Research question Is length of stay (LoS) in hospital a stable proxy for severity of injury when monitoring time trends in serious injury incidence?
Objective To investigate whether LoS metrics (mean, median and proportion exceeding several LoS thresholds) have changed over time for injury diagnoses with known severity.
Design Time series investigation.
Subjects and setting New Zealand population admitted to hospital for injury and discharged during the period 1989 to 1998.
Main outcome measures Interpolated median and geometric mean lengths of stay, as well as the proportion of cases that have an LoS greater than or equal to 3, 4, 7 and 14 days in hospital.
Methods ICD-9-CM diagnoses that are approximately homogeneous in regard to severity of injury (ICD-HS diagnoses) were identified. Trends were investigated in the LoS statistics for: injury and non-injury diagnoses combined; all injury diagnoses; major body sites of injury; severity strata; and ICD-HS diagnoses.
Results Almost without exception, there was a decline in the LoS statistics over time for all diagnoses, all injury diagnoses, each body site of injury investigated, severity strata, and the ICD-HS diagnoses.
Conclusions Reductions in median and geometric mean LoS over time, as well as reductions in the proportion exceeding selected LoS thresholds, were due to factors other than reductions in the incidence of serious injury; for example, changes in service delivery over time. An LoS threshold should not be used as a proxy for severity of injury if the goal is to monitor time trends in injury incidence.
- Length of hospitalisation
- health services
Statistics from Altmetric.com
A number of agencies have used length of stay (LoS) in hospital as a proxy for severity in their national non-fatal indicators of injury incidence (eg, New Zealand Ministry of Transport, English Department of Health).1 2 The question arises from these initiatives: is LoS a stable proxy measure for severity? It is our hypothesis that it is not.
Work we have carried out so far suggests we are correct,3 although more definitive proof has yet to be produced and is needed. The purpose of this paper is to provide that proof.
There is also anecdotal evidence that LoS is problematic. For example, in the Otago Daily Times newspaper on 25 February 2008, there was a headline: ‘Beds close as hospital services struggle’. The article described nursing shortages, bed closures, and early discharge of patients (ie, reduced length of stay) to cope with the ‘crisis’. This is far from an isolated example. These health service factors, extraneous to serious injury incidence, will affect LoS in hospital, and so are likely to affect trends in serious injury incidence based on the use of an LoS threshold to define ‘serious’.
In order to develop national indicators of injury incidence, we are reliant on the use of national, routinely collected, data sets (eg, national mortality and hospitalisations data sets). Typically, these capture anatomical response to injury through the diagnosis codes. The relationship between diagnosis and severity has been used successfully in the development of severity scales such as the Abbreviated Injury Scale (AIS) and the ICD-based Injury Severity Score (ICISS).4 5
In many national mortality and hospitalisation data systems, including that of New Zealand (NZ), diagnosis of injury is classified using the WHO's International Classification of Diseases (ICD). To investigate our research question, the approach we have taken first identified ICD injury diagnosis codes that can be regarded as homogeneous in terms of severity (ICD-HS). If LoS is a good proxy for severity, then for admitted cases with particular ICD-HS diagnoses, we would expect the proportion that exceeded a given LoS threshold to be constant over time (see appendix). Testing this expectation was the basis of our approach.
The investigation used was hierarchical, starting with ‘all diagnoses’ and ‘all injury’, injury classified to six body regions, injury categorised in threat to life severity strata, and finally investigations for ICD-HS diagnoses.
There are a number of reasons why this hierarchical approach was adopted:
To check trends in LoS overall and in body site of injury groups.
To check for consistent trends in severity strata, where severity is defined using the threat to life severity measures, AIS and ICISS.
To investigate whether the trends found for the selected ICD-HS diagnoses are likely to be consistent with other ICD-HS diagnoses.
Although this is a NZ study, the results of the study are relevant more widely since many of the drivers of admission to, and LoS in, hospital are common across countries with a similar mix of public and private hospital provision.
The source of data was the National Minimum Data Set (NMDS) of hospital discharges for the period 1989 to 1998 inclusive. We sought as long a time series as was practicable. The upper bound of 1998 was chosen since in 1999, NZ hospitals moved to classifying diagnosis of injury using the 10th revision of the ICD (ICD-10). This represented a major change in the classification system used, in which some ICD-9 codes mapped to ICD-10, but others did not. The lower bound of 1989 was chosen since, before 1989, the date of injury was not available on the NMDS; data that was crucial for calculating total days stay (see below).
