Study designs: primary research

Research can develop knowledge about the world by recording, describing and explaining events. Research can look at effects, such as what effect does sleep deprivation have on driving ability. It can also look at relationships, such as what the relationship is between quality of sleep and well-being. Research can also collect data to understand more about something without trying to answer a set question, such as collecting data on the opinions of a group of people over time.

There are two approaches to research: quantitative and qualitative. There are many overlaps between the approaches, no undisputed definition of them, and researchers may use both approaches in the same study (mixed methods). However, these approaches have different views about the aim of research. Quantitative research strives for objectivity and is often more concerned with the generalisability and replicability of research findings of research findings to other contexts [4]. Qualitative research values subjectivity and is more concerned with the data being an authentic and trustworthy reflection of the circumstances in which it was collected [5].

Researchers choose an approach and a study design based on what knowledge they are trying to develop. The following section explains some of the most common study types, including discussing what different studies can and cannot tell you. For simplicity, study designs have been divided into experimental, quasi-experimental and observational studies in the section below. However, researchers may use a combination of these types within one study and there are overlaps between the types.

Experimental

The first thing most people imagine when they think of an experiment is a study carried out in a laboratory. For example, a person might think of a chemistry experiment where there is almost complete control over everything, from the temperature of the room to the exact amount of each chemical being used. It is a lot more difficult to have this level of control in research that involves living things (such as humans), especially if the research takes place outside of a laboratory in the real world.

Experimental study designs look at the effect of a factor that has been manipulated by a researcher (the independent variable) on an outcome (the dependent variable). For example, a study could look at the effect of a new medicine (the independent variable) on cholesterol (the dependent variable). Because a researcher does not have complete control over everything, there are other things that may influence the outcome being measured (extraneous variables). Extraneous variables in the cholesterol medicine study above could include participants taking other medicines or variation in the weights of participants. The effects of these factors can be eliminated if they are held constant (or ‘controlled for’). In a study of a new cholesterol medicine the extraneous variable of other medicines could be controlled for by ensuring that participants are taking no other medicines during the study. However, controlling for these factors can create other problems. For example, a group of people taking no medicine apart from the study’s cholesterol medicine are likely to be healthier than the group of people the new medicine was made for. Additionally, other extraneous variables, such as weight differences or weight fluctuations over time, are almost impossible to hold constant.

In the world outside of the chemistry lab, a perfect experiment is impossible because a researcher cannot ensure that everything is the same apart from the independent variable that they are interested in. The perfect experiment would require running a study across identical parallel universes where the independent variable is the only factor that differs. For example, in a study of a new medicine, a chosen participant would be given the medicine in one universe but not the other. All extraneous variables would be the same in both universes, so they would all be perfectly controlled for. Cholesterol level would be measured at the end of the study in each universe and the difference between the two levels could be directly attributed to the medicine. This parallel universe where the participant receives no new medicine is the counterfactual. It reveals what would have happened if nothing had been changed. Studies use different ways to get as close as possible to comparing reality to the counterfactual.

An experiment has two key features that reduce the influence of extraneous variables and create a situation as close to the impossible perfect experiment as possible:

The independent variable is manipulated between two or more groups. In a two-group experiment, one group is the intervention (or treatment) group and one is the control group. The intervention group receives the treatment being tested (for example, a new medicine or training programme). The control group either receives no treatment or a standard treatment that is already being used (such as the current standard medicine used). When there is more than one intervention, researchers may be measuring different doses of an intervention (such as how much new medicine is most effective) or different interventions (such as comparing multiple new medicines). In an experiment, there is always a separate control group.

The individuals are randomly assigned into the groups, so every individual has a chance of being in either the control or intervention group [6-8]. This means that any one of thousands of factors that might affect the dependent variable are randomly distributed between the two groups, even if the researcher does not know what those factors are. In an experiment, individuals are always assigned randomly into control/intervention groups.

These experiments carried out in the real world are called Randomised Controlled Trials (RCTs) because they have a control group and the intervention is assigned randomly. Randomisation in RCTs can be done at an individual level, where people are randomly assigned into control or intervention groups. It can also be done at a group level, where people are already in groups, such as schools or hospitals, and these groups are randomly assigned to be control or intervention groups. This second type, where randomisation happens at a group level, is a cluster randomised controlled trial, also known as a group-randomised trial or place-randomised trial [9, 10].

