Non‐response for health‐related quality of life outcomes in ICU patients: A systematic review of the reporting in randomised trials

Health‐related quality of life (HRQoL) is frequently assessed in randomised clinical trials (RCTs) in the intensive care unit (ICU), but data are limited regarding the proportions of patients without responses or not surviving to HRQoL follow‐up and the handling of this. We aimed to describe the extent and pattern of missing HRQoL data in intensive care trials and describe how these data and deaths were handled statistically.

this systematic review, intensive care unit trial results were assessed for quality of life outcome reporting, where missing values were shown to be prominent for some studies. How missing values are managed in study analysis is a challenge, as a possible major source of systematic error (bias) influencing results.

| INTRODUCTION
More patients survive intensive care unit (ICU) admission, 1 but survivors frequently experience long-term morbidity including new or aggravated physical, psychological and mental disabilities. 2,3 As a consequence more ICU research is now using other patient-important outcomes including health-related quality of life (HRQoL). [4][5][6] In systematic reviews, it has been highlighted that missing HRQoL assessments are common and present a well-known challenge in randomised clinical trials (RCTs), particularly in long-term follow-up of HRQoL, where loss-to-follow is frequent. [6][7][8][9][10][11] A scoping review of RCTs in the ICU setting found high mortality as an additional challenge complicating analysis, reporting and interpretation of HRQoL outcomes. 6 Furthermore, interventions might affect both death and HRQoL, possibly in opposite directions (i.e., if the intervention results in survival of sicker patients, who most likely will have lower HRQoL 12 ), thus further complicating interpretations.
Loss-to-follow-up, and thereby missing data, may impact HRQoL results 7,8,13,14 and introduce attrition bias, as non-response may be associated with severity of illness and other patient characteristics. 7,8,13,14 Transparency in reporting and handling of missing data is therefore imperative for valid interpretation of trial results. 12,15,16 Complete case analysis, which disregards patients with missing data, has been reported as the most frequent approach for HRQoL outcomes, 17 however, excluding patients with missing data impacts the power and potentially the validity on the outcome if data are not missing completely at random. 18 In a systematic review of RCTs with long-term outcomes across medical specialties, 50% handled missing data using mean imputation as a sensitivity analysis supplementing the complete case analysis. 11 This is not recommended as it does not correct potential bias and gives a falsely increased precision of the result. 18 Further, one third of the RCTs did not provide reasons for missing outcome data. 11 Research into missing HRQoL data in ICU RCTs is limited with undisclosed questions considering non-response equated with methodological issues to the RCTs design. The mentioned systematic reviews concerning HRQoL, and missing data were limited in timespan, included not only RCTs, and risk of bias was not assessed. Therefore, we primarily aimed to describe the extent and pattern of missing HRQoL data in ICU RCTs and how missing data and death were handled statistically, as we hypothesised that many data on HRQoL would be missing and that higher risk of bias, larger RCTs, HRQoL being a non-primary outcome, HRQoL assessed with an extensive tool, and longer follow-up time would be associated with increased non-response.

| Study design and registration
We conducted a systematic review of RCTs in accordance with a published protocol and analysis plan 19 and the recommendations of the Cochrane Collaboration. 20 The protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO) (www. crd.york.ac.uk/prospero; registration number: CRD42019118932). Protocol deviations and rationales are described in the supplement.
Results are reported according to the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) 2020 statement (checklist included in the supplement). 21

| Objectives
Our primary objectives were to assess the proportions of missing HRQoL data due to non-response and to assess the proportion of non-survivors at the HRQoL follow-up time-point. For the primary objectives, we also assessed whether the proportion of nonrespondents was reported quantitatively and discussed, and how the proportion of non-response was affected by relevant trial or outcome characteristics.
Secondary objectives were to assess whether baseline characteristics were compared between respondents and non-respondents, and which analytic strategies were used for handling missing data and non-survivors in the HRQoL analyses.
Tertiary objectives were testing of the aforementioned hypotheses by exploring the distribution of proportion of non-respondents in the following subgroups: risk of bias (low, some concerns or high), trial size (small trials n ≤ 100 or large trials n > 100), outcome level (primary, secondary or tertiary outcome), length of tool used (short being the EQ-5D questionnaire in all variations or long being SF-36 or RAND-36), and follow-up time (≤90 days, >90 and ≤180 days or >180 days).

