4.3.10. Environmental epidemiology - I

Basic Principles and study designs

 

Authors: Eva Sugeng and Lily Fredrix

Reviewers: Ľubica Murínová and Raymond Niesink

 

Learning objectives:

You should be able to

 

 

1. Definitions of epidemiology

Epidemiology (originating from Ancient Greek: Epi -upon, demos - people, logos - the study of) is the study of the distribution and determinants of health-related states or events in specified populations, and the application of this study to the prevention and control of health problems (Last, 2001). Epidemiologists study human populations with measurements at one or more points in time. When a group of people is followed over time, we call this a cohort (originating from Latin: cohors (Latin), a group of Roman soldiers). In epidemiology, the relationship between a determinant or risk factor and health outcome - variable is investigated. The outcome variable mostly concerns morbidity: a disease, e.g. lung cancer, or a health parameter, e.g. blood pressure, or mortality: death. The determinant is defined as a collective or individual risk factor (or set of factors) that is (causally) related to a health condition, outcome, or other defined characteristic. In human health – and, specifically, in diseases of complex etiology – sets of determinants often act jointly in relatively complex and long-term processes (International Epidemiological Association, 2014).

 

The people that are subject of interest are the target population. In most cases, it is impossible and unnecessary to include all people from the target population and therefore, a sample will be taken from the target population, which is called the study population. The sample is ideally representative of the target population (Figure 1). To get a representative sample, it is possible to recruit  subjects at random.

 

Figure 1: On the left, the target population is presented and from this population, a representative sample is drawn, including all types of individuals from the target population.

 

2. Study designs

Epidemiologic research can either be observational or experimental (Figure 2). Observational studies do not include interference (e.g. allocation of subjects into exposed / non-exposed groups), while experimental studies do. With regard to observational studies analytical and descriptive studies can be distinguished. Descriptive studies describe the determinant(s) and outcome without making comparisons, while analytical studies compare certain groups and derive inferences.

 

Figure 2: The types of study designs, the branch on the left includes an exposure/intervention assigned by the researcher, while the branch on the right is observational, and does not include an exposure/intervention assigned by the researcher.

 

2.1 Observational studies

2.1.1. Cross-sectional study

In a cross-sectional study, determinant and outcome are measured at the same time. For example, pesticide levels in urine (determinant) and hormone levels in serum (outcome) are collected at  one point in time. The design is quick and cheap because all measurements take place at the same time. The drawback is that the design does not allow to conclude about causality, that is whether the determinant precedes the outcome, it might be the other way around or caused by another factor (lacking Hill’s criterion for causality temporality, Box 1). This study design is therefore mostly hypothesis generating.

 

2.1.2 Case-control study

In a case-control study, the sample is selected based on the outcome, while the determinant is measured in the past. In contrast to a cross-sectional study, this design can include measurements at several time points, hence it is a longitudinal study. First, people with the disease (cases) are recruited and then matched controls (people not affected by the disease), comparable with regard to e.g. age, gender and geographical region, are involved into the study. Important is that controls have the same risk to develop the disease as the cases. The determinant is collected retrospectively, meaning that participants are asked about exposure in the past.

 

The retrospective character of the design poses a risk for recall bias when people are asked about events that happened in the past, they might not remember them correctly. Recall bias is a form of information bias, when a measurement error results in misclassification. Bias is defined as a systematic deviation of results or inferences from the truth (International Epidemiological Association, 2014). One should be cautious to draw conclusions about causality with the case-control study design. According to Hill’s criterion temporality (see Box 1), the exposure precedes the outcome, but because the exposure is collected retrospectively, the evidence  may be too weak to draw conclusions about a causal relationship. The benefits are that the design is suitable for research on diseases with a low incidence (in a prospective cohort study it would result in a low number of cases), and for research on diseases with a long latency period, that is the time that exposure to the determinant can result in the disease (in a prospective cohort study, it would take many years to follow-up participants until the disease develops).

 

An example of a case-control study in environmental epidemiology

Hoffman et al. (2017) investigated papillary thyroid cancer (PTC) and exposure to flame retardant chemicals (FRs) in the indoor environment. FRs are chemicals which are added to household products in order to limit the spread of fire, but can leach to house dust where residents can be exposed to the contaminated house dust. FRs are associated with thyroid disease and thyroid cancer. In this case-control study, PTC cases and matched cases were recruited (outcome), and FR exposure (determinant) was assessed by measuring FRs in the house dust of the participants. The study showed that participants with higher exposure to FRs (bromodiphenyl ether-209 concentrations above the median level) had 2.3 more odds (see section Quantifying disease and associations) on having PTC, compared to participants with lower exposure to FRs (bromodiphenyl ether-209 concentrations below the median level).

