Magnitude and modifiers of the weekend effect in hospital admissions: a systematic review and meta-analysis

Objective To examine the magnitude of the weekend effect, defined as differences in patient outcomes between weekend and weekday hospital admissions, and factors influencing it. Design A systematic review incorporating Bayesian meta-analyses and meta-regression. Data sources We searched seven databases including MEDLINE and EMBASE from January 2000 to April 2015, and updated the MEDLINE search up to November 2017. Eligibility criteria: primary research studies published in peer-reviewed journals of unselected admissions (not focusing on specific conditions) investigating the weekend effect on mortality, adverse events, length of hospital stay (LoS) or patient satisfaction. Results For the systematic review, we included 68 studies (70 articles) covering over 640 million admissions. Of these, two-thirds were conducted in the UK (n=24) or USA (n=22). The pooled odds ratio (OR) for weekend mortality effect across admission types was 1.16 (95% credible interval 1.10 to 1.23). The weekend effect appeared greater for elective (1.70, 1.08 to 2.52) than emergency (1.11, 1.06 to 1.16) or maternity (1.06, 0.89 to 1.29) admissions. Further examination of the literature shows that these estimates are influenced by methodological, clinical and service factors: at weekends, fewer patients are admitted to hospital, those who are admitted are more severely ill and there are differences in care pathways before and after admission. Evidence regarding the weekend effect on adverse events and LoS is weak and inconsistent, and that on patient satisfaction is sparse. The overall quality of evidence for inferring weekend/weekday difference in hospital care quality from the observed weekend effect was rated as ‘very low’ based on the Grading of Recommendations, Assessment, Development and Evaluations framework. Conclusions The weekend effect is unlikely to have a single cause, or to be a reliable indicator of care quality at weekends. Further work should focus on underlying mechanisms and examine care processes in both hospital and community. Prospero registration number CRD42016036487


INTRODUCTION
Increased mortality rates among patients admitted to hospital during weekends have received substantial public attention. This so-called "weekend effect" has motivated policies to strengthen 7-day services in the UK but has also triggered a heated debate about how to interpret the evidence. [1][2][3][4] Hundreds of studies examining the weekend effect in different clinical areas from around the world have now been published, some focusing on unselected emergency admissions, others on elective admissions, and exploring outcomes for specific diagnostic groups. [5][6][7][8][9][10][11] More recently several systematic reviews and metaanalyses have attempted to summarise these studies. [12][13][14] However, the published reviews have been limited to describing the presence or absence, and estimating the magnitude, of the weekend effect. Few had gone beyond describing the quantitative estimates to explore possible mechanisms behind this apparently ubiquitous phenomenon. In those reviews which attempted to do so, conclusions were drawn from subgroup meta-analyses and metaregressions of a small number of variables without paying sufficient attention to potential confounding factors in study-level data and nuanced analyses reported within individual studies. 13 Understanding causation is of crucial importance for health care providers, policy makers and patients in order to take actions which are based on an accurate interpretation of the scientific evidence. We have therefore performed a comprehensive mixed methods review of the quantitative and qualitative literature. Here we report our analysis of the quantitative literature to characterise the magnitude of the weekend effect and explore potential modifiers of the effect.

Structure of the review
This paper is part of a mixed methods review incorporating a systematic review of the magnitude of the weekend effect and a framework synthesis that examines the underlying mechanisms of the effect. The protocol providing details of the overall study design and methodological approaches has been previously reported. 15 Briefly, the review aims to answer the following overarching question: We define the weekend effect as the difference in patient outcomes between weekend and weekday hospital admissions, using the definitions of 'weekend' as those given in the various publications. The research question is addressed through: (1) examination of studies providing quantitative estimates of the weekend effect and its possible modifiers; and (2) interrogation of diverse (both quantitative and qualitative, primary and secondary) evidence that sheds light on the underlying mechanisms of the weekend effect. The former is reported as a systematic review in this paper, whereas the latter will be described in a companion paper in the form of a framework synthesis. The two components of the mixed methods review shared the same initial comprehensive literature search and study screening process (described below), and were then run in parallel. Review teams of the two component reviews/syntheses shared information with each other on a regular basis, and findings from the two components were used to inform and complement each other.

Search strategy
Using MEDLINE, CINAHL, HMIC, EMBASE, EThOS, CPCI (Conference Proceedings Citation Index) and the Cochrane Library without language restriction, we limited the search to year 2000 onwards to ensure that evidence reasonably reflected contemporary health organisation and practice. Our iterative search strategy combined terms relating to Additional searches were undertaken specifically for framework synthesis, described in the companion paper.

Study selection and eligibility criteria
Records were initially screened by one reviewer. Potentially relevant records were discussed in plenary meetings by both teams to refine study eligibility criteria, and subsequently coded according to the following grouping: (1) Observational studies comparing weekday and weekend admissions with quantitative data on processes and/or outcomes; (2) Studies in which changes in service delivery and organisation at weekends were introduced and the impacts were evaluated quantitatively; (3) Studies providing qualitative evidence that could shed light on the mechanisms of the weekend effect; (4) Studies describing differences in case-mix between weekday and weekend admissions without looking into process of care or patient outcomes.  (1) above are the focus of this systematic review; studies that were classified into groups (2) to (4) were routed to framework synthesis for further consideration.
A study needed to have met the following criteria to be included in the systematic review: • Have evaluated undifferentiated admissions to acute hospitals, i.e. admissions across different conditions or specialties, rather than being limited solely to those related to specific conditions or specialties. Undifferentiated admissions included emergency and elective adult, paediatric, medical, surgical, and obstetric admissions.
For studies that reported both aggregated and condition-specific weekend effects, only the aggregated data were used in the quantitative analyses of the systematic review. We chose to focus on unselected, rather than condition-specific admissions to avoid duplicating meta-analyses 8,9,14 focusing on condition-specific admissions.
• Have compared at least one of the following outcomes of interest between weekend admissions and weekday admissions, or between patients having their critical period of care at weekends (e.g. receiving a surgical procedure just before weekend; giving birth during weekend) with those having their critical period of care on weekdays: mortality, adverse events (defined as undesirable events caused by medical management rather than the patient's underlying condition), length of hospital stay and quantitatively measured patient satisfaction. The definition of 'weekend' and the cut-points for mortality were those given in the various publications. Studies comparing out-of-hours and regular hours were included if out-of-hours included weekends. We did not study daytime-night-time comparisons alone. We excluded conference abstracts and 'grey literature' because of difficulty assessing risk of bias.
Independent duplicate coding of potentially relevant studies was performed for the first 450 (40%) of potentially relevant records to maximise consistency of approach; the remaining studies were then assessed by single reviewers. Final study selection was determined by two reviewers. Any discrepancies in study coding and selection were resolved by discussions between reviewers or by seeking further opinion from other review team members.

Data extraction and risk of bias assessment
Data extraction was carried out by one reviewer and checked by another; risk of bias was performed independently by two reviewers. Discrepancies were resolved through discussions. Data from included studies were extracted into a pre-defined and piloted spreadsheet using a detailed data extraction and coding manual (see Appendix

Bayesian meta-analysis
The primary pre-specified outcome for the meta-analysis was mortality using the end-points described in the papers; where multiple mortality end-points were given, we used mortality at hospital discharge for the main analyses. The data were meta-analysed using a Bayesian random effects model that allowed for within-study variation and between-study heterogeneity (supplementary file Appendix 3). Analyses were undertaken using (log) adjusted odds ratios, or hazard ratios or rate ratios if odds ratios were not reported. Where multiple estimates based on different reference day(s) were reported, we used the estimate based on or including Wednesday as the reference group. Where different studies appeared to have used data from the same source and period/location (see supplementary file Appendix 4), our selection criteria were based on quality of adjustment for potential confounding factors, largest sample size, and most up to date.
The main meta-analysis included all types of admissions. Exploratory subgroup analyses were performed for emergency, elective and maternity admissions. We calculated the Isquared statistic to quantify statistical heterogeneity between studies (I 2 >50% indicating a substantial degree of heterogeneity). 16 F o r p e e r r e v i e w o n l y 11 All statistical models were estimated by Hamiltonian Monte Carlo (HMC) using Stan 2. 16. 17 Four HMC chains were run for 2,000 iterations. Convergence was assessed using visual inspection of traceplots and the Rhat statistic.

Exploring potential sources of heterogeneity
We investigated whether the estimated weekend effect is influenced by various factors through a meta-regression, subgroup analyses and sensitivity analyses.
Meta-regression allows simultaneous exploration of multiple factors that could influence the magnitude of estimated weekend effects but it is susceptible to confounding. We examined the following variables: study containing emergency admissions (yes/no), containing surgical patients (yes/no), year of data collection (mid-point where multiple years were included), adequacy of case-mix adjustment (as described earlier; reference category was combined 1 and 2a, i.e. adjusted for acute physiology). The country effect is specified as a hierarchical random effect.
Subgroup meta-analyses were performed by types of admissions as described above, and we summarised additional subgroup analyses within individual studies. Sensitivity analyses that we were able to perform were limited because of insufficient data and heterogeneity between studies, increasing the risk of confounding. We focused on including or excluding studies with partially overlapping data, and examining evidence within individual studies (e.g. where a study reported both in-hospital and 30-day mortality)

Assessment of publication bias
We constructed funnel plots to assess "small study effects" (studies of smaller sample sizes tend to report larger estimated effects), for which publication bias and outcome reporting bias are among the possible causes. 18 Where funnel plot asymmetry was observed, we used a data augmentation approach to derive a pooled estimator assuming the asymmetry was caused by publication bias. 19

Patient and public involvement
Patients and the public were not involved in the design and conduct of this systematic review, which focuses on published literature. The HiSLAC project, which funded this review, received advice from patient and public representatives through their memberships in the Project Management Committee. Patients also directly contributed to the companion framework synthesis, which will be reported separately.

Literature search and study selection
After removing duplicates, 6441 records were retrieved and screened, 613 of which passed through first stage screening. Of these, 224 were routed to framework synthesis and 319 were excluded (see flow diagram in supplementary file Appendix 5). Sixty-eight studies

Mortality
Forty-nine of the included studies examined various mortality outcomes (eight of which focused on neonatal mortality).

Bayesian meta-analysis
Overall summary estimate Bayesian meta-analysis including all types of admissions (with minimal overlapping data) is shown in Figure 1. The pooled estimate suggested that weekend admissions are associated with a 17% (95% credible interval 11% to 23%) increase in the odds of death compared with weekday admissions.
[Insert Figure 1 here] Overall the level of heterogeneity is low (I 2 =16%), although the estimated weekend effect still varies widely between individual studies. Sensitivity analysis allowing for some overlapping of data between studies produced a result (OR 1.16, 95% CrI 1.09 to 1.23) that is very similar to the main analysis (OR 1.17, 95% CrI   36). Use of data augmentation methods (that assume funnel plot asymmetry was caused by publication bias and 'adjusting' for its effect) reduce the estimated weekend effect (OR 1.11, 95% CrI 1.08 to 1.13).

Meta-regression
Results from multivariate meta-regression are shown in Table 1. The main findings are: (1) Studies that included measures of acute physiology in their statistical adjustment tended to produce an estimate of the weekend effect that is closer to null and on average reported estimates that are approximately 15% lower in terms of increased odds of mortality compared with studies without adjusting for these measures.
(2) The weekend effect is significantly larger for elective admissions compared with emergency admissions, and significantly smaller (or does not exist) for maternity admissions.
(3) There is no apparent time trend in the weekend effect. However this does not necessarily agree with assessment of time trend within individual studies (see the next section).
(4) The above findings need to be interpreted with caution. For example, the finding regarding statistical adjustment relies upon data from five estimates reported in four relatively small studies 19,23,27,28,54 that adjusted for measures of acute physiology.