LoS thresholds of 3+, 4+, 7+, and 14+ days were investigated. The 3+ days threshold has been used as one of the criteria to define serious injury in the UK's Trauma Audit and Research Network database.6 The 4+ days threshold was chosen since it has been used as the basis of a definition of serious injury by some indicator developers.1 2 The other, longer thresholds were chosen to investigate whether more stringent thresholds produced consistent relationships.
The numerators and denominators, for the percentages of cases exceeding each LoS threshold, included:
first admissions only,7
with a total length of stay of at least 1 day,
whose principal diagnosis was included in the ICD-9 diagnostic code range 800–904 or 910–995,
who were not discharged dead on the first admission record.
This ICD-9 range excluded medical injury and sequelae/late effects. Deaths were excluded as this investigation relates to non-fatal indicators of injury incidence. (Note: a sensitivity analysis was carried out for all injury diagnoses, as well as the ICD-HS diagnoses, and the results were robust to the exclusion or inclusion of deaths in hospital.)
For each person-injury event, there may have been more than one hospital discharge event. For the purposes of this investigation, ‘total days stay’ was calculated as the sum of each day's stay duration relating to a person-injury event. A unique identifier was assigned to each readmission (same hospital or transfers to another hospital) relating to a person-injury event (‘readmissionid’). The ‘day stay’ field was then summed across all records with the same ‘readmissionid’. The only limit on the duration of time used for follow-up in order to identify readmissions, and to accumulate days stay relating to an injury event, was that inherent in using discharges for the period 1989 to 1998.
We tabulated interpolated median and geometric mean lengths of stay (see ‘Statistical methods’ below), as well as the proportion of cases that have an LoS of 3+, 4+, 7+, and 14+ days in hospital for:
All diagnoses (injury and non-injury diagnoses).
All injury diagnoses (in the ICD-9 diagnosis range 800–904 or 910–995).
Major body sites (from the Barell matrix)8 (http://www.cdc.gov/nchs/about/otheract/ice/barellmatrix.htm) with ‘extremities’ split into upper and lower).
Maximum AIS (MAIS) and ICISS severity strata.
ICD-HS diagnoses (see below).
The method used to calculate ICISS has been described previously.9 We used ICDMAP-90 to obtain AIS.10 The ICISS could only be used reliably after 1993—and so time trends for ICISS strata were restricted to the period 1994 to 1998.
The main properties that we were seeking for each ICD-HS diagnosis were that:
all cases captured by the particular ICD-HS diagnosis code should have approximately the same severity of injury;
there should be sufficient numbers of ICD-HS diagnosis cases that exceed the LoS thresholds chosen, to enable reasonably precise changes in the LoS statistics to be detected over time (see below).
Operationally, the candidate diagnoses were chosen if they satisfied the following criteria:
The number of cases with three or more total days stay in hospital was >100 per year.
>20% of cases with a particular diagnosis had an LoS stay of ≥14 total days;
or if the percentage of cases with LoS≥14 days in hospital was less than 20%, then the number of cases with LoS≥14 days in hospital was greater than 100 per year.
The ICD diagnosis code mapped to a single AIS severity code, using the high severity and low severity options, in ICDMAP-90.10
For each of the groupings listed above, we tabulated the interpolated median (IM) and geometric mean LoS, as well as the percentage of incident cases that had an LoS of 3+, 4+, 7+, and 14+ days in hospital.
The IM was generated using the ‘iquantile’ function in Stata V.10.1. The IM can be thought of as the calculated estimate of the true median, had there been less granularity in the scale. The geometric mean was calculated as the arithmetic mean of the logarithms of individual values, converted back by taking the antilogarithm.11
For the IM, the hypothesis of no change was tested using the non-parametric test for trend developed by Cuzack (Stata's ‘nptrend’), which is an extension of the Kruskal–Wallis test.12 For the geometric mean, the hypothesis of no trend was tested using a linear regression on logarithms of the total days stay. For the percentage of cases that have an LoS of 3+, 4+, 7+, and 14+ days, the hypotheses of no change over time were tested using χ2 tests, as well as the non-parametric tests for trend. This provided an almost equivalent test to the χ2 1-degree of freedom test for trend. No confounding factors were included, or were necessary, in the analysis, since the purpose of the hypothesis tests was to simply provide a measure of whether the trends could have occurred by chance alone.
Stata V.10.1 was used for the analysis.
For ‘all diagnoses’ (injury and non-injury diagnoses combined) and for ‘all injury diagnoses’ (excluding medical injury and late effects), there were consistent declines over the 10 years for each of the interpolated median LoS, the geometric mean LoS, as well as the percentage of admissions of 3+, 4+, 7+, and 14+ days (table 1, figure 1). Table 1 and figure 1 exclude newborns (for ‘all diagnoses’).
