Caregiver Diaries: Planning for a Change of Respondent

Consider a hypothetical scenario in late-phase clinical development: an 18-month randomized, double-blind clinical trial evaluating a novel disease-modifying therapy for mild Alzheimer's disease specifies a key secondary endpoint derived from a daily caregiver-completed electronic diary. Implemented on a provisioned handheld tablet, the instrument assesses daily living functional autonomy, instrumental activities of daily living (iADLs), and behavioral disruptions across 14 discrete observational items. The patient cannot reliably self-report due to progressive cognitive impairment. At the baseline randomization visit, the participant's 74-year-old husband is enrolled as the primary caregiver respondent, completes documented instrument training, and receives individual log-on credentials.

In month 7 of the trial, the husband suffers an acute medical emergency requiring emergency hospitalization and three consecutive weeks of inpatient rehabilitation. Determined to keep the participant's daily routine intact and prevent missing study data, the couple's 42-year-old adult daughter moves into the family home to take over daily caregiving responsibilities. Finding the study tablet on the kitchen counter with the active session prompt, she faithfully completes the electronic diary each evening for 21 consecutive days using her hospitalized father's saved login credentials. Neither the participant nor the daughter contacts the clinical trial site. The study coordinator only discovers the substitution three weeks later during a scheduled monthly safety check-in call when asking how the husband is managing the daily diary entries.

The clinical development team, trial biostatisticians, and clinical operations leads face an immediate regulatory and methodological dilemma. Three weeks of continuous daily endpoint observations have been captured, but the computer-generated electronic audit trail attributes every entry to a hospitalized individual who was physically absent from the home. The true observer—the daughter—is uncredentialed, never underwent documented protocol training on the instrument's operational definitions, and had never been registered as a diary respondent or issued her own credentials. How does the study team remediate the misattributed entries, preserve regulatory data integrity under 21 CFR Part 11, and prevent this mid-trial substitution from biasing or confounding the secondary endpoint?

The foundational requirement for any clinical trial endpoint derived from observer reporting is establishing the formal regulatory taxonomy of the outcome measure. As detailed in our analysis of clinical outcome assessment selection, the Food and Drug Administration (FDA) and international regulatory bodies maintain rigorous boundaries between Patient-Reported Outcomes (PROs), Observer-Reported Outcomes (ObsROs), Clinician-Reported Outcomes (ClinROs), and Performance Outcomes (PerfOs).

In its landmark 2009 guidance on patient-reported outcome measures, the FDA established an unambiguous rule: "A proxy-reported outcome is not a PRO." A proxy report is a report by someone other than the patient, made as if that person were the patient. The 2009 guidance discourages proxy-reported measures, particularly for symptoms only the patient can know, such as pain intensity. An observer report is different: it is limited to signs, events, or behaviors the observer can actually see.

An Observer-Reported Outcome (ObsRO) is fundamentally distinct. According to FDA's Clinical Outcome Assessment Qualification Program and consensus recommendations from the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) Rare Disease Task Force (Benjamin et al., 2017), an ObsRO is legitimately selected when a concept of interest can only be captured in daily life outside healthcare settings and the patient lacks the developmental or cognitive capacity to self-report (e.g., non-verbal pediatric populations, advanced neurodegenerative diseases, or profound intellectual disability). However, regulatory validity requires that an ObsRO be strictly restricted to observable signs, behaviors, and events—such as whether the patient required physical assistance eating breakfast, wandered outside without a coat, or exhibited vocal protest during bathing.

FDA's October 2025 PFDD Guidance 3, Appendix B, states the implementation rule for observer-reported measures. Consider predefining the minimum regular contact required before someone is selected as an observer. To the extent feasible, the same observer should complete the assessments throughout the trial, to limit variability from different reporters. If another caregiver will provide reports, that caregiver should be sufficiently trained on the ObsRO. In all cases, the reporter associated with an observation should be recorded. That guidance is the subject of this site's fit-for-purpose COA selection guide. Guidance 3 does not prescribe backup-login mechanics or a statistical model for the switch. Those decisions are in the sections below.

Because the observer is part of the measurement, the observer's identity belongs in the context of use. Before sites are activated, the protocol should pre-specify who may observe. The contact threshold below is an example a team might write; it is not an FDA-mandated hour count. A workable eligibility set includes:

• Observation Frequency Threshold: Minimum contact, written in the protocol. For example, the observer lives with the participant or spends a stated minimum of waking time with the participant on a stated number of days each week. FDA fit-for-purpose guidance recommends predefining that minimum; it does not set the hours.