RCTs are the closest researchers can get to the perfect experiment because the control group acts as the counterfactual. Random assignment is the reason researchers can treat the control group as the counterfactual. Instead of a researcher selecting which individuals receive treatment, where there may be bias in selection, a randomisation procedure (such as a lottery) is used so that assignment is entirely down to chance [11, 12]. Randomisation should ensure that known and unknown extraneous variables (for example, the age, weight, height or sex of individuals) are distributed randomly across the groups, reducing the likelihood that they will affect the outcomes.

RCTs can be used in a wide variety of research areas, including medicine (where they are the standard in Phase III clinical trials, see POSTnote 561), economics, experimental psychology and social policy [11]. For example, app and web designers often use a type of RCT (known as A/B testing) when trying out new content. Users of the app/website are randomly assigned into a group and will either see an old design or a new design. The designers then look at how the new design compares to the old design for certain outcomes (for example, how long a user stays on the page or where they click). A/B testing is used by the UK Government in the design of their webpages on gov.uk and has been used to trial new content, such as contact information or organ donor registration.

What can RCTs tell us?

When well-designed and well-run, RCTs are considered by many to be the most robust type of primary research to establish cause and effects [17, 18]. Robustness is made up of a combination of things that mean a person can be confident in drawing conclusions and making inferences from the evidence. Things that contribute to robustness include having a counterfactual, using data of adequate quality and quantity, and limiting the effect of extraneous variables on the measured outcome (see part two of this briefing for more information on interpreting research evidence [19]). Well-run RCTs can offer robust data on the effect of an intervention on an outcome because known and unknown extraneous variables are distributed randomly across the groups, reducing the likelihood that they will affect the outcomes [20-23]. This allows for a direct comparison between the intervention and the control group.

Although RCTs are used in many different areas, there are some instances where the use of an RCT is impossible, impractical or unethical [24-26]. For example, it is impossible to randomly assign participants to some variables, such as age or sex. Therefore, a study where something like age is the independent variable could never be an RCT. Additionally, some variables could be assigned randomly, but doing so would be unethical. A study of the long-term effects of malnutrition would not randomly assign individuals into groups of nourished and malnourished because of research ethics. Another ethical issue that some people raise about RCTs is the unfairness of giving treatment to one group but not another. However, first, when running an RCT it is not guaranteed that an intervention will have positive outcomes (and indeed, there may be unforeseen negative outcomes). Second, there are ways of running RCTs where eventually all participants receive the treatment, known as a stepped-wedge trial [27]. For explanations of other sub-types of RCTs, some common criticisms of RCTs and responses to these criticisms, see the Alliance for Useful Evidence’s Experimenter’s Inventory [28].

Key concept 1: Why does randomisation matter?

One way to control as many extraneous factors as possible is randomisation. Instead of a researcher selecting which individuals receive treatment, where there may be bias (selection bias), a randomisation procedure (such as a lottery) is used so that assignment is entirely down to chance [29, 30]. Randomisation allows a control group to act as the counterfactual because extraneous variables (for example, the age, weight, height or sex of individuals) have been controlled for by randomisation and, therefore, the only differences between the two groups should be the independent variable being examined. For example, imagine a secondary school has a population of 1,000 students with 200 in each of the five years (Year 7 to Year 11) and there are an equal number of boys and girls in each year group. If we randomly assign all these students into two groups, then most of the time we would expect the two groups of 500 students to reflect the distributions found in the whole school population. We would expect that girls would make around half of each group and that Year 7 pupils would make up around a fifth of each group. It would be highly unlikely for randomisation to result in one group being all girls or one group having no Year 7 students. We can also assume that characteristics that we might not think are important (such as favourite colour or being right-/left-handed), but that could be extraneous variables, have been distributed between the two groups roughly equally through randomisation. The number of individuals (sample size) in each group influences how likely it is that the groups will reflect the population. For example, if just two students (one boy and one girl) were randomly assigned into two groups then one group would be 100% female and one would be 100% male. This would not reflect the distribution found in the whole school population. However, if the group sizes were expanded to 100 people, it is far less likely that that random assignment would produce a group that was 100% male and 100% female. If the group sizes were expanded to include the entire school (1,000 people), then the likelihood that the groups would be 100% male and 100% female would reduce again. The more individuals in the sample, the more likely it is that the groups will reflect the distribution found in the population. This holds true for any other extraneous variables, such as height, weight or age.

