03 · What You Need to Know
Why Temporal Ordering and Causal Identification Are Different
What Is Temporal Precedence?
Temporal precedence means that a proposed cause occurs before the outcome it is hypothesized to influence.
For example, if teachers complete an AI training program in January and their instructional practices are assessed in June, the training precedes the measured June outcome.
This sequence is compatible with a causal effect of training on later practice.
However, the same sequence is also compatible with alternative explanations, including changes that would have occurred without training.
Temporal precedence helps rule out certain interpretations in which the later measured outcome directly causes the earlier event. It does not establish that the earlier event was independent of prior causes of the later outcome.
Why Is Temporal Order Important for Causal Inference?
For an ordinary forward causal effect, the cause must precede the consequence.
If researchers measure an exposure only after the outcome occurs, they may have difficulty establishing whether the exposure preceded the relevant disease, behavior, or institutional change.
Longitudinal measurement can clarify this sequence.
For example, measuring baseline motivation before subsequent course withdrawal provides stronger temporal evidence than asking withdrawn students to reconstruct their earlier motivation retrospectively.
Nevertheless, the measured baseline may not be the true beginning of the causal process. Earlier unobserved events could have influenced both motivation and withdrawal.
Researchers should therefore distinguish the temporal order of recorded measurements from the temporal order of the underlying mechanisms.
What Is the Difference Between Prediction and Causation?
Prediction concerns whether information about one variable helps anticipate another variable.
Causation concerns whether changing an exposure or intervention would change the outcome under a specified causal contrast.
| Question |
Primary objective |
Illustrative interpretation |
| Predictive |
Determine whether baseline AI literacy predicts later AI use |
Higher literacy is associated with greater subsequent use |
| Causal |
Determine whether increasing AI literacy changes later AI use |
AI use would differ under a specified literacy intervention compared with an alternative |
| Descriptive longitudinal |
Determine how AI use changes over time |
Observed use increased or decreased across occasions |
A variable can be a strong predictor without being a useful intervention target.
For example, baseline institutional resources may predict later AI adoption, but an association alone does not establish what would happen if a particular resource were increased.
Prediction models and causal models may use similar statistical techniques while answering fundamentally different questions.
What Is the Counterfactual Basis of Causal Inference?
Causal effects are commonly conceptualized by comparing outcomes that would occur under different exposure or intervention conditions.
For an individual, the causal contrast might compare the outcome under participation in an AI training program with the outcome under nonparticipation.
Only one of these potential outcomes can ordinarily be observed for that individual under the realized circumstances.
This is the fundamental problem of causal inference.
Research designs attempt to estimate causal contrasts by constructing or identifying suitable comparisons across individuals, groups, or periods.
Repeated observation of one individual before and after an intervention does not directly reveal the outcome that would have occurred at the later time without the intervention.
Why Doesn't a Before-and-After Difference Establish Causation?
Observed change may result from the intervention, but it may also reflect other processes occurring during the observation period.
Examples include maturation, secular trends, changes in measurement, regression to the mean, and concurrent institutional events.
Suppose teachers' AI confidence increases after a professional development program.
During the same period, the university may introduce new AI tools, colleagues may share resources, and teachers may gain experience independently.
A simple before-and-after comparison cannot determine how much of the observed change would have occurred without the program.
The limitations of pretest-posttest causal inference therefore remain even when temporal ordering is clear.
What Is Confounding in Longitudinal Research?
Confounding occurs when an exposure-outcome association is distorted by other variables or causal pathways that create a noncausal association.
Suppose institutional support influences both teachers' AI literacy and their later use of AI tools.
Teachers with greater support may receive more training and have better access to relevant technologies.
Even if literacy has no causal effect on use, the two variables could be positively associated because of shared determinants.
Measuring literacy before use does not remove this confounding.
Researchers should identify plausible common causes using substantive evidence and causal reasoning, rather than assuming that temporal precedence eliminates the need for adjustment.
What Is Time-Varying Confounding?
Time-varying confounding occurs when variables that influence exposure and outcome change during follow-up.
For example, workload may influence teachers' subsequent AI use and later burnout, while earlier AI use may itself change workload.
This creates a longitudinal causal structure in which a time-varying confounder can also be affected by prior exposure.
Ordinary regression adjustment for such variables may block part of an exposure effect or create other estimation problems.
Methods such as marginal structural models, the parametric g-formula, and related g-methods can address certain time-varying confounding structures under appropriate assumptions.
However, these methods require sufficient longitudinal information, including relevant exposure and confounder histories.
They also depend on assumptions such as exchangeability, consistency, positivity, and appropriate model specification.
Can Reverse Causation Still Occur With Longitudinal Data?
Yes, particularly when the observed measurements do not capture the beginning of the underlying process.
Suppose baseline AI confidence predicts later technology use. Teachers may already have used AI extensively before baseline, and that earlier use may have increased their confidence.
The observed relationship could therefore reflect an earlier feedback process rather than a one-directional effect beginning at baseline.
Similarly, early symptoms of a disease may influence behavior before the disease is formally diagnosed.
Researchers should distinguish reverse causation involving measured events from feedback or unobserved processes that began before the observation window.
Additional waves and historical measurements may clarify these relationships, but they do not automatically identify causal direction.
Do Cross-Lagged Panel Models Establish Causation?
