This Article Abstract Reprint (PDF) Submit a response to this article Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Services E-mail this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Add to My Personal Archive Download to Citation Manager Reprints & Permissions Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Weissert, W. Articles by Hirth, R. Search for Related Content PubMed PubMed Citation Articles by Weissert, W. Articles by Hirth, R. Related Collections Home Care

Health Tracking

MARKETWATCH

Beyond Managed Long-Term Care: Paying For Home Care Based On Risk Of Adverse Outcomes

Abstract

 
Evaluations of home care for chronically ill elderly people have shown disappointing results for many years. Improvements in outcomes have been slight and costs high. We offer a system for setting budget targets based upon effectiveness of home care in mitigating certain adverse outcomes, the risk of those outcomes occurring, and the economic value of avoiding those outcomes. We believe that such a budgeting system will encourage improved measurement of outcomes and more rigorous justification for expenditures. Moreover, such a system is designed to reallocate resources to higher-risk patients and those more likely to benefit, focusing caregiving on specific outcomes and improving those outcomes.

Several factors may contribute to these disappointing results. First, eligibility criteria are imprecise, and incentives to limit services to those most likely to benefit are often lacking. Second, there is little knowledge of effectiveness in home care practice, which results in nonstandardized practices. Third, budgets are often set so high that home care cannot be cost-effective, let alone cost-saving. Finally, expectations and evaluation measures for home care benefits may have been too heavily directed at traditional outcomes, ignoring subtle condition-specific improvements that home care may be more capable of producing, especially those related to the quality of life of the informal caregiver.

A home care budgeting model. In this paper we offer a conceptual framework for home care budgeting that we believe can address several of these problems. The essence of our model is to set a budget constraint for each patient based on three factors: risk of adverse outcomes (including failure to achieve potential benefit), the likelihood that home care can reduce the risk, and the value of avoiding the outcome. We call this approach "titrating" care—adjusting home care to the precise amount that will produce marginal returns equal to marginal investment. Titrating replaces the notion of targeting, which emphasizes whom to serve but gives little clue as to how much to serve. Titrating emphasizes adjusting care initially and then readjusting it periodically in response to additional information and response to care.

Our model is designed to be applied to patients with chronic health problems as opposed to those using home care briefly to recover from acute episodes of illness. In the case of acute illnesses, home care can be tailored to disease-specific needs, and outcomes and effectiveness can be measured more easily. In many disease areas such work is ongoing. However, many elderly persons suffer from chronic ailments with a need for home care that is not related to a short-term medical condition. For this population, appropriate outcome measures have not been fully defined, and the commonly accepted resource-allocation approach lacks an outcome-oriented focus.

Risk Estimation And Spending Data

Top Risk Estimation And Spending... Data Source Relationship Between Risks And... Conceptual Approach To Resource... Setting The Budget Advantages Of The New... NOTES

 
We begin by assessing the relationship between current home care spending and risk of adverse outcomes. We focus on four common, although admittedly incomplete and aggregate, outcome measures: nursing home admission; hospital use; functional decline; and death. This exercise, although difficult, is similar to the type of valuations inherent in cost effectiveness analysis for other medical services. In any cost-effectiveness decision, those making the choices must enumerate the outcomes of interest, measure the impact of medical care on those outcomes, and then decide if achieving those outcomes is worth the cost.

The restriction of our analysis to four broad outcome measures reflects data limitations, which in turn reflect the common lack of incentives to specify and measure home care outcomes that motivated much of this work. Consequently, while we think that an important benefit of our approach is to broaden the range of home care benefit expectations beyond traditional measures of morbidity and mortality, we are forced to use those traditional measures to outline our model here.²

Data Source

Top Risk Estimation And Spending... Data Source Relationship Between Risks And... Conceptual Approach To Resource... Setting The Budget Advantages Of The New... NOTES

 
Data to estimate risks and compare them to spending came from the Arizona Long-Term Care System (ALTCS). ALTCS is the long-term care portion of Arizona’s Medicaid-alternative capitated acute and long-term care program, now in its seventeenth year of operation. We chose this program for several reasons: (1) Its capitated payment system makes it the state of the art, and we wanted to demonstrate that shortcomings of home care practice exist even in this demanding context; (2) ALTCS has been in operation for many years (since 1989), so its current practice does not reflect start-up issues; and (3)ALTCS generously permitted us to use its data and to run a small pilot test of our model in its program.

The data set consists of 52,246 assessments of 26,754 unique clients served by the program from 1992 through 1997. Patient descriptors and measurement of change in functional status came from nurses’ comprehensive assessments. These assessments were conducted by state employees to determine program eligibility initially and annually for most patients, more frequently when there was a major change in health status, or every other year for a few patients whose health status changed little from year to year.

