 Waiting is used to make datasets more representative of the population that the data is supposed to measure. Many datasets include survey weights. This short video introduces key points about using weights for survey data. Why use weights? Surveys usually collect information from a sample of the population with the aim of inferring information about the population as a whole. To enable inference to the population, survey designers use random sampling methods to try to generate a representative sample and use survey weights to adjust sample data to make it better represent the population. Let's explore some different kinds of weights starting with design weights. One reason for weighting data is that common sampling methods mean members of a population can have different chances of being selected to take part. For example, surveys in the UK often draw a sample from a list of addresses called the postcode address file. To get a sample of individuals, we first sample addresses and then select individuals at random from each address. Using the sample design, individuals who live alone are more likely to be sampled than those who live with others. For example, if there is one person in a household, they have a one-in-one chance of being selected. However, if there are two people in a household, they have a one-in-two chance of being selected. By using these known differences in the probabilities of being selected, we can calculate design weights to adjust the sample to make it better represent the population. Weights can also help to compensate for survey non-response, and these are called non-response weights. Not everyone sampled will take part in a survey. Some people cannot be contacted and others refuse to take part. Response rates can vary systematically across groups in the population, that is, certain subgroups may be more or less likely to respond to the survey or certain questions in it. Non-response weights use information about response rates for subgroups to adjust data and limit potential bias. We can also use weights to adjust the sample to reflect key population proportions. Known as post stratification or calibration weights, they use information from sources such as the UK Census to improve the accuracy and precision of population estimates, and they're often used by statistical agencies. For example, here we can see a theoretical sample which is 56% women and 44% men. After adjusting with post stratification weights, this better represents the population with 51% women and 49% men. Let's take a closer look at how weights work. The weighting variable contains a value for each case which indicates how the case should be weighted during analysis. Underrepresented cases have higher weights to make them count more, and overrepresented cases have lower weights to make them count less. For example, we might find higher weights for individuals that live in larger households, or for those coming from population groups underrepresented in the sample, due to either chance or non-response. This image shows the weighting variables for individuals in the British Social Attitude Survey 2016 data set. The values that we can see are either over 1, such as 1.2732, or between 0 and 1, such as 0.5518. Methods for using weights vary by statistical software, but usually involve indicating the name of the weighting variable before or as part of the analysis. For example, this image shows the weight cases box in SPSS, with the data set weighted by entering the weighting variable. Why is weighting important? Results can vary between weighted and unweighted analysis. This example illustrates the difference between weighted and unweighted results from the quarterly labour force survey 2015. These charts show the proportion of individuals in each area or government office region. In particular, look at the difference between the unweighted and weighted data for the percentage of individuals in London. In this first graph, which is unweighted, the percentage in London is around 11%. After weighting, however, this figure is around 13%. Weighted results are more representative of the population because biases have been minimized by applying the weight. Weights can also be used to make the sample look the same size as the population. Known as grossing weights, they can help us describe the prevalence of social phenomena. For example, with the crime survey for England and Wales, they help us describe rates of crime. This example from the 2016-2017 data shows the number of respondents reporting bike theft in the previous 12 months. The unweighted sample gives a frequency of 393, and after weighting, this increases to over 596,000, which is an estimate of the total number of people who had a bike stolen in the previous 12 months. Now we'll look at some documentation and further reading. You will find information about the weights in the documentation that comes with the data. The user guide should give you details about which weight to use and when, and this can be found under the Documentation tab in the catalogue entry. Sometimes you will find multiple weights and will need to establish which to use for your analysis. For example, the Health Survey for England data sets contain a weight for use with the core interview data and weights for analysis of the follow-up stages. All the information about these weights and when to use them is in the user guide. This video has given an introduction to survey weights. To explore this topic further, you can read our detailed What Is Weighting Guide, and we also have guides to the main statistics software, which include instructions for applying weights.