Exploring the sources of heterogeneity
Meta-regression allows simultaneous exploration of multiple factors that could influence the magnitude of estimated weekend effects using study-level variables, but its statistical power is limited and is susceptible to confounding by study level variables. This subsection presents findings from additional subgroup and sensitivity analyses, paying particular attention to within-study comparisons to explore in more detail potential modifiers of the weekend effect.
Weekend effects by types of admission Subgroup meta-analyses by types of admissions are summarised in Table 2 [Insert Table 2 here] Among emergency admissions, one study from England 55 and another from the USA 56 demonstrated that the observed weekend effect was largely attributable to 'direct' admissions from the community (e.g. general practitioner or walk-in clinic referrals) rather than those through the ED. Another US study restricted to admissions through the ED 57 also

Weekend effect by time period and country
Although meta-regression showed no indication that the weekend effect has changed over time, analyses within individual studies showed a more varied picture (supplementary file Appendix 8.2). No time period effects were observed in studies using various databases in the UK, but a significant reduction in the weekend effect over time was reported in a large US study of emergency admissions based on the National Inpatient Sample, 58 and a small study of emergency medical admissions in a single Irish hospital. 25 Within each admission type, variation in the reported weekend effect is apparent among studies from different countries supplementary file Appendices 8.1 and 8.3); however standardised data allowing cross-country comparisons are very limited. 20 Weekend effects by disease condition Several studies provided subgroup analyses of the weekend effect based on the main diagnostic category related to the admission. The weekend effect was consistently found in admissions associated with conditions such as aortic aneurysm, pulmonary embolism and cancer, and was absent for admissions associated with conditions such as chronic airway obstruction; evidence on the presence of the weekend effect was less consistent for conditions such as myocardial infarction and Intracerebral haemorrhage (supplementary file Appendix 8.4). in the only study that was judged to have achieved comprehensive statistical adjustment, 54 the test for interaction showed no significant difference (p=0. 86  Correlation of hospital weekend effect with staffing level Two studies have attempted to correlate measures of weekend staffing (for consultants) 59 and/or weekend services 60 with observed weekend effect and/or changes in the weekend effect over time for individual hospitals in England. Neither showed an appreciable correlation (supplementary file Appendix 8.5).

Influence of statistical adjustment
Statistical adjustment was carried out in most studies in an attempt to account for different characteristics between weekday and weekend admissions. The number and nature of variables included in statistical adjustment varied widely between studies.
Only six publications reporting studies from a small number of individual hospitals or hospital groups have included measures of acute physiology in the statistical adjustment. [23][24][25]27,28,54 One of the studies 54

Adverse events
Nineteen studies compared the risk of adverse events between weekend and weekday admissions. 21,22,32,35,37,[40][41][42][43]47,48,[51][52][53][61][62][63][64][65] While some reported an increased risk for weekend admissions, overall the findings were heterogeneous across different adverse events within individual types of admissions, and the existence and magnitude of a weekend effect linked to a given adverse event were often inconsistent (supplementary file Appendix 9). None of the studies adjusted for physiological severity of illness: sicker patients (and particularly non-survivors) are more susceptible to adverse events. 66

Length of stay (LOS)
Fifteen studies compared hospital LOS between weekend and weekday admissions (supplementary file Appendix 10). 21,24 8799,25,27-30,32,33,35,37,52,58,67-69 The majority of studies show that the (unadjusted) mean or median hospital LOS was shorter (by one day or less in most cases) for admissions during weekends compared with admissions during weekdays, with a few exceptions among studies including elective and maternity admissions. 37,52,[67][68][69]  The shorter LOS associated with weekend admissions appears to be partly attributable to the higher proportion of patients who died in the hospital among weekend admissions.

Patient satisfaction
One study based on data from the 2014 English NHS adult inpatient survey reported a significantly higher level of satisfaction in the information given to them in the ED for patients admitted through this route at weekends compared with those admitted through the ED on weekdays. 70 After adjustment for potential confounders, no significant differences between weekend and weekday admissions were found in other domains covered by the inpatient survey (supplementary file Appendix 11).

DISCUSSION
This systematic review of studies reporting the weekend effect in broad ranges of admissions to hospital has found that weekend admission is associated with a 17% increase in the risk of death, but the magnitude of the effect varies by different types of admissions, case mix and illness severity, geographic location, and contextual and methodological factors.
The overall estimate of the weekend effect varies in meta-analyses published to date, e.g. a pooled adjusted odds ratio of 1.12 (95% CI 1.07 to 1.18) by Hoshijima et al., 12 1.11 (95% CI 1.10 to 1.13) by Zhou et al. 14 and a pooled relative risk of 1.19 (95% CI 1.14 to 1.23) by Pauls et al. 13 Our meta-analysis covers by far the largest number of admissions; our pooled adjusted odds ratio of 1.17 (95% credible interval 1.11 to 1.23) is broadly in line with other studies, whereas the wider credible interval may, in part, reflect the use of Bayesian Considering that each of the above meta-analyses covers at least tens of millions of admissions, and yet the estimated weekend effects could differ nearly two-folds, a clear message is that such an estimate is subject to a large amount of noise due to the myriad of contextual factors and different underlying mechanisms associated with different studies and admissions, which need to be examined more closely -and this is the key contribution of our review.
Weekend admissions differ from weekdays: fewer patients are admitted at weekends despite similar weekend-weekday ED attendance rates (thus creating a reduction in the denominator of the weekend mortality ratio) 55 and those that are admitted are sicker (case mix). 27,28,54 There is scant evidence to support the contention that hospital care is of inferior quality at weekends: adverse events may be more common but confounding by illness severity has not been excluded. In stroke care different patterns of variation in timeliness and adherence to best practice standards have been reported across the week, with no difference in weekend and weekday admission mortality rates. 71 In one study, vital signs were recorded more reliably at weekends than on weekdays. 28 The finding that mortality risk is higher for elective than for emergency admissions at weekends might be explained by inadequate case mix adjustment, but is also consistent with the hypothesis that hospitals are configured to care for emergencies at weekends, while elective admissions might be overlooked. Determining the proximate causes for these observations requires detailed study of patient pathways, health service provision, care processes and patient experience in the community, at the interface between community and hospital, and in hospital following admission on weekdays and at weekends. The paucity of published literature on quantitatively measured patient satisfaction is surprising, 70 as patient's, carer's, and service provider's experience must be at the centre of the design and delivery of health services.
We will fill in these important evidence gaps through our companion framework synthesis, and other components of the HiSLAC project. 72,73 While our estimation of the overall association between weekend hospital admission and mortality is broadly in line with those reported previously, [12][13][14] our review has several unique strengths. First, previous reviews have either examined only mortality, [12][13][14] or mortality with a small number of care process or outcome measures for specific disease conditions. 6,9 Our review covers institution-wide and/or nationwide samples of hospital admissions and examined adverse events, LOS and patient satisfaction in addition to death.
Secondly, previous reviews have focused on using study level data to generate pooled estimates of the weekend effect. We have extended this by examining the more nuanced analyses available within individual studies. Our companion framework synthesis will look in more detail at possible underlying mechanisms of the weekend effect.
This systematic review was limited by the exclusion of condition-specific admissions, although others have extensively reviewed these separately. Nevertheless, we believe the limitation is not a major threat for the validity of our conclusions as we have carefully Most studies included in this review utilise routinely collected administrative data. Our review suggests the need for caution in the analysis and interpretation of these information sources. For example, data on important confounders such as severity of illness are often unavailable, and undiscriminating adjustment of other variables such as hospital teaching status and bed size could risk "adjusting away" some of the weekend effect attributable to care quality. Differential data quality between weekend and weekday admissions is another potential contributor to the weekend effect. 22 28 We recommend a shift of focus from final adjusted mortality rates to considering how different pathway factors influence these estimates ( Figure 2), using configurative analyses (pattern identification) to supplement aggregative (pooled) approaches. 74 [Insert Figure 2 here]

CONCLUSION
Weekend admissions are associated with a 17% increase in the risk of mortality. Increasing evidence suggests that the weekend effect on mortality may be largely attributable to case mix and contextual factors surrounding admissions, and therefore the cause may lie upstream of the care pathway, in the community. In addition, the magnitude of estimated weekend effect can be influenced by methodological approaches and data quality. These

PATIENT CONSENT AND ETHICAL APPROVAL
Patient consent and ethical approval are not required for this systematic review.

AUTHOR CONTRIBUTION
YFC led the preparation of the review, contributed to all stages from conceptualisation to drafting the manuscript of the review, and is the guarantor; XA contributed to all stages (except data analysis) of the review; CH, NJC and RB contributed to data extraction and quality assessment; SIW planned and carried out Bayesian analyses; AB contributed to literature search, study screening and coordination; CT and ES contributed to development of review methods and study screening; CWW contributed to data checking, CA contributed to development of review methods and managerial support, AG contributed to study coding; RJL contributed to development of review methods and provided senior advice; JB contributed to development of review methods, provided senior advice and is the principal investigator for the HiSLAC project. All authors commented on draft manuscripts and approved the final version.

ACKNOWLEDGMENTS
We thank study authors who kindly responded to our queries.      *This analysis focuses on best adjusted studies that include mixed (both emergency and elective admissions within the same study, with or without including maternity admissions); it thus differs from the main Bayesian meta-analysis (pooled mean 1.17, 1.11 to 1.23) which, in addition to studies included in this meta-analysis, also includes individual types or sub-types of admissions provided that they do not overlap with studies that cover mixed types of admissions. Overall I 2 =16% (95% CrI for I 2 0 to 60%). Individual studies can contribute to multiple estimates where the weekend effect was presented for different sub-populations; some of the included studies were not included in the meta-analysis due to overlap of data between studies. 'Posterior predictive' indicates the predictive interval (see main text) obtained from the Bayesian meta-analysis.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59   Below is a list of data items that we want to extract from included studies into the Excel spreadsheet provided. Please follow the instructions/examples as closely as possible when you go through each item.
Some items require free text while others require some sort of classification / coding. For the latter the codes are listed in the table below. If none of the codes seems to be appropriate, you can always code it as 'Other' (when none of the codes is suitable) or 'Unclear' (when you are not sure which codes to choose) and then put further details using the 'Comment' function (you can do this by firstly select the relevant cell, then right click and choose the 'Insert comment' option).
When the desired information was not described/reported in the paper, please code as 'NR' (not reported). Sometimes an item is not relevant for a particular study, in which can you can enter 'NA' (not applicable).

Item
Free text to enter / codes Explanation

Study characteristics and methods
Author year (Free text) First author and year of publication e.g. Albright 2009 The first author's last name and year of publication of the paper.

ID (Number)
The record number for the EndNote database This is provided in the file name of the paper.

Further comments (Free text)
Free space This is a free space for you to add any comments and observations not captured in the extracted data or raise any questions to be discussed Country (Free text) Name of country or region e.g. USA; six Middle Eastern countries The name of the country/countries where the study was conducted. Further information on region/location can be entered as comments

Study period (free text)
Year(s) for which data were collected e.g. 2001; 1994 -2002 Please record year(s) and months (where reported) List the source of the data, e.g. HES, HCUP NIS; or code as  ad hoc Please record the name of the database/registry/audit -either abbreviation (if available -please record full name in the cell comment) or full name; or code as 'ad hoc' which indicates that the data was collected specifically for a study without a study name Type of data source (code) Code as 'Administrative' if the data came from a routine database such as HES in England and NIS in the US; code as 'clinical' if the data came from ad hoc registry or audit in which clinical information was also collected. Accuracy of data source (free text) List information concerning the accuracy and completeness of the data source This is usually in the form of previous studies (e.g. comparison of coding accuracy). If no information was provided, state "NR".

Inclusion/exclusion criteria (free text)
Enter (copy & paste) the criteria for selecting patients / admissions into the study and the rationale behind the criteria (if provided) Record "NR" where applicable.

Cross-sectional or longitudinal (type of data)
Code as 'Cross-sectional' if the data were analysed as one period (irrespective of whether it spanned over several years); code as 'Longitudinal' if data were collected and analysed for more than one year (e.g. repeated cross-sectional data by years) and allowed the observation of changes over years. Can  If more than one reference day or time period (against which weekend admissions were compared) was used, please record all (and where reported, which was used in the primary analysis, the rationale and whether this was prespecified). Sensitivity analyses by using different reference day/time (code)

 Yes  No
Code 'Yes' if the study had estimated weekend effects using more than one reference day/time Otherwise code 'No'

 Yes  No
Code 'Yes' if the study reported weekend effects for specific conditions/diagnoses in addition to an estimate for all admissions Otherwise code 'No' Additional analyses (free text) List any other comparisons or analyses that were carried out For example additional comparisons between night time vs day time; analyses based on different definitions of weekends or outcomes (e.g. 7-day mortality vs 30-day mortality);analyses of mortality risk by number of days since admission; etc Final sample size (number) List the total sample size in terms of number of admissions Final sample size is defined here as the number of admissions included in the analysis. If the unit was the number of patients, highlight this in the cell comment. Initial sample size (number) List the initial sample size before any exclusions were made; or code  No exclusion  NR Initial sample size is defined as the number of admissions included in the initial sample before any exclusion (e.g. due to incomplete data) was made.  Adverse events (AEs) are defined here as any undesirable events (other than death) that may be caused by medical management rather than the underlying condition of the patient, e.g. surgical complications. This definition does not imply preventability.

