Administrative data is the term used to describe everyday data about individuals collected by government departments and agencies. Examples include exam results, benefit receipt and National Insurance payments.
Attrition is the discontinued participation of study participants in a longitudinal study. Attrition can reflect a range of factors, from the study participant not being traceable to them choosing not to take part when contacted. Attrition is problematic both because it can lead to bias in the study findings (if the attrition is higher among some groups than others) and because it reduces the size of the sample.
Body mass index is a measure used to assess if an individual is a healthy weight for their height. It is calculated by dividing the individual’s weight by the square of their height, and it is typically represented in units of kg/m2.
Cohort studies are concerned with charting the lives of groups of individuals who experience the same life events within a given time period. The best known examples are birth cohort studies, which follow a group of people born in a particular period.
Complete case analysis is the term used to describe a statistical analysis that only includes participants for which we have no missing data on the variables of interest. Participants with any missing data are excluded.
Conditioning refers to the process whereby participants’ answers to some questions may be influenced by their participation in the study – in other words, their responses are ‘conditioned’ by their being members of a longitudinal study. Examples would include study respondents answering questions differently or even behaving differently as a result of their participation in the study.
Confounding occurs where the relationship between independent and dependent variables is distorted by one or more additional, and sometimes unmeasured, variables. A confounding variable must be associated with both the independent and dependent variables but must not be an intermediate step in the relationship between the two (i.e. not on the causal pathway).
For example, we know that physical exercise (an independent variable) can reduce a person’s risk of cardiovascular disease (a dependent variable). We can say that age is a confounder of that relationship as it is associated with, but not caused by, physical activity and is also associated with coronary health. See also ‘unobserved heterogeneity’, below.
Cross-sectional surveys involve interviewing a fresh sample of people each time they are carried out. Some cross-sectional studies are repeated regularly and can include a large number of repeat questions (questions asked on each survey round).
Data harmonisation involves retrospectively adjusting data collected by different surveys to make it possible to compare the data that was collected. This enables researchers to make comparisons both within and across studies. Repeating the same longitudinal analysis across a number of studies allows researchers to test whether results are consistent across studies, or differ in response to changing social conditions.
Data imputation is a technique for replacing missing data with an alternative estimate. There are a number of different approaches, including mean substitution and model-based multivariate approaches.
Data linkage simply means connecting two or more sources of administrative, educational, geographic, health or survey data relating to the same individual for research and statistical purposes. For example, linking housing or income data to exam results data could be used to investigate the impact of socioeconomic factors on educational outcomes.
Dummy variables, also called indicator variables, are sets of dichotomous (two-category) variables we create to enable subgroup comparisons when we are analysing a categorical variable with three or more categories.
General ability is a term used to describe cognitive ability, and is sometimes used as a proxy for intelligent quotient (IQ) scores.
Heterogeneity is a term that refers to differences, most commonly differences in characteristics between study participants or samples. It is the opposite of homogeneity, which is the term used when participants share the same characteristics. Where there are differences between study designs, this is sometimes referred to as methodological heterogeneity. Both participant or methodological differences can cause divergences between the findings of individual studies and if these are greater than chance alone, we call this statistical heterogeneity. See also: unobserved heterogeneity.
Household panel surveys collect information about the whole household at each wave of data collection, to allow individuals to be viewed in the context of their overall household. To remain representative of the population of households as a whole, studies will typically have rules governing how new entrants to the household are added to the study.
Kurtosis is sometimes described as a measure of ‘tailedness’. It is a characteristic of the distribution of observations on a variable and denotes the heaviness of the distribution’s tails. To put it another way, it is a measure of how thin or fat the lower and upper ends of a distribution are.
Longitudinal studies gather data about the same individuals (‘study participants’) repeatedly over a period of time, in some cases from birth until old age. Many longitudinal studies focus upon individuals, but some look at whole households or organisations.
Non-response bias is a type of bias introduced when those who participate in a study differ to those who do not in a way that is not random (for example, if attrition rates are particularly high among certain sub-groups). Non-random attrition over time can mean that the sample no longer remains representative of the original population being studied. Two approaches are typically adopted to deal with this type of missing data: weighting survey responses to re-balance the sample, and imputing values for the missing information.