Many of the trends for each major body site of injury were consistent with those for all diagnoses and all injury (see table 1S in the supplementary file).
For the AIS severity strata (MAIS=1, 2, 3, or 4+), there were downward trends over time (with a few exceptions) in the LoS statistics investigated (see table 2S, supplementary file). This is illustrated, by way of an example, for 7+ total days' stay (figure 2). This was true in all cases except that it was less obvious for trends in the percentage of cases with 3+ days and 4+ days in hospital where the maximum AIS was 3 (‘serious’) or 4+ (‘severe’).
For each of the ICISS severity strata, there were observed downward trends over time in the percentage of cases with 3+, 4+, 7+, and 14+ total days stay (see table 3S, supplementary file). (Note: ICISS gives a measure of likelihood of survival, and so the smaller the ICISS score, the greater the severity of injury, as measured by threat to life.) This is illustrated, by way of an example, for 7+ total days stay (figure 3). This was true in all cases except for trends for strata with an ICISS ≤0.93 (high threat to life injury) (see supplementary file). For these, the trends were more erratic, in the sense that most showed a decline, but the observed trend was not monotonic in all cases. There were statistically significant trends for the interpolated medians and the geometric means for all of the ICISS strata.
Table 2 shows the 11 diagnosis codes that satisfied the criteria for selection as ICD-HS diagnoses (see Methods). By way of illustration, figure 4 shows trends in the interpolated median LoS for all diagnosis codes satisfying the criteria.
With the exception of ICD-9-CM=81201 (fracture of humerus: upper end, closed—surgical neck), the interpolated medians and the geometric means showed strongly significant downward trends. The same was true for the trends in percentage with LoS of 7+ and 14+ days—with the exception of 14+ days for ICD-9-CM=82300 (upper end closed fracture of the tibia). It was true also for the percentages with 3+ and 4+ days stay, except in the instances where the percentages staying for at least 4 days were close to 100% for the whole period.
Over the 10 years included in this study, lengths of stay in hospital for admissions for all diagnoses combined (injury and non-injury), as well as for all injury diagnoses, reduced. These reductions were shown using all of the LoS metrics investigated.
If LoS was a good proxy for severity and was stable over time, then for a particular group of injuries of constant severity, it would be expected that these LoS statistics would have remained constant over time. For example, for the group of injuries that have an AIS severity score of 3, it would be expected that the geometric mean and interpolated median LoS would be constant over time. This was found not to be the case. The results for the LoS thresholds, and for each of the AIS and ICISS severity strata, suggest that LoS is not a stable proxy for severity.
We considered 11 ICD-9-CM diagnoses (ICD-HS diagnoses) for which cases classified to a particular diagnosis mapped to a single AIS severity code, using the high severity and low severity options, in ICDMAP-90. We have shown in this paper that, for each ICD-HS diagnosis considered, almost all of the LoS statistics reduced (statistically significantly) over time.
These results are consistent with the results from a 2001 publication, for which it was found that there were declines in LoS over time for specific diagnoses, including for fractured neck of femur, within each of the seven OECD countries considered.13 The declines are likely to be associated with changes in treatment and care services. For example, the UK government has regarded reducing LoS as a component of increased efficiency,14 thus resulting in pressure to reduce diagnosis-specific LoS. In the recent Clarke and Rosen review, the authors' state: ‘In many health systems, there are managerial and financial incentives to reduce LOS’.13 They indicate that from a healthcare provider's viewpoint, reasons to reduce LoS include: to tailor care to the individual and their preferences (often home care, if it can be provided safely); and to reduce resources spent on an individual so that the health dollar can stretch further.13 These trends give further evidence against using LoS as a stable proxy for severity of injury.
Although agencies have used LoS as a proxy for severity in their national indicators of injury incidence,1 2 we are unaware of any published work that gives either theoretical or empirical justification for this. Rather, we have heard verbal justifications using the following sort of argument: ‘If a person sustains an injury and they are admitted to hospital, their injury must by definition be serious. The longer that they stay in hospital, the more serious the injury must be.’
In this paper, we have produced evidence for this in the context of a cross-sectional analysis, since each of the LoS metrics that we used increases with increased threat to life, that is, with increasing AIS level and decreasing ICISS value. Nevertheless, this current paper focuses on the use of LoS as a proxy for severity when monitoring serious injury over time. In this context, we have provided strong evidence against LoS being a good and stable proxy for the severity of injury.
