1.1 Why Researchers Draw Causes Before Collecting Data

Lesson 2 turned a broad topic into a structured research question. A question names an exposure and an outcome, but it does not tell you what else to measure. Every outcome in health has many causes, and those causes are tangled together. A study that measures only the exposure and the outcome cannot tell whether an association between them reflects a real effect or the influence of something else, and a study that measures everything becomes long, expensive and intrusive. Researchers therefore need a way to decide, before any data are collected, which variables belong in the study and why.

This lesson teaches two drawings that support that decision. The first is a causal web, a broad diagram of the causes that the literature and experience suggest are connected to the outcome. The second is a directed acyclic graph (DAG), a more disciplined diagram that states exactly which variable is assumed to affect which. In this lesson the word cause has a plain meaning: a variable is a cause of an outcome if changing it would change how likely, or how large, the outcome is. Smoking is a cause of lung cancer in this sense, because a person who smokes more has a higher chance of developing it than the same person would have had after smoking less.

1.2 The Web of Causation

The idea that disease has a single cause served public health well in the era of infectious disease, when identifying the organism behind cholera or tuberculosis pointed directly to prevention. It served less well for heart disease, cancer and other chronic conditions, which arise from many factors acting together over long periods. In their 1960 textbook Epidemiologic Methods, Brian MacMahon, Thomas Pugh and Johannes Ipsen described the causes of disease as a web of causation, in which many factors are linked to the outcome and to one another (MacMahon et al., 1960). Each cause in a web has causes of its own, and an intervention can act at any strand that is open to change.

The web of causation has been criticized as well as used. In a widely cited paper, Nancy Krieger (1994) asked whether anyone had seen the spider: if disease arises from a web, what force spins it? She argued that webs drawn in epidemiology tended to place individual behaviours and biological factors in the foreground, while the social, economic and political conditions that shape those factors were left out or treated as background. Krieger went on to develop ecosocial theory. For a student drawing a first causal web, her critique has a practical consequence. A web that contains only individual risk factors will lead to a study that measures only individual risk factors, so the outer levels of the web, such as income, housing, transport and access to services, deserve deliberate attention.

HSCI 341 Lesson 1 Section 3 returns to the web of causation to separate direct from indirect causes and to link the web with the component-cause model.

1.3 From a Literature Summary to a List of Links

A causal web is built from evidence and from the experience of people who know the problem. Lesson 2 showed how to find three to five key papers on a topic. Those papers, together with the knowledge of people affected by the problem, supply the raw material. The process has five steps, and each step produces a record that the team keeps with its study files.

Step 1: Start from the question

Write the exposure and the outcome at the top of a page. Everything that follows is chosen because it may be connected to one or both of them. For Cedar Valley, the exposure is loneliness and the outcome is the number of emergency department (ED) visits in the 12 months after the survey.

Step 2: Build an evidence table

For each source, record what kind of evidence it is and what it suggests about causes. Write each finding in your own words beside its citation. The table below shows part of the Cedar Valley team's evidence table. The published studies are real and their general findings are summarized briefly; the last two rows record knowledge from the team's advisory group and partner clinics.

Step 3: Write each link as a short statement

Rewrite the last column as a list of link statements of the form "A affects B", one link per line, with the source in brackets. A link statement makes the direction explicit. "Loneliness and depression are related" is too vague to draw, whereas "loneliness leads to later depressive symptoms (Cacioppo et al., 2006)" names a direction and a source. Where sources disagree about direction, write both statements and note the disagreement, because Section 2 shows how to handle a two-way relationship.

Step 4: Group the factors by level

Sort the factors into rough levels: individual health (chronic conditions, hearing loss, depression), relationships (living alone, bereavement), and community or structural conditions (income, rural residence, transport). Then ask, with Krieger's critique in mind, whether any level is thin. The Cedar Valley team found its first list thin on transport and income and returned to the advisory group's notes.