 

2.1.3 Cohort study

A cohort study, another type of a longitudinal study, includes a group of individuals that are followed over time in the future (prospective) or that will be asked about the past (retrospective). In a prospective cohort study, the determinant is measured at the start of the study and the incidence of the disease is calculated after a certain time period, the follow-up. The study design needs to start with people who are at risk for the disease, but not yet affected by the disease. Therefore, the prospective study design allows to conclude that there may be a causal relationship, since the health outcome follows the determinant in time (Hill’s criterion temporality). However, interference of other factors is still possible, see paragraph 3 about confounding and effect modification. It is possible to look at more than 1 health outcome, but the design is less suitable for diseases with a low incidence or with a long latency period, because then you either need a large study population to have enough cases, or need to follow the participants for a long time to measure cases. A major issue with this study design is attrition (loss to follow-up), it means to what extent do participants drop out during the study course. Selection bias can occur when a certain type of participants drops out more often, and the research is conducted with a selection of the target population. Selection bias can also occur at the start of a study, when some members of the target population are less likely to be included in the study population in comparison to other members and the sample therefore is not representative of the target population.

 

An example of a prospective cohort study

De Cock et al. (2016) present a prospective cohort study investigating early life exposure to chemicals and health effects in later life, the LInking EDCs in maternal Nutrition to Child health (LINC study). For this, over 300 pregnant women were recruited during pregnancy. Prenatal exposure to chemicals was measured in, amongst others, cord blood and breast milk and the children were followed over time, measuring, amongst others, height and weight status. For example, prenatal exposure to dichlorodiphenyl-dichloroethylene (DDE), a metabolite of the pesticide dichlorodiphenyl-trichloroethane (DDT), was assessed by measuring DDE in umbilical cord blood, collected at delivery. During the first year, the body mass index (BMI), based on weight and height, was monitored. DDE levels in umbilical cord blood were divided into 4 equal groups, called quartiles. Boys with the lowest DDE concentrations (the first quartile) had a higher BMI growth curve in the first year, compared to boys with the highest concentrations DDE (the fourth quartile) (De Cock et al., 2016).

 

2.1.4 Nested case-control study

When a case-control study is carried out within a cohort study, it is called a nested case-control study. Cases in a cohort study are selected, and matching non-cases are selected as controls. This type of study design is useful in case of a low amount of cases in a prospective cohort study.

 

An example of a nested case-control study

Engel et al. (2018) investigated attention-deficit hyperactivity disorder (ADHD) in children in relation to prenatal phthalate exposure. Phthalates are added to various consumer products to soften plastics. Exposure occurs during ingestion, inhalation or dermal absorption and sources are for example plastic packaging of food, volatile household products and personal care products (Benjamin et al., 2017). Engel et al. (2018) carried out a nested case-control study within the Norwegian Mother and Child Cohort (MoBa). The cohort included 112,762 mother-child pairs of which only a small amount of cases with a clinical ADHD diagnosis. A total of 297 cases were randomly sampled from registrations of clinically ADHD diagnoses. In addition, 553 controls without ADHD were randomly sampled from the cohort. Phthalate metabolites were measured in maternal urine collected at midpregnancy and concentrations were divided into 5 equal groups, called quintiles. Children of mothers in the highest quintile of the sum of metabolites of the phthalate bis(2-ethylhexyl) phthalate (DEHP) had 2.99 (95%CI: 1.47-5.49) more odds (see chapter Quantifying disease and associations) of an ADHD diagnosis in comparison to the lowest quintile.  

 

2.1.5 Ecological study design

All previously discussed study designs deal with data from individual participants. In the ecological study design data at aggregated level is used. This study design is applied when individual data is not available or when large-scale comparisons are being made, such as geographical comparisons of the prevalence of disease and exposure. Published statistics are suitable to use which makes the design relatively cheap and fast. Within environmental epidemiology, ecological study designs are frequently used in air pollution research. For example, time trends of pollution can be detected using aggregated data over several time points and can be related to the incidence of health outcomes. Caution is necessary with interpreting the results: groups that are being compared might be different in other ways that are not measured. Moreover, you do not know whether, within the groups you are comparing, the people with the outcome you are interested in are also the people who have the exposure. This study design is, therefore, hypothesis-generating.

 

2.2 Experimental studies

A randomized controlled trial (RCT) is an experimental study in which participants are randomly assigned to an intervention group or a control group. The intervention group receives an intervention or treatment, the control group receives nothing, usual care or a placebo. Clinical trials that test the effectiveness of medication are an example of an RCT. If the assignment of participants to groups is not randomized, the design is called a non-randomized controlled trial. The latter design provides less strength of evidence.

 

When groups of people instead of individuals, are randomized, the study design is called a cluster-randomized controlled trial. This is, for example, the case when classrooms with children at school are randomly assigned to the intervention- and control group. Variations are used to switch groups between the intervention and control group. For example, a crossover design makes it possible that people are both intervention group and control group in different phases of the study. In order to not restrain the benefits of the intervention to the control group, a waiting list design makes the intervention available to the control group after the research period.