Risk data. Incidence of the four adverse outcomes (death, hospitalization, nursing home admission, and functional decline) in this population came from encounter data files provided by each managed care plan contractor to the state’s program administrators who pay the plan contractor their capitated monthly payments. They were as follows: death, 18.9 percent; hospitalization, 18.1 percent; nursing home admission, 31.5 percent; and functional decline, 52.1 percent. Functional decline was measured among persons who had a second assessment eleven to thirteen months after the first assessment and a function status measure that aggregates activities of daily living (ADLs) and instrumental activities of daily living (IADLs). The data have been extensively tested in previous evaluation studies for reliability and usefulness in producing risk models that closely track the aggregate actual annual occurrence of events being predicted.³

An extensive effort was undertaken as part of this study to profile risks of the four adverse outcomes in the population. Our statistical model was a discrete time-hazard model that allowed for time-varying covariates. Unconditional models were estimated for each risk, but our approach was robust to correlations among the adverse outcomes. ⁴ From these models we computed a predicted risk for each patient, for each of our four outcomes, for each month in which a client was receiving home care at the beginning of the month.⁵

Major covariates of risk of adverse outcome identified in the models were found to be physical limitations, age, some diseases, and recent discharge from a hospital or nursing home. Our findings were generally consistent with the predominant findings in the existing literature (which itself is equivocal). We are less concerned with coefficients on specific covariates than with the predicted risks for groups of home care clients. Tests of model fit using split-sample techniques and comparisons of actual to predicted outcomes for subsamples of the population indicated that the models fit very well for each outcome. Data limitations necessitated making our estimates on a treated population, however, which is less than ideal, and required adjustments to compensate for possible treatment effects.⁶

Spending data We obtained spending data for all Medicaid-covered services from the claims and encounter files of ALTCS. Although the system is capitated to the managed care plan, individual providers still file claims or encounter reports for services rendered to individual clients. Claims or encounter reports were aggregated to get monthly spending for each home care patient. Our comparison of home care spending and predicted risk is based on the subset of patient months in which no adverse event occurred (because such an event would often censor home care spending).⁷ The spending variable captures the actual consumption of home care services for the clients studied during the estimation period.

Relationship Between Risks And Home Care Spending

Top Risk Estimation And Spending... Data Source Relationship Between Risks And... Conceptual Approach To Resource... Setting The Budget Advantages Of The New... NOTES

 
Our findings confirm the expectation of little relationship between our four specific risks and home care spending. That is, risks and spending were not significantly correlated, showing correlations of only 0.0567 for hospitalization; –0.1708 for nursing home admission; –0.0324 for death; and –0.0454 for physical functioning decline. The latter was measured on a thirty-seven-point scale combining ADLs and IADLs. A decline represented a drop of three or more points. These very low correlations suggest that the level of risk faced by a patient played little role in how much home care the patient used.

Moreover, service costs vary widely among patients facing similar risks. The standard deviation of spending within each of ten risk deciles approximated the mean, suggesting coefficients of variation of about 1. These results suggest that even in this well-designed, cost-conscious program, spending is not well targeted toward risks.

Conceptual Approach To Resource Allocation

Top Risk Estimation And Spending... Data Source Relationship Between Risks And... Conceptual Approach To Resource... Setting The Budget Advantages Of The New... NOTES

 
We believe that it is possible to design a payment method that encourages home care case managers to adjust the home care prescription for patient risks, the value of mitigating those risks, and the effectiveness of home care in mitigating those risks, even though current knowledge is limited for some of these key elements. Conceptually, the payment method would ensure that each patient’s budget target is based upon his or her risk of adverse outcomes (where adverse outcome is defined to include failure to achieve benefits that might have been achieved with appropriate treatment). Budgets would be calculated for each outcome and summed across all risks. These budgets, set for a month of care, could be used as a target by capitated programs, permitting case managers to vary above and below them for individual patients, but staying within a budget equal to the sum of all individuals’ monthly budgets.

Our paradigm is based on the idea that the expected benefits (pecuniary and other) from home care should equal or exceed its costs.⁸ We conceptualize the benefit of home care to be the reduction in the risk (and associated costs) of various adverse outcomes (including failure to maintain or improve current levels of satisfaction, functioning, and so forth). Total benefit will be the aggregation of the benefit associated with risk reduction across all adverse outcomes. An analogous approach could be taken for positive outcomes that home care might achieve.

Without home care, the expected cost associated with each adverse outcome equals the risk of suffering the adverse event without home care, R, multiplied by the value (pecuniary and other) associated with avoiding the outcome, V. The benefit of home care is conceptualized to be the reduction in this pre– home care expected cost attributable to home care. This will depend on the effectiveness of home care, E, defined as the percentage reduction in the risk of the adverse outcome attributable to home care. Effectiveness is a function of home care spending. However, because the literature lacks evidence on the relationship between episodes of home care and outcomes, we assume a constant effectiveness of home care, independent of spending. Moreover, effectiveness is also likely to be a function of patient characteristics. However, without convincing evidence, we assume for expository purposes that all patients would experience the same relative risk reduction from home care. Thus, patients at a higher absolute risk experience a greater absolute risk reduction in our model. More advanced models would recognize the possible variation in relative risk reduction across patients, but no data are currently available to identify such variation in effectiveness. In theory, with appropriate information, the marginal benefit of home care spending should be equated to its marginal cost.

