In brief

Scientists develop theoretical models which aim to describe the true relationships between variables. Statistical models are used to link their theories with data. They allow us to make predictions about future events, and our confidence in these predictions. But accurate predictions alone aren’t enough: We also want to understand our data and the process that generated it.

Causal diagrams are a useful tool for thinking about causes and effects. For psychological phenomena the diagrams can become complicated. But good research simplifies: We focus on small parts of a large network of causes and effects to make incremental progress.

The quality of evidence for different parts of a causal diagram can vary a lot. We are much more sure about some links in the network than others. But causal diagrams can represent both our knowledge and our hypotheses explicitly; this is useful as we build statistical models to check how well our theories perform in the real world.

1. Tasks and slides

This worksheet contains explanatory notes and some additional reading. For quick reference, the following are shortcuts to the tasks you must complete during, or as homework between, the two workshops:

Main workshop

• Identify variables
• Draw a causal digram
• Moderation in your model

Preparation

• Find evidence

TARA led session

• Journal club
• Identify confounders

The slides which accompanied this workshop are available on the module DLE site.

Variables and constructs

Our first job is to try and identify the variables or constructs involved in the phenomena we are interested in.

We’ll talk more about variable and constructs in the session on measurement, but for the moment we’ll use the following definition for a variable:

• Things we can measure (observations)
• Things that cause other things, even if we can’t measure them (constructs)

For example, let’s imagine we were trying to explain attendance rates for a 9am lecture.

• We might be interested in the time the cohort of students went to bed. This is (in theory) measurable, and we could make observations of students’ bedtimes.

• We might also hypothesise that students’ mood or affect would influence their attendance. If someone was depressed they might skip the lecture. We can’t measure depression directly, but we do think of it as something which causes other things to happen. For this reason we would talk about depression as a construct: we infer that it is present from other indicators (e.g. reports of poor mood, hopelessness, fatigue, trouble concentrating, etc. in a psychometric scale).

Causal diagrams

This section in intended as notes to the lecture/background reading. If you are in the workshop you can skip to the next section for the task.

Once we have identified the variables we think might be involved, we can try and describe the network of causes and effects between them and the phenomena we’re interested in. This is our model.

To represent this model we can use a special type of diagram called a directed acyclic graph. That’s a fancy name for a boxes-and-arrows diagram, with a few special rules (we can come to those later).

This diagram represents a simple situation in which one variable (“Predictor”) causes another variable (“Outcome”). In this diagram we are saying nothing else causes the outcome. That’s important to note because it’s quite a strong assumption that this model makes.

This diagram says our footwear and exam grades AREN’T related AT ALL, because we didn’t draw a line between them:

Diagrams can also describe how variables are related in a particular causal sequence. For example, in our research we might ask questions like:

• does childhood poverty reduce academic achievement by delaying brain development?
• does weaker childhood attachment reduce academic performance by reducing the motivation to study?

This pattern – where variables are linked in a series – is called ‘mediation’:

Finally, if you don’t know which direction the arrow should point — that is, you don’t know which is the cause and which is the effect — we can (temporarily) draw an arrowhead at both ends like this:

This represents a correlation (see the stage 1 notes on correlation and relationships).

Our hope is that — as we learn more, by collecting data or running experiments — we can decide which way the arrow should point.

‘Moderation’

Another common question for researchers is whether relationships between variables are true all the time, or if they vary depending on the context or person.

In concrete terms, we might ask questions like:

• “does low self esteem hurt academic performance more for women than for men?”
• “does social media cause less anxiety for people who have high emotional intelligence (EQ)?”

There is no universally agreed-upon way to represent effect modification in DAGs¹, but for the moment, you should draw it like this:

The arrow from gender points at the relationship between self esteem and grades. We mean that the effect of self esttem on grades depends on whether you are a woman.

This pattern is called moderation or effect modification. Checking to see if a relationship is the same for different groups is also called stratification.

Confounding and bias

2. TARA-led session

3. Homework and assessment progress

After this session you should be able to answer (or at least make a start on) questions 1 and 2 from part 1 of the ‘data’ assignment.

1. Save a digital version of the causal diagram you drew in the workshop. This image can be either be a powerpoint file, a screen-grab or a photo of a hand drawn sketch. It doesn’t have to look neat — it’s the ideas that matter here.

Feel free to update/amend the diagram from the one you drew in the session. Make the model reflect what you would predict/expect to happen, based on what you know.

To be explicit: It’s fine if your model diverges from the one you developed as a group. You can each submit your own model if you like, or submit the one you developed together.

2. Write a short summary of the evidence for the three most important relationships described by your model (providing a citation for each).