Quasi-experimental

It is sometimes impossible, impractical or unethical to use an RCT to answer a research question [31-33]. Sometimes researchers are interested in the effect of a factor (independent variable) on an outcome (dependent variable), but the independent variable is impossible to assign people to randomly (such as age). Sometimes it would be unethical to assign people to the independent variable randomly (such as malnutrition). For such situations, quasi-experimental designs are used instead. A quasi-experiment differs from an experiment because it does not meet both of the two key features of an experiment, meaning:

1. There may not be a separate control group
2. If there is a control group, the independent variable is not assigned randomly between groups

The lack of a randomised control group in quasi-experiments makes the data from them more liable to be influenced by extraneous variables. Some of the most common quasi-experimental study designs are explained below. The first type (non-equivalent groups) has a control group but does not use randomisation to assign participants into groups. The second type (one-group) does not have a separate control group.

Key concept 2: Is it correlation or causation?

Research studies usually try to answer questions. These questions often boil down to ‘does A cause B?’ However, not all study designs can support a causal relationship. Instead, they may only reveal if there is a correlation between two factors. Correlation is the first step towards proving causation, but it is not sufficient by itself. Showing that A causes B requires studies to show that A and B are correlated, that A happens before B, and that all other likely causes for B (apart from A) have been ruled out. Many studies that show correlation (meeting the first step towards causation) do not show that A happens before B and/or eliminate all other likely causes. Indeed, just because there is a correlation does not mean there is any association between two factors whatsoever. There is a correlation between the amount of cheese eaten year and the number of deaths per year from becoming entangled in bedsheets [34]. Does that mean eating cheese causes death by bedsheet entanglement? Unless researchers had a pre-existing hypothesis as to why there might be a link between cheese-eating and bedsheet-death, it is highly unlikely that there is genuine association between these two factors. Instead, it is probably a spurious correlation (a coincidental similar pattern between two unrelated factors). There is a correlation between the number of umbrellas sold in a month and the average rainfall for the month. Does this mean that umbrellas make it rain? Clearly, this is not the case. Looking at the data on a smaller timescale (for example, daily or hourly) it would be clear that A does not happen before B. Instead, B (it raining) happens before A (people buying umbrellas). There is also a correlation between the amount of ice cream sold in a month and the number of deaths by drowning. Does this mean that ice cream causes death by drowning? Again, this is clearly not the case. Instead, there is likely to be another factor that happens to correlate with both A (ice cream sales) and B (deaths by drowning). Higher average monthly temperature is correlated with higher ice cream sales and also with higher drowning rates (because people are more likely go swimming when it is warm). In this case there is another probable causative factor at play. Well-run observational studies are able to find evidence of associations. For example, observational studies have found strong associations between cannabis use and psychosis [35, 36]. However, as the evidence comes from observational studies, it would not be possible to say from just one study showing correlation whether there is a causal relationship, or which way the causal relationship might go. Any of the following may be true: Using cannabis increases the chance of developing psychosis. Being prone to developing psychosis increases the chance of using cannabis. Cannabis use and likelihood to develop psychosis have no relationship to each other but are simply correlated with another third causative factor, such as another health condition or certain life circumstances (which itself may or may not be causative). Although some study types cannot prove causation by themselves, they can create a body of evidence that begins to point to causation. For example, nowadays it is widely accepted that smoking causes various types of cancer. The evidence for this causality came from observational studies, mainly case-control studies.

Non-equivalent groups

A non-equivalent groups design is like an RCT, in that it has an intervention group and a control group being compared on a particular outcome (dependent variable). However, unlike RCTs, the independent variable is not randomly assigned. This may be because it is impossible to assign the variable randomly (such as sex or age) or unethical to assign the variable randomly (such as malnutrition or smoking status). Non-equivalent groups may also be used when an RCT is deemed too impractical, costly or time-consuming; although advocates of RCTs note that there are ways to run RCTs cost- and time-efficiently [37, 38]. Because the participants have not been assigned to groups randomly, the researcher cannot assume that extraneous variables (other things that may influence the outcome being measured) have been randomly distributed across groups, meaning their effect on outcomes has not been limited.

Where the independent variable is an inherent characteristic (such as sex or height), there may be other factors related to this characteristic that bias the results. For example, if a study wanted to compare the effect of a new reading programme on children with English as a first language and those with English as an additional language then there are several extraneous variables that could bias the results. For example, children with English as an additional language are more like to live in certain geographical areas [39]. Children with English as an additional language are also more likely to have been born abroad or have parents who have been born abroad, both of which are associated with an increased likelihood of living in poverty [40]. These are examples of known extraneous variables that could influence outcomes. However, there are also many other unknown factors that might influence outcomes besides the independent variable of English as first or additional language. Because of these challenges, researchers using a non-equivalent groups design try to limit the influence of extraneous variables as much they can by trying to keep the two groups of participants as similar as possible. For example, in the study above they might limit the study to one geographical area. However, it is not possible for researchers to control for all the potential unknown extraneous variables.

