EBH R3 Populations and Samples

Objectives
After studying this section you will be able to define a population define the sorts of sample commonly used in research give examples of each give the advantages and drawbacks of each say how bias can be introduced by some sorts of sample

What is a population?
When we use the word population in everyday speech, it usually means the population of a region a country or province, or even just a village. But in research it has a wider meaning; applied to people it means any group of people who have a common characteristic. The population of a city share the characteristic of living in the same city, but we also use the word population to describe groups like working mothers preschool children patients being treated for hypertension rst-time admissions to coronary care. As you can see, we would have to be careful to specify exactly what we meant by "working" (How many hours a week? Does unpaid work count?) or any of the other words we have used to describe the populations. A population need not be people. You will encounter the term used in research papers in the statistical sense, which is wider than the meaning of population in everyday language. In research, a population is a definable set, thats all. It can be people, animals, things or even events. So if you are interested in researching the fat content of snack foods, your population is snack foods; and if you are interested in studying the precipitants of cardiac arrest, your population is a population of eventscardiac arrests.

Learning about populations by taking samples

In medicine, we commonly use samples. We measure lipids or blood gases using samples. When you think about it, a biopsy is obviously a sample, but so too is an ECG, which R3 Samples and sampling

takes a sample of your hearts activity. A blood pressure reading is also a sample. So many of the ideas of sampling will already be intuitively familiar to you, and you will be able to appreciate the principles of good sampling if you simple remember that sampling in the statistical sense is the same sort of activity as sampling in the clinical setting. Whether you are sampling a persons blood pressure to decide if they need treatment or sampling the children of a rural community in the Rift Valley to see if they meet the target level of immunisation, you will have the same concerns: 1. You will want to make sure that your sample is big enough to allow you to measure things precisely (without too much error). Clearly one blood pressure reading of 160/100 isnt enough to warrant starting on antihypertensive treatment. Because blood pressure is quite variable from moment to moment, you would like a bigger sample. Similarly, if you are going to begin a vaccination campaign, you would like to make sure that your sample was big enough to give you a precise enough idea of the present level of immunisation to allow you to set a target for the programme. One of the important questions in sampling, then, is how precise do my results need to be? The size of the sample will determine the precision of the results.

2. You want, above all, to make sure that the sample is representative of the population. In clinical terms, you are careful to take the blood pressure when the person is resting, with a cuff of appropriate size, and so on. Likewise, when we are drawing a sample of children, we want to make sure that they are representative, so that conclusions based on the sample will not be biased. You will meet bias a lot on this course, but for the moment you should just note that a representative sample is an unbiased sample. An unrepresentative sample is a biased sample. And no matter how big it is, a biased sample gives biased results. So the most important thing about a sample is making sure it is representative.

What makes a sample representative?

Probability and non-probability samples There are two ways of thinking about representativeness. A group of people may be represented by, for example, elected representatives, or community leaders who speak on their behalf. So if we are researching the health needs of a group of people, it is important to talk to representatives of the group, who will articulate the needs and concerns of the people they represent. In this sense, a representative sample is made up of people who can speak on behalf of the group they represent. Such people may be community leaders or key informants or participants in focus groups. They are used in qualitative research, and the information they provide is vital in understanding the needs and experiences of the population. These samples are called nonprobability samples.

R3 Samples and sampling

For research aimed at using a sample to generalise to the population, a representative sample is somewhat different. Here, a representative sample is a sample taken in such a way that there is no systematic difference between the sample and the population. These samples are drawn so that each member of the population has a known probability of being included in the sample, hence their nameprobability samples

Random sample: an equal probability sample

The best sort of representative sample is a random sample. This is a subset of the population drawn so that every member of the population had an equal chance of being selected, and where the selection of one person does not affect the chances that another person will be selected. A random sample is hard to achieve unless you have a list of everyone in the population. Such a list is called a sample frame. If you have not got a sample frame, you may have to draw some other sort of sample.

Cluster (multi-stage) sample

Sometimes a random sample is possible, but not practical, because the area you are working in is simply too large. A national survey of 400 schoolchildren using a random sample might have two children from one village, another one from a village miles away and so on. In this case, cluster sampling is short-cut to random sampling. You choose units (schools in this case) at random from the area you are studying. Then you carry out a random sample within each unit. For example, you might sample ten districts, and within each choose five schools at random and nally select twenty pupils at random from each school. This sort of sampling is also known as multi-stage sampling. If you use cluster sampling, the problem is that participants within clusters are more alike than participants randomly sampled from the population. Ten children from a particular school are more alike than ten children randomly sampled from the whole country. This will give you a falsely low estimate of the amount of variation in the population. To compensate for this, you need special statistical methods that estimate the extent of the similarity within clusters and then correct for it. Cluster samples are not only useful when the size of the area to be studied makes a simple random sample impractical. They are also useful when there is no overall sample frame. In the case of schoolchildren, there is no national register of all the names of schoolchildren, so you cannot take a random sample. However, you will be able to obtain a list of all the schools in each district, and use these to select districts and schools at random. Each school will have a list of pupils which you can use for the final stage of the sample selection. This also applies to studies of patients. There is no national register of patients being treated for depression in general practice. But you can get a sample frame of general practitioners and use it to take a random sample. Within each general practice you will have to identify the patients who are eligible for your study and then sample them, but this is less of a task than drawing up a national register! R3 Samples and sampling