• Familiarity with Patient Baseline: Enough familiarity with the participant to apply the instrument's definitions of observable behavior. This is a protocol choice, not a regulatory formula.

• Operational and Cognitive Capability: Ability to use the diary device, meet the reporting window, and understand the items in the language of administration.

• Absence of Disqualifying Conflict: A named plan for who reports if the primary observer cannot. The failure mode is an unmonitored handoff among unnamed people, not the existence of a backup.

Data origination follows the person who actually enters the observation. FDA's September 2013 electronic source data guidance says that when a subject transmits a patient-reported outcome directly, the subject is listed as the data originator. An ObsRO diary is a different record: the caregiver who observes and enters the diary is the originator who should be named on that entry. A login that still names a hospitalized spouse records the wrong person.

A common vulnerability in clinical protocol design is treating caregiver continuity as an assumed constant. Clinical operations teams frequently assume that the family member enrolled at baseline will remain available, healthy, and willing to complete daily diaries for the entire 12, 18, or 24 months of a pivotal trial. Empirical research demonstrates that informant and caregiver replacement is not an isolated protocol exception, but a predictable demographic reality.

Grill and colleagues (2015) analyzed 1,818 eligible Alzheimer disease dementia participants in the National Alzheimer's Coordinating Center (NACC) Uniform Data Set. Of those participants, 282 (15.5%) had an informant replaced at least once. Replacement typically occurred after the first or second research visit. Among participants replaced once, 42% of switches fell between baseline and the first follow-up and 31% between the first and second follow-up. The rate also depended on who the baseline informant was:

• Spousal Informants: About 10% of participants with a spouse informant at baseline (9.7% in the tabled results) later had that informant replaced. This was the lowest of the three groups.

• Adult Child Informants: About 24% of participants with an adult-child informant at baseline (23.7% in the tabled results) had a replacement.

• Other Relatives and Non-Family Informants: About 38% of participants with another informant at baseline (37.9% in the tabled results) had a replacement.

To determine whether similar attrition occurs in strictly monitored clinical trial environments, Nishida and colleagues (2023) analyzed pooled data from 1,336 participants across four Alzheimer's Disease Cooperative Study (ADCS) randomized trials. They documented that 63 participants (4.7%, or approximately 5%) experienced an informant replacement during the trial. Subsequently, Nishida et al. (2024) investigated 2,637 randomized participants across two large, industry-sponsored Phase 3 trials in mild-to-moderate Alzheimer's disease (the semagacestat clinical development program). Across these two commercial trials, 69 participants (2.6%) experienced 78 distinct informant-replacement events.

The table below synthesizes empirical replacement frequencies across observational registries and controlled clinical trials, illustrating the relationship between study duration, trial oversight, and informant volatility:

When an observer respondent changes mid-trial, how does the transition affect the clinical data? Biostatisticians and outcome assessment specialists must distinguish between two separate questions: First, does changing the respondent introduce acute measurement error and score instability at the transition? Second, does that instability introduce structural bias into overall trial conclusions?

The adjusted analyses show a disturbance at the visit of replacement. They do not separate a new observer's calibration from real clinical change. In the four ADCS trials in Nishida et al. (2023), the mean between-visit change on the ADCS-ADL was 2.44 points lower for replacement versus stable informants (95% CI -3.91 to -0.98). Concurrently, the mean absolute between-visit change was 2.38 points larger (95% CI 1.24 to 3.52). Most strikingly, the variance ratio of 18-month change-from-baseline ADCS-ADL scores between replaced and stable informants was 1.80 (95% CI 1.19 to 2.99). A variance ratio of 1.80 means the variance of that 18-month change was 80% higher in the replacement group. The paper did not report a power calculation.

In the two industry-sponsored trials analyzed by Nishida et al. (2024), this acute friction was corroborated: the mean absolute between-visit ADCS-ADL change was 2.02 points larger for replacement versus stable informants (95% CI 0.34 to 3.70; P = 0.019). The signed mean between-visit change in those same trials was about 1.61 points lower and was not statistically significant (95% CI -3.79 to 0.57; P = 0.147). The replicated finding is larger absolute movement around the switch, not a consistent downward jump.

Furthermore, even when a single respondent remains in place, rater characteristics fundamentally modify score levels. In a rigorous evaluation published in Nature Aging, Vargas-Gonzalez and colleagues (2024) reported associations between informant characteristics and Clinical Dementia Rating Sum of Boxes (CDR-SB) scores in a NACC analysis. The published abstract does not give a sample size, so none is stated here. The associations they report are:

• Informant Kinship: CDR-SB was 0.39 points higher when the informant was the patient's child than when the informant was a spouse or partner, and 0.20 points higher when the informant was female. These are adjusted associations, not a measured difference in tolerance.

