 00 0
1 50 0
City/Country
0
5
10
Temperature
Sporadic cases only
Outbreaks removed
Kovats et al. 2004
15
20
Adelaide
M
No
4.9%
3.4, 6.4
Perth
M
No
4.1%
3.1, 5.2
Brisbane
M
No
11.0%
7.7, 11.2
Melbourne
M
No
5.1%
3.8, 6.5
Sydney
M
No
5.6%
4.3, 7.0
Canada
W
Poland
M
6 (. , 7)
8.7%
4.7, 12.9
Scotland
W
3 (., 12)
5.0%
2.2, 7.9
Denmark
W
15 (., .)
0.3%
- 1.1, 1.8
England & Wales
W
5 (5, 6)
12.5%
11.6, 13.4
Estonia
W
13 (3, 14)
9.2%
- 0.9, 20.2
Netherlands
W
7 (7, 8)
8.8%
8.0, 9.5
Czech Republic
W
-2 (-6, -1)
9.2%
7.8, 10.7
Switzerland
W
3 (., 3)
9.1%
7.9, 10.4
Slovak Republic
2W
6 (., .)
2.5%
- 2.6, 7.8
Spain
W
6 (., 8)
4.9%
3.4, 6.4
Results by age: Relative risks for 5 countries, same
threshold, by age group
1.2
1.15
1.1
1.05
1
0.95
(014)
(15- (65+) (064)
14)
(15- (65+) (064)
14)
(15- (65+) (064)
14)
(15- (65+) (064)
14)
(15- (65+)
64)
SC
DK
EW
NL
CH
Time lags/time windows

Acute events


Cause before effect (temporality)
Use literature to hypothesise the time lags (days)




Need to address incubation period for infectious diseases
1-2 days salmonellosis, 7-14 days typhoid fever
Delays in reporting process
Critical time windows

Aetiological relevant exposure windows


E.g. childhood exposures to UV, in utero exposures
Need to address latency periods (?years) between exposure and
outcome.
ENSO and health





Large scale climate phenomenon
Irregular occurrence
Climate variability can be important driver of year to year variation in
disease.
 ?driven by precipitation
Insight into effects not evident at local scales

rainfall, predator balance (Venezuela)
Applications
 Epidemic prediction using seasonal forecasts
 Effects of increased frequency of ENSO events under climate
change
 But cannot directly assess effects of progressive warming from
direct extrapolation of ENSO-health relationships
Systematic review – ENSO and health

Criteria for inclusion.





Published in peer reviewed journal
Original research article using epidemiological data.
Quantified association with an ENSO parameter (e.g. El Niño
year, SST, SOI or other index).
The outcome was an infectious disease in humans.
The time series included more than one El Niño event.
Systematic review – ENSO and health
District or country
Outcome
Time period
ENSO parameter
Brazil
Colombia, Antioquia
Colombia
Colombia
Ecuador
French Guiana
Guyana
India + Pakistan (Punjab)
Kenya, Kericho in western
highlands
Pakistan (northern region)
Peru
Sri Lanka, South west
region
Surinam
Venezuela
Venezuela, coastal region
Venezuela
Annual incidence
Monthly cases
Annual cases
Annual incidence
Annual incidence
Annual incidence
Annual incidence
Monthly cases
1956–1998
1980–1997
1960–1992
1959–1998
1956–1998
1971–1998
1956–1998
1867–1943
1966–1998
SOI, El Niño year
El Niño year
El Niño year / SST
SOI, El Niño year
SOI, El Niño year
SOI, El Niño year
SOI
El Niño year / SST
MENSOI
Annual incidence
Annual incidence
Epidemic years
1970–1993
1972–1999
1870–1945
SST
SOI
El Niño year / SST
Annual incidence
Annual incidence
Annual deaths
Annual cases
1956–1998
1956–1998
1910–35
1975–90
SOI
SOI
El Niño year / SST
El Niño year / SST
Evaluating ENSO-health studies



Need to identify correct climate “driver”
Biological mechanisms
Alternative explanations,

Hay et al. Inter-epidemic periods in mosquito-borne diseases
 Dengue – new serotypes on population
Limited data series  need more than 1 event..
Most appropriate geographical aggregation
Disease data is of uncertain quality (and may not be diseasespecific)




e.g. cyclical changes in immunity
Tick-borne Encephalitis, Sweden: 1990s vs 1980s:
winter warming trend
Early1
980s
Mid1990s
White dots indicate locations where ticks were reported. Black line indicates study region.
(Lindgren et al., 2000)
Evaluating early effects: Criteria..

What constitutes evidence of early effects?

To detect changes in distribution or phenology/seasonality, sample sizes
should be maximised by studying multiple species/diseases/populations.
To detect polewards or altitudinal shifts in vector or disease distributions,
studies should extend across the full range (Parmesan 1996), or at least the
extremes of the range. (Parmesan et al. 2000), so as to exclude simple
expansions or contractions.




Given the natural variability in both climate and biological responses, long
data series are needed (i.e. > over 20 years).
Variability in the climate series (e.g. year to year) should correspond to
variability in the health time series.
Analyses should take into account, as far as possible, other changes that
have occurred over the same time period which could plausibly account for
any observed association with climate.
Kovats et al. 2001
Surveys up to 1940
Surveys up to 1980
Surveys up to 1960
Surveys up to 2000
Summary I: Get the study right

1. Correct design

2. As accurate a measure of exposure and
outcome as possible

3. Control confounding
Summary II: Evaluating



Reviews must be systematic and thorough
Epidemiological literature must be evaluated
Climate and health studies should have..



clear hypotheses
plausible biological mechanisms
reported validity and precision
Summary III: Criteria

Good studies…………….






measure and control confounders;
describe the geographical area from which the health data are
derived;
use appropriate observed meteorological data for population of
interest (the use of reanalysis data may give spurious results for
studies of local effects);
have plausible biological explanation for association between
weather parameters and disease outcome;
remove any trend and seasonal patterns when using time-series
data prior to assessing relationships;
report associations both with and without adjustments for spatial or
temporal autocorrelation.
Thank you!