Step 5: Draw the web and review it with others

Place the exposure near the centre and the outcome to its right, add each factor, and draw an arrow for each link statement. Then show the draft to people who know the problem. Lesson 4 develops interest holder engagement in detail; at this stage, a conversation with a few people who live with the problem or work with it often reveals missing factors. A more formal and reproducible version of this process, called evidence synthesis for constructing directed acyclic graphs (ESC-DAGs), has been published by Ferguson and colleagues (2020). This lesson uses a simplified form suited to a first project, and systematic searching itself is taught in HSCI 241.

1.4 The Cedar Valley Causal Web

The figure below shows the team's first causal web. Read it from left to right. The factors on the left (bereavement, living alone, hearing loss, limited transport and rural residence) are drawn as causes of loneliness. Loneliness sits in the centre, and the outcome, ED visits, sits on the right. Chronic conditions and low income appear as factors linked to loneliness, and two pathways from loneliness toward ED visits pass through depression and physical inactivity.

Three features of the web deserve attention. First, every arrow traces back to a row of the evidence table, so a reader can ask where each link came from. Second, the web contains two loops, shown in red. Loneliness and depression are drawn as affecting each other, and a second loop runs from chronic conditions to loneliness to physical inactivity and back to chronic conditions. Third, the web does not say whether chronic conditions come before loneliness or after it, or which factors the team can measure. A causal web is a thinking tool for gathering possibilities, and it is deliberately broad.

1.5 From a Causal Web to a DAG

A causal web is good at showing how much is going on. It is less good at telling the team exactly what to measure, because its boxes can be broad, its arrows can run in both directions, and it does not distinguish what happened first. A directed acyclic graph imposes three disciplines. Each box, called a node, is a single variable that could in principle be measured at a particular time. Each arrow points one way only, from cause to effect, which is what "directed" means. And the diagram contains no loops, which is what "acyclic" means: if you start at any node and follow the arrows, you can never return to where you started. "Graph" is the mathematical name for a set of nodes joined by lines.

DAGs grew out of path diagrams, which the geneticist Sewall Wright developed in the 1920s, and were given a formal footing by the computer scientist Judea Pearl in the 1990s (Pearl, 1995). Sander Greenland, Judea Pearl and James Robins introduced them to epidemiologists in an influential paper (Greenland et al., 1999), and they are now widely used in health research to plan which variables to measure and how to analyze them (Hernán & Robins, 2020; Tennant et al., 2021). The table below summarizes how the two diagrams differ.

The web and the DAG are two stages of the same work. The web keeps the team honest about the breadth of causes, including the structural ones that Krieger warned are easily dropped. The DAG then forces the team to decide, for each pair of variables, whether one affects the other and in what order. Section 2 shows how to make those decisions and draw the result in DAGitty.

2.1 The Parts of a DAG

A directed acyclic graph has two kinds of parts. A node is a variable, drawn as a box or a name. An arrow (also called an edge) joins two nodes and points from the variable assumed to be a cause to the variable assumed to be its effect. When an arrow runs directly from A to B, A is said to have a direct effect on B, meaning an effect that does not pass through any other variable shown in the diagram. Two more words help when reading larger diagrams. A node's ancestors are all the nodes from which you can reach it by following arrows forward, and its descendants are all the nodes you can reach from it by following arrows forward. DAGitty uses these words in its legend, so they are worth learning now.

The figure below shows a first, small DAG drawn by the Cedar Valley team. It contains four nodes: age, hearing loss, loneliness and ED visits. Read aloud, it says that age affects hearing loss, loneliness and ED visits directly; that hearing loss affects loneliness; and that loneliness affects ED visits. Age is therefore an ancestor of every other node, and ED visits is a descendant of every other node.

2.2 What an Arrow Claims, and What a Missing Arrow Claims

An arrow from A to B is a modest claim. It says that A might have a direct effect on B, and that the team is not prepared to rule this out. It does not say how large the effect is, whether it raises or lowers B, or whether it is the same for everyone. Adding an arrow is the cautious choice when the evidence is unclear, because an arrow allows for an effect without assuming one.