• Contact Frequency: CDR-SB was 0.38 points higher with at least weekly contact and 0.65 points higher with daily contact, against the study's lower-contact reference. The result does not show that more contact simply uncovers true impairment. Contact can also track how severe the participant's illness is.

• Cohabitation Status: CDR-SB was 0.57 points higher when the informant lived with the participant. CDR-SB is a clinician-integrated staging scale, not a caregiver diary, and the finding is cross-sectional.

Crucially, however, trial teams must examine the counter-evidence regarding trial-level outcomes. In the two industry Phase 3 trials (Nishida et al., 2024), despite the significant acute between-visit score instability (P = 0.019), there were no statistically significant difference in end-of-study change from baseline between participants with a replacement and those with a stable informant (estimate -0.70 points; 95% CI -5.88 to 4.48; P = 0.790). That estimate is the association of replacement with the participant's own change score, not a treatment-effect contrast. The variance ratio was 1.13 (95% CI 0.67 to 2.28; P = 0.600).

The methodological takeaway is profound: an informant replacement is associated with larger between-visit movement and, in the academic trials, with higher variance of the 18-month change. It has not been shown to shift end-of-study ADCS-ADL change by a statistically significant amount, and the industry-trial interval is too wide to treat that null as proof that a treatment effect is unaffected. Teams do not need to panic and scrap data; instead, they must implement disciplined rater onboarding and pre-specify statistical adjustments to preserve endpoint interpretability, drawing lessons from established frameworks in clinician rater drift and calibration.

In the Alzheimer's cohorts above, informant replacement ranged from 2.6% of randomized participants in two industry trials to 15.5% of eligible registry participants. Those figures are not daily-diary rates and do not transfer automatically to other chronic diseases. They are common enough that the protocol should define the replacement path before the first participant is randomized. Waiting until a caregiver is incapacitated leaves the site choosing between a shared login and a gap in the diary.

Under Principle 6 of ICH E6(R3) (consolidated guideline, 16 June 2026), critical-to-quality factors are identified prospectively because they affect whether results can be interpreted. Who may report, and how a replacement is trained and recorded, belong on that list when the endpoint is an observer diary:

1. Pre-Identification of Backup Caregivers: During screening and informed consent, study sites should identify not only the Primary Caregiver but also a Designated Backup Caregiver who meets identical eligibility criteria (living in the home or maintaining regular observation). While the backup caregiver does not enter data during baseline, their identity, contact information, and eligibility are pre-qualified.

2. Standardized Instrument Training: As emphasized in the Journal of the Society for Clinical Data Management's (JSCDM) eCOA guidance (Amatya et al., 2025), all individuals entering data into an eClinical system must receive documented, standardized training. Caregivers must be instructed on the operational definitions of each diary question—specifically understanding that they are recording strictly observable occurrences during the specified recall period (e.g., "within the last 24 hours"), not interpreting internal motivation or pain.

3. Identity Verification: In accordance with 21 CFR 11.100(b), the organization assigning an electronic signature or login credential must verify the individual's identity before issuing system access. Sites must document that the individual receiving credentials is the verified family member, not an unvetted third party.

4. Protocol Delegation Documentation: Under ICH E6(R3) Annex 1, section 2, the investigator is responsible for trial conduct at the site, including the qualifications of people performing trial-related duties (sections 2.1 and 2.3). A family caregiver who completes an ObsRO is a study partner and, when that person enters the diary, a data originator. Completing the diary does not by itself make the caregiver a trial participant or a member of site staff. The site should still keep a subject-level caregiver log in the investigator site file naming each respondent, the training date, and the date credentials were issued.

The technical core of caregiver diary operations is compliance with electronic record attribution. In our detailed evidence guide on 21 CFR Part 11 for eCOA, electronic observations entered by a caregiver often have no separate paper chart to check them against; the electronic record is the source for that entry. The audit trail is what lets a reviewer reconstruct who created or changed the record. It is not, by itself, proof that the named user was the person in the room.

Under Title 21 of the Code of Federal Regulations, two specific provisions govern caregiver credentialing:

• 21 CFR 11.10(e) — Audit Trails: Systems must maintain secure, computer-generated, time-stamped audit trails that independently record the date and time of operator entries and actions that create, modify, or delete records. Crucially, audit trail documentation must be fully attributable to the specific individual performing the action.

