What is a Dataset?

• Data is typically collected into tables where rows are observations, and columns are variables
• Variables may be numerical or categorical
• Numerical variables may be continuous or discrete
• Categorical variables may have an implicit ordering (“ordinal”), or no implicit ordering (“factors”)
  • Categorical data is often “coded” into number representations
• The kind of data you have should guide your choice of statistical and visualisation approaches.

1 Introduction

1.1 What is a dataset?

It might be an unhelpful definition, but a dataset is just a set of data.

In this notebook, we will explore features of a dataset that make it useful for scientific analyses, and characteristics of data that may guide a choice of statistical approach.

Data comes in many forms, but in science we nearly always want our data to be well-organised. To organise our data, we typically collect it into tables (in the R language, these are referred to as dataframes, you may also see them called matrices or tables). These are two-dimensional data structures, with rows and columns, and this allows us to deal with two important and distinct concepts about data: observations and variables (variables may also be referred to as features, or measurements)

1.2 Observations

An observation is the act of collecting one or more measurements of, or from, an “experimental unit.”

What constitutes an experimental unit depends on the nature of the experiment. For example, if you measured the BMI of each patient in a GP’s surgery, the experimental unit would be the patient (and the variable would be their BMI). On the other hand, if you were tracking the concentration of a compound in blood samples taken from a single patient over a period of time, each blood sample could be an experimental unit.

1.3 Variables, Features and Measurements

In an experiment, we measure properties of each “experimental unit.” The things that are measured are usually referred to as variables, or features. In the first example above, the BMI would be termed a variable; in the second example, the compound concentration would be called a variable.

The terms variable, feature and measurement may often be used interchangeably, but they are sometimes used to imply slightly different things:

• variable:
  • a property that varies across the experimental units
• feature:
  • 1. a property that varies across the experimental units, but is expected to be constant for each of the experimental units (like eye colour)
  • 2. a generic term for a variable
• measurement:
  • 1. a generic term for a variable
  • 2. the actual value that was measured

The use of these terms is not always consistent in the literature, and can sometimes be confusing.

1.4 Data Tables

By convention, we represent data in tables so that rows represent observations, and columns represent variables.

Example table showing the layout of observations, variables, and measurements

The interactive tables and visualisations below show two example datasets that are similar to those you may meet in your own work.

1.4.1 The iris dataset

The iris dataset describes morphological variation of the flowers (the lengths and widths of sepals and petals, measured in cm), of three related iris species: Iris setosa, I. virginica and I. versicolor. The dataset was introduced in 1936 papers by the botanis Edgar Anderson and the statistician R.A.Fisher, and is a common example dataset in statistics packages. The version of the dataset we are using is a standard variable in R.

Anderson, E. (1936) “The species problem in iris” Annal of the Missouri Botanical Garden 23: 457-509 https://doi.org/10.2307%2F2394164

Fisher, R.A. (1936), “The Use of Multiple Measurements in Taxonomic Problems.” Annals of Eugenics, 7: 179-188. https://doi.org/10.1111/j.1469-1809.1936.tb02137.x

We can visualise relationships between the variables with a pairs plot, which gives us a rich visual overview of the relationships in our dataset.

We will explore data visualisation further, later in the workshop.

Figure 1.1: Pairs plot of iris data, providing an overview of relationships between variables.

1. Which variables appear to distinguish between iris species?
2. What features of the data and visualisation indicate that variables discriminate between species?
3. Which variables look like they are correlated?
4. What features of the data and visualisation indicate correlation?

1.4.2 The Pima Diabetes Dataset

The pima dataset describes a set of females aged at least 21 and having Akimel O’odham (formerly known as Pima Indian) heritage. These are a Native American people living in an area of Central and Southern Arizona. This group have a high rate of diabetes mellitus. The dataset was collected by the National Institute of Diabetes and Digestive and Kidney Diseases (in the US), and has been used as an example dataset for machine learning, e.g. in

Smith, J.W., Everhart, J.E., Dickson, W.C., Knowler, W.C., & Johannes, R.S. (1988). Using the ADAP learning algorithm to forecast the onset of diabetes mellitus. In Proceedings of the Symposium on Computer Applications and Medical Care (pp. 261–265). IEEE Computer Society Press. Online at PubMedCentral

Again, we can visualise relationships with a pairs plot. The class variable has values 0 (indicating did not test positive for diabetes) and 1 (indicating did test positive for diabetes)

Figure 1.2: Pairs plot of pima data, providing an overview of relationships between variables.

1. Do any variable appear to distinguish between individuals who did and did not test positive for diabetes?
2. Why do you think this?
3. How many observations are there in the pima dataset?
4. How many variables are there in the iris dataset?
5. Do the visualisations appear to suggest any potential problems with either the iris or pima datasets?

2 Types of variable

When analysing an experiment, especially when critically reviewing someone else’s work, it is very important to understand the dataset, and the data represented in it. There are two main types of data: numerical and categorical

2.1 Numerical Data

Numerical data is, unsurprisingly, data that can reasonably be represented as numbers. Values like 3.141, 0, -7, \(10^8\) and so on are all numbers, and could be found in numerical data.

But it’s a little more complicated than that.

Numerical data can be subdivided into two categories: Discrete and Continuous data

2.1.1 Continuous Data

This might be the first kind of numerical data that comes to mind for you. Measurements like length, mass, concentration, and time, can be represented by real numbers. The important characteristic of real numbers and therefore continuous data is that, in principle, any number can be represented to an arbitrary level of precision.