What can non-equivalent groups designs tell us?

Non-equivalent groups designs are the closest that researchers can get to the robustness of RCTs when randomisation is impossible, unethical or impractical [41]. The presence of a control group creates more robust evidence than designs without a control group. However, as participants have not been randomly assigned to groups, there is a higher possibility of extraneous variables influencing the data. This means that a researcher cannot be as certain of causation (that the independent variable has caused the measured change in the dependent variable) as they could be if they had used an RCT.

One-group designs

One-group designs do not have separate control and intervention groups. This may be because they have no control group at all, as found in the one-group before-and-after study design. For example, if a study was looking at the effect of a new reading intervention for children, a researcher might measure child’s reading ability before and after the reading intervention then compare the results. If children’s reading ability was higher after the intervention, a researcher might want to conclude that the reading intervention was effective. However, there are issues with the one-group before-and-after study design that means it is difficult to draw such conclusions. The main problem is without a control group, there is no way of knowing what might have happened without the intervention. The children may have just got better at the reading test over time for reasons that are unrelated to the reading intervention. Perhaps they improved naturally as the test got more familiar, for example. The researcher cannot confidently conclude that the changes measured are due to the reading intervention. There are particular issues with before-and-after studies that are conducted when there is an extreme measure (such as being carried out in an extremely low-performing school or hospital) as they are vulnerable to regression to the mean.

Key concept 3: If measures are extreme, is it regression to the mean?

When there is an extreme measure (one that is particularly high or particularly low), it is highly likely that the next measure will be closer to average. This phenomenon is called regression to the mean. Sometimes there are random highs and random lows in data and, when these happen, researchers should not put too much emphasis on them. Interventions that specifically target high- or low-performing individuals, areas or organisations may be at greater risk of ‘successes’ being partly or wholly due to regression to the mean. For example, a local police force might decide to site mobile speed cameras in an ‘accident blackspot’ to reduce the number of car crashes. They look at the data on accidents in the local area for three months of the year and find one junction had 25 crashes, far more than anywhere else. The police force put mobile speed cameras at the junction and over the next three months, the number of accidents falls to 11. They consider this a success and proof that mobile speed cameras reduce accidents. However, by choosing a location that is an outlier (because it had far more crashes than anywhere else) with an extreme value, they have increased the likelihood that their results are affected by regression to the mean. The differences between the number of accidents before and after the speed cameras might just be because of natural variation and might not be attributable to the intervention. Researchers can try to compensate for regression to the mean by having collected data for longer before and after the intervention. This is important because it reduces the likelihood that changes after the intervention are just random movements. For example, if the local police force had looked at the quarterly data from the last two years (table below), they would have seen that the 25 accidents in a quarter was a random extreme and that the ‘fall’ in accidents after the intervention was just a return to normal.

Another form of one-group design (known as a within-subject design or repeated-measures design) aims to overcome issues caused by extraneous variables by using the same individuals for intervention and control. For example, a researcher might want to know whether participants prefer the taste of the most popular instant coffee on the market (control) or a new instant coffee (intervention). A researcher could just ask participants to try the intervention coffee and rate it out of ten, but without a control the researcher cannot make inferences about how good the coffee is. If someone rated the coffee two out of ten, it might be because they hate coffee or it might be because the new coffee is very unpleasant. Therefore, in a within-subject study, participants are given every intervention and control to produce a comparison. In this instance, they would taste the control coffee and rate it out of ten. They would also taste the intervention coffee and rate it out of ten. This way, extraneous variables (such as how much each participant likes coffee) are controlled for. The researcher can directly compare which coffee each participant prefers because the participants provide their own counterfactual.

In some circumstances, within-subject designs can be preferable to an RCT. For example, a coffee tasting study as an RCT would require that participants be randomly assigned to taste either the control or the intervention coffee. The randomness is what allows a researcher to accept the control group as the counterfactual and assume the characteristic of how much individuals like coffee is balanced across the groups. However, the number of participants needed to ensure that there are no group differences (such as one group just happening to prefer coffee more than the other) would be higher than required for a within-subject design.