Observational studies focus on observing the characteristics of a particular sample without attempting to influence any aspects of the participants’ lives. They can be contrasted with experimental studies, which apply a specific ‘treatment’ to some participants in order to understand its effect.
Panel studies follow the same individuals over time. They vary considerably in scope and scale. Examples include online opinion panels and short-term studies whereby people are followed up once or twice after an initial interview.
A percentile is a measure that allows us to explore the distribution of data on a variable. It denotes the percentage of individuals or observations that fall below a specified value on a variable. The value that splits the number of observations evenly, i.e. 50% of the observations on a variable fall below this value and 50% above, is called the 50th percentile or more commonly, the median.
In prospective studies, individuals are followed over time and data about them is collected as their characteristics or circumstances change.
Recall error or bias describes the errors that can occur when study participants are asked to recall events or experiences from the past. It can take a number of forms – participants might completely forget something happened, or misremember aspects of it, such as when it happened, how long it lasted, or other details. Certain questions are more susceptible to recall bias than others. For example, it is usually easy for a person to accurately recall the date they got married, but it is much harder to accurately recall how much they earned in a particular job, or how their mood at a particular time.
Record linkage studies involve linking together administrative records (for example, benefit receipts or census records) for the same individuals over time.
A reference group is a category on a categorical variable to which we compare other values. It is a term that is commonly used in the context of regression analyses in which categorical variables are being modelled.
Residuals are the difference between your observed values (the constant and predictors in the model) and expected values (the error), i.e. the distance of the actual value from the estimated value on the regression line.
Respondent burden is a catch all phrase that describes the perceived burden faced by participants as a result of their being involved in a study. It could include time spent taking part in the interview and inconvenience this may cause, as well as any difficulties faced as a result of the content of the interview.
In retrospective studies, individuals are sampled and information is collected about their past. This might be through interviews in which participants are asked to recall important events, or by identifying relevant administrative data to fill in information on past events and circumstances.
Sample is a subset of a population that is used to represent the population as a whole. This reflects the fact that it is often not practical or necessary to survey every member of a particular population. In the case of birth cohort studies, the larger ‘population’ from which the sample is drawn comprises those born in a particular period. In the case of a household panel study like Understanding Society, the larger population from which the sample was drawn comprised all residential addresses in the UK.
A sampling frame is a list of the target population from which potential study participants can be selected.
Skewness is the measure of how assymetrical the distribution of observations are on a variable. If the distribution has a more pronounced/longer tail at the upper end of the distribution (right-hand side), we say that the distribution is negatively skewed. If it is more pronounced/longer at the lower end (left-hand side), we say that it is positively skewed.
Study participants are the individuals who are interviewed as part of a longitudinal study.
Survey weights can be used to adjust a survey sample so it is representative of the survey population as a whole. They may be used to reduce the impact of attrition on the sample, or to correct for certain groups being over-sampled.
The term used to refer to a round of data collection in a particular longitudinal study (for example, the age 7 sweep of the National Child Development Study refers to the data collection that took place in 1965 when the participants were aged 7). Note that the term wave often has the same meaning.
The population of people that the study team wants to research, and from which a sample will be drawn.
Tracing (or tracking) describes the process by which study teams attempt to locate participants who have moved from the address at which they were last interviewed.
Unobserved heterogeneity is a term that describes the existence of unmeasured (unobserved) differences between study participants or samples that are associated with the (observed) variables of interest. The existence of unobserved variables means that statistical findings based on the observed data may be incorrect.
Variables is the term that tends to be used to describe data items within a dataset. So, for example, a questionnaire might collect information about a participant’s job (its title, whether it involves any supervision, the type of organisation they work for and so on). This information would then be coded using a code-frame and the results made available in the dataset in the form of a variable about occupation. In data analysis variables can be described as ‘dependent’ and ‘independent’, with the dependent variable being a particular outcome of interest (for example, high attainment at school) and the independent variables being the variables that might have a bearing on this outcome (for example, parental education, gender and so on).
