Original Article
Length of comorbidity lookback period affected regression model performance of administrative health data

https://doi.org/10.1016/j.jclinepi.2005.12.013Get rights and content

Abstract

Background and Objective

The impact of different comorbidity ascertainment lookback periods on modeling posthospitalization mortality and readmission was examined.

Methods

Index cases comprised medical (n = 326,456) and procedural (n = 349,686) patients with a hospital admission from 1990–1996. Administrative hospital data were extracted for 102 comorbidities, ascertained at index admission and for 1-, 2-, 3-, and 5-year lookback periods. Deaths and readmissions were identified within 12 months and 30 days of separation, respectively. Hierarchically nested and nonnested Cox regressions as well as Receiver Operator Characteristic Area Under the Curve (ROC-AUC) were used to determine model-fit and predictive ability of lookback period models.

Results

The 1-year lookback period provided the best model-fit for both patient groups when modeling mortality. A similar model-fit was seen at index admission for procedural but not medical patients. The superior readmission model employed 5 years of lookback for both patient groups. With one exception, all lookback period models were superior to those abstracting comorbidity from index admission only. Similar results were evident from ROC-AUC, although greater predictive ability was seen with modeling of mortality (0.847–0.923) compared with readmission (0.593–0.681).

Conclusion

The explanatory power of regression models, when adjusting for comorbidity, is influenced by length of lookback, outcome investigated and clinical subgroup. Shorter periods (∼1 year) appear appropriate for modeling posthospitalization mortality, whereas longer lookback periods are superior for readmission outcomes.

Introduction

It is well recognized that patient comorbidity plays an important role in determining mortality and morbidity outcomes [1], [2]. As a result, it has become common practice to adjust for comorbidity in analyses of health outcome data [3], [4], [5].

With the emergence of administrative databases, the ability to access longitudinal patient data to adjust for comorbidity has improved considerably. This raises the issue of the most appropriate lookback period to determine patients' disease status for risk estimation, given that the effects of different comorbid conditions may vary depending on their recency and duration. Most research employing comorbidity adjustment has used relatively short lookback durations (<1 year), based primarily on practical considerations such as convenience, clinical judgement, and availability of data, rather than empirical evidence [6]. Longer lookback periods are likely to capture more conditions per patient, as well as assign comorbidities to a greater proportion of patients. However, this may come at a cost of weakening the effect of comorbid conditions deemed to be present, and degrade the fit of the risk adjustment model.

The influence of lookback period on risk estimation has been demonstrated with certain primary diagnoses such as ischaemic stroke [7] and cancer [8]. However, appropriate lookback period length for ascertainment of comorbidity for risk adjustment has received limited attention. To date, research has mostly ascertained comorbid disease status either at index admission only or within a relatively short period of 1 year [2], [3], [9]. Isolated research has employed lookback periods as long as 2 years for evaluating posthospitalization mortality [6]. However, it is unclear what effect longer lookback periods would have for such outcomes in terms of statistical modeling and predictive ability. Further, the appropriateness of different lookback durations on health outcomes such as hospital readmission is yet to be investigated. Also, it remains unclear whether the suitability of different lookback periods differs depending on the clinical population investigated.

Our aim was to determine, using routinely collected administrative health data, the effects of comorbidity ascertainment with different lookback durations for modeling (1) death within 1 year of hospital admission; and (2) readmission within 30 days of hospital separation, for two clinically distinct populations.

Section snippets

Western Australian Data Linkage System (WADLS)

The WADLS systematically links administrative health data within a single Australian State (population 1.95 million) [10], [11]. It combines seven core databases; births, midwives' notifications, cancer registrations, inpatient hospital morbidity, mental health services, deaths, and electoral registrations, variously dating back to 1966. The system is updated on a continuous basis, and includes working links with >30 external health service and research databases. The Western Australian (WA)

Results

Longer lookback resulted in more comorbidity being identified. For the entire sample (n = 676,142), 46.8% of comorbidity observed across the 5-year lookback period was recorded at index hospitalization. This increased to 68.6, 79.1, and 89.5% with the addition of 1, 2, and 3 years of lookback, respectively. Similar 5-year trends were seen for both patient groups, although a greater occurrence of comorbidity was present in the medical group (163.0% of procedural records).

Discussion

This study is the first to investigate effect of lookback durations >2 years for comorbidity ascertainment on model fit and predictive ability. Further, it is the only investigation to focus on this area of modeling methodology for both mortality and readmission, and is the first to report differences in lookback period performance between clinical groups.

Our results demonstrated that <50% of comorbidity recorded in the preceding 5 years was present at index hospitalization. As anticipated,

Conclusion

Comorbidity adjustment with health outcome modeling is influenced by lookback period length, outcome of interest, and patient type. For posthospitalization mortality, while the addition of comorbidity up to 5 years preindex admission does not degrade model fit, little improvements are seen above that with approximately 1 year of lookback. Further, ascertainment of comorbidity (for adjustment purposes) at index admission only appears appropriate for procedural patients but not for medical

Acknowledgments

The initial construction of the Western Australian Data Linkage System (WADLS) was funded by the WA Lotteries Commission. This study was supported by a project grant from the Australian National Health and Medical Research Council.

References (25)

  • R.A. Silliman et al.

    Comparison of interview-based and medical-record based indices of comorbidity among breast cancer patients

    Med Care

    (1999)
  • J.X. Zhang et al.

    The performance of different lookback periods and sources of information for Charlson comorbidity adjustment in Medicare claims

    Med Care

    (1999)
  • Cited by (180)

    View all citing articles on Scopus
    View full text