A missing arrow is a strong claim. If there is no arrow from A to B, the diagram asserts that A has no direct effect on B at all, so that any influence of A on B must travel through other variables shown in the diagram. In the four-node DAG above, the missing arrow between hearing loss and ED visits asserts that hearing loss affects ED visits only by making people lonelier. When Dr. Hart showed this draft to the advisory group, one member pointed out that hearing loss can make it harder to hear traffic, alarms or a clinician's instructions, and that several people she knew had been injured in falls they connected with poor hearing. That is a route from hearing loss to ED visits that does not pass through loneliness. The team added an arrow from hearing loss to ED visits. Section 3 shows that this single decision changed the role hearing loss plays in the study.

2.3 Four Conventions for Drawing a First DAG

Experienced researchers follow a few conventions that keep a DAG honest and readable. They are drawing habits, and they do not require the formal rules for reading a DAG. HSCI 230 Lesson 7 Section 2 introduces those rules through the three basic structures and collider bias, and HSCI 341 Lesson 7 Section 3 states them in full as the backdoor criterion and a procedure for choosing an adjustment set.

Convention 1: One node is one variable at one time

A node should be something that could be measured, at least in principle, at a stated time. "Social factors" fails this test. "Living alone at the time of the survey" passes it. When a factor could be measured at two different times and the timing matters, it becomes two nodes.

Convention 2: Arrows point forward in time

Arrange the nodes so that time runs from left to right. Factors fixed early in life or measured before the survey go on the left, the exposure in the middle, and the outcome and anything measured during follow-up on the right. If an arrow points from right to left, check whether it describes something that happened earlier.

Convention 3: Unroll loops into time

The causal web in Section 1 showed loneliness and depression affecting each other. A DAG cannot contain that loop. The solution is to split the factor into time-stamped nodes: depression before the survey can affect loneliness at the survey, and loneliness at the survey can affect depression during the following year. The figure below shows the change.

Convention 4: Include shared causes, even unmeasured ones

If a variable causes two or more of the variables already in the diagram, it belongs in the diagram, even if the team cannot measure it. Frailty, for example, could make it hard for an older adult to leave the house (raising loneliness) and could also raise the risk of injury (raising ED visits). The Cedar Valley survey does not measure frailty. Leaving it out of the DAG would hide a problem the team needs to discuss. DAGitty lets you mark such a node as unobserved. A variable that affects only one node, and is of no interest in itself, can usually be left out to keep the diagram readable.

2.4 Drawing the DAG in DAGitty, Step by Step

The steps build a six-node version of the Cedar Valley DAG with the nodes Loneliness, ED_visits, Age, Chronic_conditions, Living_alone and Later_depression. Use the _ character in place of spaces in node names. DAGitty can handle names with spaces by placing them in double quotation marks in its code, but names without spaces are easier to type and can later serve as variable names in a dataset (Section 4).

1. Open the program. Go to dagitty.net and select the link labelled "Launch DAGitty online in your browser". The screen shows a menu bar along the top, a drawing area (the canvas) in the middle, and panels of options and information at the sides.
2. Start a new model. From the Model menu, choose New model. DAGitty asks for the name of the exposure and then the name of the outcome. Type Loneliness and then ED_visits. DAGitty draws the two nodes with an arrow from the exposure to the outcome. Keep this arrow, because it represents the effect your question asks about.
3. Add the other variables. Double-click an empty spot on the canvas, or move the mouse pointer to an empty spot and press the n key. Type the variable's name in the box that appears and press Enter or click OK. Add Age, Chronic_conditions, Living_alone and Later_depression in this way.
4. Add arrows. Double-click the node that is the cause; it becomes highlighted. Then double-click the node that is the effect. An arrow appears between them. As an alternative, move the pointer over the cause and press c, then move it over the effect and press c again. Add each arrow from your link list in turn.
5. Correct mistakes. To remove an arrow, repeat the same two double-clicks that created it. To delete a variable, move the pointer over it and press d or the Delete key; its arrows disappear with it. To rename a variable, move the pointer over it and press r.
6. Arrange the layout. Drag the nodes with the mouse so that time runs from left to right and arrows cross as little as possible. Put the exposure toward the left of centre and the outcome toward the right.
7. Set each variable's status. Click a node and use the checkboxes in the panel labelled Variable, or move the pointer over the node and press e for exposure, o for outcome or u for unobserved. The exposure and outcome are already set from step 2. If you add Frailty, mark it as unobserved.
8. Read the model code. Find the text box labelled Model code. DAGitty writes your diagram there as text, with each arrow on a line such as Age -> Loneliness. Compare the list with your link statements to check that nothing is missing or reversed.
9. Save your work. Copy all the text in the Model code box and paste it into a plain text file in the study folder, with a name and version number such as CedarValley_DAG_v1.txt. To reopen the diagram later, paste the text back into the Model code box and click Update DAG.
10. Export a figure. The Model menu offers export as a PDF, SVG, PNG or JPEG file. A PNG file is the simplest to insert into a document, and a PDF or SVG file stays sharp when enlarged.

