2027 Choosing Adjustment Variables With a Causal Diagram for MPH and MS Students
Co-Founder and Chief Data Scientist
A regression model can include many variables and still produce a misleading estimate of a public health intervention's effect. For MPH and MS students, the difficult step is deciding which variables belong in the model before running it.
Causal diagrams make that decision explicit: they show the assumptions connecting an exposure, an outcome, and possible sources of bias. This guide explains how to select an adjustment set, recognize variables that can make bias worse, and find coursework that lets you practice these decisions with real research questions.
Key Things You Should Know
- Adjust for an appropriate set of pre-exposure variables that blocks noncausal paths between the exposure and outcome; do not select covariates solely because they predict the outcome or have small p-values.
- For a total-effect question, generally leave mediators out of the adjustment set, and avoid conditioning on colliders that can create a misleading association.
- A useful training sequence combines study design, directed acyclic graphs, regression, and supervised analysis; a degree or software course alone does not establish causal inference proficiency.
What are adjustment variables in causal diagrams for MPH and MS research?
Adjustment variables are measured characteristics included in an analysis - or addressed through another design or analysis strategy - to reduce bias in an estimate of an exposure's effect on an outcome. A causal diagram, often called a directed acyclic graph or DAG, represents assumptions about which variables cause which others. Its arrows indicate proposed causal direction, not correlations established by a dataset.
Start with a hypothetical public health question: among adults eligible for smoking-cessation coaching, what is the total effect of participating in coaching, compared with not participating, on abstinence six months later? Participation is the exposure; abstinence is the outcome. Baseline nicotine dependence could affect both participation and abstinence, making it a candidate adjustment variable. Skills acquired through coaching occur afterward and may transmit part of its effect.
The population, comparison, timing, and desired effect matter as much as the diagram. Estimating the effect of being offered coaching is different from estimating the effect of participating. Students choosing between practice-oriented and research-intensive training can use a guide to choosing a public health master's specialization, but should inspect actual syllabi for opportunities to formulate and analyze causal questions.
How do causal diagrams identify confounders for public health studies?
A confounder helps create a noncausal association between an exposure and an outcome. In the coaching example, baseline nicotine dependence might influence who seeks coaching and who quits. The diagram contains the path coaching participation ← baseline dependence → abstinence. Comparing participants with nonparticipants without addressing dependence could mix its influence with the effect of coaching.
Draw the diagram using substantive knowledge before inspecting associations in the analysis dataset. Specify when each variable was measured: "motivation" assessed before enrollment may be a common cause, while motivation measured after coaching begins could be affected by coaching. A variable's name alone does not determine its role. Ask clinicians, community partners, or researchers whether the proposed arrows and missing arrows are plausible, and document disagreements.
Students entering graduate methods work from another discipline may need a foundation in epidemiology and statistics before tackling these judgments. Short-session courses for online public health majors describe one undergraduate course format; their pace is not, by itself, evidence of preparation for graduate causal analysis.
Which variables should MPH and MS students adjust for in analyses?
Choose a sufficient adjustment set: a set that blocks the diagram's backdoor paths from exposure to outcome without blocking the effect you intend to estimate or opening a path through a collider. For the hypothetical coaching study, baseline dependence belongs in a candidate set if it affects both participation and abstinence. Other pre-enrollment factors, such as access to coaching and prior quit attempts, require the same causal assessment - not automatic inclusion.
A workable selection process keeps the causal question ahead of the statistical model:
- Write down the eligible population, exposure, comparison, outcome, follow-up period, and whether the target is a total or direct effect.
- Draw plausible arrows among exposure, outcome, pre-exposure characteristics, post-exposure variables, and factors affecting inclusion in the dataset.
- Identify each path that enters the exposure through an arrow pointing toward it; select measured variables that block those paths without conditioning on inappropriate variables.
- Check measurement quality, missingness, and whether comparable exposed and unexposed people exist at relevant covariate levels; then document the proposed adjustment set and alternatives.
Adjustment may be implemented through regression, standardization, weighting, matching, or another justified method. None repairs an omitted, unmeasured common cause simply by adding more unrelated predictors. If foundational preparation is the obstacle, compare the cost of a needed course or supervised project with a full degree; a public health bachelor's affordability guide is relevant chiefly to students who also need undergraduate degree completion, not to someone seeking one advanced methods skill.
