03 · What You Need to Know
Understanding Temporal Precedence and Its Limits in Causal Research
What Is Temporal Order in Research?
Temporal order refers to the sequence in which events, exposures, interventions, and outcomes occur.
For a proposed causal relationship, the cause must precede the effect it produces.
For example, feedback received after a writing assessment cannot cause performance on that earlier assessment.
By contrast, feedback received before the assessment could plausibly influence performance, although additional evidence would be required to establish that it did.
Temporal order is therefore a necessary feature of a coherent causal explanation.
However, it is only one part of the causal argument.
Why Does Temporal Precedence Matter?
Temporal precedence helps rule out causal explanations that are inconsistent with the sequence of events.
For example, if AI feedback use is measured before a later writing assessment, the later score cannot cause the earlier recorded exposure.
However, students' earlier writing ability may influence both their AI use and their later performance.
Similarly, students may anticipate future academic difficulties and begin using AI tools before the final assessment.
Thus, prospective measurement can clarify timing without eliminating the possibility that earlier characteristics or expectations influenced exposure selection.
Researchers should distinguish temporal precedence from exchangeability and other causal identification requirements.
What Is the Difference Between Temporal Order and Causation?
Temporal order establishes that one event occurred before another. Causation concerns whether changing the earlier event would change the later outcome under a specified intervention contrast.
For example, students may begin using AI-supported feedback before their writing scores improve.
The sequence is compatible with a causal effect of AI feedback.
However, the same sequence could arise if motivated students were more likely to use the system and improve through additional practice.
The causal question requires evidence about what would have happened to comparable students under an alternative feedback condition.
Temporal ordering does not provide that counterfactual evidence by itself.
How Can a Relationship Have Correct Temporal Order but Still Be Noncausal?
Several processes can generate an exposure-outcome association even when exposure precedes the measured outcome.
| Alternative explanation |
Illustration |
Why temporal order is insufficient |
| Confounding |
Prior achievement influences both AI use and later writing performance. |
An earlier common cause may generate the observed association. |
| Selection bias |
Participation depends on characteristics related to exposure and outcomes. |
The analyzed sample may contain a distorted relationship. |
| Concurrent influences |
Additional tutoring occurs after AI adoption and before assessment. |
Other changes may contribute to the outcome. |
| Anticipatory behavior |
Students expecting difficulties begin using AI before final assessment. |
Earlier expectations may influence exposure selection. |
| Measurement error |
Recorded AI use does not represent actual feedback engagement. |
Correct timing does not establish valid exposure measurement. |
| Historical trends |
Writing scores improve over time alongside increasing AI adoption. |
Both may change because of broader educational developments. |
These explanations differ methodologically. Some concern confounding, others selection, measurement, or temporal changes. A defensible causal design should identify which are plausible in the specific setting.
What Is the Post Hoc Ergo Propter Hoc Fallacy?
The phrase post hoc ergo propter hoc means, approximately, "after this, therefore because of this."
It describes the mistaken inference that because one event follows another, the earlier event must have caused the later event.
For example, a university introduces an AI-supported feedback system, and average writing scores increase the following semester.
The sequence alone does not establish that the system produced the improvement.
Changes in teaching, student composition, assessment difficulty, or other conditions may explain part or all of the observed increase.
The fallacy is particularly relevant to simple before-and-after evaluations.
However, recognizing it should not lead researchers to dismiss temporal evidence entirely. Timing remains necessary for identifying plausible causal directions.
Can Confounding Exist Even When Exposure Is Measured First?
Yes. Confounding can arise from characteristics that precede both exposure and outcome.
Suppose students with higher prior writing ability are more likely to use AI-supported feedback.
Prior ability may also predict later writing performance.
Even when AI use is recorded before the final assessment, the observed association may partly reflect prior ability.
In a simplified causal structure, prior ability influences both AI use and writing performance.
Measuring exposure prospectively does not remove that common cause.
Researchers should identify relevant confounders using substantive knowledge and consider whether the design supports appropriate adjustment or another identification strategy.
Does Temporal Order Rule Out Reverse Causality?
It can rule out certain forms of reverse causality, but not every temporal complication.
If AI use is recorded before the final writing score, that future score cannot directly cause the earlier recorded AI use.
However, earlier writing performance may influence AI adoption and also predict later performance.
Students may also change their behavior because they anticipate future academic outcomes.
Consequently, prospective measurement reduces some ambiguity about direction without necessarily establishing that the exposure is independent of prior outcome-related processes.
Researchers should distinguish the timing of the measured outcome from the history of the underlying construct.
Can Longitudinal Research Establish Causation?
Longitudinal research can strengthen causal reasoning by observing variables across time.
Repeated observations may clarify temporal ordering, describe trajectories, and support particular causal designs.