No. Cross-lagged panel models estimate associations between variables measured at different occasions under specified statistical assumptions.
For example, baseline confidence may predict later AI use after accounting for earlier use.
Such coefficients can be informative about temporal prediction, but they do not independently eliminate confounding or establish intervention effects.
Traditional cross-lagged panel models may also mix stable between-person differences with within-person processes.
Alternative specifications, including random-intercept cross-lagged panel models, attempt to distinguish these components under additional assumptions and sufficient repeated measurements.
Even then, causal interpretation requires a defensible identification strategy beyond fitting a longitudinal structural equation model.
Why Is the Measurement Interval Important?
The spacing between observations determines which temporal relationships can be examined.
If a hypothesized effect develops within days but variables are measured annually, the relevant sequence may be obscured.
Conversely, measuring every hour may be uninformative for a process that develops over several years.
Lagged coefficients are conditional on the chosen measurement interval and model. They should not be treated as universal estimates of causal delay.
Researchers should align measurement spacing with the expected process and acknowledge unobserved changes between waves.
Can Selection Bias Affect Longitudinal Causal Claims?
Yes. Participants may enter or remain in a study for reasons associated with exposure and outcome.
Suppose students with low motivation are more likely to leave university and stop responding to follow-up surveys.
The retained sample may disproportionately contain students with higher motivation and more favorable academic outcomes.
Associations estimated among those retained may differ from the relationships in the original cohort.
Statistical methods may address certain missing-data mechanisms under assumptions, but they cannot guarantee recovery of unobserved outcomes.
Designing for participant attrition is therefore important for causal as well as descriptive longitudinal research.
What Is Measurement Error's Role in Temporal Inference?
Measurement error can distort longitudinal associations, especially when variables are measured with different reliability across occasions.
For example, teachers may interpret an AI literacy questionnaire differently after gaining practical experience.
Apparent change could partly reflect a shift in the meaning of the measurement rather than change in the underlying construct.
Similarly, measurement error in baseline predictors can affect estimated lagged associations.
Researchers should evaluate construct validity, reliability, and longitudinal measurement comparability.
Temporal ordering does not compensate for invalid measurement.
Can Randomization Strengthen Longitudinal Causal Inference?
Yes. Random assignment can help create exchangeability between intervention groups at baseline, subject to proper implementation and the intended causal estimand.
When participants are randomly assigned to an intervention or control condition and followed over time, differences in subsequent outcomes may support causal inference more strongly than an uncontrolled observational comparison.
However, randomization does not automatically eliminate problems involving nonadherence, attrition, interference, measurement error, or deviations from the protocol.
The analysis must reflect the assignment mechanism and estimand.
Longitudinal outcome measurement can complement randomization by describing when effects emerge and whether they persist.
Can Quasi-Experimental Designs Support Causal Inference?
Yes, under defensible identification assumptions.
Interrupted time-series designs, difference-in-differences approaches, regression discontinuity designs, and other quasi-experimental methods can sometimes estimate causal effects without random assignment.
However, their validity depends on design-specific assumptions.
For example, difference-in-differences analysis commonly relies on an appropriate parallel-trends assumption for the untreated potential outcomes.
Interrupted time-series analysis requires consideration of underlying trends, concurrent events, autocorrelation, and other temporal influences.
Collecting longitudinal data is not equivalent to satisfying these assumptions.
What Does a Causal Diagram Contribute?
A directed acyclic graph can represent hypothesized causal relationships among exposures, outcomes, confounders, mediators, and selection variables.
It can help researchers determine which variables should be measured and which adjustment strategies may be inappropriate.
For example, a diagram may reveal that controlling a mediator would remove part of the total effect or that conditioning on a collider could introduce bias.
However, a causal diagram is only as credible as the assumptions used to construct it.
It does not independently establish that the proposed relationships are true.
Researchers should use diagrams to clarify assumptions and evaluate identification, not as substitutes for empirical evidence.
What Assumptions Are Needed for Causal Interpretation?
The relevant assumptions depend on the estimand and design, but observational causal analyses commonly require consideration of exchangeability, consistency, positivity, and appropriate measurement and model specification.
Exchangeability concerns whether the comparison can be interpreted without uncontrolled confounding under the relevant conditions.
Consistency concerns the relationship between observed outcomes and potential outcomes under the exposure or intervention actually received, requiring sufficiently well-defined treatment conditions.
Positivity requires that relevant exposure or intervention alternatives be possible within the covariate patterns needed for estimation.
These assumptions are not guaranteed by repeated measurement.
Researchers should identify which assumptions support the particular causal claim and whether the available evidence makes them plausible.
How Should Researchers Word Longitudinal Findings?
Language should reflect the evidence and design.
If baseline AI literacy is associated with later technology use in an observational study, researchers can report that higher baseline literacy predicted or was associated with greater subsequent use, subject to the meaning of prediction in the analysis.
They should avoid claiming that literacy increased adoption unless the design and assumptions support a causal interpretation.
Similarly, observing improvement after an intervention does not establish that the intervention produced the improvement.
Researchers should distinguish temporal association, prediction, and estimated causal effects throughout the abstract, results, and discussion.
The objective is not to avoid causal language categorically. It is to use causal language only when the identification strategy and evidence justify it.