For the total effect of coaching on abstinence, do not routinely adjust for a mediator such as skills gained through coaching. The path coaching → skills → abstinence is part of the effect being studied. Controlling for skills would change the question and could introduce additional bias.
A collider is a variable caused by two other variables. Suppose coaching participation and baseline social support both affect attendance at a follow-up visit, while social support also affects abstinence. Conditioning on attendance can open the path coaching → attendance ← social support → abstinence, creating an association that was not present on that path before conditioning. Restricting analysis to people with observed follow-up can have this effect even if attendance never appears as a regression covariate.
Also avoid adding a variable merely because it is statistically significant, strongly predicts abstinence, or is available in an administrative file. Its position in the diagram is the first consideration. Advanced research training can deepen these judgments; finding public health doctorates that fit your background is a separate decision from fixing an adjustment set for a master's project.
A backdoor path connects exposure and outcome but begins with an arrow pointing into the exposure. In coaching ← baseline dependence → abstinence, adjusting appropriately for dependence blocks that path. If several backdoor paths exist, the selected set must block all relevant paths in the proposed diagram. A variable need not appear on every path, and adjusting for every measured variable is not the goal.
Blocking paths in a DAG is an assumption-based criterion, not proof that an observational estimate is causal. The result also depends on adequate measurement, no consequential unaddressed confounding, a well-defined exposure, and sufficient overlap between comparison groups. Survey sampling, missing outcomes, and errors in recorded variables may require separate attention. If plausible alternative diagrams imply different adjustment sets, report that sensitivity rather than presenting one diagram as established fact.
A thesis may justify more advanced work on these limitations, but the method should follow the research problem. Even when comparing weighing tuition for a public health doctorate, students should not assume that a longer credential is necessary to learn backdoor-path reasoning; a rigorous master's course with feedback may meet their immediate need.
Where offered, introductory epidemiology courses may use DAGs to distinguish confounding, mediation, and selection bias. More advanced causal inference courses may connect diagrams to identification assumptions, potential outcomes, standardization, inverse-probability weighting, and questions involving time-varying exposures. These are possible curricular components, not universal MPH or MS requirements.
The useful distinction is between drawing a diagram and defending it. Strong training asks students to specify a target effect, justify arrows using subject-matter knowledge, identify a sufficient set, explain what available data cannot identify, and communicate uncertainty. An MPH may place that work in program evaluation or applied epidemiology; a research-oriented MS may require greater methodological depth. Examine assignments and supervision rather than inferring depth from the degree label.
If admission logistics are also part of a student's decision, MPH admissions without faculty recommendations addresses one application constraint. It should not substitute for checking whether the program actually teaches and assesses causal analysis.
Either format can support causal inference training through diagram exercises, analysis labs, peer critique, and faculty feedback. The meaningful differences are practical: access to live discussion, opportunities to question assumptions, computing support, and supervision for a project involving real public health data. An online course with asynchronous lectures may still require scheduled group meetings; a campus course may offer limited individual feedback.
Ask to see a recent syllabus and a description of assessment. Does the course require students to defend an adjustment set in writing? Can remote students attend office hours and access approved datasets and software? For a thesis or applied project, who will review the diagram before modeling begins? These questions matter more than whether instruction is delivered by video or in a classroom.
Students considering less competitive online MPH admission options should assess admissions fit separately from methods rigor. Ease of entry does not establish the depth of a program's causal inference curriculum.
A practical sequence begins with epidemiologic study design and introductory statistics, then moves to regression and causal reasoning, followed by a supervised analysis or research project. Study design teaches students to define a population and comparison; statistics supports estimation and uncertainty; causal inference links the question to an appropriate adjustment strategy. Programming helps implement and reproduce the analysis but cannot decide which arrows belong in a diagram.
Look for assignments that require both a diagram and a written explanation of why variables are included or excluded. Work with a mentor on one modest, answerable question before attempting complex longitudinal methods. If a student already has quantitative preparation, a targeted course and supervised project may be more efficient than another degree. Conversely, a course in DAG terminology alone may not be enough for someone who has never evaluated measurement, selection, or regression assumptions.
