Tomas J. Philipson, PhD; Julia Thornton Snider, PhD; Darius N. Lakdawalla, PhD; Benoit Stryckman, MA; and Dana P. Goldman, PhD
Malnutrition is a serious and underappreciated problem among hospitalized patients. Malnourished patients face heightened risks of poor outcomes, including increased length of stay (LOS),1-3 healthcare costs,1-4
Estimates of malnutrition prevalence in the inpatient population range from 8% to 62%, depending on the location and the specific patient population considered.13-17
Groups at highest risk include elderly as well as oncology and gastroenterology patients.16
Despite evidence documenting the deleterious effects of malnutrition in the inpatient setting, studies suggest it is a common problem that often goes unrecognized and undertreated.15,18
A growing body of evidence suggests that oral nutrition supplements (ONS), which deliver both macronutrients and micronutrients for special medical purposes in addition to normal food, might improve outcomes among hospitalized patients. A variety of benefits have been found for ONS use, including reduced LOS,3
inpatient episode cost,3,19
and readmission rates,22,23
and improved lean body mass recovery.24
However, previous studies suffer from limitations, including modest sample sizes, narrowly selected patient populations, and in observational studies, possible selection bias. Consequently, questions remain regarding the robustness and generalizability of existing findings and the size of gains and healthcare costs associated with ONS use in hospitalized patients.
This retrospective data analysis was conducted to assess the association and causal impacts of ONS on health outcomes for hospitalized patients, focusing on 3 key outcomes: LOS, episode cost, and probability of 30-day readmission.METHODSSetting, Subjects, and Data Sources
This analysis was conducted using the Premier Perspectives Database. This database contains diagnostic and billing information on 44.0 million adult inpatient episodes at 460 sites during the years 2000 to 2010. Premier estimates that these data cover 20% of all US inpatient episodes. The sample was restricted to adults 18 years or older and excluded terminal episodes and all episodes involving tube feeding, leaving only oral feeding for examination. All monetary figures were reported in 2010 dollars and inflation-adjusted based on the Bureau of Labor Statistics medical Consumer Price Index (http://www.bls.gov/ cpi/#tables).Measures
The study’s 3 key outcome variables were LOS, episode cost, and probability of 30-day readmission. Length of stay was defined as the number of days of direct patient care (minimum 1 day) from admission to discharge. Episode cost was defined as the actual costs to treat the patient during the hospitalization. Thirty-day readmission probability was defined as a return hospitalization for any diagnosis. For patient confidentiality purposes, the Premier database only contains the month and year of an inpatient episode. Therefore, the 30-day readmission window was calculated by identifying admissions later the same month or during the following month. Given that there are no International Classification of Diseases, Ninth Revision
or Current Procedural Terminology codes that identify ONS use, ONS was defined as a “complete nutritional supplement, oral,” as indicated by the Premier data, and coded as a binary variable, indicating any ONS used during the inpatient episode.Statistical Analysis
Naïve ordinary least squares (OLS) regression analyses were performed on the full matched sample. Analyses controlled for a variety of patient, episode, and provider characteristics. Demographic covariates included age, age squared, insurance type, marital status, race, and sex. Comorbidity covariates included all components of the Charlson Comorbidity Index.25,26
Health history covariates included whether the patient had been admitted to any Premier network hospital in the previous 6 months, and whether the patient was admitted from the emergency department, by physician referral, or by inter-facility transfer. Hospital- specific covariates included number of beds, urban location, whether the site was a teaching hospital, and region (Northeast, Midwest, West, or South as defined by US Census data). Time trends were controlled for using year and quarter dummies.
Treatment is assigned randomly in clinical trials to avoid confounding; however, there is potential for selection bias in observational research from unobserved factors that may influence study outcomes.27,28
Because it is likely to be administered to individuals who are less healthy, ONS use could be spuriously associated with increased LOS. Additionally, only certain patient health-related risk factors were directly observable with available data. Therefore, further methods were used to remove these sources of potential selection bias: propensity score matching and instrumental variables analysis.Propensity Score Matching.
To diminish the potential for confounding due to differences in observed personal characteristics and to identify nutritionally at-risk patients, propensity score matching29
was used to match ONS episodes to similar non-ONS episodes. The probability of receiving ONS was estimated using a logistic regression of ONS based on the covariates noted above. After removing all episodes involving children (<18 years) and tube feeding, each ONS episode was matched to its nearest non-ONS episode neighbor.Instrumental Variables Analysis.
Instrumental variable analysis was used to specifically address potential bias due to nonrandomized treatment selection, which could not be addressed with propensity matching alone. Instrumental variables can remove the effect of selection bias and identify the causal effect of a treatment on outcomes.30-32
Using this method requires an instrument that correlates with the treatment of interest but does not affect the outcome, except through its influence on the likelihood of receiving treatment.
For this analysis, the selected instrument was the fraction of episodes involving any ONS use in a given hospital in a given quarter. By looking at changes in ONS use based on a hospital’s inclination to prescribe it, rather than underlying patient characteristics, the unbiased identification of the effect of ONS was made feasible. Since instrumental variable properties are best understood in linear settings,33
this instrument was applied to linear models of the 3 outcomes. Several tests of the instrument’s validity were performed.
To control further for unobserved patient heterogeneity, the model included fixed effects for groups based on how long patient data remained observable prior to loss to follow-up. These fixed-effects “follow-up” groups were no patient follow- up data; 1 day through 1 year of follow-up; 1 to 2 years of follow-up; 2 to 3 years of follow-up; and more than 3 years of follow-up. Because life expectancy cannot generally be observed in the Premier database (except when individuals die in a Premier network hospital), follow-up duration served as a proxy for underlying health status. Observed follow-up using hospital-based data may be a preferable measure of overall patient frailty, because the diagnostic codes present in a single episode (vs multiple follow-up episodes) are unlikely to reflect the full range of patient comorbidities.Additional Modeling of Readmissions.
For the readmission outcome, the matched sample was restricted to episodes where the patient was known to be at risk of readmission following discharge. If patients did not die following hospitalization, it could be assumed ONS had 2 potential benefits: it prevented readmissions by making people healthy, or it delayed readmissions among those eventually readmitted. Because the Premier data did not distinguish between patients not readmitted due to recovery and those not readmitted due to death, the current study could only measure the effect of delayed readmission (by calculating the change in 30-day readmission probability among patients eventually readmitted). This approach also provided a conservative estimate of the total impact of ONS on readmission.Return on Investment Calculations.
Next, estimates for the effect of ONS on LOS, episode cost, and readmission probability were used to calculate a return on investment (ROI) for ONS use, using the following formula:
ROI = savings generated through ONS use – amount spent on ONS
amount spent on ONS
The above formula yields the “episode cost” ROI of ONS use through reduced episode cost. Savings generated from ONS use were defined as the average reduction in episode cost due to ONS use. The amount spent on ONS was the average episode cost of ONS use. For readmission ROI, the average episode cost among the readmitted population was multiplied by the reduction in the probability of readmission to calculate the savings generated through ONS use. The amount spent on ONS was defined as for the episode cost ROI.
PDF is available on the last page.