Although our ability to measure things is limited in practice (e.g. a standard 30cm ruler might not be able to measure a difference in length smaller than 1mm accurately), it’s the ability of the property to take on one of these values that matters. For instance, with a kitchen timer we might not be able to time a 100m sprint more accurately than the nearest second; but the runners could finish at any time between consecutive ticks of the clock (e.g. at 14.461723884756325347s exactly, even if we only measure 14.46s on a stopwatch).

It is not our ability to measure precisely that makes the data continuous, it is the nature of the thing being measured.

Measurements are almost always taken as discrete values, such as mass to the nearest gram, but we model the measurements on their underlying continuous scale.

2.1.2 Discrete Data

Not all numerical data is continuous. For instance, count data (e.g. security staff ticking off how many people are in a pub, club or concert hall, for fire safety) is not continuous: you can only count whole number, integer values. We call this kind of numerical data discrete numerical data.

With discrete data, every state of the property is different and separate; intermediate values are either not allowed or not possible. Because the values are restricted, we need to treat this data differently than continuous numerical data, and interpret analyses differently.

For example, the difference between discrete and continuous data is why it’s sometimes reported that “the average person in the UK has fewer than two legs”. An individual may have zero, one or two legs, and many more people have two legs than not. Most people you will meet (“the average person”) are likely to have exactly two legs, rather than one. But calculating an average number of legs using the formula

\[\textrm{average number of legs} = \frac{\textrm{total number of legs}}{\textrm{total number of people}}\]

may well give an answer somewhere between 1 and 2.

The problem here is not so much mathematical, but that the wrong choice of model was made. When we calculate an average we’re really calculating a central value or location. There’s more than one way to calculate a central value - this kind of mean is just one of several methods. Alternative values, such as the median or mode number of legs in the population will most likely be exactly 2.

XKCD 2435: Geothmetic Meandian; this is satire

It can be a warning sign that something is not right with an analysis if a calculation gives apparently nonsensical values, such as fractional values for entities that can only be whole numbers. That suggests that an inappropriate model may have been used somewhere.

2.2 Categorical Data

Categorical data is data that can be divided up into a finite set of categories. There may be few categories (e.g. “has diabetes” and “does not have diabetes”) or many (such as nationality, sex, age group, make of car, and so on).

But, as with numerical data, it’s a little more complicated than that.

Catgorical data can be subdivided into two categories: Factors and Ordinal data

Sometimes, as with continuous numerical data, the category is a property of the experimental unit that falls into one or more groups, such as eye colour, or nationality. Here there is no implicit ranking or ordering (blue eyes are not “more than” green eyes, for instance). These are often referred to as factors.

Other categories however may have an implicit ranking - like “large”, “medium”, and “small” drinks at a fast food outlet - and these are called ordinal data.

2.2.1 Ordinal Data

In a competition, individuals might finish first, second, third and so on. These rankings are often treated as numbers, but they are in fact categories, and they are categorical data. We might consider that those who finish first are “better” or “more” in some way than those who finish second or third. There is a natural ordering to these categories, which we want to consider in our analyses - but we can’t rightly treat 1, 2 and 3 as the whole numbers \(1\), \(2\) and \(3\) - they are instead treated as labels.

Similarly, we often divide populations into age ranges, rather than using an accurate age. These ranges are categories too, and there is an implicit ordering: “70+” is older than “55-70” is older than “45-55”. Here, the categories are ordered, but they are not the same size, and we can’t treat them as numerical values without taking more care.

Likewise, we might divide disease symptoms into “none”, “mild”, “severe”, and “critical”. Again, there is an ordering of sorts, but there is no natural numerical representation.

2.2.2 Coding

Categorical data is often coded. This means that the categories are assigned numbers as labels. Each discrete category gets a different number (a bit like the Dewey decimal system in a library, matriculation numbers, or IP addresses on a computer network). The numbers don’t necessarily relate to the size, value, or measure of a category. The codes are just convenient ways of representing different things in a table.

By convention we often take 1 implicitly to mean “true” or “present”, and 0 to mean “false” or “absent”.

1. What kind of data is in the iris dataset?
2. What kind of data are the variables pregnancies, BMI, and class in the pima dataset?

3 R code used in this notebook

Click to toggle R code for visualising iris data

The iris data exists as a variable in R, and we visualise it with the ggpairs() function from the GGally package.

p = ggpairs(iris)                           # create pairs plot for all iris data
p = p + theme(plot.background = element_rect(fill = figbg,    # colour background
                                             color = figbg))
p  

Click to toggle R code for visualising pima data

The pima dataset is visualised in the same way, but must be loaded in from file, first. The data is read in using the read_csv() function from the readr package, defining useful column/field names for the variables, and specifying the data types for each variable in the col_types string:

• i: integer (numerical)
• f: factor (categorical)
• d: double (numerical, continuous)

pima = read_csv("../assets/data/pima/pima-indians-diabetes.csv",
                col_names=c("pregnancies", "conc_Glc",        # Load data from file
                "diastolic", "skin_fold",
                "insulin", "BMI", "diabetes_pedigree",
                "age", "class"),
                col_types="iiiiiddif")
p = ggpairs(pima)                                             # create pairs plot
p = p + theme(plot.background = element_rect(fill = figbg,    # colour background
                                             color = figbg))
p                                                             # show plot