7.3 Measuring variables, exposures, and outcome measures

In a cohort study, exposures and other factors are measured at baseline, and outcome measures are ascertained over the subsequent months or years. The most common type of cohort study analysis is of the association between the baseline information and the outcomes. However, some exposures, for example, diet and lifestyle (Box 7.3), could change over time, especially if there is a long follow-up period. If this is likely to occur, the exposures should be measured during follow-up (called longitudinal cohort study). Interest would be in how changes in diet and lifestyle over time influence the outcome measures (such as cardiovascular disease). Figure 10.1 is an example in which using only baseline exposure information could lead to a diluted effect if the exposure status significantly changes during a long follow-up period. Principles covered in Section 5.3 also apply to cohort studies.

Cohort studies are commonly used to estimate risk, that is, the chance of having a defined event (e.g. stopped smoking), of developing a specific disorder, or of death. Studies of risk such as these could use two types of outcome measures: ‘counting people’ and time-to-event data. However, there are also cohort studies that aim to examine how certain outcomes change over time (e.g. measures of blood pressure or body weight), and for these, there is no concept of risk.²

Exposures and outcome measures should be measured using the same methods for all participants, and when determining the outcome measure, it is best to use standard and validated criteria where possible. An established method for cause of death is the World Health Organisation ICD. Non-standard methods developed specifically for a study require careful explanation and justification, and if the method has been developed and then tested on the same dataset, there is a possibility of bias. Independent assessment of the outcome measure for all individuals may strengthen the reliability of the findings, that is, assessment by someone who does not know the exposure status of the participant or is not directly involved in the research study team (but this can be expensive to do).

7.4 Collecting the data

Data for cohort studies can be collected in a similar way to case–control studies (Section 6.4) or cross-sectional studies (Section 5.4). Unlike some case–control and cross-sectional studies, in which a proxy could provide data for the study instead of the participant, it is usual for cohort studies to obtain information directly from the participants (particularly the baseline information). The three examples in this chapter show that data were collected in a variety of ways (Figure 7.2). Prospective cohort studies require significant resources, so efficient methods are needed to capture data, especially during follow-up which can last for several years. Attempting to collect data regularly requires a balance of objectives:

• Frequent collection would be expensive, and participants may become less interested, though this approach would be less affected by having to recall information from memory.
  
• Infrequent collection is cheaper, but there is a loss of information in how the exposures and/or outcome measures change over time, and people may forget exposures/outcomes since the last assessment.

Self-completed questionnaires during follow-up are an efficient way of requesting large amounts of information but non-responders and incomplete responders are likely. Telephone surveys, where the researcher works through the questionnaire items with the study participant, could overcome this, but these require staff (increasing costs) and often a shorter questionnaire, because participants are unlikely to want to spend much more than an hour going through a survey by telephone.

In the future, cohort studies are likely to involve more information technology (see page 19).

There are two major problems with prospective cohort studies, which often increase with longer follow-up and can have an impact on the study size available for the statistical analyses and therefore reliability of the results:

1. People who are lost to follow-up: They often move address or do not respond to correspondence, and the researchers cannot trace them. This problem is difficult to overcome, particularly if people move abroad. Participants could be reminded during follow-up to inform the research team if they move.
   
2. People who withdraw, that is, those who decide to not continue in the study: All participants have a right to withdraw at any time; however, reasons for this could be ascertained. It could be lack of interest, lack of perceived personal benefit, or the assessments have become burdensome. If the latter arises, it is important to still be allowed access to medical records (and registry data) if used to provide outcome data, because individuals often agree to this.

These two features are different. Obtaining outcome measures on those lost to follow-up is usually impossible, but this is not necessarily the case for those who withdraw.

7.5 Sample size

The three examples used in this chapter are each based on many study participants. Although it is possible to conduct relatively smaller studies, for example, from a single centre in which there may only be 100 individuals to follow up, the size and duration of follow-up must still be sufficient to address the study objectives.

Many cohort studies are already ongoing, with fixed numbers of participants and events, so a sample size calculation for a specific research objective might have limited value or not done, as in two examples covered in this chapter [8,9]. However, if a sample size were estimated, and the target number of participants and/or events far exceeded that observed in the study, the researchers may still wish to proceed with the analysis.

Information needed for sample size estimation when examining associations

The principles of sample size estimation for cohort studies are similar to those for case–control studies (see Section 6.5). When the outcome measure is based on ‘counting people’ or time-to-event data, the number of events is important, often more so than the total number of participants in the study. Therefore, although it might be easy to conduct a small cohort study, finding no or few events would not provide useful information (wide 95% confidence intervals and large p-values). If the number of participants recruited is limited, extending the follow-up may increase the number of events. There is, therefore, a relationship between target sample size, expected event rate, length of follow-up, and available resources that need to be considered by researchers.