Number of hospitals
Interventions and procedures that are carried out mainly to deal with AEs rather than as part of the routine management of a condition are sometime used as indicators for the occurrence of AEs, such as some of the items included in the Patient Safety Indicators. These will also be considered as AEs for this review.

Other outcomes of potential interest
Record any other outcomes not listed above that were reported and might be useful e.g. any process measures or costs information. Record 'None' where appropriate. List ALL variables that have been explored and/or included in the final multivariate model; or code as  None These could include: Patient demographics and clinical conditions, such as age/age group, sex, race/ethnicity, insurance type, diagnosis/diagnosis-related group (DRG), comorbidity etc. Physiological measures that reflect the severity/frailty/instability of patients' conditions, such as blood oxygen saturation, pulse rates and other blood biochemistry. Provider characteristics, defined as features of health care organisations or health care professionals that could influence the capacity to provide high quality health care, such as hospital teaching status, hospital sizes, specialist centre designation, level of staffing (e.g. presence of consultants, nurse to patient ratio) and training or qualification of the doctors. Other variables, such as measures of clinical processes (e.g. guideline adherence) or length of stay etc.

Risk of Bias Assessment
The items below are modified from the Newcastle-Ottawa quality assessment scale Selection  below, but did not adjust for acute physiology 3) partial adjustment: study adjusted for important patient factors including age, main diagnosis, comorbidity/frailty indices AND treatment pathway (elective vs urgent/ emergency, operative vs non-operative) but did not adjust for factors listed in a) and b) above 4) inadequate adjustment: study did not adjust for some important factor(s) listed in c) above or did not control for any factor at all Outcome Assessment of outcome Select ONE option from a) to d) a) independent blind assessment b) record linkage (e.g. information obtained from hospital records) c) self report d) no description Was follow up of outcomes beyond hospital stay?
Select ONE option from a) to b) a) yes b) no Exclusion due to missing data Select ONE option from a) to e) a) no or very few of exclusions due to missing data -unlikely to affect study results b) some level of missing data, but admissions with missing data were retained in analyses using imputed data; c) some level of exclusion due to missing data, but authors demonstrated that admissions with missing data were similar to admissions included in analyses d) excluded a substantial proportion (5%) of admissions due to missing data from analyses e) no statement concerning missing data  For studies with a large sample size, trivial differences between weekday and weekend admission can still be statistically significant. Please add comments to describe if this is the case.

Quantitative results
These can be classified into two groups according to the types of outcome:  Dichotomous (binary) variables such as deaths or occurrence of complications  Continuous variables such as length of hospital stay

Technical details of the Bayesian meta-analysis
Analyses were undertaken using (log) adjusted odds ratios. For studies that only reported adjusted hazard ratios or rate ratios, we used these figures as approximations of adjusted odds ratios as results for these effect measures were very similar where they had been estimated in the same study (see Appendix 7.3.3).
As several studies provided multiple estimates of the weekend effect from different sub-samples (e.g. different time periods or different locations), we specified a three level Bayesian randomeffects model to take into account the correlation of results from different sub-samples within the same study while allowing for within sample variation and between study heterogeneity. In particular, for analysis or sub-sample = 1, … , from study = 1, … , with effect size estimate and estimated standard deviation : Weakly informative priors were specified for the model parameters in order to constrain the parameter to realistic values without providing any further information to influence the posterior estimates, 4 while also provide a degree of regularisation to facilitate computation. In particular halfnormal(0,1) priors for standard deviation terms and normal(0,1) for mean effects. We calculated the I-squared statistic, 5 which is the proportion of total variance attributable to between-study heterogeneity taking into account variance at three levels. Convergence was assessed by visual inspection of traceplots of MCMC chains and the Rhat statistic. Models were estimated in Stan. 6

Technical details of the Bayesian meta-regression
The model described in Equation (1) is extended to allow for varying mean effects according to characteristics of the sample, : where are a set of parameters to be estimated.
The following variables were included in a planned, exploratory meta-regression: Two pre-specified variables were not included in the meta-regression due to lack of data: type of population (few studies focused on children) and country income category (none of the included studies was conducted in low and middle income countries). Instead we included an indicator variable for each country. The country effect is specified as a hierarchical 'random' effect.

Appendix 4. Examination of potential overlap in the coverage of admissions between different studies
Many studies included in this systematic review utilised data from routine administrative databases, most prominently the Hospital Episode Statistics (HES) from England and the National Inpatient Sample (NIS) from USA. Inclusion of studies that cover data related to the same or overlapping admissions in a meta-analysis results in double-counting and therefore needs to be avoided.

Mortality -subgroups by country
As illustrated in Figure 1 the main text and other forest plots in Appendix 8.1 above, the weekend effect appears to vary between studies undertaken in different countries. Two studies provided data of cohorts from different countries. In a Global Comparators Project, Ruiz et al. investigated 30-day in-hospital mortality for emergency admissions and elective surgical admissions using data from four countries ( Table 8 in Appendix 8.1.5 above). 18 Weekend effect was found across the countries and type of admissions, but there were notable variations between the countries and no apparent

Mortality -subgroups by disease conditions
Although systematic reviews of the weekend effect for individual disease conditions have been published, [78][79][80][81] comparisons of the weekend effect between different disease conditions could be confounded by differences in study-level characteristics between studies. Several studies included in this review reported weekend effects by selected, individual disease conditions and they provide a chance to make such a comparison that is less susceptible to confounding by study-level variables.
The data are presented in this section. We selected conditions for which the mortality is likely to be affected by hospital staffing level (ruptured abdominal aortic aneurysm, acute epiglottitis, and pulmonary embolism) and those for which mortality is unlikely to be influenced by staffing level as originally hypothesised by Bell and colleagues in their seminal paper, 51 as well as other conditions that commonly contribute to death during hospital admissions.
Overall the estimated weekend effect from different studies are fairly consistent for most of the conditions, but discrepancies exist and the findings do not necessarily agree with hypotheses initially set out by Bell and colleague. A finding worth highlighting is that in the only study (Walker et    (2) the in-hospital mortality rate among patients with the condition is high; (3) the first few days of hospitalisation are critical; (4) the condition is treatable; (5) care involves logistic difficulties; (6) death can be rapid; (7) patients with the condition typically receive a substantial amount of care in clinical settings other than a critical care unit or A&E.
^Conditions hypothesised by Bell et al. for which a weekend effect is less likely to be observed: The first was acute myocardial infarction, which is usually managed in a critical care setting, where fluctuations in staffing levels are minimal. The second was acute intracerebral haemorrhage, for which effective treatment is generally unavailable. The third was acute hip fracture, a condition that is sometimes treated more promptly on weekends than on weekdays, because operating rooms are more available on weekends.  76 While an increased risk of experiencing various adverse events was observed for weekend admissions in some studies, the findings were heterogeneous and inconsistent (e.g. risk of was increased for some measures of adverse events but not increased or even decreased for other measures within individual studies; inconsistent findings with regard to the existence and magnitude of the weekend effect for a given adverse event between different studies). Findings for different measures of adverse events are presented below.

Composite measures of adverse events
In a large study using the Nationwide Inpatient Sample, Attenello and colleagues 34    Four studies examined the weekend effect on various measures of readmission, reoperation, ICU admission and A&E visits following the initial admission (or following discharge from the initial admission). The results are heterogeneous: one study focusing on children 42 with inadequate adjustment of potential confounding factors reported a 9% increase in the odds for 30-day readmissions associated with weekend admissions; Khanna and colleagues, based on data from a single hospital in Chicago in which there were no difference in physician level between weekdays and weekends, found that weekend admissions were associated with significantly lower risk of ICU transfer during hospitalisation and were not associated with an increased risk of readmissions or re-visit to the A&E; 60 Dubois and colleague found an increase odds of 7% for ICU admission among elective surgeries carried out on Fridays compared with those carried out on Mondays, but no increase in 30-day reoperation or readmission was observed. 65 Walker et al. reported significantly higher risk of admitting to ICU for weekend admissions when factors available from administrative database were adjusted for, but this weekend effect was substantially attenuated when laboratory test results were also adjusted for. 12    Many studies examined the weekend effect on patient safety indicators or other measures related to adverse events during or following surgery. While an increased risk of surgical adverse events was found in several studies, the findings were not consistent across different outcomes within individual studies and were also heterogeneous across studies for a given outcome measure.

Perinatal and neonatal adverse events
With a few exceptions (some of which were subgroup analyses), studies of maternity admissions generally reported relatively small or no weekend effect for perinatal and neonatal adverse events.

Maternal adverse events
Weekend effect was found for several adverse events in some studies, although the statistical adjustment was only judged to be partial or inadequate in many cases.       Appendix 11. Evidence on the weekend effect related to patient satisfaction Only one study compared quantitative measures of patient satisfaction between weekend and weekday admissions. 45 Based on data from the 2014 NHS adult inpatient survey (154 trusts, with 59,083 respondents representing a 47% response rate) and accident and emergency (A&E) department surveys (142 trusts, with 39,320 respondents representing a 34% response rate) and the adult inpatient survey, Graham compared the reported satisfaction of patients who attended A&E departments, admitted to hospital or discharged from hospital at weekends (including public holidays) with those who experienced these events during weekdays. Patients who died following the A&E visits/admissions were excluded from the surveys. Patients admitted at weekends were less likely to respond compared to those admitted during weekdays, but this was accounted for by patient and admission characteristics (e.g. age groups, emergency vs elective admissions and ethnicity).
The findings, which adjusted for patient age group, sex, ethnicity, use of proxy response (self-completed or supported), limiting long-term conditions, NHS trust, route of admission (emergency or planned, for the inpatient survey only) and destination post discharge (admitted or discharged, for the A&E survey only), show that patients who attended A&E at weekends were significantly more satisfied about 'doctors and nurses' and 'care and treatment' compared with those who attended during weekdays. Patients admitted to hospital via A&E at weekends were also more positive about the information given to them in A&E. There were no significant differences in other dimensions of care covered in the surveys. 45 Page 107 of 115 For peer review only -http://bmjopen.bmj.com/site/about/guidelines.xhtml BMJ Open 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59         Study selection 9 State the process for selecting studies (i.e., screening, eligibility, included in systematic review, and, if applicable, included in the meta analysis).
p. 5-7 Data collection process 10 Describe method of data extraction from reports (e.g., piloted forms, independently, in duplicate) and any processes for obtaining and confirming data from investigators. for each meta analysis. p.8-9 Page 1 of 2

Section/topic # Checklist item Reported on page #
Risk of bias across studies 15 Specify any assessment of risk of bias that may affect the cumulative evidence (e.g., publication bias, selective reporting within studies).

RESULTS
Study selection 17 Give numbers of studies screened, assessed for eligibility, and included in the review, with reasons for exclusions at each stage, ideally with a flow diagram.
p.11, online Appendix 5 Study characteristics 18 For each study, present characteristics for which data were extracted (e.g., study size, PICOS, follow-up period) and provide the citations.
Online Appendix 6 Risk of bias within studies 19 Present data on risk of bias of each study and, if available, any outcome level assessment (see item 12).
Online Appendix 6 (first column) Results of individual studies 20 For all outcomes considered (benefits or harms), present, for each study: (a) simple summary data for each intervention group (b) effect estimates and confidence intervals, ideally with a forest plot.     The review only focuses on hospital-wide sample of admissions and does not include condition-specific admissions.
 Quantitation of the weekend effect does not explain underlying mechanisms.

