Factors associated with missed assessments in a 2-year longitudinal study of acute respiratory distress syndrome survivors

Mechanically ventilated patients, meeting the American-European Consensus Conference criteria for acute lung injury (ALI) that were in effect during the time of enrollment [25], were enrolled from 13 intensive care units from 4 teaching hospitals in Baltimore, MD (October 2004 – October 2007) [19]. Hereafter, we use the term ARDS, rather than ALI, to be consistent with the more recent Berlin definition [26]. Exclusion criteria included having > 96 h between ARDS diagnosis and enrollment, > 5 days mechanical ventilation before enrollment, pre-existing ARDS when transferred to a study ICU, pre-existing illness with a life expectancy of less than 6 months, any limitation of care at the time of enrollment (e.g. no cardiopulmonary resuscitation status), previous lung resection, inability to speak or understand English, and no fixed address. Prior to hospital discharge, study participants or proxies were administered a retrospective questionnaire on pre-hospitalization health status. Additionally, at hospital discharge, participants were assessed for independence in activities of daily living (ADLs, includes continence, toileting, and feeding), select health symptoms (e.g. shortness of breath), and discharge disposition including any health services required if discharged to home. Lastly, participants completed a battery of patient-reported and performance-based assessments (assessments listed under “Primary Outcome” section) of their physical and psychological status at 3, 6, 12, and 24 months after ARDS.

Follow-up patients from all 4 sites was conducted centrally by the coordinating center (Johns Hopkins University). The research staff collecting follow-up data underwent rigorous training and on-going quality assurance evaluations for conducting all participant assessments. Loss to follow-up in this cohort was minimized using published retention methods [27‐30]. Retention strategies included: sending participants letter and magnet with study name/logo and phone number; reminder phone calls and letters for upcoming visits; meal vouchers, free parking or taxi rides; home visits to those unable to come to research clinic; thank you letters after visit; and newsletters and birthday cards to maintain contact between visits [19]. We also offered flexible visit hours (e.g. early or late in the day, and weekend) and home visits.

At each of the 3, 6, 12 and 24 month follow-up visits, there were 15 participant assessments of physical and psychological status: 1) Activities of Daily Living (ADLs), 2) Instrumental Activities of Daily Living (IADLs, activities that require more complex thought, e.g. using telephone, managing finances) [31], 3) Hearing Handicap Inventory for Adults-Screening (HHIA-S) [32], 4) EQ-5D [33], 5) Short-Form 36 Questionnaire v2 (SF-36) [34], 6) Hospital Anxiety and Depression Scale (HADS) [35], 7) Impact of Event Scale-Revised (IES-R) [36], 8) 6-min walk distance [37], 9) manual muscle testing (MMT) [38], 10) hand grip strength [39], 11) maximal inspiratory pressure (MIP) [40], 12) Telephone Interview of Cognitive Status (TICS) [41], 13) Sydney Swallowing Questionnaire (SSQ) [42], 14) anthropometric measurements, and 15) a collection of miscellaneous questions about employment, caregiver, etc. There were a small number of assessments that were not applicable to some participants (e.g. contraindications, comatose/cognitive status, amputated limbs or digits), and the number of possible assessments were reduced from maximum of 15. For the purposes of this analysis, assessments that were missed for reasons unrelated to participant factors (e.g. staff or equipment unavailable to conduct assessment) were considered “not applicable” and the total number of possible assessments was modified. Partially completed assessments (i.e. individual surveys or tests) were not considered missed. Reasons for missed or incomplete visits were categorized as due to the physical status of the participant (poor physical condition cited as the reason for not completing the assessment, although no explicit contraindication was present), refusal, lost contact, and other.

The outcome of interest was the number of missed assessments out of the number of possible assessments at each follow-up visit. Participants who missed an entire visit are included in analyses and considered to have missed 100% of the assessments at that visit.