As you work, DAGitty colours the nodes and arrows and fills a panel labelled Causal effect identification with suggestions. These features apply the formal rules for reading a DAG, which HSCI 230 Lesson 7 Section 2 introduces and HSCI 341 Lesson 7 Section 3 states in full as the backdoor criterion. For this course, treat DAGitty as a drawing and record-keeping tool and leave those panels for later.

Typing the model code directly

Some people find it faster to type the diagram than to click it, and typing also works well with a keyboard or a screen reader. The model code for the six-node DAG is shown below. Each line inside the braces either declares a variable's status in square brackets or states one arrow. Paste the code into the Model code box and click Update DAG, and DAGitty draws the diagram. The nodes may land in an untidy arrangement, and dragging them into place adds position information to the code automatically.

dag {
Loneliness [exposure]
ED_visits [outcome]
Age -> Loneliness
Age -> ED_visits
Age -> Chronic_conditions
Age -> Living_alone
Chronic_conditions -> Loneliness
Chronic_conditions -> ED_visits
Living_alone -> Loneliness
Living_alone -> ED_visits
Loneliness -> Later_depression
Later_depression -> ED_visits
Loneliness -> ED_visits
}

2.5 Documenting the Arrows

A DAG is easier to review when each arrow, and each important missing arrow, comes with a one-line justification. The justification can cite a study, a theory, the knowledge of interest holders, or a logical point such as the order of events. The Cedar Valley team kept a justification table beside its DAG. Part of it is shown below.

The last row illustrates a missing arrow that is easy to defend, because the timing makes it impossible. Most missing arrows are harder to defend, and a justification table makes the weaker ones visible. Tennant and colleagues (2021), reviewing published health studies that used DAGs, recommended that researchers report their full diagrams and explain how they were built. A justification table is a simple way to meet that recommendation in a small study. Section 3 now uses the finished diagram to name the role each variable plays.

3.1 Roles Are Defined by the Question

Once a DAG is drawn, each variable can be given a role. The roles are defined relative to the research question, so a variable has no role on its own. The exposure is the variable whose effect the question asks about. It may be a risk factor (loneliness), a treatment (a medication), a program (a community connector service) or a condition of life (living in a rural area). The outcome is the variable on which that effect is measured. In the Cedar Valley PECO question, loneliness is the exposure and the number of ED visits in the following 12 months is the outcome. The same variable can play different roles in different studies. Loneliness is the exposure in this study, and it would be the outcome in a study asking whether a connector program reduces loneliness.

Three further roles describe how other variables relate to the exposure and the outcome. A confounder is a common cause of both. A mediator lies on a causal pathway from the exposure to the outcome. A collider is a common effect of two variables in the diagram. The figure below shows the three patterns in their simplest form.

3.2 Confounders

A confounder is a common cause of the exposure and the outcome, or a proxy for such a cause, that is not on the causal pathway between them. The word comes from the Latin for "to mix together", and the name describes the problem: a confounder mixes its own effect into the association between the exposure and the outcome. A classic teaching example concerns carrying matches and lung cancer. People who carry matches develop lung cancer more often than people who do not. Matches do not cause cancer. Smoking causes people to carry matches, and smoking causes lung cancer, so smoking is a confounder of the association between carrying matches and lung cancer. A study that compared match carriers with non-carriers and ignored smoking would wrongly conclude that matches are harmful.