INTRODUCTION
Increased mortality rates among patients admitted to hospital during weekends have received substantial public attention. This so-called "weekend effect" has motivated policies to strengthen 7-day services in the UK but has also triggered a heated debate about how to interpret the evidence. [1][2][3][4] Hundreds of studies examining the weekend effect in different clinical areas from around the world have now been published, some focusing on unselected emergency admissions, others on elective admissions, and exploring outcomes for specific diagnostic groups. [5][6][7][8][9][10][11] More recently several systematic reviews and metaanalyses have attempted to summarise these studies. [12][13][14] However, the published reviews have been limited to describing the presence or absence, and estimating the magnitude, of the weekend effect. Few had gone beyond describing the quantitative estimates to explore possible mechanisms behind this apparently ubiquitous phenomenon. In those reviews which attempted to do so, conclusions were drawn from subgroup meta-analyses and metaregressions of a small number of variables without paying sufficient attention to potential confounding factors in study-level data and nuanced analyses reported within individual studies. 13 Understanding causation is of crucial importance for health care providers, policy makers and patients in order to take actions which are based on an accurate interpretation of the scientific evidence. We have therefore performed a comprehensive mixed methods review of the quantitative and qualitative literature. Here we report our analysis of the quantitative literature to characterise the magnitude of the weekend effect and explore potential modifiers of the effect.

Structure of the review
This paper is part of a mixed methods review incorporating a systematic review of the magnitude of the weekend effect and a framework synthesis that examines the underlying mechanisms of the effect. The protocol providing details of the overall study design and methodological approaches has been previously reported. 15 Briefly, the review aims to answer the following overarching question: What is the magnitude of the weekend effect associated with hospital admission, and what are the likely mechanisms through which differences in structures and processes of care between weekdays and weekends contribute to this effect?
We define the weekend effect as the difference in patient outcomes between weekend and weekday hospital admissions, using the definitions of 'weekend' as those given in the various publications. The research question is addressed through: (1) examination of studies providing quantitative estimates of the weekend effect and its possible modifiers; and (2) interrogation of diverse (both quantitative and qualitative, primary and secondary) evidence that sheds light on the underlying mechanisms of the weekend effect. The former is reported as a systematic review in this paper, whereas the latter will be described in a companion paper in the form of a framework synthesis. The two components of the mixed methods review shared the same initial comprehensive literature search and study screening process (described below), and were then run in parallel. Review teams of the two Additional searches were undertaken specifically for framework synthesis, described in the companion paper.

Study selection and eligibility criteria
Records were initially screened by one reviewer. Potentially relevant records were discussed in plenary meetings by both teams to refine study eligibility criteria, and subsequently coded according to the following grouping: (2) Studies in which changes in service delivery and organisation at weekends were introduced and the impacts were evaluated quantitatively; (3) Studies providing qualitative evidence that could shed light on the mechanisms of the weekend effect; (4) Studies describing differences in case-mix between weekday and weekend admissions without looking into process of care or patient outcomes.
Studies that fell under (1) above are the focus of this systematic review; studies that were classified into groups (2) to (4) were routed to framework synthesis for further consideration.
A study needed to have met the following criteria to be included in the systematic review:  Have evaluated undifferentiated admissions to acute hospitals, i.e. admissions across different conditions or specialties, rather than being limited solely to those related to specific conditions or specialties. Undifferentiated admissions included emergency and elective adult, paediatric, medical, surgical, and obstetric admissions.
For studies that reported both aggregated and condition-specific weekend effects, only the aggregated data were used in the quantitative analyses of the systematic review. We chose to focus on unselected, rather than condition-specific admissions to avoid duplicating meta-analyses 8,9,14 focusing on condition-specific admissions.  Have compared at least one of the following outcomes of interest between weekend admissions and weekday admissions, or between patients having their critical period of care at weekends (e.g. receiving a surgical procedure just before weekend; giving birth during weekend) with those having their critical period of care on weekdays: mortality, adverse events (defined as undesirable events caused by medical management rather than the patient's underlying condition), length of hospital stay and quantitatively measured patient satisfaction. The definition of 'weekend' and the cut-points for mortality were those given in the various publications.
Studies comparing out-of-hours and regular hours were included if out-of-hours included weekends. We did not study daytime-night-time comparisons alone. We excluded conference abstracts and 'grey literature' because of difficulty assessing risk of bias.
Independent duplicate coding of potentially relevant studies was performed for the first 450 (40%) of potentially relevant records to maximise consistency of approach; the remaining studies were then assessed by single reviewers. Final study selection was determined by two reviewers. Any discrepancies in study coding and selection were resolved by discussions between reviewers or by seeking further opinion from other review team members.

Data extraction and risk of bias assessment
Data extraction was carried out by one reviewer and checked by another; risk of bias was performed independently by two reviewers. Discrepancies were resolved through discussions. Data from included studies were extracted into a pre-defined and piloted

Data synthesis
Our pre-specified primary outcome is mortality. The quantitative synthesis methods described below were used to analyse mortality data, which form the main part of this article. Data related to adverse events, length of hospital stay and patient satisfaction were tabulated and presented in the supplementary file, with a brief narrative summary provided in this article.

Bayesian meta-analysis
The primary pre-specified outcome for the meta-analysis was mortality using the end-points described in the papers; where multiple mortality end-points were given, we used mortality The primary meta-analysis included all types of admissions. Exploratory subgroup analyses were performed for mixed, emergency, elective and maternity admissions. We calculated the I-squared statistic to quantify statistical heterogeneity between studies (I 2 >50% indicating a substantial degree of heterogeneity). 16 All statistical models were estimated by Hamiltonian Monte Carlo (HMC) using Stan 2.16. 17 Four HMC chains were run for 10,000 iterations including 2,000 warm-up/burn-in iterations or more iterations in the same proportion if convergence was judged not to have been achieved. Convergence was assessed using visual inspection of trace-plots and the Rhat statistic.

Exploring potential sources of heterogeneity
We investigated whether the estimated weekend effect is influenced by various factors through a meta-regression, subgroup analyses and sensitivity analyses. Meta-regression allows simultaneous exploration of multiple factors that could influence the magnitude of estimated weekend effects but it is susceptible to confounding. We examined the following variables: study containing emergency admissions (yes/no), containing surgical patients (yes/no), year of data collection (mid-point where multiple years were included), adequacy of case-mix adjustment (as described earlier; reference category was combined 1 and 2a, i.e. adjusted for acute physiology). The country effect is specified as a hierarchical random effect.
Subgroup meta-analyses were performed by types of admissions as described above, and we summarised additional subgroup analyses within individual studies. Sensitivity analyses that we were able to perform were limited because of insufficient data and heterogeneity between studies, increasing the risk of confounding. We focused on including or excluding studies with partially overlapping data, and examining evidence within individual studies (e.g. where a study reported both in-hospital and 30-day mortality) to determine the potential impact of methodological differences on the estimated weekend effect.

Assessment of publication bias
We constructed funnel plots to assess "small study effects" (studies of smaller sample sizes tend to report larger estimated effects), for which publication bias and outcome reporting bias are among the possible causes. 18 Where funnel plot asymmetry was observed, we used

Patient and public involvement
Patients and the public were not involved in the design and conduct of this systematic review, which focuses on published literature. The HiSLAC project, which funded this review, received advice from patient and public representatives through their memberships in the Project Management Committee.

Literature search and study selection
After removing duplicates, 6441 records were retrieved and screened, 613 of which passed through first stage screening. Of these, 224 were routed to framework synthesis and 319 were excluded (see flow diagram in supplementary file Appendix 5). Sixty-eight studies (reported in 70 articles) met our inclusion criteria. Altogether, these studies included over 640 million admissions (with some overlap between studies).

Characteristics of included studies
Key characteristics of the selected 68 studies are shown in supplementary file Appendix 6.

Risk of bias
Only one study 56

Mortality
Fifty-six of the included studies examined various mortality outcomes (eight of which focused on neonatal mortality).

Bayesian meta-analysis
Results of planned analyses are presented below. All estimated models show good convergence of the chain. HMC trace-plots for the primary analysis, and Rhat statistic and effective sample sizes for all meta-analyses can be found in supplementary file, Appendix 7.
Overall summary estimate Bayesian meta-analysis including all types of admissions (with minimal overlapping data) is shown in Figure 1. The pooled estimate suggested that weekend admissions are associated with a 16% (95% credible interval [CrI] 10% to 23%) increase in the odds of death compared with weekday admissions.

Meta-regression
Results from multivariate meta-regression are shown in Table 1. The main findings are: (1) Studies that included measures of acute physiology in their statistical adjustment tended to produce an estimate of the weekend effect that is closer to null and on average reported (2) The weekend effect is significantly larger for elective admissions compared with emergency admissions, and significantly smaller (or does not exist) for maternity admissions.
(3) There is no apparent time trend in the weekend effect. However this does not necessarily agree with assessment of time trend within individual studies (see the next section).
(4) The above findings need to be interpreted with caution. For example, the finding regarding statistical adjustment relies upon data from five estimates reported in four relatively small studies 56,58,62,63 that adjusted for measures of acute physiology. Therefore there is still a substantial level of uncertainty, and the apparent effect of adjustment of acute physiology could have been confounded by other patient or service features associated with the availability of these measures.

Exploring the sources of heterogeneity
Meta-regression allows simultaneous exploration of multiple factors that could influence the magnitude of estimated weekend effects using study-level variables, but its statistical power is limited and is susceptible to confounding by study level variables. This subsection presents findings from additional subgroup and sensitivity analyses, paying particular  36 and/or weekend services 48

Influence of other methodological features
Included studies used different definitions of the weekend; most defined the weekend as Saturday and Sunday (n=28) or referred to "weekend" without defining the term (n=14).

Adverse events
Nineteen studies compared the risk of adverse events between weekend and weekday admissions. 21 26,28,29,72,87 The shorter LOS associated with weekend admissions appears to be partly attributable to the higher proportion of patients who died in the hospital among weekend admissions.

Patient satisfaction
One study based on data from the 2014 English NHS adult inpatient survey reported a significantly higher level of satisfaction in the information given to them in the ED for patients admitted through this route at weekends compared with those admitted through the ED on weekdays. 31 After adjustment for potential confounders, no significant differences between weekend and weekday admissions were found in other domains covered by the inpatient survey (supplementary file Appendix 13).

DISCUSSION
This systematic review of studies reporting the weekend effect in broad ranges of admissions to hospital has found that weekend admission is associated with a 16% increase in the risk of death, but the magnitude of the effect varies by different types of admissions, case mix and illness severity, geographic location, and contextual and methodological factors. need to be examined more closely -and this is the key contribution of our review.
Weekend admissions differ from weekdays: fewer patients are admitted at weekends despite similar weekend-weekday ED attendance rates (thus creating a reduction in the denominator of the weekend mortality ratio) 47 and those that are admitted are sicker (case mix). 56,62,63 There is scant evidence to support the contention that hospital care is of inferior quality at weekends: adverse events may be more common but confounding by illness severity has not been excluded. In stroke care different patterns of variation in timeliness and adherence to best practice standards have been reported across the week, with no difference in weekend and weekday admission mortality rates. 91 In one study, vital signs were recorded more reliably at weekends than on weekdays. 63 The finding that mortality mortality with a small number of care process or outcome measures for specific disease conditions. 6,9 Our review covers institution-wide and/or nationwide samples of hospital admissions and examined adverse events, LOS and patient satisfaction in addition to death.
Secondly, previous reviews have focused on using study level data to generate pooled estimates of the weekend effect. We have extended this by examining the more nuanced analyses available within individual studies.
This systematic review was limited by the exclusion of condition-specific admissions, although others have extensively reviewed these separately. Nevertheless, we believe the limitation is not a major threat for the validity of our conclusions as we have carefully triangulated the findings by examining subgroups both across and within studies and by carrying out sensitivity analyses. We have attempted to focus on more recent evidence by restricting our inclusion to studies published from year 2000 onwards. However some of the included studies (14/68) covered admissions pre 2000. This is unlikely to have substantial impacts on our findings as our meta-regression did not identify a significant time trend. Due to resource constraint (and paucity of data in the case of patient satisfaction) we were unable to carry out more sophisticated analyses for non-mortality outcomes.
Most studies included in this review utilise routinely collected administrative data. Our review suggests the need for caution in the analysis and interpretation of these information status and bed size could risk "adjusting away" some of the weekend effect attributable to care quality. Differential data quality between weekend and weekday admissions is another potential contributor to the weekend effect. 22 63 We recommend a shift of focus from final adjusted mortality rates to considering how different pathway factors influence these estimates (Figure 2), using configurative analyses (pattern identification) to supplement aggregative (pooled) approaches. 94 [Insert Figure 2 here]

CONCLUSION
Weekend admissions are associated with a 16% increase in the risk of mortality. Increasing evidence suggests that the weekend effect on mortality may be largely attributable to case mix and contextual factors surrounding admissions, and therefore the cause may lie upstream of the care pathway, in the community. In addition, the magnitude of estimated weekend effect can be influenced by methodological approaches and data quality. These suggest that the weekend effect is not a good measure of care quality in hospitals at weekends. Future research and interpretation of research findings on the weekend effect must go beyond the narrow focus of case mix adjustment of routine hospital data and attempt to examine the broader issues related to the whole care pathway both within and outside the hospital; the quality and availability of data that can allow measurement of care

DATA SHARING
All data relevant to the study are included in the article or uploaded as supplementary information.

