Informative Visit Processes: When Who Shows Up Starts Writing the Results
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
Real-world longitudinal data rarely arrive on a tidy schedule. Patients return when they are worried, when clinicians are worried, when a portal alert fires, when a treatment is started, when symptoms flare, or when everyone forgets for a while and the chart goes quiet. That is life. It is not neutral measurement.
An informative visit process exists when the timing of observation depends on latent health status, treatment, or both. Once that happens, the rows in your repeated-measures dataset are no longer just data points. They are selected opportunities to observe the outcome. If you ignore that, the apparent effect can start borrowing signal from who showed up and when.
The Core Design Rule
If encounter timing depends on prior symptoms, clinician concern, test results, alert systems, or treatment assignment, then visit times are part of the data-generating mechanism and need design or analytic attention rather than polite omission.
Decision rule:
When your outcome can only be observed at visits and the visit schedule is endogenous, do not trust a naive longitudinal estimate until you have asked who gets seen more often, why, and whether that mechanism differs by exposure.
Or less ceremoniously: if worse patients show up more, and one arm manufactures extra showing up, the outcome curve is already carrying methodological makeup before you model it.
How This Differs from Plain Surveillance Bias
Surveillance bias is the broad idea that more looking can create more recorded disease. Informative visit processes are the longitudinal, repeated-measures version of that problem. The issue is not merely that one group gets more tests. The issue is that the observation schedule itself is responding to evolving health status and often to the intervention.
Health state drives visits
Sicker patients come back sooner, call more often, and trigger more labs, imaging, or chart review.
Intervention drives visits
Treatment initiation, remote-monitoring programs, AI alerts, and safety workflows create extra observation opportunities.
Outcome capture depends on visits
Symptoms, laboratory worsening, toxicity grades, functional status, and even some diagnoses only become visible when someone looks.
Put those three together and you get an outcome process entangled with an encounter process. That is a very different creature from "we had a few irregular timestamps and mixed models handled it."
A Concrete Clinical Example
Imagine a heart-failure remote-monitoring program that prompts extra nurse calls and laboratory checks after symptom alerts. The study compares the intervention with usual care and tracks worsening renal function over 90 days using creatinine-based criteria recorded in the electronic health record.
What the abstract says
The intervention arm had more renal deterioration events, suggesting a possible safety tradeoff.
What may really differ
The intervention arm generated more contact, more repeat creatinine tests, and more opportunities to convert transient abnormalities into coded outcomes.
What a reviewer should ask
Were encounter intensity, test frequency, and symptom-triggered follow-up treated as design features or silently outsourced to the EHR?
That does not prove the program is harmless. It proves the paper has not yet separated worsening kidney function from worsening opportunity to notice kidney function.
Interactive visit-process explorer
The true deterioration risk stays fixed. The recorded risk does not.
This toy model assumes both groups have the same underlying probability of clinical deterioration. The only thing that changes is how often deterioration generates encounters and how many extra monitoring visits one arm receives. More observation can manufacture more recorded events before biology changes at all.
Think remote-monitoring prompts, protocol callbacks, symptom portals, or denser post-initiation safety checks. The treatment effect stays at zero here. Only the visit process moves.
Control arm
18%
Recorded deterioration risk under routine follow-up and symptom-driven visits.
Intervention arm
21%
Higher recorded deterioration created by extra monitoring rather than extra underlying harm.
Detection gap
3%
Apparent absolute risk increase generated by observation timing alone.
| Quantity | Routine arm | Extra-monitoring arm | Why it matters |
|---|---|---|---|
| Visits among patients who are deteriorating | 4.8 | 7.7 | Once visit schedules depend on symptoms and intervention, the observation process becomes part of the estimator. |
| Recorded deteriorations per 1,000 patients | 180 | 206 | Same biology, different counts, because one arm has more chances to turn latent worsening into a recorded outcome. |
| Hidden deteriorations still uncaptured | 4% | 1% | The lower-intensity arm can look better simply because more worsening stays off the record. |
| Apparent risk ratio | 1.14x | If visit timing is ignored, this can be sold as treatment harm or treatment sensitivity depending on the abstract writer's mood. | |
How to read the toy model
This is not a validated inferential tool. It isolates one stubborn problem: when worsening health creates visits and one arm also creates more visits, the observed data are denser exactly where clinical state is changing.
In the real world this stacks with confounding, missing data, and measurement error. The visit process is rarely the only problem. It is simply one that gets ignored with suspicious regularity.
Decision rule
If encounter timing depends on prior symptoms, clinician concern, alert systems, or treatment assignment, then observation times are informative and should not be treated like neutral calendar marks.
The more your outcome depends on being seen, the less comfortable you should be with naive repeated-measures summaries.