 

An example of an experimental study

An example of an experimental study design within environmental research is the study of Bae and Hong (2015). In a randomized crossover trial, participants had to drink beverages either from a BPA containing can, or a BPA-free glass bottle. Besides BPA levels in urine, blood pressure was measured after exposure. The crossover design included 3 periods, with either drinking only canned beverages, both canned and glass-bottled beverages or only glass-bottled beverages. BPA concentration was increased with 1600% after drinking canned beverages in comparison to drinking from glass bottles.

 


3. Confounding and effect modification

Confounding occurs when a third factor influences both the outcome and the determinant (see Figure 3). For example, the number of cigarettes smoked is positively associated with the prevalence of esophageal cancer. However, the number of cigarettes smoked is also positively associated with the amount of standard glasses alcohol consumption. Besides, alcohol consumption is a risk factor for esophageal cancer. Alcohol consumption is therefore a confounder in the relationship smoking and esophageal cancer. One can correct for confounders in the statistical analysis, e.g. using stratification (results are presented for the different groups separately).

 

Effect modification occurs when the association between exposure/determinant and outcome is different for certain groups (Figure 3). For example, the risk of lung cancer due to asbestos exposure is about ten times higher for smokers than for non-smokers. A solution to deal with effect modification is stratification as well.

 

Figure 3: Confounding and effect modification in an association between exposure and outcome. A confounder has associations with both the exposure/determinant and the outcome. An effect modifier alters the association between the exposure/determinant and the outcome.

 

Box 1: Hill’s criteria for causation

With epidemiological studies it is often not possible to determine a causal relationship. That is why epidemiological studies often employ a set of criteria, the Hill’s criteria of causation, according to Sir Austin Bradford Hill, that need to be considered before conclusions about causality are justified (Hill, 1965).

  1. Strength: stronger associations are more reason for causation.
  2. Consistency: causation is likely when observations from different persons, in different populations and circumstances are consistent.
  3. Specificity: specificity of the association is reason for causation.
  4. Temporality: for causation the determinant must precede the disease.
  5. Biological gradient: is there biological gradient between the determinant and the disease, for example, a dose-response curve?
  6. Plausibility: is it biological plausible that the determinant causes the disease?
  7. Coherence: coherence between findings from laboratory analysis and epidemiology.
  8. Experiment: certain changes in the determinant, as if it was an experimental intervention, might provide evidence for causal relationships.
  9. Analogy: consider previous results from similar associations.

 

References

Bae, S., Hong, Y.C. (2015). Exposure to bisphenol a from drinking canned beverages increases blood pressure: Randomized crossover trial. Hypertension 65, 313-319. https://doi.org/10.1161/HYPERTENSIONAHA.114.04261

Benjamin, S., Masai, E., Kamimura, N., Takahashi, K., Anderson, R.C., Faisal, P.A. (2017). Phthalates impact human health: Epidemiological evidences and plausible mechanism of action. Journal of Hazardous Materials 340, 360-383. https://doi.org/10.1016/j.jhazmat.2017.06.036

De Cock, M., De Boer, M.R., Lamoree, M., Legler, J., Van De Bor, M. (2016). Prenatal exposure to endocrine disrupting chemicals and birth weight-A prospective cohort study. Journal of Environmental Science and Health - Part A Toxic/Hazardous Substances and Environmental Engineering 51, 178-185. https://doi.org/10.1080/10934529.2015.1087753

De Cock, M., Quaak, I., Sugeng, E.J., Legler, J., Van De Bor, M. (2016). Linking EDCs in maternal Nutrition to Child health (LINC study) - Protocol for prospective cohort to study early life exposure to environmental chemicals and child health. BMC Public Health 16: 147. https://doi.org/10.1186/s12889-016-2820-8

Engel, S.M., Villanger, G.D., Nethery, R.C., Thomsen, C., Sakhi, A.K., Drover, S.S.M., … Aase, H. (2018). Prenatal phthalates, maternal thyroid function, and risk of attention-deficit hyperactivity disorder in the Norwegian mother and child cohort. Environmental Health Perspectives. https://doi.org/10.1289/EHP2358

Hill, A.B. (1965). The Environment and Disease: Association or Causation? Journal of the Royal Society of Medicine 58, 295–300. https://doi.org/10.1177/003591576505800503

Hoffman, K., Lorenzo, A., Butt, C.M., Hammel, S.C., Henderson, B.B., Roman, S.A., … Sosa, J.A. (2017). Exposure to flame retardant chemicals and occurrence and severity of papillary thyroid cancer: A case-control study. Environment International 107, 235-242. https://doi.org/10.1016/j.envint.2017.06.021

International Epidemiological Association. (2014). Dictionary of epidemiology. Oxford University Press. https://doi.org/10.1093/ije/15.2.277

Last, J.M. (2001). A Dictionary of Epidemiology. 4th edition, Oxford, Oxford University Press.