PATIENT CONSENT AND ETHICAL APPROVAL
Patient consent and ethical approval are not required for this systematic review.

AUTHOR CONTRIBUTION
YFC led the preparation of the review, contributed to all stages from conceptualisation to drafting the manuscript of the review, and is the guarantor; XA contributed to all stages

ACKNOWLEDGMENTS
We thank study authors who kindly responded to our queries.

ACKNOWLEDGMENT OF FUNDING
The

COMPETING INTERESTS STATEMENT
None declared     Note: Mohammed 2012 and Ruiz 2015 contributed to two estimates for each country as the weekend effect was estimated separately for different sub-populations (e.g. emergency and elective admissions). 'Posterior predictive' indicates the predictive interval (see main text) obtained from the Bayesian meta-analysis. I 2 =16% (95% [credible interval] CrI for I 2 0 to 62%). The I 2 represents the ratio of between-study variance to total variance in this threelevel model. The apparently low I 2 could be attributed to the between-study variance being relatively small compared with the between-estimate variance within individual studies. As the wide CrI indicates, the I 2 was estimated with substantial uncertainty. Several studies included in the review were not included in this meta-analysis due to substantial overlap of data between studies; in this case, studies which were judged to have adopted the most comprehensive statistical adjustment were selected.   Below is a list of data items that we want to extract from included studies into the Excel spreadsheet provided. Please follow the instructions/examples as closely as possible when you go through each item.

Figure 2: Factors that may contribute to or modify the weekend effect
Some items require free text while others require some sort of classification / coding. For the latter the codes are listed in the table below. If none of the codes seems to be appropriate, you can always code it as 'Other' (when none of the codes is suitable) or 'Unclear' (when you are not sure which codes to choose) and then put further details using the 'Comment' function (you can do this by firstly select the relevant cell, then right click and choose the 'Insert comment' option).
When the desired information was not described/reported in the paper, please code as 'NR' (not reported). Sometimes an item is not relevant for a particular study, in which can you can enter 'NA' (not applicable).

Study characteristics and methods
Author year (Free text) First author and year of publication e.g. Albright 2009 The first author's last name and year of publication of the paper.

ID (Number)
The record number for the EndNote database This is provided in the file name of the paper. If more than one reference day or time period (against which weekend admissions were compared) was used, please record all (and where reported, which was used in the primary analysis, the rationale and whether this was pre-specified). List the initial sample size before any exclusions were made; or code  No exclusion  NR Initial sample size is defined as the number of admissions included in the initial sample before any exclusion (e.g. due to incomplete data) was made. Adverse events (AEs) are defined here as any undesirable events (other than death) that may be caused by medical management rather than the underlying condition of the patient, e.g. surgical complications. This definition does not imply preventability.

Number of hospitals
Interventions and procedures that are carried out mainly to deal with AEs rather than as part of the routine management of a condition are sometime used as indicators for the occurrence of AEs, such as some of the items included in the Patient Safety Indicators. These will also be considered as AEs for this review.

Other outcomes of potential interest
Record any other outcomes not listed above that were reported and might be useful e.g. any process measures or costs information. Record 'None' where appropriate.

Risk of Bias Assessment
The items below are modified from the Newcastle-Ottawa quality assessment scale Selection  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59    Exclusion due to missing data Select ONE option from a) to e) a) no or very few of exclusions due to missing data -unlikely to affect study results b) some level of missing data, but admissions with missing data were retained in analyses using imputed data; c) some level of exclusion due to missing data, but authors demonstrated that admissions with missing data were similar to admissions included in analyses d) excluded a substantial proportion (5%) of admissions due to missing data from analyses e) no statement concerning missing data These can usually be found in a table or in the first couple of paragraphs in the Results section.

Significant differences observed in characteristics between weekday and weekend admissions (free text)
Record the characteristics for which weekend admissions were found out to be significantly different from weekday admissions List the names of the variables for which significant differences between weekday and weekend admissions were found. This can be defined statistically (i.e. p<0.05) or numerically (i.e. ≥5% difference between weekday/weekend admissions). No need to record numerical results at this stage.
For studies with a large sample size, trivial differences between weekday and weekend admission can still be statistically significant. Please add comments to describe if this is the case.

Quantitative results
These can be classified into two groups according to the types of outcome:  Dichotomous (binary) variables such as deaths or occurrence of complications  Continuous variables such as length of hospital stay

Appendix 2. Risk of bias assessment
Risk of bias assessment was embedded within the data extraction form shown in Appendix 1. We initially used the Newcastle-Ottawa scale 1 with modification of some of the items and wording because the included studies were mostly population database studies rather than the conventional cohort study for which the scale was designed.
However during the review process it became apparent that results of the risk of bias assessment using this modified scale were either unreliable (due to difficulties in judging the Discrepancies between reviewers in the classification were resolved by discussions between reviewers, and where queries remained, other review team members were supplied with information concerning statistical adjustment made in individual studies in the absence of study identity and outcome data to reach consensus prior to data analysis.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59

Technical details of the Bayesian meta-analysis
Analyses were undertaken using (log) adjusted odds ratios. For studies that only reported adjusted hazard ratios or rate ratios, we used these figures as approximations of adjusted odds ratios as results for these effect measures were very similar where they had been estimated in the same study (see Appendix 7.3.3).
Convergence was assessed by visual inspection of traceplots of MCMC chains and the Rhat statistic.
Models were estimated in Stan. 7

Technical details of the Bayesian meta-regression
The model described in Equation (1) is extended to allow for varying mean effects according to characteristics of the sample, :~( where are a set of parameters to be estimated.
The following variables were included in a planned, exploratory meta-regression:  Categorical variable indicating adequacy of case-mix adjustment as described earlier.
Reference category was combined 1 and 2a (with adjustment of measures of acute physiology).

Sensitivity analyses for the primary meta-analysis
Our primary meta-analysis was conducted using the best adjusted, non-overlapping data from individual studies to avoid double counting. As shown in Appendix 4, data from many studies included in this review were potentially overlapping (i.e. they were based on the same admissions) and these were excluded. As the degree of overlapping between studies varies, the primary analysis may have discarded some useful information. We therefore performed a sensitivity analysis that included these additional data by relaxing our rule and allowing for some overlapping of data between studies. For studies/articles that are based on entirely overlapping or the same dataset, the rule of using the best adjusted effect estimate still applies here. The result of the sensitivity analysis is shown in the table below.
To explore potential small study effects (i.e. studies of smaller sample sizes reporting larger effects), we constructed a funnel plot, which is shown in Figure 4 below. Some level of asymmetry was observed in the plot.

Figure 4 Funnel plot for the weekend effect on mortality for all types of admissions
In view of the apparent asymmetry of the funnel plot, we used data augmentation to explore the potential impact on the estimated weekend effect if the funnel plot asymmetry was caused by publication bias. 78 Data augmentation is a method that can be used to 'adjust for ' potential publication bias by assuming that observation of a study is determined by its p-value alone. P-values are divided into different categories, e.g. [0 to 0.1], [0.1 to 0.5]… and within each category the probability of observing a study (identifying the study and including it in a systematic review) can be different, for example studies that fall into a small p-value category (i.e. studies with a statistically highly significant result) are more likely to be published (and hence be 'observed') than those fall into a larger p-value category (i.e. studies with statistically non-significant results). Findings from repeating our primary meta-analysis and sensitivity analysis using data augmentation are presented in Table 12 below, and statistical outputs from these analyses are provided in Table 13 and Table 14.

Emergency admissions
Note: some of the studies reported two separate estimates of the weekend effect for a given country, for example Saturday vs. weekday(s) and Sunday vs. weekday(s). Both estimates were included in the meta-analysis as they provided additional information while the correlation between the estimates within individual studies was accounted for in the multi-level Bayesian model. The study by Ruiz

Mortality -subgroups by time period
This is partly dealt with in meta-regression but this could be confounded by study-level variables, Below is a summary of within-study observations as triangulation of this finding from metaregression.

Mortality -subgroups by country
As illustrated in Figure 1 the main text and other forest plots in Appendix 9.1 above, the weekend effect appears to vary between studies undertaken in different countries. Two studies provided data of cohorts from different countries. In a Global Comparators Project, Ruiz et al. investigated 30-day in-hospital mortality for emergency admissions and elective surgical admissions using data from four countries ( Table 15 in Appendix 9.1.5 above). 19 Weekend effect was found across the countries and type of admissions, but there were notable variations between the countries and no apparent weekend effect was observed in Australia for emergency admissions in their primary analysis. By contrast, Freemantle and colleagues obtained very similar estimates for two independent datasets from England and USA ( Figure 1 in the main text)

Mortality -subgroups by disease conditions
Although systematic reviews of the weekend effect for individual disease conditions have been published, [79][80][81][82] comparisons of the weekend effect between different disease conditions could be confounded by differences in study-level characteristics between studies. Several studies included in this review reported weekend effects by selected, individual disease conditions and they provide a chance to make such a comparison that is less susceptible to confounding by study-level variables.
The data are presented in this section. We selected conditions for which the mortality is likely to be affected by hospital staffing level (ruptured abdominal aortic aneurysm, acute epiglottitis, and pulmonary embolism) and those for which mortality is unlikely to be influenced by staffing level as originally hypothesised by Bell and colleagues in their seminal paper, 52 as well as other conditions that commonly contribute to death during hospital admissions.
Overall the estimated weekend effect from different studies are fairly consistent for most of the conditions, but discrepancies exist and the findings do not necessarily agree with hypotheses initially set out by Bell and colleague. A finding worth highlighting is that in the only study (Walker et 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59   (2) the in-hospital mortality rate among patients with the condition is high; (3) the first few days of hospitalisation are critical; (4) the condition is treatable; (5) care involves logistic difficulties; (6) death can be rapid; (7) patients with the condition typically receive a substantial amount of care in clinical settings other than a critical care unit or A&E.
Conditions hypothesised by Bell et al. for which a weekend effect is less likely to be observed: The first was acute myocardial infarction, which is usually managed in a critical care setting, where fluctuations in staffing levels are minimal. The second was acute intracerebral haemorrhage, for which effective treatment is generally unavailable. The third was acute hip fracture, a condition that is sometimes treated more promptly on weekends than on weekdays, because operating rooms are more available on weekends.
A & E: accident & emergency; NR: not reported.  (Sunday) and a weekday (Wednesday) in 2014. 24 Specialist intensity was defined as the selfreported estimated number of specialist hours per ten emergency admissions between 08:00 hour and 20:00 hour in each trust. Trust-specific weekend effect on mortality was calculated using the

Different measures for mortality
In general, weekend effect is more profound for short-term mortality than longer-term mortality but there are exceptions.   52 stated that "analyses of deaths within two days after admission, rather than total in-hospital deaths, generally showed larger relative differences in mortality between weekend and weekday admissions." Perez Concha et al. 2014 showed very similar pattern between in-hospital deaths & post-discharge deaths at 7 days. 56   found that "in unadjusted models, excess risks associated with weekend admission were greater at shorter timescales; however, after adjusting for administrative factors excess risks associated with emergency admission on Saturdays or Sundays vs Wednesdays were similar for 7-day to 30-day mortality ( Supplementary Figure 9(a)). Similarly, adjusting for test results attenuated these excess

Different effect measures
Two studies used different effect measures for a given mortality outcome measure. The numerical estimates of the weekend effect were very similar between odds ratio and hazard ratio, and between odds ratio and risk ratio in the respective study.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  This section shows that even based on the same data source, estimation of the weekend effects can still vary substantially, indicating the potentially poor signal-to-noise ratio in using the weekend effect on hospital mortality as a reliable measure of care quality. *Multiple estimates were reported across the three papers based on the same dataset (with exact the same number of admissions), but the effect measures and corresponding methods (e.g. type of statistical techniques/models and variables being adjusted for) were poorly described.  77 While an increased risk of experiencing various adverse events was observed for weekend admissions in some studies, the findings were heterogeneous and inconsistent (e.g. risk of was increased for some measures of adverse events but not increased or even decreased for other measures within individual studies; inconsistent findings with regard to the existence and magnitude of the weekend effect for a given adverse event between different studies). Findings for different measures of adverse events are presented below.