When there is a single exposure factor of interest and a single outcome measure (e.g. disorder), several pieces of information are needed for the sample size calculation (Figure 7.3). Items such as the percentage of participants who are expected to be exposed and the percentage of the unexposed group who would have the outcome should ideally come from prior information but other times are simply best guesses. Statistical packages have sample size facilities, and there are software programmes [10, 11], including those freely available for observational studies [12].

For example, in the folic acid study, Box 7.2, the study size of 85,176 had 93, 73 and 45% power to detect odds ratios of 0.50, 0.60 and 0.70 respectively, assuming 68% of the participants were exposed (ie. mothers took folic acid) and ASD prevalence of 0.13%.

Drop outs could be allowed for in the sample size. For example, if the calculation produced a study size of 1000 participants and 15% are expected to be lost during follow-up (there is no measure of the endpoint on them), the target size could be inflated to 1180 [85% of 1180 is 1000, where 1180 = 1000/(1 − 0.15)].

7.6 Analysing data and interpreting results

When the outcome measure is based on ‘counting people’ or time-to-events, the concept of risk is used. Unlike cross-sectional or case–control studies, cohort studies can be used to estimate the incidence of a disorder. When considering the influence of potential confounding factors, the same approaches are used as covered in Chapter 6 (including Figure 6.3).

7.7 Outcome measures based on ‘counting people’ endpoints: Folic acid and ASD (Box 7.2)

7.8 Outcome measures based on ‘taking measurements on people’ endpoints: Lifestyle habits and body weight (Box 7.3)

Analysing data and interpreting results

In this study the outcome measure, body weight, can be analysed using linear regression (Chapter 3, page 66).

What are the main results?

Unlike the other two examples in this chapter, in which the exposure was a single, categorical factor, in this study, there were many exposures, so it was not possible to report the baseline characteristics according to exposed and unexposed groups. Instead, the baseline characteristics table in the published paper provides a summary of the key dietary and lifestyle factors in each of the groups used in the paper.

Table 7.2 shows the main results for selected exposures (dietary items). Each item is included in the regression as a continuous measure, without being put into categories, and so a linear relationship is assumed between the dietary item and change in body weight. The effect size for each factor (obtained from the linear regression) was the change in body weight over a 4-year period. As expected, there were many potential confounding factors, and each effect size for a dietary item was adjusted for all other items. Allowing for so many factors is usually only reliable for large datasets; in smaller datasets, the regression model could ‘break down’, which could be indicated by very large or small effect sizes or CI limits (may appear as infinity). The results in this study can be assessed by examining whether the effect size materially changes after adjustment, and whether it moves closer or further away from the no effect value (Figure 6.3).

The results in Table 7.2 can be interpreted as follows:

• For every extra serving of fruit consumed each day, body weight reduced by 0.69 pounds over a 4-year period after allowing for age only. The effect after allowing for all other factors was a weight reduction of 0.49 pounds (0.22 kg), indicating that some, but not all, of the effect of the 0.69 pound loss was due to other (confounding) factors.
  
• The largest weight reduction was associated with increasing the intake of nuts by 1 serving per day: –0.57 pounds (−0.26 kg).
  
• The largest weight increase was seen for french fries. Each extra serving, per day, led to a weight increase of 6.59 pounds (3.0 kg) over 4 years, but about half of this effect was due to confounding factors; the adjusted effect size was 3.35 pounds (1.5 kg). There was a similar finding for potato chips (i.e. the effect was approximately halved, from 3.01 to 1.69 pounds.).
  
• A small/moderate weight loss was seen for consumption of low-fat dairy foods (−0.17 pounds), but this became almost no effect after allowing for other factors (−0.05).
  
• There was little/no association for cheese, either before or after adjustment.

These changes in weight might appear relatively small, in the context of gains or losses over a 4-year period, but they were associated with only one extra serving, per day, of one food item. Additional servings of the same food item, and the effects of other food/drink items, are therefore likely to have a cumulative effect on weight. It was unexpected that eating more fruit or nuts could lead to weight loss, because both add calories. It is possible that people who ate fruit and nuts reduced their intake of other foods with high calorific content, thus decreasing their total energy intake, which led to the small/moderate weight loss.

Figure 7.4 shows the independent effects of diet and physical activity on changes in body weight. Diet was quantified as an overall score, based on changes in each food/drink item. As expected, people with the ‘worst’ diet and the least amount of physical activity had the largest weight gain, on average almost 6 pounds over 4 years, compared with those in the highest category. The figure shows that food consumption and exercise had independent effects. By holding one category of diet constant (so that people had similar diet scores, and therefore diet would not act as a confounder), weight increased with decreasing physical activity (i.e. as the quintile moved from 5 to 1). The same pattern was seen when holding a category of physical activity constant.