Several baseline and pre-hospitalization variables were considered including participant demographics (age, sex, race, and education level), unemployment due to health condition, whether or not the participant resided at home without healthcare services, inability to walk at least 5 min, Charlson Comorbidity Index (CCI) [43], Functional Comorbidity Index (FCI) [44], and retrospectively collected baseline ADLs, IADLs, EQ-5D and SF-36. History of excessive alcohol use, illicit drug use, and any psychiatric comorbidity were collected from medical record. At hospital discharge, patients were evaluated for shortness of breath, ADLs, and discharge location. At each follow-up visit the following variables were evaluated: ADLs, IADLs, shortness of breath, participant living location, HHIA-S score, unemployment due to health, EQ-5D Visual Analogue Scale (VAS) (range: 0 to 100; higher score is better) and utility scores (range: − 0.11 to 1.0; higher score is better), SF-36 Physical Component Score (PCS) and Mental Component Score (MCS) (mean of 50, SD = 10; higher score is better), HADS anxiety and depression subscales scores (for each, range: 0 to 21; lower score is better, with scores ≥8 indicating substantial symptoms), IES-R score (range: 0 to 4; lower score is better ≥1.6 indicating substantial symptoms), 6-min walk test (percent of predicted value), MMT strength (score range: 0 to 60; higher score is better), hand grip strength (percent of predicted value), MIP (percent of predicted value), missing at least 1 assessment, and whether all data was collected at the research clinic (vs. via phone or mail, or visit to the participant’s home or long-term care facility). For all time-points, ADLs variable was defined as number of ADLs dependencies (out of a possible six activities) or dichotomized as ≥1 vs. 0 ADL dependencies. Similarly, the IADLs variable was defined as number of IADLs dependencies (out of a possible eight activities) or dichotomized as ≥2 vs. < 2 IADLs dependencies.

For all patients at all follow-ups, we assumed that the outcome, the number of missed assessments, followed a Binomial distribution with required parameters: the total number of possible assessments and the mean, the probability of a missed assessment. Assuming the outcome follows the Binomial distribution implies that the probability of a missed assessment is the same for all possible assessments for the patient at the given follow-up. To quantify the association between the probability of a missed assessment (i.e. the mean of the outcome variable) with the a priori identified exposure variables, Binomial logistic regression models were used that accounted for variation in the total number of possible assessments across patients and follow-ups [45, 46]. More details on the Binomial logistic regression model can be found in the Additional file 1. In the Binomial logistic regression model, associations were quantified using odds ratio, i.e. the relative odds of a missed assessment per unit change in the exposure variable of interest. Standard errors for the odds ratios were estimated using robust variance estimates to account for the potential over- or under-dispersion in the assumed Binomial variance. First, pre-ARDS baseline and hospital discharge variables were correlated with the number of missed assessments at the initial follow-up at 3-months via bivariable Binomial regression models. For evaluating missed assessments across 6-, 12-, and 24-months, longitudinal Binomial logistic regression models fit with generalized estimating equations with an exchangeable correlation structure were used. The longitudinal models included main effects for follow-up time, exposure and their interaction. For variables where the relationship did not vary over time, the interaction term was dropped from subsequent models. Exposure variables were selected for inclusion in final multivariable models based on p < 0.20 for their univariable association with the outcome. If two definitions of the same variable were significant at p < 0.20 (for example, ≥1 ADL and number of ADLs), the one with the stronger association (i.e., smaller p-value) was used in the multivariable model.

There was minimal missing data for baseline and discharge characteristics. However, in the longitudinal models, exposures measured at the prior visit were included (i.e. 3-month IADLs as an exposure for missed assessments at 6-months) and could contain missing data. In these models, an indicator for whether or not the exposure was assessed at the prior visit was included as well as the interaction between this indicator and the exposure. Linearity of the association of each continuous exposure variable was assessed using locally weighted scatterplot smoothing (LOWESS) and restricted cubic splines, and there were no continuous exposures for which the linearity assumption was strongly violated. Standard regression diagnostics were used to assess model fit (evaluated by comparing predicted versus observed values and comparing quasi-information criteria (QIC) between multivariable models), influential data points (evaluated by Cooks D), and multicollinearity (evaluated by Variance Inflation Factors (VIFS)) (See Additional file 1: Table S2 for results). Logit of proportion of missed assessments was calculated for each person at each time-point in order to visualize the associations between the outcome variable and significant variables. Logit of proportion of missed assessments was undefined when probability was 0 or 1; therefore these logits were set to − 4 and 4, respectively, for the purpose of these illustrations. Figures were then created by calculating the univariable means of the logit of proportion of missed assessment with corresponding 95% confidence intervals or regression line, where appropriate.