In Cedar Valley, chronic conditions diagnosed before the survey are a confounder. Conditions such as heart failure, arthritis or chronic lung disease can limit how often a person leaves home and sees others, which raises loneliness. The same conditions produce acute episodes that bring people to the ED. If the team compared ED visits between lonely and non-lonely respondents without accounting for chronic conditions, some of the difference would reflect the greater burden of illness among lonely respondents. The team's DAG contains several other confounders: age, income, living alone, rural residence, earlier depression, frailty and hearing loss.

Hearing loss deserves a second look. In the first draft in Section 2, hearing loss had an arrow into loneliness and no arrow into ED visits, so it was a cause of the exposure only and was not a confounder. After the advisory group's comment, the team added an arrow from hearing loss to ED visits. Hearing loss then became a common cause of loneliness and ED visits, and therefore a confounder. Its role changed because an arrow changed, which is why the justification table in Section 2 matters.

3.3 Mediators

A mediator is a variable that lies on a causal pathway from the exposure to the outcome: the exposure affects the mediator, and the mediator affects the outcome. A mediator carries part of the exposure's effect. A simple example outside the case concerns a school breakfast program. Providing breakfast may improve children's attention in morning classes, and better attention may raise test scores. Attention is a mediator of the effect of the breakfast program on test scores.

In Cedar Valley, depressive symptoms that develop after the survey are a mediator. Loneliness predicts later depressive symptoms (Cacioppo et al., 2006), and depression can lead to ED visits through self-harm, neglect of chronic conditions or physical symptoms that prompt a visit. If loneliness raises ED visits partly because it leads to depression, then depression is one of the routes by which the effect travels. That route is part of the effect the PECO question asks about.

This has a practical consequence. If the team treated later depression as though it were a confounder and removed its influence in the analysis, it would remove part of the very effect it set out to estimate. The analysis would then answer a narrower question about the effect of loneliness through routes other than depression. Studying how much of an effect travels through a mediator is called mediation analysis, which HSCI 410 Lesson 8 teaches. In HSCI 207 the task is simpler: recognize a mediator in the DAG, record it in the variable list, and avoid treating it as a confounder.

3.4 Telling a Mediator from a Confounder

Students confuse these two roles more than any others. In a recent survey of HSCI 341 students in this series, about one in four (23 percent) chose "a variable on the causal pathway" as the definition of a confounder, which is the definition of a mediator. The confusion is understandable. A confounder and a mediator are both "third variables" that are associated with the exposure and with the outcome, and in a dataset they can look alike. The difference lies in the direction of the arrow between the third variable and the exposure, and that direction comes from knowledge about timing and mechanism, which the data alone cannot supply. Three questions settle most cases.

Depression in Cedar Valley shows why the three questions matter. Depression can affect loneliness, and loneliness can affect depression. As Section 2 showed, a DAG handles this by splitting depression into two time-stamped nodes. The figure below places them on a time line.

Apply the three questions to each node. Earlier depression points into loneliness (question 1), was present before the survey (question 2), and could not be changed by changing loneliness at the survey (question 3), so it is a confounder. Later depression is pointed to by loneliness, develops after the survey, and would change if loneliness changed, so it is a mediator. The practical lesson for the team is that it must measure depression at two times, or at least know when each measurement was taken, because a single measure of "depression" with no date cannot be assigned a role.

3.5 Colliders

A collider is a variable that is caused by two other variables in the diagram, so that two arrowheads meet, or collide, at it. Colliders are less intuitive than confounders and mediators, because the problem they cause appears only when a study restricts its sample to one level of the collider or adjusts for it in the analysis. A worked example with simple numbers shows what can happen.