However, longitudinal data do not automatically establish causation.
Participants may differ in stable or changing characteristics that influence both exposure and outcomes.
Time-varying confounding may also arise when earlier exposure affects later covariates that influence subsequent exposure and outcomes.
Hernán and Robins (2020) discuss why longitudinal causal inference may require methods that explicitly address treatment and confounding histories.
The number of measurement occasions does not by itself determine causal validity.
What Is Time-Varying Confounding?
Time-varying confounding occurs when characteristics that influence exposure and outcomes change over time.
For example, students' academic confidence may influence whether they continue using AI feedback and may also influence subsequent writing performance.
Earlier AI feedback use may itself affect confidence.
In this situation, confidence may be both a consequence of earlier exposure and a confounder of later exposure-outcome relationships.
Conventional regression adjustment may be inappropriate for some longitudinal causal questions because conditioning on treatment-affected confounders can introduce bias or remove parts of the effect of interest.
Methods such as the parametric g-formula, inverse probability weighting, or other g-methods may be relevant under appropriate assumptions.
These methods require careful specification and do not eliminate the need for credible measurement and identification assumptions.
Why Is a One-Group Pretest-Posttest Design Limited?
A one-group pretest-posttest design measures outcomes before and after an intervention among the same participants.
For example, students complete writing assessments before and after receiving AI-supported feedback.
The design can describe observed change.
However, improvement may reflect regular instruction, maturation, practice, regression to the mean, or other events occurring during the intervention period.
Without an appropriate counterfactual comparison, the researcher cannot determine how much change would have occurred without the intervention.
Shadish, Cook, and Campbell (2002) discuss these threats in experimental and quasi-experimental research.
The presence of temporal order does not resolve them.
What Is the Difference Between Change Over Time and a Causal Effect?
Observed change compares outcomes measured at different times.
A causal effect compares potential outcomes under alternative intervention conditions.
For example, students' average writing score may increase by eight points during a semester.
That eight-point change is not necessarily the effect of AI-supported feedback.
Students might have improved by six points through ordinary instruction even without the intervention, in which case the intervention-related contrast would differ from the observed change.
However, the hypothetical six-point counterfactual cannot simply be assumed. It must be supported through an appropriate design and identification assumptions.
Researchers should avoid treating the observed before-and-after difference as equivalent to a treatment effect.
Does Measuring More Time Points Solve the Problem?
Additional measurements may help characterize trends and distinguish some temporal patterns.
For example, repeated writing assessments before and after an intervention may reveal whether performance changed abruptly or gradually.
However, more observations do not automatically establish causation.
Changes in curriculum, assessment procedures, student composition, or other interventions may coincide with the focal intervention.
The appropriate interpretation depends on the design, statistical model, and assumptions concerning the counterfactual outcome trajectory.
Repeated observations can strengthen a causal argument when incorporated into a defensible identification strategy, not simply because the dataset contains more time points.
How Can Interrupted Time-Series Designs Help?
An interrupted time-series design uses repeated outcome observations before and after a defined intervention or event.
It may estimate changes in outcome level or trend associated with the intervention point.
For example, a university could examine monthly indicators of student writing performance before and after introducing a new feedback policy.
Bernal, Cummins, and Gasparrini (2017) discuss interrupted time-series regression and its methodological considerations.
However, causal interpretation requires assumptions about the pre-intervention trajectory, model specification, seasonality, autocorrelation, and other events occurring around the intervention.
A concurrent comparison series may strengthen the design when an appropriate reference exists.
The intervention's position in time is not sufficient by itself.
Can Difference-in-Differences Strengthen Temporal Causal Inference?
Difference-in-differences compares changes over time between groups experiencing different intervention conditions.
For example, a university department introducing AI-supported feedback may be compared with another department that does not, using observations before and after implementation.
The design can account for certain fixed group differences and common time-related influences under appropriate assumptions.
However, causal interpretation depends on the credibility of counterfactual trend assumptions and other design conditions.
Parallel trends concerns how the groups' outcomes would have changed in the absence of treatment, not whether their observed baseline outcomes were identical.
Researchers should not assume that subtracting two observed changes automatically produces a causal effect.
What About Cross-Lagged Panel Models?
Cross-lagged panel models examine relationships among variables measured repeatedly over time.
For example, a researcher may investigate whether earlier AI use predicts later academic confidence after accounting for previous confidence measurements.
Such models can provide information about temporal associations.
However, a statistically significant cross-lagged coefficient does not automatically establish a causal effect.
Unmeasured confounding, measurement error, stable individual differences, and model assumptions may influence the estimated relationships.
Hamaker, Kuiper, and Grasman (2015) demonstrate important limitations of the traditional cross-lagged panel model when stable between-person differences are not adequately distinguished from within-person processes.