| Eligibility
We included RCTs of adult ICU patients (as defined in the original RCT) where HRQoL (by any score as defined by the original trial) at any follow-up time point(s) was reported as an outcome. Trials assessing interventions that started or were administered in the ICU were included. We only excluded RCTs that we were unable to obtain in full text.

| Search methods for identification of studies
We applied a systematic and sensitive search strategy designed to include RCTs without limitations to language, publication year or journal. 22 (Table S1, supplement) after the first 5 and subsequent 10 papers.

| Extracted characteristics
We extracted both trial characteristics and baseline characteristics according to our protocol. 19 For the primary and secondary aim, we For RCTs reporting HRQoL outcomes at multiple time-points and/or using multiple tools, these were included as separate outcomes.
For RCTs where we could not locate a protocol or statistical analysis plan, we contacted the corresponding author twice per e-mail, to make sure we did not miss these supplemental materials.
Data extraction was performed using Excel (Microsoft Corporation, Redmond, WA).

| Definitions
1. Non-respondents were defined as all eligible patients invited to answer the HRQoL questionnaire at the follow-up time point (i.e., non-survivors are not considered non-respondents), but who did not respond due to emigration, withdrawal of consent, loss to follow-up, 24 or failure to answer the questionnaire for any other reason. Those who did not receive the intervention were not considered non-respondents (modified intention-to-treat principle).
2. Respondents were defined as those who replied to the questionnaire and includes both patients and proxies' responses.
3. Non-survivors were all patients that died before the follow-up time-point of the HRQoL outcome.

| Risk of bias
We assessed risk of bias using the Risk of Bias tool 2 (RoB 2) for the HRQoL outcomes, 25 according to the treatment assignment (i.e., the intention-to-treat principle). The assessment was performed independently and in duplicate by two of three authors (Maj-Brit Nørregaard T A B L E 1 Trial-level characteristics.

| Statistical analyses
Trial-level and outcome-level characteristics (including the proportion of outcomes for which the number of non-respondents was reported) T A B L E 2 Outcome-level. Missing for 52 outcomes (28%). More than one method could be used. d Either by face-to-face interview or telephone interview. e Retention strategy is when trying more than one way to obtain HRQoL assessment and includes gift vouchers (6.1%) and a reminding phone call (26.5%) among others.
f Letting the patient choose what is best for them-either mail or phone. g No missing data due to the category 'not reported'. h The total percentage will not match 100% as some outcomes used more than one method to handle missing data. i Last assessment carried forward. j Sensitivity analyses used for these outcomes were best-worst/worst-best (or the other way around) case scenarios. k The total percentage will not match 100% as some outcomes used more than one method to handle non-survivors. l Other ways to handle non-survivors were imputation of dead using 'survivor average causal effect' (2 outcomes), better worst case scenario (one outcome), sensitivity analyses (one outcome) and composite score ("Lachin 3") where (a) survivors were ranked higher than non-survivors with nonsurvivor ranking according to time to death with early mortality considered worse than late mortality and (b) the EQ-5D-3L health state scale, lower health scores were considered a worse outcome than higher scores (one outcome).
were summarised descriptively using medians with interquartile ranges (IQRs) and full ranges for numeric data and numbers with percentages for categorical data. The proportions of missing data are presented for all variables.
The proportions of non-survivors and non-respondents across all outcomes were similarly summarised descriptively using medians with IQRs and full ranges. Proportions were meta-analysed using random effects models, which assume that the underlying proportions vary across RCTs, but follows normal distributions of which the mean values along with their 95% confidence intervals (CIs) were estimated. 20 In addition to the primary analyses, separate subgroup analyses were conducted, with proportions in each subgroup analysed similarly and absolute risk differences (RDs) between subgroups with 95% CIs calculated based on the separate subgroup results. 19 Overall subgroup differences were assessed with test-of-interaction P-values. In cases where multiple outcomes from the same RCT were included in one of the overall analyses or in a specific subgroup, the mean numerators/denominators across those outcomes from the same RCT were meta-analysed, to avoid double counting and

| RESULTS
After removing duplicates, we found 10,075 unique records that were screened ( Figure S1) and included 196 outcomes from 88 RCT reports published from 2002 to 2022. In the included RCTs, sample sizes ranged from 20 to 7000 participants and numbers of HRQoL outcomes from 1 to 9.