However, the within-subject design is vulnerable to other extraneous variables. For example, an order effect is where differences in dependent variables (such as coffee scores out of ten) are influenced by what number in an order they are (for example, first, second, third and so on). This is an issue for the study described above because perhaps people generally rate the first coffee they try higher or lower than subsequent ones. Order effects can be controlled for in within-subject designs by counterbalancing. Counterbalancing ensures that all possible orders of presenting interventions and controls are used. In the coffee study, the researcher would have half of participants taste the control first and half the participants taste the intervention coffee first. This way, any effect of the order in which the coffees are presented will be balanced out.

There are also occasions where researchers may want to combine a non-equivalent groups design with a within-subjects design. This is called a mixed-factorial design. For example, if the researcher in the study above wanted to know whether there was a difference between which coffee men preferred and which coffee women preferred, then they would use a mixed-factorial design. Men and women would be in separate groups but both would try the control coffee and intervention coffee. The researcher could then conclude which coffee was preferred by men and which coffee was preferred by women.

What can one-group designs tell us?

One-group designs (especially one-group before-and-after) generally require fewer participants than non-equivalent groups designs. Sometimes it is the only study design available because of time pressures or policy changes. For example, one-group before-and-after designs are often used when evaluating public health interventions, especially when an intervention is being applied to a national population, such as a smoking ban in public spaces [42]. Because there is no control group in a one-group before-and-after study, it is impossible to know for certain what would have happened without an intervention. This means that researchers cannot be certain that the independent variable (such as policy intervention) has caused changes in a dependent variable. However, there are various techniques used in policy evaluation to increase the confidence a researcher has in attributing changes in measures after an intervention to the invention itself [43]. These can include taking measurements over a longer time frame.

Within-subjects designs allow direct comparison of a control and an intervention without concerns that individual differences will bias the results. However, unless within-subject designs are carefully counterbalanced, there is a risk that effects (such as order effects) will bias the results. Counterbalancing to reduce effects can require high numbers of participants, especially when there are numerous different interventions being tested.

Observational

Observational studies do not meet either of the criteria of an experiment, meaning:

1. They do not have a control group
2. They do not assign an independent variable

Researchers are not intervening in this study design but are instead observing what is happening [44]. There are three main types of observational studies described below: cross-sectional studies, natural experiments, and case designs. These study designs differ in how they select the participants they will be observing.

Cross-sectional studies

Cross-sectional studies look at different factors within a particular population (such as a country or region). They are observational studies that measure factors (such as health or lifestyle outcomes) without researchers intervening. Cross-sectional studies take measurements from individuals across two or more different sub-populations to get a ‘snapshot’ of certain factors. Cross-sectional studies collect data from a group of participants to describe a certain population, such as recording the alcohol consumption of different university students. They provide a snapshot of a certain factor in a certain population at a particular point in time. These studies can be repeated at a future point with a new set of participants (such as looking at alcohol consumption of university students every 5 years, with a new group of students each time). These are called repeat cross-sectional studies. For example, since 1983 the British Social Attitudes survey has been asking questions to a different group of 3000 participants each year chosen at random to be representative of the UK population [45].

What can cross-sectional studies tell us?

Cross-sectional studies offer more detail on a ‘snapshot’ of a particular population at a certain time than other one-off studies tend to. Repeat cross-sectional studies provide information about trends over time and can explore whether age or generation is a factor in certain outcomes, as it is possible to compare multiple cohorts and see if trends emerge [46, 47]. However, data must be collected over a long time frame for repeat cross-sectional studies to gather relevant, long-term results. This requires that participants not drop out. Because there is no randomly assigned independent variable in cross-sectional studies, it is not possible to be certain of causation [48].

Natural experiments

Natural experiments are situations where individuals (such as two different sub-populations) have been exposed to different factors without a researcher intervening. Researchers can look at the association between different factors and a dependent variable. For example, researchers have explored the drafting of individuals into the armed forces to look at the effect of military service on lifetime earnings [49]. Natural experiments where sub-populations are exposed to factors that may affect health are called ecological studies. Ecological studies have included looking at the effect of radiation, air pollution and famine on various outcomes [50-52].

What can natural experiments tell us?

Natural experiments can investigate the effects of some factors that would be unethical or impossible to assign randomly to individuals. However, there is no control over the variables that participants are exposed to and the level of the variable may differ between individuals (for example, how much air pollution a person is exposed to). As assignment is not random, researchers cannot be confident that any differences seen are not related to extraneous variables. For example, a study might show a correlation between worse health and living near a main road. However, it may be that areas around main roads tend to be poorer and that poverty (as opposed to proximity to a main road) may be the crucial factor in having worse health.