Where Informative Visit Processes Create Trouble
| Setting | Why visits are informative | Failure mode |
|---|---|---|
| EHR symptom or lab trajectories | Sicker patients generate denser measurements and clinicians order repeats when values look bad. | The observed trend over-represents periods of instability and can make one arm look more volatile than it truly is. |
| Remote monitoring or AI triage programs | The intervention explicitly creates additional review points, callbacks, and confirmatory testing. | Higher event capture can be misread as toxicity, superior detection, or both, depending on the author's ambitions. |
| Comparative effectiveness after treatment start | Initiation and titration create asymmetric follow-up early, especially for newer or riskier therapies. | Early harms and early biomarker changes become easier to find in one arm before true risk has had a fair chance to differ. |
| Pragmatic or hybrid trials | Flexible care pathways allow adherence, concern, and worsening symptoms to influence return visits. | Repeated-measures outcomes can drift away from the protocol question and toward a care-intensity question. |
What Good Practice Looks Like
| Design task | Minimum credible move | Why it helps |
|---|---|---|
| Describe the observation schedule | Report visit frequency, lab cadence, alert-trigger rules, missed visits, and whether the intervention changes any of them. | You cannot diagnose an informative visit process if the manuscript treats follow-up mechanics like stage props. |
| Prefer fixed assessment windows when possible | Define outcomes on prespecified windows or comparable time anchors rather than on whatever encounter happened to occur. | Fixed windows reduce the direct dependence of outcome measurement on ad hoc visit timing. |
| Model the visit process | Use explicit visit-intensity modeling, inverse-intensity weighting, joint modeling, or well-justified sensitivity analyses when observation is clearly endogenous. | The point is not to perform a ritual method. The point is to acknowledge that outcome observation is selectively generated. |
| Pressure-test harder endpoints | Compare the pattern against outcomes less sensitive to encounter timing, such as hospitalization, procedure, or death when clinically relevant. | If only soft or visit-dependent outcomes move, measurement dynamics deserve suspicion before mechanism gets the credit. |
Reviewer Red Flags
- The outcome is only observable at visits, but the visit schedule is not described.
- The intervention itself creates callbacks, portal prompts, safety labs, or confirmatory testing.
- The paper reports person-time but not encounter intensity, laboratory density, or missed-visit patterns.
- Worsening symptoms plausibly increase visit probability, yet the analysis treats observation times as ignorable.
- The main signal lives in mild events, laboratory abnormalities, symptom scores, or short-term fluctuations that require active measurement.
- Mixed models or repeated-measures regression appear in the methods section as if they automatically dissolve endogenous observation.
What to Ask Before You Trust the Result
1. Does the exposure change how often patients are seen?
If yes, outcome capture may differ before any biology does.
2. Does worsening health itself trigger visits?
If yes, observation times are carrying prognostic information and cannot be shrugged off as random irregularity.
3. Could the outcome be assessed on fixed windows instead?
If yes, ask why the design preferred encounter-driven observation over a more stable assessment rule.
4. Did the analysis explicitly address informative observation?
If not, the result may still be useful, but it belongs in the exploratory bucket rather than the confident-causal bucket.
The Practical Bottom Line
In longitudinal clinical data, observation schedules are often part of the mechanism, not a clerical afterthought. When who shows up depends on how people are doing, and the intervention also changes who shows up, the dataset starts selecting its own drama.
That is exactly the sort of thing Aqrab should help research teams notice early: before the analysis plan hardens, before the methods section becomes decorative, and before reviewers have to explain that "more measurements" is not the same thing as "more disease." If you are building or auditing this kind of study, theAqrab workspaceis a good place to pressure-test visit schedules, outcome definitions, and reviewer red flags while the design is still willing to improve.
Keep reading
Don't stop at one method.
Good methods judgment comes from contrast. Read the neighboring guides, see where the assumptions diverge, and avoid treating every observational problem like it needs the same hammer.
Differential Misclassification: When One Study Arm Gets More Chances to Be Wrong
A practical guide to differential misclassification for clinical researchers. Covers arm-specific outcome detection, adjudication asymmetry, false positives, missed events, and what reviewers should demand before trusting an effect estimate.
Channeling Bias: When the Newer Treatment Inherits the Easier Patients
A practical guide to channeling bias for clinical researchers. Covers preferential prescribing, formulary-era drift, specialist selection, and what reviewers should demand before trusting observational comparisons of newer therapies.
Confounding by Contraindication: When the Untreated Group Is Too Fragile for the Therapy
A practical guide to confounding by contraindication for clinical researchers. Covers how treatment avoidance in high-risk patients can make therapies look safer or more effective than they are, and what reviewers should demand instead.