In Cedar Valley, the team identified referral to the health authority's community connector program as a collider. Clinicians refer older adults who appear lonely, and ED staff refer patients who return to the ED frequently, so loneliness and ED visits both cause referral. If the team analyzed only respondents who had been referred, or adjusted for referral in its analysis, it could produce a misleading association between loneliness and ED visits, much as the falls program produced a misleading association between balance and vision. The variable list in Section 4 therefore records referral as a variable that the team will not restrict on or adjust for. HSCI 230 Lesson 7 and Lesson 8 discuss the biases that colliders produce, and HSCI 341 Lesson 7 Section 3 states the formal rules that explain them.

3.6 Labelling the Full Cedar Valley DAG

The team's final DAG for this stage contains twelve nodes. The figure below groups the eight confounders into a single box for legibility. In DAGitty each confounder is a separate node with its own arrows into loneliness and ED visits, as the model code below the figure shows.

dag {
Loneliness [exposure]
ED_visits [outcome]
Frailty [latent]
Age -> Loneliness
Age -> ED_visits
Age -> Chronic_conditions
Age -> Hearing_loss
Age -> Living_alone
Age -> Frailty
Income -> Loneliness
Income -> ED_visits
Income -> Chronic_conditions
Living_alone -> Loneliness
Living_alone -> ED_visits
Rural_residence -> Loneliness
Rural_residence -> ED_visits
Hearing_loss -> Loneliness
Hearing_loss -> ED_visits
Chronic_conditions -> Loneliness
Chronic_conditions -> ED_visits
Chronic_conditions -> Frailty
Frailty -> Loneliness
Frailty -> ED_visits
Earlier_depression -> Loneliness
Earlier_depression -> ED_visits
Earlier_depression -> Later_depression
Loneliness -> Later_depression
Later_depression -> ED_visits
Loneliness -> ED_visits
Loneliness -> Connector_referral
ED_visits -> Connector_referral
}

In DAGitty's model code, the word latent in square brackets marks a variable as unobserved, which is what pressing u over a node does. Pasting this code into DAGitty reproduces the full diagram, and you can drag the nodes into the arrangement shown in the figure.

With every node labelled, the DAG now tells the team which variables matter and why. Section 4 turns that information into the variable list and data dictionary that guide data collection.

4.1 Why the DAG Needs a Variable List

A DAG states what the team believes about causes. It does not yet say what the team will collect, from where, or when. The bridge between the diagram and data collection is a variable list, a table with one row for every node in the DAG. Each row records the variable's role, the time at which it takes its value, the source that will supply it and what the team plans to do with it. A variable list has three uses. It turns the DAG into a practical plan for the survey, the data request and the chart review. It shows which nodes cannot be measured, so that the team can discuss them as limitations before the study begins. And it justifies every piece of information the team proposes to collect from participants.

The third use matters for ethics as well as for efficiency. Lesson 5 shows that research ethics boards expect researchers to explain why they need each piece of personal information they collect. A variable list linked to a DAG answers that question variable by variable: the team collects income because income is a common cause of loneliness and ED visits, and it can point to the arrows that say so. Variables with no node in the DAG and no descriptive purpose are candidates for removal.

4.2 Building the Variable List

The variable list is built in five steps, working from the labelled DAG in Section 3.

1. List every node. Copy each node name from the DAGitty model code, including unobserved nodes such as Frailty. Using the exact node names keeps the list and the diagram linked.
2. Record the role. Write exposure, outcome, confounder, mediator or collider for each node, as in Section 3. A node that has none of these roles, such as a cause of the outcome only, is recorded as "other".
3. Record the timing. State when the variable takes its value relative to the exposure: before the survey, at the survey or during follow-up. Timing is what separated earlier from later depression in Section 3.
4. Identify a source. Name where the data will come from: the survey, linked administrative records, a chart review, or nowhere. Lesson 8 describes surveys, administrative data and chart reviews, and Lesson 9 describes data linkage.
5. State the planned use. Write what the team intends to do with the variable, using plain phrases such as "account for as a confounder", "keep for a later mediation analysis", "do not restrict or adjust" or "record as a limitation". HSCI 230 Lesson 7 Section 2 introduces the structures behind the formal rules that decide which variables an analysis should adjust for, and HSCI 341 Lesson 7 Section 3 states the backdoor criterion and the procedure for choosing an adjustment set. At this stage the plan is provisional and follows from the roles.