• 21 CFR 11.100(a) — Signature Uniqueness: When an electronic signature is used, 21 CFR 11.100(a) requires that signature to be unique to one individual and not reused or reassigned. Where the signature is an identification code combined with a password, 21 CFR 11.300(a) requires that no two individuals have the same combination.

JSCDM eCOA guidance (Amatya et al., 2025) says that only trained, authorized people should have access, each with their own credentials, and that two people should not share an identification or password. An audit trail that names a hospitalized husband for entries made by his daughter records the wrong operator. That is an attribution failure under 11.10(e), and it is the kind of record an inspector can question.

To operationalize compliant transitions, modern eCOA systems must support a structured respondent handover workflow. When a caregiver transition occurs, a free-text note is not enough for the analysis dataset. The data plan should carry sponsor-defined fields such as these. The names are examples, not CDISC standard variables:

• RESPONDENT_ID: A unique, immutable system identifier assigned to the specific caregiver.

• RESPONDENT_TYPE: Categorical classification (e.g., PRIMARY_SPOUSE, BACKUP_ADULT_CHILD, PROFESSIONAL_NURSE).

• HANDOVER_START_DT: The date and time the new respondent finished training and began reporting, stored so the analysis dataset can align scores to the switch.

• HANDOVER_REASON: Controlled terminology capturing the cause of transition (e.g., PRIMARY_ILLNESS, CAREGIVER_BURDEN, PERMANENT_RELOCATION).

• TRANSITION_FLAG: A sponsor-defined indicator in the analysis dataset, for example an ADaM covariate or a supplemental qualifier, marking observations inside a pre-specified window after the switch. It is not a required SDTM variable.

flowchart TD
  subgraph Onboard["1. Caregiver Qualification & Credentialing"]
    A["Protocol Authoring
Specify Caregiver Eligibility"] --> B["Screening & Consent
Identify Primary & Backup Caregiver"]
    B --> C["Identity Verification
(21 CFR 11.100(b))"]
    C --> D["Documented Training on
Observable Signs (ICH E6(R3))"]
    D --> E["Issue Unique Individual Login
(21 CFR 11.100(a))"]
  end
  subgraph DailyOps["2. Daily Diary Operational Routine"]
    E --> F["Daily eCOA Completion
Under Verified Credential"]
    F --> G["Audit Trail Records Date, Time,
& Distinct User ID (11.10(e))"]
    G --> H{"Is Primary Caregiver
Available & Active?"}
    H -- "Yes (Stable)" --> F
  end
  subgraph Transition["3. Compliant Handover Event"]
    H -- "No (Incapacitated)" --> I["Trigger Handover Workflow
Site Contacted Immediately"]
    I --> J{"Was Backup Caregiver
Pre-Qualified at Baseline?"}
    J -- "Yes" --> K["Activate Backup User Account
Deactivate Primary Account"]
    J -- "No" --> L["Site Performs Identity Verification
& Accelerated Instrument Training"]
    L --> K
    K --> M["Structured Transition Record Logged
(RESPONDENT_ID, HANDOVER_DT)"]
    M --> N["New Daily Observations Resume
Under Backup's Unique Credential"]
  end

Returning to the opening failure case: the adult daughter completed three weeks of daily diary entries under her hospitalized father's credentials before the clinical site discovered the substitution. What should the study team do with these 21 days of misattributed data?

In December 2023, the Critical Path Institute (C-Path) eCOA Consortium, eClinical Forum, and SCDM published consensus recommendations in JSCDM governing eCOA data changes. Their framework establishes foundational data management principles directly applicable to this crisis:

1. Principle 1: Preserve Raw Clinical Observations: Data entered into an eCOA system reflect real-time observations of the patient's daily functional state. The daughter actually observed the patient; she completed 21 evenings of genuine assessments. Under C-Path guidelines, clinical data values should reflect the respondent's chosen response without alteration, deletion, or manipulation by site staff or sponsors. Deleting the records destroys authentic observational evidence.

2. Principle 2: Transparent, Reconstructable Audit Trails: Every action taken within the system must be completely reconstructable. The electronic record must clearly state what occurred: entries recorded between Day X and Day Y were physically entered by Person B while authenticated as Person A.

3. Principle 3: Site Control with Documented Justification: Data clarification and re-attribution must occur under investigator control with transparent sponsor visibility. The site investigator must execute a comprehensive File Note and Data Clarification Request detailing the timeline, the father's hospitalization dates, the daughter's identity, and the absence of intentional fraud.