The term used to refer to a round of data collection in a particular longitudinal study (for example, the age 7 wave of the National Child Development Study refers to the data collection that took place in 1965 when the participants were aged 7). Note that the term sweep often has the same meaning.
We are now going to extend our model to consider variables that may influence or confound the association between our predictor and outcome variables. These new variables being considered are: sex, parents’ education and family social class.
The sex variable has already been recoded to be binary (see the Preparing the data for modelling section) and in this regression analysis we are using the category ‘male’ as the reference group.
In addition, we are going to include a few family background factors in the model. These include two parental education measures that denote whether the participant’s mother (n016nmed) and father (n716dade) left school at the minimum age or not; these are also binary variables. For both of these variables, we are using the ‘left school at the minimum age’ as the reference group.
The final potential confounder we are including is the social class of the study participant’s father (n1171_2). This is a categorical variable with 5 values. In Stata you can automatically create dummy variable(s) for each of the values in a multi-category variable by appending the prefix of ‘i.’ to the variable name, e.g. i.n1171_2. In this instance, it means that the model will compare each of ‘III Skilled non-manual’, ‘III Skilled manual’, ‘IV Partly skilled’ and ‘V unskilled’ against the ‘I/II Prof & Managerial’ category. Stata will use ‘I/II Prof & Managerial’ as the reference category simply because it is the first category in the variable.
From the output table above, we can see that including the study participant’s sex and family background factors have not markedly changed the model. A small proportion, 4%, of the variance of BMI at age 42 is accounted for by family background, general ability at age 11 and the sex of the study participant. The participant’s general ability is still significant; for a 1 unit increase in general ability, we can expect a .03 decrease in BMI at age 42. The average BMI for females at age 42 is 1.14 lower than males, taking account of general ability at age 11. If the participant’s mother did not leave school at the minimum age, on average the participant’s BMI at age 42 was .47 lower than a participant whose mother left school. The father staying on at school was not significant, as this was explained by the father’s social class which was also included in the model. Social class and education are highly correlated; an individual’s educational attainment will in part reflect later occupational status which determines social class (You can explore this yourself as the syntax for the model above with social class excluded has been provided in the Stata .do file that accompanies this module). Compared to a participant whose father was in the highest social classes (I and II), having a father in the skilled and partly skilled manual social classes increased a participant’s BMI by .63 and .57 respectively (if all other factors remained equal). If the participant’s father was instead in the unskilled class, the increase in BMI was on average higher by 1.
In our final model we add bmi11, the BMI of the study participants when they were aged 11. By adding BMI at age 11 we adjust for earlier measures of BMI, thereby focusing on the change in BMI from age 11 to age 42. This allows us to measure BMI and general ability over a comparable duration from the age of 11 to 42 years.
The R-squared value in the output table above tells us that a quarter (25.8%) of the variance of BMI at age 42 is accounted for when we include BMI at age 11, as well as family background, general ability at age 11 and the sex of the participant, in the model. We can infer from the fact that mother’s education is no longer a significant predictor in this updated model that childhood BMI explains its significance in the earlier model. However, all other factors that were significant in the earlier less-adjusted model remain significant in this updated model, including our ‘general ability’ predictor variable. It may be that the influence of mother’s education on the participant’s midlife BMI, for example, reflects the family’s early eating habits, physical activity and health behaviours, which would be more influential in a child’s early life and therefore be reflected in their childhood BMI. For a 1 unit increase in general ability, we would expect a .02 decrease in BMI at age 42. In other words, a participant with a general ability score of 60 at age 11 would have a .63 lower BMI score at age 42 than a study participant with a general ability score of 30 at age 11, after controlling for BMI at age 11 and other factors.
However, we have still only explained a quarter (25.8%) of the variance in BMI at age 42. There are other factors, not included in this analysis which may play a role in that unexplained variance as they are known to be associated with BMI, such as physical activity, diet, sleep duration, socio-economic factors in later life, parent’s BMI and genetic factors.