Composite measures of adverse events
In a large study using the Nationwide Inpatient Sample, Attenello and colleagues 35

Needs for further hospital care following initial admission
Four studies examined the weekend effect on various measures of readmission, reoperation, ICU admission and A&E visits following the initial admission (or following discharge from the initial admission). The results are heterogeneous: one study focusing on children 43 with inadequate adjustment of potential confounding factors reported a 9% increase in the odds for 30-day readmissions associated with weekend admissions; Khanna and colleagues, based on data from a single hospital in Chicago in which there were no difference in physician level between weekdays and weekends, found that weekend admissions were associated with significantly lower risk of ICU transfer during hospitalisation and were not associated with an increased risk of readmissions or re-visit to the A&E; 61 Dubois and colleague found an increase odds of 7% for ICU admission among elective surgeries carried out on Fridays compared with those carried out on Mondays, but no increase in 30-day reoperation or readmission was observed. 66

Patient safety indicators and surgical adverse events
Many studies examined the weekend effect on patient safety indicators or other measures related to adverse events during or following surgery. While an increased risk of surgical adverse events was found in several studies, the findings were not consistent across different outcomes within individual studies and were also heterogeneous across studies for a given outcome measure.

Perinatal and neonatal adverse events
With a few exceptions (some of which were subgroup analyses), studies of maternity admissions generally reported relatively small or no weekend effect for perinatal and neonatal adverse events.

Maternal adverse events
Weekend effect was found for several adverse events in some studies, although the statistical adjustment was only judged to be partial or inadequate in many cases.  Fifteen studies compared hospital LOS between weekend and weekday admissions. 8,15,23,28,29,39,40,42,45,[59][60][61][62][63][64]66 Data reported in individual studies are shown in      Data collection process 10 Describe method of data extraction from reports (e.g., piloted forms, independently, in duplicate) and any processes for obtaining and confirming data from investigators. p.9-10

RESULTS
Study selection 17 Give numbers of studies screened, assessed for eligibility, and included in the review, with reasons for exclusions at each stage, ideally with a flow diagram.
p.13, online Appendix 5 Study characteristics 18 For each study, present characteristics for which data were extracted (e.g., study size, PICOS, follow-up period) and provide the citations.
Online Appendix 6 Risk of bias within studies 19 Present data on risk of bias of each study and, if available, any outcome level assessment (see item 12).
Online Appendix 6 (first column) Results of individual studies 20 For all outcomes considered (benefits or harms), present, for each study: (a) simple summary data for each intervention group (b) effect estimates and confidence intervals, ideally with a forest plot.     1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  at weekends fewer patients are admitted to hospital, those that are admitted are more severely ill, and there are differences in care pathways before and after admission. Evidence regarding the weekend effect on adverse events and LoS is weak and inconsistent, and on patient satisfaction is sparse. The overall quality of evidence for inferring weekend/weekday difference in hospital care quality from the observed weekend effect was rated as 'very low' based on the GRADE framework.

Conclusions:
The weekend effect is unlikely to have a single cause, or to be a reliable indicator of care quality at weekends. Further work should focus on underlying mechanisms and examine care processes in both hospital and community.

INTRODUCTION
Increased mortality rates among patients admitted to hospital during weekends have received substantial public attention. This so-called "weekend effect" has motivated policies to strengthen 7-day services in the UK but has also triggered a heated debate about how to interpret the evidence. [1][2][3][4] Hundreds of studies examining the weekend effect in different clinical areas from around the world have now been published, some focusing on unselected emergency admissions, others on elective admissions, and exploring outcomes for specific diagnostic groups. [5][6][7][8][9][10][11] More recently several systematic reviews and metaanalyses have attempted to summarise these studies. [12][13][14] However, the published reviews have been limited to describing the presence or absence, and estimating the magnitude, of the weekend effect. Few had gone beyond describing the quantitative estimates to explore possible mechanisms behind this apparently ubiquitous phenomenon. In those reviews which attempted to do so, conclusions were drawn from subgroup meta-analyses and metaregressions of a small number of variables without paying sufficient attention to potential confounding factors in study-level data and nuanced analyses reported within individual studies. 13 Understanding causation is of crucial importance for health care providers, policy makers and patients in order to take actions which are based on an accurate interpretation of the scientific evidence. We have therefore performed a comprehensive mixed methods review of the quantitative and qualitative literature. Here we report our analysis of the quantitative literature to characterise the magnitude of the weekend effect and explore potential modifiers of the effect.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59

Structure of the review
This paper is part of a mixed methods review incorporating a systematic review of the magnitude of the weekend effect and a framework synthesis that examines the underlying mechanisms of the effect. The protocol providing details of the overall study design and methodological approaches has been previously reported. 15 Briefly, the review aims to answer the following overarching question: What is the magnitude of the weekend effect associated with hospital admission, and what are the likely mechanisms through which differences in structures and processes of care between weekdays and weekends contribute to this effect?
We define the weekend effect as the difference in patient outcomes between weekend and weekday hospital admissions, using the definitions of 'weekend' as those given in the various publications. The research question is addressed through: (1) examination of studies providing quantitative estimates of the weekend effect and its possible modifiers; and (2) interrogation of diverse (both quantitative and qualitative, primary and secondary) evidence that sheds light on the underlying mechanisms of the weekend effect. The former is reported as a systematic review in this paper, whereas the latter will be described in a companion paper in the form of a framework synthesis. The two components of the mixed methods review shared the same initial comprehensive literature search and study screening process (described below), and were then run in parallel. Review teams of the two  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y   7 component reviews/syntheses shared information with each other on a regular basis, and findings from the two components were used to inform and complement each other.

Search strategy
Using MEDLINE, CINAHL, HMIC, EMBASE, EThOS, CPCI (Conference Proceedings Citation Index) and the Cochrane Library without language restriction, we limited the search to year 2000 onwards to ensure that evidence reasonably reflected contemporary health organisation and practice. Our iterative search strategy combined terms relating to 'weekend/weekday' or 'out-of-hours' with terms relating to 'hospital admissions'. Additional searches were undertaken specifically for framework synthesis, described in the companion paper.

Study selection and eligibility criteria
Records were initially screened by one reviewer. Potentially relevant records were discussed in plenary meetings by both teams to refine study eligibility criteria, and subsequently coded according to the following grouping:  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y   8 (1) Observational studies comparing weekday and weekend admissions with quantitative data on processes and/or outcomes; (2) Studies in which changes in service delivery and organisation at weekends were introduced and the impacts were evaluated quantitatively; (3) Studies providing qualitative evidence that could shed light on the mechanisms of the weekend effect; (4) Studies describing differences in case-mix between weekday and weekend admissions without looking into process of care or patient outcomes.
Studies that fell under (1) above are the focus of this systematic review; studies that were classified into groups (2) to (4) were routed to framework synthesis for further consideration.
A study needed to have met the following criteria to be included in the systematic review:  Have evaluated undifferentiated admissions to acute hospitals, i.e. admissions across different conditions or specialties, rather than being limited solely to those related to specific conditions or specialties. Undifferentiated admissions included emergency and elective adult, paediatric, medical, surgical, and obstetric admissions.
For studies that reported both aggregated and condition-specific weekend effects, only the aggregated data were used in the quantitative analyses of the systematic review. We chose to focus on unselected, rather than condition-specific admissions to avoid duplicating meta-analyses 8,9,14 focusing on condition-specific admissions.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y   9  Have compared at least one of the following outcomes of interest between weekend admissions and weekday admissions, or between patients having their critical period of care at weekends (e.g. receiving a surgical procedure just before weekend; giving birth during weekend) with those having their critical period of care on weekdays: mortality, adverse events (defined as undesirable events caused by medical management rather than the patient's underlying condition), length of hospital stay and quantitatively measured patient satisfaction. The definition of 'weekend' and the cut-points for mortality were those given in the various publications.
Studies comparing out-of-hours and regular hours were included if out-of-hours included weekends. We did not study daytime-night-time comparisons alone. We excluded conference abstracts and 'grey literature' because of difficulty assessing risk of bias.
Independent duplicate coding of potentially relevant studies was performed for the first 450 (40%) of potentially relevant records to maximise consistency of approach; the remaining studies were then assessed by single reviewers. Final study selection was determined by two reviewers. Any discrepancies in study coding and selection were resolved by discussions between reviewers or by seeking further opinion from other review team members.

Bayesian meta-analysis
The primary pre-specified outcome for the meta-analysis was mortality using the end-points described in the papers; where multiple mortality end-points were given, we used mortality at hospital discharge for the main analyses. The data were meta-analysed using a Bayesian random effects model that allowed for within-study variation and between-study heterogeneity (supplementary file Appendix 3). Analyses were undertaken using (log) adjusted odds ratios (or hazard ratios or rate ratios if odds ratios were not reported) and the reported standard errors or equivalence. Studies were therefore implicitly weighted by the estimated variance of individual effect estimates. Where multiple estimates based on different reference day(s) were reported, we used the estimate based on or including Wednesday as the reference group. Where the weekend effect was reported separately for Saturday and Sunday, we used the estimate for Sunday in the primary analysis and included both estimates in subgroup and sensitivity analyses (described below). Where different studies appeared to have used data from the same source and period/location (see supplementary file Appendix 4), our selection criteria were based on quality of adjustment for potential confounding factors, largest sample size, and most up to date.

Exploring potential sources of heterogeneity
We investigated whether the estimated weekend effect is influenced by various factors through a meta-regression, subgroup analyses and sensitivity analyses. Meta-regression allows simultaneous exploration of multiple factors that could influence the magnitude of estimated weekend effects but it is susceptible to confounding. We examined the following variables: study containing emergency admissions (yes/no), containing surgical patients (yes/no), year of data collection (mid-point where multiple years were included), adequacy of case-mix adjustment (as described earlier; reference category was combined 1 and 2a, i.e. adjusted for acute physiology). The country effect is specified as a hierarchical random effect.

Assessment of publication bias
We constructed funnel plots to assess "small study effects" (studies of smaller sample sizes tend to report larger estimated effects), for which publication bias and outcome reporting bias are among the possible causes. 23 Where funnel plot asymmetry was observed, we used a data augmentation approach to derive a pooled estimator assuming the asymmetry was caused by publication bias. 24

Assessment of overall quality of evidence
We followed the GRADE framework to rate the overall quality of evidence for each of the four outcomes examined in this review. Based on this framework, evidence from observational studies starts with a baseline quality rating of "low". The rating for each outcome is then downgraded or upgraded according to our assessment against each of the eight criteria (risk of bias, inconsistency, indirectness, imprecision and publication for potential downgrading; 25 large magnitude of effect, dose response and direction of effect of plausible confounding factors for potential upgrading). 25,26

Literature search and study selection
After removing duplicates, 6441 records were retrieved and screened, 613 of which passed through first stage screening. Of these, 224 were routed to framework synthesis and 319 were excluded (see flow diagram in supplementary file Appendix 5). Sixty-eight studies (reported in 70 articles) met our inclusion criteria. Altogether, these studies included over 640 million admissions (with some overlap between studies).

Characteristics of included studies
Key characteristics of the selected 68 studies are shown in supplementary file Appendix 6.

Risk of bias
Only one study 16

Bayesian meta-analysis
Results of planned analyses are presented below. All estimated models show good convergence of the chain. HMC trace-plots for the primary analysis, and Rhat statistic and effective sample sizes for all meta-analyses can be found in supplementary file, Appendix 7.
Overall summary estimate Bayesian meta-analysis including all types of admissions (with minimal overlapping data) is shown in Figure 1. The pooled estimate suggested that weekend admissions are associated with a 16% (95% credible interval [CrI] 10% to 23%) increase in the odds of death compared with weekday admissions.

Meta-regression
Results from multivariate meta-regression are shown in Table 1. The main findings are: (1) Studies that included measures of acute physiology in their statistical adjustment (adequacy of statistical adjustment group 1 or 2a) tended to produce an estimate of the weekend effect that is closer to null and on average reported estimates that are approximately 15% lower in terms of increased odds of mortality compared with studies without adjusting for these measures (groups 2b, 3 and 4).
(2) The weekend effect is significantly larger for elective admissions compared with emergency admissions, and significantly smaller (or does not exist) for maternity admissions.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  (4) The above findings need to be interpreted with caution. For example, the finding regarding statistical adjustment relies upon data from five estimates reported in four relatively small studies 16,19,61,65 that adjusted for measures of acute physiology, and the 95% credible intervals include zero (Table 1). Therefore there is still a substantial level of uncertainty, and the apparent effect of adjustment of acute physiology could have been confounded by other patient or service features associated with the availability of these measures.