| Characteristics
Trial-level characteristics are summarised in Table 1

| Primary objectives
The number of non-respondents was reported for 149 of 196 (76%) outcomes. The crude median proportion of non-respondents was 20% with an IQR of 9%-38%. The meta-analysis estimated mean proportion was 15% with a 95% CI of 11%-19% among 149 outcomes from 76 RCTs. The crude median proportion of non-survivors was 27% (IQR 14%-39%), while the meta-analysis estimated mean was 19% (95% CI 15%-25%) among 167 outcomes from 81 RCTs. Data on the primary objective for all RCTs and outcomes are presented in Table S4 in the supplement.

| Secondary objectives
For 33 outcomes (17%) a comparison of baseline characteristics for respondent's versus non-respondents was presented, although the potential of attrition bias was discussed in 47% (Table 2). In 17 of 88 RCTs (20%) it was discussed whether non-response might have caused bias and affected the result; in five (6%) the difficulty of obtaining long-term data from the critically ill population was stressed, and in two (2%) it was suggested that non-respondents were in poorer health than respondents.
Analysis methods for missing HRQoL outcomes were not described for 54% of the 196 outcomes, and therefore assumed to be complete case analysis, while complete case analysis was explicitly mentioned as the primary analysis method for 28% of outcomes. Multiple imputation was used for 20% of outcomes ( Table 2). Handling of non-survivors in the analyses was reported for 46% of outcomes and in 26% of outcomes assigned to a value of zero or worst possible score.
HRQoL outcomes were adjudicated as overall low and high risk of bias in 22% and 61% of the RCTs respectively. Details on risk of bias assessments are presented in Figure 1 and Estimated mean proportions and risk-differences (with 95% CIs) obtained from meta-analyses of proportions using random effects models across all outcomes, assuming that the proportions follow a normaldistribution to account for between-trial heterogeneity in proportions. More detailed description in the 'statistical analysis' plan in the methods. d Approximate number of patients calculated as described in the methods section, by using the highest number of patients represented for a HRQoL outcome in each trial (to avoid double counting). e EQ-5D, EQ-5D-3L and EQ-5D-5L.
where the trial participants were not blinded to the intervention was the most frequent methodological source of bias, as it is the participant who is the outcome assessor. 26 For 26 out of 88 (30%) trials the blinding procedure was not transparent and for 32 out of 88 (36%) the trial outcome assessor was blinded, but the participant unblinded (further details in Table S5).

| Tertiary objectives
The five subgroup analyses are presented in Table 3.
The main finding from the subgroup analyses were that nonresponse was statistically significant according to time to follow-up with a RD for non-response was 24% (95% CI 16%-28%) for >180 days and 3% (À5%-8%) for >90 and ≤180 days compared with ≤90 days ( p < .001).