4. Principle 4: Systematic Database Annotation: In accordance with clinical data reconciliation standards, data managers must not alter the primary raw database table. Instead, a supplemental metadata variable (e.g., ACTUAL_RESPONDENT_ID) is populated in the analytical data pipeline, accompanied by a protocol deviation flag.

The table compares three ways a team might handle those 21 days. None of them is named as a required procedure in regulation. They differ in whether the original record stays reconstructable.

Preserving the raw entries and recording who actually observed them keeps the original record inspectable. It also leaves the statistical analysis plan a choice about whether to include the transition window, censor a short stretch of it, or examine it in a sensitivity analysis.

A critical conceptual trap in clinical endpoint design is misclassifying a caregiver replacement within the estimand framework. Under the international standard ICH E9(R1) (Statistical Principles for Clinical Trials: Addendum on Estimands and Sensitivity Analysis, 2019), an intercurrent event is an event after treatment initiation that affects either the interpretation or the existence of the measurements associated with the clinical question of interest. The addendum's examples are patient-level events such as discontinuation, rescue medication, and death.

Classic intercurrent events include: patient discontinuation of investigational product due to an adverse event, administration of prohibited rescue medications, or patient mortality. In each of these situations, the patient's biological exposure or clinical status has fundamentally shifted.

A change of caregiver respondent is fundamentally different. A caregiver replacement is best handled as a measurement-process event. The participant can remain on the assigned treatment while the person recording observations changes. The switch can disturb the measurement. It does not, by itself, change exposure to the investigational product.

Treatment-policy, while-on-treatment, and hypothetical strategies answer questions about the participant's treatment course. They are a poor default for a change of observer. Setting the new observer's entries to missing is also a poor default. As discussed in missing data and sensitivity analyses, recoding those observations as missing manufactures a gap. Whether that gap is missing at random or missing not at random is a separate assumption. The deletion itself does not identify the mechanism, and it reduces the information in the longitudinal model.

Instead, the trial's Statistical Analysis Plan (SAP) must address rater transitions through pre-specified analytical modeling:

• Primary Longitudinal Model: Incorporate all valid diary observations into the primary Mixed-Effects Model for Repeated Measures (MMRM) or Generalized Linear Mixed Model (GLMM). The model can account for the rater transition by including a time-varying covariate for respondent type or incorporating a random intercept for RESPONDENT_ID nested within participant.

• Pre-Specified Transition Window Sensitivity Analysis: Because empirical data (Nishida et al., 2023, 2024) prove that acute score instability and variance inflation occur during the immediate handover, the SAP should pre-specify a sensitivity model that censors a narrow window (e.g., 7 to 14 days) immediately following a respondent change. Comparing this model against the primary MMRM demonstrates whether transient rater-calibration friction influenced overall findings.

• Informant-Stable Subgroup Sensitivity Analysis: Re-run the primary efficacy model restricted exclusively to participants with continuous, stable caregiver respondents throughout the trial. If the estimate in the stable-respondent subset is similar to the overall estimate, that argues against a large influence of turnover. It is not proof, particularly when few participants switch.

• Informant Kinship Adjustment: In trials where baseline rater shifts are pronounced (such as spousal vs adult-child informants documented by Vargas-Gonzalez et al., 2024), include baseline informant relationship as a fixed-effect adjustment factor in the primary efficacy analysis.

At a GCP inspection under ICH E6(R3) or a data-integrity review under 21 CFR Part 11, reviewers reconstruct how endpoint records were created. For a caregiver diary, the file should show who was authorized to enter data, whose credential was used, and how a deviation was documented.

The Trial Master File and investigator site file, organized with the TMF Reference Model specifications, should let a reviewer follow that chain. Zone numbers below follow the TMF Reference Model's zone list. Companies subdivide artifacts inside those zones.

When an unauthorized transition occurs—such as the three-week period in our opening failure case—the Principal Investigator and Medical Monitor must conduct a formal protocol deviation assessment. ICH E6(R3) Annex 1, section 2.5.3 says the investigator documents all protocol deviations and, for those deemed important, explains the deviation and implements measures to prevent recurrence. A sponsor procedure may still sort deviations into minor and major categories. In the hypothetical case, the file should record what happened, whether the team judges the deviation important, and the action taken, such as training the daughter and issuing her own credential before further entries.

Pre-specifying who may report, keeping each credential unique, logging the switch, and saying in the statistical analysis plan how the switch is examined makes the diary's authorship reconstructable. It does not guarantee that a regulator will accept the endpoint.