Systematic samples
In some cases, there is an easy short cut to getting a representative sample. Suppose you are surveying10% of the staff of a hospital by putting a questionnaire in their pay notifications. You could begin by drawing a random sample from the payroll and then putting the questionnaire in each. But an easier way would be to put a questionnaire in every tenth envelope. This gives you a representative sample, since it is not likely that there is something unusual about every tenth person on the payroll. It is not strictly a random sample, since the selection of the first participant affects the chances of everyone else being selected, but this is not a worry, since it is unlikely to introduce bias. There are many tricks to draw systematic samples that save time and effort. You can get a systematic sample of 20% of hospital patients by taking patients whose chart number ends in, say, a 3 or a 5. You can get a sample of about a third of students by getting those whose birthday falls between the first and the tenth of the month.

Stratified samples
One more sampling technique needs mentioning. This one comes about when you want to compare several groups within the population but these groups are not equally common. For example, you might want to compare job stress in male and female medical consultants. A simple random sample of medical consultants will have fewer women than men. If you want to compare male and female consultants, you may decide to pick half your sample from the women and half from the men. This gives you a better ability to detect differences between the two groups. A stratified random sample is a random sample from each of a number of subgroups in the population. You can also stratify other kinds of samples. A stratified consecutive series (see below), for instance, would be where you decided to study the next 100 people with cardiac chest pain and the next 100 people with non-cardiac chest pain presenting at the emergency department.

Nonprobability samples in quantitative research

Consecutive series
In some cases, the sample frame will not be available because the sample happens in 'real time'. There is no list of the people who will call the emergency services in the next year, or who will attend their doctor. To study these kinds of populations, we have to use a consecutive series. We might decide to document all calls to the emergency services or all patients attending casualty with chest pain for a period of time. There is one obvious problem with this sampling method. If the period of time is unrepresentative, so is the sample. A sample of people attending emergency services with chest pain will look different in Winter, when chest pain due to heart disease is commoner than in the R3 Samples and sampling

Summer. Seasonal variation in the makeup of the population can make the sample unrepresentative. There is also a less obvious problem. If you decide to survey general practice attenders by sampling everyone who attends in the next month, the most frequent attenders have the highest chance of getting into your sample. A person who attends the doctor once a year has only one chance in twelve of being in your sample, while a person who attends twelve times a year is pretty likely to attend during the sample month. These problems mean that we have to be very careful about generalizing from consecutive series samples. You can also combine systematic sampling and a consecutive series by taking every Nth person. This is usually done to reduce the workload. If you had to document every patient in casualty, the work of the department might grind to a halt! But if you only have to document every tenth person, it's less burdensome. However, staff have been known to manipulate the waiting list to make sure that bothersome or difficult patients don't get into the study! A consecutive series isn't a probability sample: it contains an unknown amount of bias because you cannot tell how representative the series is.

Snowball samples
In some populations it is hard to access the members because they have legitimate reasons for not wanting to be identified. Typical cases would be researching sex workers, illegal migrants, substance abusers etc. The researcher may start by recruiting a number of people and ask them to pass copies of the questionnaire to others whom they know, who pass it on to yet others. In this way, people receive copies of the questionnaire from someone they know and are more inclined to trust. This allows researchers to find out about the normally 'invisible' sectors of these populations. While the sample method will contain bias to an unknown degree, the method is valuable in telling us about people who otherwise would not appear in research at all. By comparing the initial participants with those recruited through the social network, researchers can try to understand similarities and differences between the 'known' and 'unknown' members of the population.

Quota samples
If you cannot get a sample using any of the methods above, you can try getting a quota sample. This is where the people collecting the data are told to get a certain number of subjects in each of a number of categories. For instance, you might be interviewing shoppers and asking them about nutrition labelling on food. A quota sample approach would stipulate that each interviewer had to run 50% of their interviews on men, and within each sex, 20% within each of ve age-bands. This makes sure that you dont end up with lots of data on people who are easiest to interview, but its otherwise not very recommendable. You have no way of knowing how representative your sample is.

Grab samples (convenience samples)

Finally, you can simply send out interviews and pick whatever people you nd. This is called a grab sample. Its only use is in a procedure called Lot Quality Assurance, which is used in quality control in industry and in programme monitoring in developing countries. It is an easy R3 Samples and sampling

and cheap method of getting a sample but it requires special statistical methods and can answer only a very limited number of questions.

R3 Samples and sampling