Strengths and limitations of the study
The strengths of the methods include the use of several LoS statistics and an investigation of several different groups of ‘cases’. It was a national study, was based on a large number of cases, and so provided estimates and trends with good precision. The data we used were based on discharges from all public hospital providers in New Zealand, thus our results are not subject to selection bias.
Our results are based on data from the period 1989 to 1998. We have investigated trends in LoS metrics during the period 2001 to 2008. The trend changes during this period compared with that during the 1990s. Instead of a decline, there is a small, statistically significant, increase in interpolated median LoS for injury cases during this period. The nature of the change is irrelevant, however; these further results also provide evidence against LoS being a stable proxy for the severity of injury.
A potential weakness is due to the inaccuracies of diagnosis coding that have been identified in New Zealand.15 These inaccuracies were mainly at the most specific level of ICD-9-CM, however. We investigated the likely effect that this would have on the trends we produced, by considering the trends in the LoS statistics for the ‘families’ of diagnostic codes related to the ICD-HS diagnoses—for example, for all codes relating to closed fracture of the neck of femur. The trends in the LoS statistics were consistent across all codes within each family—thus suggesting that there would be minimal effect of any such coding inaccuracies.
One of our assumptions is that the ICD-HS severity level is constant over time. The damage to the body represented by the ICD-HS code may be constant over time, but does the level of severity change? For example, there is evidence that case-fatality rates have been improving.16 Consequently it could be argued that for a particular injury diagnosis the threat to life severity is reducing over time. Nevertheless, the reduction in LoS has been occurring across the board: the LoS trends for disease diagnoses are similar to those for injury diagnoses. Although there will be some drift in injury diagnosis-specific level of severity over time, that drift would not account for the magnitude in trends in the LoS measures observed.
Another assumption is that ICD-HS diagnosis codes are homogeneous in severity. Even though the ICD-10-AM rubric includes 1783 injury diagnosis codes, injuries captured by any particular code are unlikely to be totally homogeneous in severity. It is possible, therefore, that the mix of severity of cases captured by a particular code changes over time. It is highly unlikely, however, that it will occur in a consistent way for each ICD-HS code, and for each of the severity categories, and body site of injury groups, that were considered. Heterogeneity of the ICD-HS codes is a very unlikely explanation for the consistent downward trends in the LoS metrics that we have found.
The results show that length of stay is not a stable proxy for severity. An LoS threshold should not be used as a proxy for severity of injury if the goal is to monitor time trends in injury incidence. The concern is that using an LoS threshold to identify serious injuries will result in misleading trends; and this applies where there are diagnosis-specific changes in LoS over time, in any country, any period of time, and to any coding frame, be it ICD-9-CM, ICD-10-AM, or other.
What is already known on the subject
Theoretically, length of stay (LoS) in hospital is an unstable proxy for severity of injury, because factors other than severity of injury influence LoS in hospital (eg, health services factors).
There is limited empirical evidence to support this proposition.
There is evidence that LoS is affected by health service changes.
What this study adds
The results show that LoS is not a stable proxy for severity of non-fatal injury.
Using LoS as a proxy for severity of non-fatal injury will lead to misleading trends over time.
We thank Caroline Finch for her very insightful comments, and Ian Civil and John Campbell for their helpful comments, on the penultimate draft of this manuscript. This work is based on data supplied to the Injury Prevention Research Unit at the University of Otago by the Ministry of Health.
If an LoS threshold was a good proxy for a severity threshold (severity measured by threat-to-life, threat-of-disability, or any other relevant measure), then each year the number of cases exceeding the LoS threshold should be proportional to the number of cases exceeding the severity threshold; that is,where ∼=‘proportional to’. Consider an ICD-HS code that exceeds a severity threshold of interest. Then, for such an ICD-HS diagnosis code, if an LoS threshold is a good proxy for a severity threshold, each year:where n(LoS+|ICD-HS) means the number of cases exceeding the severity threshold among those classified to the particular ICD-HS code.
In other words, over time:
This measure is ‘the percentage of cases within the diagnostic group that exceed the length of stay threshold’. So if LoS is a good proxy for severity, this percentage will be constant over time.
Funding This research was supported by the Accident Compensation Corporation of New Zealand, Molesworth Street, Wellington, New Zealand. The views expressed in this paper are those of the authors and do not necessarily reflect those of the ACC.
Competing interests None.
Ethics approval This study was conducted with the approval of the Multiregion Ethics Committee, New Zealand.
Provenance and peer review Not commissioned; externally peer reviewed.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.