The figure below sorts the twelve Cedar Valley nodes by when they take their values and shows the source of each. The colour of each dot shows the variable's role.

The completed variable list is shown below. Notice that every node appears, including the two that the team will not collect.

4.3 From the Variable List to a Data Dictionary

A data dictionary is a document that describes every variable in a dataset so that anyone using the data knows exactly what each column contains. Where the variable list is organized around the DAG, the data dictionary is organized around the dataset the team will build. It gives each variable a short name, a readable label, a definition, its possible values and codes, its type, its source and any notes a future analyst would need. Lesson 11 shows how a data dictionary becomes a codebook used for data entry and cleaning. Here the aim is a first draft written before data collection begins.

Conceptual and operational definitions

Each entry needs two definitions. A conceptual definition says what the idea means in words. Perlman and Peplau (1981) defined loneliness as the unpleasant experience that occurs when a person's network of social relationships is deficient in quantity or quality, and that is the concept the Cedar Valley team intends to study. An operational definition says exactly how the idea will be measured in this study. For Cedar Valley, loneliness is operationalized as the total score on the three-item UCLA Loneliness Scale (Hughes et al., 2004), and "lonely" means a score of 6 or higher. Lesson 7 develops conceptual and operational definitions, levels of measurement and the choice of validated instruments in detail.

Naming variables

Variable names will later be typed into statistical software, so they follow simple conventions. Use short names in lower case, start each name with a letter, join words with the _ character, and avoid spaces and special characters. Make names informative enough to be read without the dictionary, such as ed_visits_12m for ED visits in the 12 months after the survey. Keep a clear link between DAG node names and dataset names; the Cedar Valley team simply converted its node names to lower case and added detail where a node needed more than one dataset variable.

4.4 Checking the DAG Against the Data

The last step before data collection is a gap check, in which the team compares its DAG with the data it can actually obtain. The Cedar Valley gap check produced four findings, which show the kinds of problems such a check reveals.

The gap check also works in the other direction. Some variables the team wants to collect have no node in the DAG, such as gender and the language spoken at home. These variables are still useful, because they describe who took part and help readers judge to whom the findings apply. Lesson 11 shows how such descriptive variables appear in a Table 1. They are recorded in the data dictionary with the role "descriptive", which keeps the reason for collecting them visible.

4.5 Putting It Together

The work of this lesson produces a set of linked documents: an evidence table, a causal web, a DAG with a justification table, a variable list and the start of a data dictionary. Together they explain what the team believes about causes, what it will measure and why. The Cedar Valley team will use these documents again when it maps interest holders in Lesson 4, drafts its consent form in Lesson 5 and builds its survey in Lesson 8.

The same chain applies to a qualitative question, with one adjustment. For a question in SPIDER form, the causal web serves as a map of the factors that shape the experience under study, and the DAG is drawn for a related quantitative question, as the Cedar Valley team did with its PECO question.

Bringing It All Together

This lesson followed the fictional Cedar Valley Social Connection Study from a broad literature summary to a plan for data collection. The team began with an evidence table and link statements, drew a causal web that included community and structural causes as well as individual ones, and then disciplined that web into a directed acyclic graph in which every node is a measurable variable at a stated time and every arrow points forward in time.

With the DAG drawn, the team named the role of each variable relative to its PECO question. Age, income, living alone, rural residence, hearing loss, chronic conditions, earlier depression and frailty are confounders; depression that develops after the survey is a mediator; and referral to the connector program is a collider. The contrast between earlier and later depression showed that the same condition can play different roles depending on when it is measured, which is the key to telling a mediator from a confounder.

Finally, the team turned its DAG into a variable list and a first data dictionary, and a gap check exposed an unmeasured confounder and two timing problems before any data were collected. The reflection and the 15-item assessment below ask you to apply these steps to new questions.

The final reflection asks you to build and label a DAG for a research question and to explain the decisions it implies for data collection.

Final Knowledge Assessment

This glossary defines the terms, tools and people introduced in this lesson on causal webs and directed acyclic graphs.