Case designs

Case designs can be used when there are individuals who have been exposed to a certain factor [53]. Case designs involve carrying out an in-depth study looking at outcomes related to a particular factor, such as the relationship between sustaining a brain injury and changes in language, memory or personality [54-57]. Case reports present a detailed description of one particular individual or incident. For example, a case report might explore the symptoms and prognosis of a person with a rare illness or the economic effects on an individual following a one-off natural disaster. Case series (also known as clinical series) present descriptions of participants who have been exposed to a particular factor, such as having had a particular medical treatment or medical condition [58].

Case-control studies (also known as case-referent studies) compare the medical and lifestyle history of a larger number of participants with a particular medical condition (cases) alongside participants without that medical condition (controls) who are matched for other variables. A study might compare the medical and lifestyle histories of people who have suffered a stroke before age 50 and those who have not. By comparing their histories, potential causative factors may be revealed. For example, a case-control study in 1950 reported evidence that lifetime cigarette smoking was associated with lung cancer [59].

What can case designs tell us?

Case designs can investigate the effects of some factors that would be unethical or impossible to assign randomly to individuals. They can be the only way to research rare conditions or occurrences. They can provide in-depth information and help to identify potential consequences of certain factors and interventions [60]. There is no control over variables being examined in case designs, and levels may differ across individuals (for example, the area and extent of brain damage in case reports or the level of smoking in case-control studies). As assignment is not random, researchers cannot be confident that any differences seen are not related to extraneous variables not considered. Although case-control studies may provide evidence of a correlation between two factors that suggest possible causation, no case study design can prove causality [61].

Key concept 4: What if we give it more time?

One way researchers can collect a larger and more convincing body of evidence is to collect data over longer periods of time. A design bolstered with more observations taken over a longer period and often with a large sample size, is an interrupted time series. In this design, the dependent variable is measured over time, then an intervention is introduced and the dependent variable continues to be measured. The dependent variable is compared between pre-intervention and post-intervention to explore if there is a difference. For example, a city might want to know what effect a drink driving campaign (independent variable) has had on the number of people caught drink driving (dependent variable). If there are police records of the number of people arrested before the drink driving campaign, a researcher could look at the trends before the campaign and the trends afterwards. If there is a decrease after the intervention then the researcher might want to conclude that the campaign successfully reduced drink driving. Although this design shows correlation and indicates possible causation, without a control group, a researcher cannot eliminate the possibility that these apparently positive results are caused or influenced by extraneous variables [62, 63]. For example, perhaps the drink driving campaign required extra resources and caused a reduction in the number of police doing drink driving spot-checks. In this case, the change of trend would be due to fewer drink drivers being caught, not a reduction in drink driving. Studies may also use a longitudinal design to collect lots of data over time. Longitudinal studies measure factors in a group of people over a period of time, revisiting the same participants during their lives [64]. There are a number of different longitudinal study designs that differ in the groups they collect data from, how the data are collected and when the data are collected. Prospective longitudinal studies recruit participants at the start of the research process and collect data from these individuals over time. Panel Studies collect data over time from participants who are selected based on certain characteristics. For example, household panel studies look at how chosen households change over time. These households may be chosen to be representative of a whole population (such as an entire country) or may be chosen to represent a sub-population (such as single-parent households). Data are collected from the same participants at intervals over time. One particular type of panel study is the cohort study. Cohort studies collect a sample that is representative of individuals with a specific characteristic (such as year of birth or having a particular health condition [65, 66]). These include birth cohort studies like the National Survey of Health and Development, the National Child Development Study, the British Cohort Study, and the Millennium Cohort Study, which track the lives of a large group of people born in 1946, 1958, 1970 and 2000–2001 respectively [67, 68]. Prospective cohort studies follow a group over time to investigate the role that certain factors play on a particular independent variable. These independent variables tend to be health conditions or health behaviours, such as having diabetes or smoking. For example, in the USA the Framingham Heart Study collected data from over 5000 residents of one town aged 30–62 years to look at factors associated with cardiovascular disease [69]. Other forms of longitudinal study use official data, such as census records, to look at different factors in a particular population across time. For example, record linkage studies use data from different types of official record to track outcomes for participants [70]. Retrospective longitudinal studies begin the research process by finding data that have already been collected over time and analysing it to help answer a particular research question [71]. For example, the Hertfordshire Cohort Study used county midwife and health visitor records for babies born 1911–1939 and linked these data to death certificates to explore the role of early childhood on cardiovascular disease [72, 73].