Methods

In the simplest study, two population-based measures are required, one for the exposure of interest and the other for the health outcome.

A classic example

Ecological studies are generally the first step in exploring whether there is a differential distribution of disease among people with different risk profiles. For example, ecological comparisons showed that economically developed countries with a higher intake of dietary fat had much higher coronary heart disease (CHD) rates than countries with lower dietary fat consumption. This evidence was based on an early study analysing diets from groups of men in seven different countries (Keys et al. 1986); see Figure 2.1. These results have been challenged over the years because of difficulties in characterising the dietary intakes of the different country populations. Other types of study are needed to show causation.

A recent example

Diet features very strongly as a risk factor for top adverse health outcomes in the recently published Global Burden of Disease Study 2010 (Lim et al. 2012); see Figure 2.2. This study used published and unpublished secondary sources of data to calculate the relationships between 67 different risk factors in 21 regions and linked them with deaths or disease burden for each region between 1990 and 2010. Out of the top 20 leading risk factors contributing to the burden of disease in 2010, 6 are dietary factors (diet low in fruit, nuts and seeds, whole grains, vegetables, seafood and omega-3 fatty acids, and high in sodium) and another 7 are directly linked to diet (high blood pressure, high body mass index, high fasting plasma glucose, childhood underweight, iron deficiency, suboptimal breastfeeding and high total cholesterol). An ecological approach was employed to link risk factors to disease outcomes, using data collected via different epidemiological methods. The data do not directly link individual exposures to risk factors with the diseases of interest. Limitations include variable quality of exposure data across countries and the possibility of residual confounding (see Section 2.6), meaning that some associations could be the result of other factors that have not been considered or taken into account in the analysis.

Analysis of ecological data

The most straightforward analysis would be the calculation of a correlation coefficient between the exposure of interest and the outcome. This is a measure of the strength and direction of the linear relationship between two different continuous variables, for example energy intake and body mass index. The correlation coefficient, denoted by ‘r’, can have values between +1 (a perfect positive linear relationship) and –1 (a perfect inverse linear relationship). A value of 0 indicates no linear relationship between the two variables. An ecological analysis of 21 wealthy countries (Pickett et al. 2005) found that income inequality was positively correlated with the percentage of obese men (r = 0.48, p = 0.03). The relationship was even stronger for obese women in these countries, with a positive correlation coefficient of 0.62 (p = 0.003).

Further analysis of ecological studies could include multiple regression modelling to estimate the magnitude of associations, taking into account other factors of relevance that may otherwise confound the analysis. Confounding factors may include age and other lifestyle factors. Regression modelling can be undertaken using continuous variables as the dependent variable or outcome, such as height or weight. In this case linear regression modelling would be undertaken. When the outcome is categorical or dichotomous, such as the presence or absence of a disease, then logistic regression is appropriate. A study of routine data from South Australia used logistic regression analysis to assess factors that might affect food security, a dichotomised outcome (Foley et al. 2010). Food insecurity was highest in households with low levels of education or limited capacity to save money, and in Aboriginal households and those with three or more children.

Problems with ecological analyses

The ‘ecological fallacy’ is the major trap for the unsuspecting researcher. This occurs when relationships that are observed for groups are assumed to hold for individuals. For example, ecological analysis has shown that countries with more fat in the diet have higher rates of breast cancer, suggesting that women who eat fatty foods would be more likely to develop breast cancer. This assumption is only weakly supported by case-control and cohort data. Correlations found in ecological analyses may be due to confounding by other related factors that have not been controlled for, some of which may be difficult to measure at the population level. Age standardisation often needs to be undertaken, since countries may have very different age profiles. This process adjusts disease rates to a standard population, allowing comparisons to occur. When disease rates are age standardised, any differences in the rates over time or between geographical areas will not simply reflect variations in the age structure of the populations. This is important when looking at disease rates because some conditions, such as cancer, can predominantly affect the elderly. So if rates are not age standardised, a higher disease rate in one country may simply reflect the fact that it has a greater proportion of older people. Additionally, the quality of diagnostic data can differ widely between countries and over time.

A recent example

A cross-sectional analysis of data from older people in the Singapore Longitudinal Ageing Study found that higher measures of fasting homocysteine and low folate were negatively associated with measures of performance-oriented mobility and activities of daily living (Ng et al. 2012). Although these results are suggestive of a relationship in the direction of poorer nutrition to poorer physical function in older people, it is not possible to claim causality, primarily because temporal relationships between exposure and disease were not examined. It is equally plausible that older people with poorer physical functioning have a poorer diet and therefore a worse nutritional status. In order to prove cause-and-effect relationships a different type of study, a randomised controlled trial, would be needed.

Methods

Describing population characteristics

The major nutrition survey conducted in the UK is the National Diet and Nutrition Survey (NDNS). It is a rolling programme that began in 2008 and collects nationally representative dietary data from 1000 individuals per year aged 18 months and over from private households. The National Health and Nutrition Examination Survey (NHANES) is a major rolling programme of survey data collection in the USA that began in the early 1960s. About 5000 individuals are surveyed each year. The sample is selected to represent the US population of all ages. To produce reliable statistics, NHANES over-samples people 60 and older, African Americans, Asians and Hispanics.

There are two major aspects of national nutrition surveys that are important with respect to data collection: cost and organisation. Data should be as nationally representative as possible and also be as accurate and complete as possible (Stephen et al. 2013). In the NDNS, national representation in terms of age, gender and region is achieved by randomly selecting postcodes and addresses from the UK population as a whole (Figure 2.3). The NDNS currently uses the four-day estimated diary to assess diet. This is a compromise between detail and respondent burden. Respondent burden is particularly important to consider in large-scale surveys of this kind. A high level of low-energy reporting has been found in a previous national survey of older British adults that used four-day weighed diaries, which was considered to be a result of the weighed intake method and reluctance to report consumption of unhealthy food.

Three large cross-sectional data sets from the USA, including NHANES, were used to explore causes of changing energy intake in children from 1977 to 2010. Changes in the number of eating/drinking occasions per day and portion size per eating occasion were the major contributors to changes in total energy intake per day (Duffey and Popkin 2013).

Migrant studies

Cross-sectional analyses of migrants, comparing populations migrating from rural to urban areas or migrating between countries, have been undertaken to explore the associations between genetic background and environmental exposures in relation to risk of disease. Rural–urban migrants experience rapid environmental changes associated with urbanisation, enabling epidemiological transitions to be examined. Changes seen in migrants over relatively short time periods may therefore provide insights into wider population health changes. The Indian Migration Study (Bowen et al. 2011) explored the impact of migration to urban areas on dietary patterns, comparing migrants with their rural siblings. Migrant and urban participants reported up to 80% higher fruit and vegetable intake than rural participants (p = 0.001) and up to 35% higher sugar intake (p = 0.001). Meat and dairy intake were higher in migrant and urban participants than in rural participants (p = 0.001); see Figure 2.4.