| DISCUSSION
In this systematic review, we explored the reporting, analysis, and interpretation of missing HRQoL data in 88 ICU RCTs reporting 196 different HRQoL outcomes. We found that the proportions of non-respondents were substantial and in half of the RCTs the impact of missingness was not interpreted in relation to the results. Further, the statistical handling of non-response and non-survival was poorly reported (for half of the outcomes, not reported at all). Longer followup was associated with higher proportions of non-respondents.
A recent scoping review of ICU RCTs they found a lower number of missing HRQoL data. 6 The reason for this difference was primarily due to the denominator including non-survivors in their calculation and we excluded the non-survivors resulting in higher missing, also findings may be explained by different timespan for including studies, and the inclusion of interventions commenced before ICU admission in the scoping review. Our results also varies from other reviews due to differences in settings, populations, study design, follow-up durations, and timespan for including studies. 5,[7][8][9][10]17 For this reason results are difficult to compare across reviews.  Figure 2A.
Missing HRQoL data are a challenge, which is aggravated when information regarding handling of incomplete data and comparison of baseline characteristics are unavailable, 7,12,15 making it difficult to interpret whether or not missing data could bias the results. 5,15,16 Non-response is a valid concern because of risk of selection bias, 8,13,14,25 particular as non-response in ICU survivors has been shown to be associated with patient characteristics such as being male, younger, sicker and without a partner. 14 Three of the RCTs included in the present study suggested that non-respondents could be in poorer health, more difficult to contact, or unable to complete the questionnaire. [27][28][29] This is supported by other systematic reviews in ICU populations. 5,[7][8][9] In the present study, complete case analysis was used for almost 81% of outcomes.
Another review covering different medical specialties and study designs, described that missing data were handled by complete case analysis in most cases (79%) or single imputation techniques (19%), for example, either mean imputation, single regression imputation or last observation carried forward. 17 Sensitivity analyses (not further detailed) were conducted in 11% of the studies to investigate the influence of missing data on their results. 17 Another review concerning reporting and handling of missing values showed that 46% of ICU RCTs analysed missing HRQoL data using complete cases only, 36% did not provide information, 14% reported no missing data, while 5% used more advanced methods. 10 Our findings are similar to these results as most outcomes in our study were analysed using complete case analysis, some using multiple imputation (20%), and a few single imputation (2%) or alternative methods (3%).
Non-survivors were not described for 54% of outcomes, which we interpreted as survivors only being included in the analyses. In the previously mentioned scoping review, they also found a fifth of the analyses being restricted to survivors only, and a quarter of the analyses assigned non-survivors zero or the worst score possible. 6 Nonsurvivors are a challenge in RCTs with high mortality 12 and restricting analyses to survivors only may hamper interpretation of the results, especially if interventions have differential effects on mortality and HRQoL. 12 Mortality may be more appropriately handled by using methods with a precise value anchored to death, for example, as for EQ-5D index values, 30 where zero is considered as bad as being dead and values below zero correspond to health states worse than death. 31 Our subgroup analysis confirmed our hypothesis that longer follow-up time was associated with higher non-response rate. Shorter HRQoL tool and trials using HRQoL for non-primary outcome had higher non-response rates, but level of risk of bias and trial size did not statistically significantly influence the response rate; however, this does not rule out that associations exist and the proportion of nonrespondents seemed to be higher with increasing risk of bias.
The findings of this systematic review emphasise the challenges with reporting and interpretation of HRQoL outcomes and that the proportion of missing HRQoL data has not decreased over time. We found that inadequate reporting and handling of missing data is a significant concern in ICU RCTs reporting HRQoL outcomes as missing data are probably not missing completely at random. There are patient reported outcome (PRO) checklists outlining the importance of considering how to minimise and report missing data. 32,33 Retention strategies are important for long-term follow-up and should be pre-planned, 34,35 together with pre-defined analyses including non-survivors in the primary HRQoL analyses along with survivors only, and finally, describing how missing data will be handled.

| Strengths and limitations
This systematic review has several strengths. We followed the methodology outlined in the published protocol, 19 including adherence to the Cochrane Handbook's recommendations, 20 and reporting in accordance with the PRISMA statement. 21 Further, we had no limitation of time or language in our search beside being able to obtain a full text manuscript of the RCT.
This systematic review also has some limitations, including several protocol deviations. One was to use the new RoB2 tool, 25 which is more nuanced regarding missing data than its pre-decessor, and became available after the protocol was written. Further, we deviated from the original analysis plan to properly estimate differences and uncertainty, and to improve our handling of between-study variation and the inclusion of multiple outcomes from some studies. Another limitation was that our search strategy included search terms for HRQoL, which may not have been indexed in search databases in all cases, potentially leading to relevant RCTs not being identified. Our search did not include the term 'health status'. Of note, we found 18 additional RCTs outside the search through hand search of other reviews.

| CONCLUSIONS
In this systematic review of RCTs conducted in the ICU with HRQoL reported as an outcome, we found that both non-response and non-survival was frequent. The proportions of non-respondents and non-survivors were insufficiently reported with sparse information regarding the statistical handling in most RCTs. Non-respondents and non-survivors were excluded from most analyses. Finally, longer follow-up times were associated with increased number of non-respondents. Together these findings call for improvements in the planning and design of RCTs in the ICU setting.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request. All data are available in the supplement.