When comparing the time and expense of educational routes, public health degree cost and duration comparisons provide broader context. Verify the courses available in the term you would take them; a short program is not necessarily the shortest route to this particular skill.
Accreditation and methods training answer different questions. Verify the institution's accreditation and, if CEPH accreditation matters to your goals, confirm the exact school, unit, or program covered. Then inspect the curriculum for actual causal inference instruction. CEPH status does not mean every degree offers an advanced DAG course, and institutional accreditation alone does not establish public health-specific accreditation.
Request current syllabi, elective schedules, prerequisites, and information about faculty or project supervision. Ask whether students analyze observational data, receive feedback on diagrams and adjustment choices, and learn to explain findings without overstating causality. For an MS thesis, clarify advisor availability and data-access expectations; for an MPH applied project, ask whether a suitable evaluation question can be pursued within its approved scope. Neither arrangement guarantees access to restricted health data.
An online public health master's accreditation guide can help frame the accreditation check. Confirm current scope with the accreditor and program before enrolling, then make the training decision from course-level evidence.
Which public health careers use causal inference and epidemiologic methods?
Epidemiologists and research analysts may use causal diagrams to plan observational studies, evaluate confounding, and explain limitations. Program evaluators may use the same reasoning to decide whether an observed change can plausibly be attributed to an intervention. Biostatisticians may develop or apply estimation methods alongside subject-matter collaborators. The balance of study design, coding, community knowledge, and advanced methods varies by employer and role.
For an early-career student, a credible portfolio example is a clearly labeled, hypothetical diagram paired with a reproducible analysis of appropriately accessible data and a candid limitations statement - not a claim to have established an intervention's effect. Supervised project experience can demonstrate judgment that a course title cannot. Some specialist or independent research roles require additional quantitative training or experience; an MPH or MS alone does not confer them.
If compensation affects an education decision, use occupation- and location-specific postings alongside average earnings after an MPH. Degree-wide earnings summaries cannot predict pay for a causal inference role or establish the return on a particular methods course.
Other Things You Should Know About Public Health
No. A DAG makes causal assumptions visible and helps determine what those assumptions imply for analysis. Data and a well-drawn diagram alone cannot prove that its arrows are correct.
No. A legible hand-drawn diagram can support an initial discussion. Software can help check complex paths, but it does not supply the subject-matter knowledge needed to draw defensible arrows.
State what is missing and how it might affect the estimate. Consider feasible alternative data, study designs, or sensitivity analyses; do not describe an adjusted association as an identified causal effect if a consequential backdoor path remains open.
Not necessarily. Requirements depend on the activity, data, institution, and intended use. Check institutional review guidance, data-use terms, and any community or site permissions before obtaining or sharing data.
- Principles of confounder selection - PMC - NIH https://pmc.ncbi.nlm.nih.gov/articles/PMC6447501/
- Methods in causal inference. Part 1: causal diagrams and confounding | Evolutionary Human Sciences | Cambridge Core https://www.cambridge.org/core/journals/evolutionary-human-sciences/article/methods-in-causal-inference-part-1-causal-diagrams-and-confounding/E734F72109F1BE99836E268DF3AA0359
- Causal Mediation | Columbia University Mailman School of Public Health https://www.publichealth.columbia.edu/research/population-health-methods/causal-mediation
- Principles of confounder selection - Springer Nature https://link.springer.com/article/10.1007/s10654-019-00494-6
- Handling time varying confounding in observational research | jstor.org https://www.jstor.org/stable/26951608
- Directed Acyclic Graphs | github.io https://soc114.github.io/topics/12_DAGs.html
- 3 Directed Acyclic Graphs | scunning.com https://mixtape.scunning.com/03-directed_acyclical_graphs
- Through the lens of causal inference: Decisions and pitfalls of covariate selection | Published in Aperture Neuro https://apertureneuro.org/article/124817-through-the-lens-of-causal-inference-decisions-and-pitfalls-of-covariate-selection
- 10 Causal estimands | r-causal.org https://www.r-causal.org/chapters/10-estimands