Researchers should therefore avoid treating temporal prediction as equivalent to causal identification.
Does Granger Causality Establish Causation?
Granger causality concerns whether past values of one time series improve prediction of another series beyond the information provided by relevant past observations.
Despite its name, this predictive relationship does not automatically establish causal effects in the interventionist or counterfactual sense.
For example, earlier platform activity may improve prediction of later assessment performance without establishing that increasing platform activity would improve performance.
Unmeasured common influences and model specification can affect predictive relationships.
Researchers should distinguish temporal predictive usefulness from causal identification.
Can Randomization Strengthen the Temporal Argument?
Yes. Random assignment to well-defined intervention conditions can strengthen causal inference by addressing systematic baseline selection in expectation.
For example, students may be randomly assigned to AI-supported or conventional feedback before the instructional period begins.
Outcomes are then measured after exposure to the assigned conditions.
The combination of appropriate temporal ordering and random assignment provides stronger causal evidence than a voluntary exposure comparison with the same measurement schedule.
However, randomization does not eliminate every threat involving missing outcomes, nonadherence, interference, or measurement.
The causal claim should remain aligned with the intervention and assignment structure.
What If Randomization Is Impossible?
Researchers should investigate whether the exposure or intervention assignment process supports a credible observational or quasi-experimental design.
For example, an institutional policy may be introduced according to a threshold, a predetermined schedule, or another rule that creates potentially informative variation.
Regression discontinuity, interrupted time series, and difference-in-differences may be appropriate in particular circumstances.
However, each depends on specific assumptions and data requirements.
Researchers should not choose a quasi-experimental label simply because the study has measurements before and after an intervention.
The design must explain why the comparison provides credible information about the relevant counterfactual.
How Does Temporal Measurement Affect Exposure Definition?
Researchers should distinguish when an exposure begins, how long it continues, and when the outcome is measured.
For example, AI feedback use recorded at the beginning of a semester may not represent use throughout the entire instructional period.
Some students may stop using the system, while others may increase their engagement later.
Classifying students according to future exposure histories without accounting for timing can introduce biases in some observational designs.
Researchers should define eligibility, time zero, exposure strategies, follow-up, and outcome assessment consistently.
Hernán and Robins (2016) discuss the value of specifying a hypothetical target trial when using observational data to investigate causal effects.
Why Is Time Zero Important?
Time zero is the point at which eligibility, treatment assignment or strategy initiation, and follow-up are aligned for a causal comparison.
For example, comparing students who eventually become frequent AI users with those who never use the system may be problematic if the groups are defined using behavior observed throughout follow-up.
Students must remain enrolled and observed long enough to qualify as frequent users, potentially introducing selection or immortal-time-related bias.
Researchers should avoid defining exposure groups using future information in ways that create unequal opportunities to experience the outcome.
Clear temporal alignment is therefore important beyond simply recording that exposure occurred before the final outcome.
How Do Researchers Distinguish Temporal Evidence From Causal Evidence?
Temporal evidence establishes when variables or events occurred.
Causal evidence requires an argument about how the observed comparison identifies the outcome difference under alternative conditions.
Researchers should examine exposure selection, confounding, consistency, positivity, measurement, and other assumptions relevant to the design.
They should also consider whether the outcome could be influenced by earlier values of the same construct or by shared causes.
A temporal association may be valuable for prediction or hypothesis generation even when it does not support causal attribution.
The distinction should be maintained throughout the research question, methodology, results, and discussion.
What If the Temporal Association Is Very Strong?
A strong association with clear temporal ordering may strengthen interest in a causal hypothesis, but it does not independently eliminate alternative explanations.
For example, students who begin using AI feedback may subsequently improve substantially.
However, the improvement may coincide with additional tutoring, intensive revision practice, or other instructional changes.
The magnitude of the association does not establish which process caused it.
Researchers should examine whether the design distinguishes the focal intervention from plausible competing causes.
Strong temporal associations can motivate further investigation without being treated as conclusive causal evidence.
How Should Temporal Claims Be Reported?
Researchers should specify when exposure, relevant covariates, and outcomes were measured.
They should explain whether the data represent a single baseline measurement, repeated observations, or a more complex exposure history.
The methodology should identify how temporal alignment was handled and whether important time-varying confounders were considered.
Results should distinguish prediction and association from causal effect estimates.
The discussion should explain which causal directions are ruled out by the timing and which alternative explanations remain plausible.
Appropriate causal language should follow from the full identification argument rather than temporal precedence alone.
Watch Out
Do not conclude that an exposure caused an outcome merely because the exposure was measured first. Earlier common causes, selection processes, time-varying confounding, and concurrent events may still explain the association. Temporal order is necessary for the proposed causal direction, but it is not a substitute for counterfactual identification.