Exploring the sources of heterogeneity
Meta-regression allows simultaneous exploration of multiple factors that could influence the magnitude of estimated weekend effects using study-level variables, but its statistical power is limited and is susceptible to confounding by study level variables. This subsection presents findings from additional subgroup and sensitivity analyses, paying particular attention to within-study comparisons to explore in more detail potential modifiers of the weekend effect.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  [Insert Table 2 here] Among emergency admissions, one study from England 17 and another from the USA 20 demonstrated that the observed weekend effect was largely attributable to 'direct' admissions from the community (e.g. general practitioner or walk-in clinic referrals) rather than those through the ED. Another US study restricted to admissions through the ED 57 also showed a substantially smaller weekend effect compared with other studies including all emergency admissions (supplementary file Appendix 9, p.48).

Adverse events
Nineteen studies compared the risk of adverse events between weekend and weekday admissions. 28,29,31,39,41,56,69,72,74,[77][78][79][80]84,85,[88][89][90][91] While some reported an increased risk for weekend admissions, overall the findings were heterogeneous across different adverse events within individual types of admissions, and the existence and magnitude of a weekend effect linked to a given adverse event were often inconsistent (supplementary file Appendix 11). None of the studies adjusted for physiological severity of illness: sicker patients (and particularly non-survivors) are more susceptible to adverse events. 92

Patient satisfaction
One study based on data from the 2014 English NHS adult inpatient survey reported a significantly higher level of satisfaction in the information given to them in the ED for patients admitted through this route at weekends compared with those admitted through the ED on weekdays. 37 After adjustment for potential confounders, no significant differences between weekend and weekday admissions were found in other domains covered by the inpatient survey (supplementary file Appendix 13).

GRADE assessment of overall quality of evidence
The overall quality of evidence was rated as "very low" for each of the outcomes (mortality, adverse events, length of hospital stay and patient satisfaction) examined in this review primarily due to the observational nature of evidence and inadequate or complete lack of adjustment for potential confounding factors in the majority of included studies. Further details on the GRADE assessment are presented in supplementary file Appendix 14.

DISCUSSION
This systematic review of studies reporting the weekend effect in broad ranges of admissions to hospital has found that weekend admission is associated with a 16% increase in the risk of death, but the magnitude of the effect varies by different types of admissions,  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  The overall estimate of the weekend effect varies in meta-analyses published to date, e.g. a pooled adjusted odds ratio of 1.12 (95% CI 1.07 to 1.18) by Hoshijima et al., 12 1.11 (95% CI 1.10 to 1.13) by Zhou et al. 14 and a pooled relative risk of 1.19 (95% CI 1.14 to 1.23) by Pauls et al. 13 Our meta-analysis covers by far the largest number of admissions; our pooled adjusted odds ratio of 1.16 (95% credible interval 1.10 to 1.23) is broadly in line with other studies, whereas the wider credible interval may, in part, reflect the use of Bayesian methods which appropriately account for both within-and between-study variations. Each of the above meta-analyses covers at least tens of millions of admissions, and yet the estimated weekend effects could differ by nearly two-fold. A clear message is that such an estimate is subject to a large amount of noise due to the myriad of contextual factors and different underlying mechanisms associated with different studies and admissions, which need to be examined more closely -and this is the key contribution of our review.
Weekend admissions differ from weekdays: fewer patients are admitted at weekends despite similar weekend-weekday ED attendance rates (thus creating a reduction in the denominator of the weekend mortality ratio) 17 and those that are admitted are sicker (case mix). 16,19,65 There is scant evidence to support the contention that hospital care is of inferior quality at weekends: adverse events may be more common but confounding by illness severity has not been excluded. In stroke care different patterns of variation in timeliness  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y   25 and adherence to best practice standards have been reported across the week, with no difference in weekend and weekday admission mortality rates. 93 In one study, vital signs were recorded more reliably at weekends than on weekdays. 19 The finding that weekend mortality effect is larger among elective than emergency admissions might be explained by a greater case mix difference between weekends and weekdays that is unaccounted for by statistical adjustment among elective admissions compared with emergency admissions. For example, for procedures that are often carried out during elective admissions, such as hip and knee replacement and surgery for large bowel, 75 switching the timing of admission from weekdays to weekends due to change in urgency (which is unlikely to be captured by administrative database) or delay in admission during weekday due to capacity issues ('overflow') is fairly plausible. On the other hand, a greater weekend effect associated with elective admissions is also consistent with the hypothesis that hospitals are configured to care for emergencies at weekends, while elective admissions might be overlooked.
Our review clearly illustrates the old wisdom that large volumes and advanced statistical techniques cannot make up for the inherent limitation within the data. Our assessment of overall quality of evidence using the GRADE framework reinforces the need to appreciate the weakness in available evidence when using observed weekend effect to make an inference on quality of hospital care at weekends. Nonetheless, careful examination of the data may help pin point areas for further investigation. For example, our findings show that the observed weekend effect is substantially larger among elective admissions compared with emergency admissions. Identifying specific types of elective admissions associated with  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  Determining the proximate causes for the weekend effect requires detailed study of the whole care pathways including health service provision, care processes and patient experience in the community, at the interface between community and hospital, and in hospital following admission on weekdays and at weekends. The paucity of published literature on quantitatively measured patient satisfaction is surprising, 37 as patient's, carer's, and service provider's experience must be at the centre of the design and delivery of health services. We will fill in these important evidence gaps through our companion framework synthesis, and other components of the HiSLAC project. 94,95 While our estimation of the overall association between weekend hospital admission and mortality is broadly in line with those reported previously, 12-14 our review has several unique strengths. First, previous reviews have either examined only mortality, 12-14 or mortality with a small number of care process or outcome measures for specific disease conditions. 6,9 Our review covers institution-wide and/or nationwide samples of hospital admissions and examined adverse events, LOS and patient satisfaction in addition to death.
Secondly, previous reviews have focused on using study level data to generate pooled estimates of the weekend effect. We have extended this by examining the more nuanced analyses available within individual studies.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y   27 This systematic review was limited by the exclusion of condition-specific admissions, although others have extensively reviewed these separately. Nevertheless, we believe the limitation is not a major threat for the validity of our conclusions as we have carefully triangulated the findings by examining subgroups both across and within studies and by carrying out sensitivity analyses. We have attempted to focus on more recent evidence by restricting our inclusion to studies published from year 2000 onwards. However some of the included studies (14/68) covered admissions pre 2000. This is unlikely to have substantial impacts on our findings as our meta-regression did not identify a significant time trend. Due to resource constraint (and paucity of data in the case of patient satisfaction) we were unable to carry out more sophisticated analyses for non-mortality outcomes.
Most studies included in this review utilise routinely collected administrative data. Our review suggests the need for caution in the analysis and interpretation of these information sources. For example, data on important confounders such as severity of illness are often unavailable, and undiscriminating adjustment of other variables such as hospital teaching status and bed size could risk "adjusting away" some of the weekend effect attributable to care quality. Differential data quality between weekend and weekday admissions is another potential contributor to the weekend effect. 29 19 We recommend a shift of focus from final adjusted mortality rates to considering how different pathway factors influence these estimates (Figure 2), using configurative analyses (pattern identification) to supplement aggregative (pooled) approaches. 96 [Insert Figure 2 here]  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59

CONCLUSION
Weekend admissions are associated with a 16% increase in the risk of mortality. However the overall quality of evidence is very low. Increasing evidence suggests that the weekend effect on mortality may be largely attributable to case mix and contextual factors surrounding admissions, and therefore the cause may lie upstream of the care pathway, in the community. In addition, the magnitude of estimated weekend effect can be influenced by methodological approaches and data quality. These suggest that the weekend effect is not a good measure of care quality in hospitals at weekends. Future research and interpretation of research findings on the weekend effect must go beyond the narrow focus of case mix adjustment of routine hospital data and attempt to examine the broader issues related to the whole care pathway both within and outside the hospital; the quality and availability of data that can allow measurement of care quality with minimal bias; and importantly, take into account the experience of patients, carers and care providers.

DATA SHARING
All data relevant to the study are included in the article or uploaded as supplementary information.

ACKNOWLEDGMENTS
We thank study authors who kindly responded to our queries and peer reviewers for their comments which helped to improve our manuscript.

Supplementary file: online appendices
Please note that references cited in this supplementary file are listed at the end of this document. Therefore the reference numbers cited in-text correspond to reference numbers of the reference list within this supplementary file, and they are different from the reference numbers quoted in the main paper.  When the desired information was not described/reported in the paper, please code as 'NR' (not reported). Sometimes an item is not relevant for a particular study, in which can you can enter 'NA' (not applicable).

Study characteristics and methods
Author year (Free text) First author and year of publication e.g. Albright 2009 The first author's last name and year of publication of the paper.

ID (Number)
The record number for the EndNote database This is provided in the file name of the paper. If more than one reference day or time period (against which weekend admissions were compared) was used, please record all (and where reported, which was used in the primary analysis, the rationale and whether this was pre-specified). List the initial sample size before any exclusions were made; or code  No exclusion  NR Initial sample size is defined as the number of admissions included in the initial sample before any exclusion (e.g. due to incomplete data) was made.

Number of hospitals (number)
List the number of hospitals from which the admissions were sampled Record the number of hospitals and put additional information (such as the number of NHS Trusts) in Comment Adverse events (AEs) are defined here as any undesirable events (other than death) that may be caused by medical management rather than the underlying condition of the patient, e.g. surgical complications. This definition does not imply preventability.
Interventions and procedures that are carried out mainly to deal with AEs rather than as part of the routine management of a condition are sometime used as indicators for the occurrence of AEs, such as some of the items included in the Patient Safety Indicators. These will also be considered as AEs for this review.

Other outcomes of potential interest
Record any other outcomes not listed above that were reported and might be useful e.g. any process measures or costs information. Record 'None' where appropriate.

Risk of Bias Assessment
The items below are modified from the Newcastle-Ottawa quality assessment scale Selection  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59   Select ONE option from a) to b) a) yes b) no Exclusion due to missing data Select ONE option from a) to e) a) no or very few of exclusions due to missing data -unlikely to affect study results b) some level of missing data, but admissions with missing data were retained in analyses using imputed data; c) some level of exclusion due to missing data, but authors demonstrated that admissions with missing data were similar to admissions included in analyses d) excluded a substantial proportion (5%) of admissions due to missing data from analyses e) no statement concerning missing data These can usually be found in a table or in the first couple of paragraphs in the Results section.

Significant differences observed in characteristics between weekday and weekend admissions (free text)
Record the characteristics for which weekend admissions were found out to be significantly different from weekday admissions List the names of the variables for which significant differences between weekday and weekend admissions were found. This can be defined statistically (i.e. p<0.05) or numerically (i.e. ≥5% difference between weekday/weekend admissions). No need to record numerical results at this stage.
For studies with a large sample size, trivial differences between weekday and weekend admission can still be statistically significant. Please add comments to describe if this is the case.

Quantitative results
These can be classified into two groups according to the types of outcome:  Dichotomous (binary) variables such as deaths or occurrence of complications  Continuous variables such as length of hospital stay

Appendix 2. Risk of bias assessment
Risk of bias assessment was embedded within the data extraction form shown in Appendix 1. We initially used the Newcastle-Ottawa scale 1 with modification of some of the items and wording because the included studies were mostly population database studies rather than the conventional cohort study for which the scale was designed.
However during the review process it became apparent that results of the risk of bias assessment using this modified scale were either unreliable (due to difficulties in judging the "representativeness" of the study sample for diverse types of admissions and lack of reported information about handling of missing data) or uninformative (e.g. all the included studies derived their control group [weekday admissions] from the same source and using the same inclusion criteria as with the exposure group [weekend admissions]). Therefore we subsequently only focus on adequacy of statistical adjustment, which was the key item stated a priori in our protocol. 2 The classification of statistical adjustment stated in the protocol needed to be refined during the review in view of emerging evidence indicating the importance of including measures of severity and urgency of the patients in the adjustment.

Technical details of the Bayesian meta-analysis
Analyses were undertaken using (log) adjusted odds ratios. For studies that only reported adjusted hazard ratios or rate ratios, we used these figures as approximations of adjusted odds ratios as results for these effect measures were very similar where they had been estimated in the same study (see Appendix 7.3.3).
As several studies provided multiple estimates of the weekend effect from different sub-samples (e.g. different time periods or different locations), we specified a three level Bayesian randomeffects model to take into account the correlation of results from different sub-samples within the same study while allowing for within sample variation and between study heterogeneity. In particular, for analysis or sub-sample from study with effect size estimate = 1,…, = 1,…, and estimated standard deviation :~(  . 5 We therefore used half-normal(0,1) priors for standard deviation terms and normal(0,1) for mean effects. We calculated the I-squared statistic, 6 which is the proportion of total variance attributable to between-study heterogeneity taking into account variance at three levels.
Convergence was assessed by visual inspection of traceplots of MCMC chains and the Rhat statistic.
Models were estimated in Stan. 7

Technical details of the Bayesian meta-regression
The model described in Equation (1) is extended to allow for varying mean effects according to characteristics of the sample, :~( where are a set of parameters to be estimated.
The following variables were included in a planned, exploratory meta-regression:  Categorical variable indicating adequacy of case-mix adjustment as described earlier.
Reference category was combined 1 and 2a (with adjustment of measures of acute physiology).
Two pre-specified variables were not included in the meta-regression due to lack of data: type of population (few studies focused on children) and country income category (none of the included studies was conducted in low and middle income countries). Instead we included an indicator variable for each country. The country effect is specified as a hierarchical 'random' effect. admissions in a meta-analysis results in double-counting and therefore needs to be avoided.
In the tables below we summarise characteristics of studies based in England and USA and illustrate the extent of potential overlap of data between these studies. Attention was paid to the hierarchical nature of the data; for example a study that included all emergency admissions would have included the same data from another study that focused on emergency medical admissions if they used the same data source and covered the same period of time, even though the former may not have provided an estimate of the weekend effect specific to emergency medical admissions.

Mortality -subgroups by country
As illustrated in Figure 1 the main text and other forest plots in Appendix 9.1 above, the weekend effect appears to vary between studies undertaken in different countries. Two studies provided data of cohorts from different countries. In a Global Comparators Project, Ruiz et al. investigated 30-day in-hospital mortality for emergency admissions and elective surgical admissions using data from four countries ( Table 15 in Appendix 9.1.5 above). 19 Weekend effect was found across the countries and type of admissions, but there were notable variations between the countries and no apparent weekend effect was observed in Australia for emergency admissions in their primary analysis. By contrast, Freemantle and colleagues obtained very similar estimates for two independent datasets from England and USA ( Figure 1 in the main text). 18  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59

Mortality -subgroups by disease conditions
Although systematic reviews of the weekend effect for individual disease conditions have been published, [79][80][81][82] comparisons of the weekend effect between different disease conditions could be confounded by differences in study-level characteristics between studies. Several studies included in this review reported weekend effects by selected, individual disease conditions and they provide a chance to make such a comparison that is less susceptible to confounding by study-level variables.
The data are presented in this section. We selected conditions for which the mortality is likely to be affected by hospital staffing level (ruptured abdominal aortic aneurysm, acute epiglottitis, and pulmonary embolism) and those for which mortality is unlikely to be influenced by staffing level as originally hypothesised by Bell and colleagues in their seminal paper, 52 as well as other conditions that commonly contribute to death during hospital admissions.
Conditions hypothesised by Bell et al. for which a weekend effect is less likely to be observed: The first was acute myocardial infarction, which is usually managed in a critical care setting, where fluctuations in staffing levels are minimal. The second was acute intracerebral haemorrhage, for which effective treatment is generally unavailable. The third was acute hip fracture, a condition that is sometimes treated more promptly on weekends than on weekdays, because operating rooms are more available on weekends.

Different effect measures
Two studies used different effect measures for a given mortality outcome measure. The numerical estimates of the weekend effect were very similar between odds ratio and hazard ratio, and between odds ratio and risk ratio in the respective study.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  This section shows that even based on the same data source, estimation of the weekend effects can still vary substantially, indicating the potentially poor signal-to-noise ratio in using the weekend effect on hospital mortality as a reliable measure of care quality. *Multiple estimates were reported across the three papers based on the same dataset (with exact the same number of admissions), but the effect measures and corresponding methods (e.g. type of statistical techniques/models and variables being adjusted for) were poorly described.  77 While an increased risk of experiencing various adverse events was observed for weekend admissions in some studies, the findings were heterogeneous and inconsistent (e.g. risk of was increased for some measures of adverse events but not increased or even decreased for other measures within individual studies; inconsistent findings with regard to the existence and magnitude of the weekend effect for a given adverse event between different studies). Findings for different measures of adverse events are presented below.

Composite measures of adverse events
In a large study using the Nationwide Inpatient Sample, Attenello and colleagues 35 found that weekend admissions are associated with a 25% increase in the odds of experiencing a hospital acquired condition, which is considered by the US Centers for Medicare & Medicaid Services as  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46 F o r p e e r r e v i e w o n l y 63 a condition that "could reasonably have been prevented through the application of evidence-based guidelines". Falls and trauma within hospitals were the most common hospital acquired condition, which constituted of 88% of the events.
Based on data collected in a voluntary incident-reporting system, Buckley & Bulger 2012 77 suggested that the risk of clinical management incidents (including both errors leading to harm and near misses and errors that did not cause harm) was higher among patients admitted during weekends (OR 2.738, 2.552 to 2.937). However the increased risk was most pronounced for incidents that were less serious, and no adjustment was made for severity of illness.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46 F o r p e e r r e v i e w o n l y 64

Needs for further hospital care following initial admission
Four studies examined the weekend effect on various measures of readmission, reoperation, ICU admission and A&E visits following the initial admission (or following discharge from the initial admission). The results are heterogeneous: one study focusing on children 43 with inadequate adjustment of potential confounding factors reported a 9% increase in the odds for 30-day readmissions associated with weekend admissions; Khanna and colleagues, based on data from a single hospital in Chicago in which there were no difference in physician level between weekdays and weekends, found that weekend admissions were associated with significantly lower risk of ICU transfer during hospitalisation and were not associated with an increased risk of readmissions or re-visit to the A&E; 61 Dubois and colleague found an increase odds of 7% for ICU admission among elective surgeries carried out on Fridays compared with those carried out on Mondays, but no increase in 30-day reoperation or readmission was observed. 66 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46 F o r p e e r r e v i e w o n l y 67

Patient safety indicators and surgical adverse events
Many studies examined the weekend effect on patient safety indicators or other measures related to adverse events during or following surgery. While an increased risk of surgical adverse events was found in several studies, the findings were not consistent across different outcomes within individual studies and were also heterogeneous across studies for a given outcome measure.    1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46 F o r p e e r r e v i e w o n l y 71

Indirectness
None (could have been downgraded) Timing of the admissions was used as a proxy for hospital care quality. Therefore inference was made indirectly. As the rating is already downgraded to "very low", no further downgrading is possible.

Imprecision
None This is not a particular concern given the large volume of evidence included. The 95% credible interval (1.10 to 1.23) is reasonably narrow.

Publication Bias
None Although funnel plot asymmetry was observed, our sensitivity analysis using data augmentation methods showed that adjustment for the asymmetry only had a small impact on the pooled estimate. 6. Large magnitude of effect

None
The magnitude of effect was not large and could plausibly be attributed to confounding. 7. Dose response None Limited evidence from two studies examining the relationship between staffing level/7-day service provision and mortality did not show correlation between them. 24,27 8. Effect of plausible confounding factors

None
Plausible confounding factors would produce an effect in the same direction as the observed weekend effect Final rating Very low This is in relation to using the estimated weekend effect on mortality to infer weekday/weekend difference in care quality in the hospital.  Inconsistency was observed between studies and between different adverse events within studies. However as the rating is already downgraded to "very low", no further downgrading is possible.

Indirectness
None (could have been downgraded) Timing of the admissions was used as a proxy for hospital care quality. Therefore inference was made indirectly. In addition, measures such as hospital acquired conditions and patient safety indicators are proxy measures of adverse events arising from suboptimal care. There is therefore some level of indirectness in the evidence. However as the rating is already downgraded to "very low", no further downgrading is possible.

Imprecision
None The level of precision varied by individual adverse events. As the rating is already downgraded to "very low", no further downgrading is required.

Publication Bias None
We were unable to assess publication bias due to lack of study registry, and we did not carry out meta-analyses or construct funnel plots for this outcome given the diverse measures used. 6. Large magnitude of effect

None
The magnitude of effect was not large and was inconsistent, and could plausibly be attributed to confounding. 7. Dose response None Limited evidence from one study 61 showed no difference in physician level between weekday and weekend. 8. Effect of plausible confounding factors

None
Plausible confounding factors would produce an effect in the same direction as the observed weekend effect.

Final rating
Very low This is in relation to using the estimated weekend effect on adverse events to infer weekday/weekend difference in care quality in the hospital.  There is notable heterogeneity between studies, not just in the magnitude but in the direction (i.e. some found the length of stay for weekend admissions was longer than weekday admissions while others found the opposite). However as the rating is already downgraded to "very low", no further downgrading is possible.

Indirectness
None (could have been downgraded) Timing of the admissions was used as a proxy for hospital care quality. Therefore inference was made indirectly. Length of stay was directly measured from administrative records.

Imprecision
None This is not a particular concern given the large number of admissions examined in the included studies.

Publication Bias None
We were unable to assess publication bias due to lack of study registry, and we did not carry out meta-analyses or construct funnel plots for this outcome. Publication bias may not be a major concern given the diverse findings reported in the included studies. 6. Large magnitude of effect

None
The magnitude of effect was not large and was inconsistent. 7. Dose response None Evidence was lacking to allow assessment of dose response. 8. Effect of plausible confounding factors

None
Different confounding factors may produce effects in the same or opposite direction as the weekend effect (where observed).

Final rating
Very low This is in relation to using the estimated differences in the length of stay to infer weekday/weekend difference in care quality in the hospital.  Downgrade one level Only one study provided data. 46 Adjustment for potential confounding factors was very limited.

Inconsistency
None We were unable to assess this domain as only one study provided data for this outcome. 46

Indirectness
None (could have been downgraded) Timing of the admissions was used as a proxy for hospital care quality. Only one study provided data on this outcome, 46 and the analysis of inpatient survey had to focus on questions related to admission and discharge processes rather than the period of stay as inpatient.

Imprecision
None The sample size was reasonably large, although only one study provided data on this outcome.

Publication Bias
None Given that only one study was found and that there is no study registry available, we were unable to assess the potential impact of publication bias. 6. Large magnitude of effect

None
The magnitude of effect was not large.

Dose response None
Evidence was lacking to allow assessment of dose response. 8. Effect of plausible confounding factors

None
Different confounding factors may produce effects in the same or opposite direction as the weekend effect (where observed).
p. 7-9 Data collection process 10 Describe method of data extraction from reports (e.g., piloted forms, independently, in duplicate) and any processes for obtaining and confirming data from investigators. p.9-10

Section/topic # Checklist item Reported on page #
Risk of bias across studies 15 Specify any assessment of risk of bias that may affect the cumulative evidence (e.g., publication bias, selective reporting within studies).

p.13
Additional analyses 16 Describe methods of additional analyses (e.g., sensitivity or subgroup analyses, meta-regression), if done, indicating which were pre-specified. p.12

RESULTS
Study selection 17 Give numbers of studies screened, assessed for eligibility, and included in the review, with reasons for exclusions at each stage, ideally with a flow diagram.
p.14, supplementary file Appendix 5 Study characteristics 18 For each study, present characteristics for which data were extracted (e.g., study size, PICOS, follow-up period) and provide the citations.
Online Appendix 6 Risk of bias within studies 19 Present data on risk of bias of each study and, if available, any outcome level assessment (see item 12).
Online Appendix 6 (first column) Results of individual studies 20 For all outcomes considered (benefits or harms), present, for each study: (a) simple summary data for each intervention group (b) effect estimates and confidence intervals, ideally with a forest plot.

DISCUSSION
Summary of evidence 24 Summarize the main findings including the strength of evidence for each main outcome; consider their relevance to key groups (e.g., healthcare providers, users, and policy makers).