The expected cost associated with the risk of adverse outcomes after home care is thus the product of the level of risk that remains under home care (1–E) and the pre–home care expected cost of a given adverse outcome (R x V). Subtracting the post–home care expected cost from the pre–home care expected cost yields the expected benefit of home care associated with mitigating the risk of any given adverse outcome, which we term the effectiveness and risk-weighted value (ERV).

For example, if (1) a particular patient has a 25 percent monthly chance of hospitalization, (2) home care has the potential to lower that risk by 20 percent, and (3) the value of avoiding a hospitalization is $10,000, then the monthly ERV of hospitalization is $500 (500 = .20 x .25 x 10,000).

Once an ERV has been estimated for each patient for each identified adverse outcome, the total expected benefit of home care for that patient is calculated by summing the ERVs for the patient across all of his or her risks. This allows the benefit to reflect the value of home care in mitigating the risk of each outcome. The conceptual framework would be the same even if the outcome were maintenance of ADL function, skin condition, hygiene, or the informal caregiver’s feelings of reduced anxiety over her ability to endure in her caregiving role.

A major benefit of this approach is the information it conveys to the case manager: the patient’s risks and how they compare to each other and to the risks faced by other patients. The approach also forces providers to think about the likelihood that home care can mitigate specific risks and the value of avoiding adverse outcomes.

Setting The Budget

Top Risk Estimation And Spending... Data Source Relationship Between Risks And... Conceptual Approach To Resource... Setting The Budget Advantages Of The New... NOTES

 
For expository purposes, we estimated the ERV for each patient in our ALTCS sample by entering values into each term of the ERV formula for all four of the outcomes mentioned. Using the estimates of effectiveness, discussed below, we adjusted our risk estimates to represent pretreatment risk (Exhibit 1). The value and effectiveness estimates are essentially ways of converting the risks of different adverse outcomes into a single dollar value. The ERV represents a monthly budget target for each patient. Given the relatively poor performance of home care in the literature and lack of dose-response information, we made optimistic assumptions based on existing literature about home care effectiveness and assumed effectiveness to be insensitive to the level of home care spending. We believe that these estimates are at the upper bound of what the literature could justify.⁹

Estimation of the value of avoiding each outcome was also based on existing literature. Because our conceptualization of value includes nonpecuniary benefits, much of our analysis was based on the value-of-life, cost-effectiveness, and quality-of-life literature. Adapting the results from this literature to our purposes required us to make a wide variety of assumptions.

To value death and functional decline, we used life tables to estimate life expectancy, the literature on cost-effectiveness to derive a range of accepted values per quality-adjusted life year (QALY) gained, and the quality-of-life literature to estimate quality of life in various functional states. To value nursing home placement, we estimated the additional cost relative to home care and the average length of- stay. We valued hospitalization using average Medicare-allowable charges for the institutional and physician payments per admission. Additionally, for hospitalization and functional decline, the "value" was adjusted to reflect the impact of these events on future probabilities of death, nursing home admission, further functional decline, or additional hospitalization. The adjustment was based on the coefficients for recent hospitalization and recent functional decline in the risk models.

Given the number of assumptions that were necessary, the values we used should not be viewed as "hard" numbers. Again, these assumptions were necessary because the field has not devoted much attention to valuing the various outcomes from home care. Yet we are confident that it could, and if forced to by our budgeting approach, it would. Essentially, values must be set according to the willingness of public and private payment systems to pay for a wide range of home care outcomes. Public debate is required. Here we make assumptions of value based upon reasonable methods and available data.

We related those ERV budgets to actual spending in Arizona.¹⁰ Using what we believe are optimistic assumptions about home care’s effectiveness at mitigating risks, on average the actual spending was 2 percent above the dollar value that our equation suggests would be appropriate. However, the relationship between ERV budgets and spending was weak, with a correlation actually estimated to be slightly negative (–0.04), although not statistically different from zero.

The distribution of spending illustrates home care’s lack of response to the client’s potential to benefit in terms of risk reduction, at least for the specific risks assessed. We split the sample into deciles by ERV. In each decile we reported the mean ratio of actual home care spending to ERV. A ratio of 1 would indicate that spending approximated potential benefit. Spending equaled potential benefit at about the fifth decile. For home care clients with higher ERVs (more potential benefit), spending fell below ERV. For example, at the ninth decile of risk, clients received 36 percent of ERV. Clients at lower risk received more care than ERV could justify. Specifically, at the second decile, clients received about 50 percent more home care than ERV would suggest.

Less optimistic assumptions about effectiveness or value would change the assessment of the level of spending, but the pattern of spending across ERV groups would remain the same. For example, less generous assumptions suggest that home care spending might exceed potential benefit by 60 percent, with potential benefit exceeding spending only in the ninth and tenth deciles of ERV. Regardless of the assumptions, the basic pattern suggesting that high-risk clients do not receive enough care and low-risk clients receive too much care persists.

Advantages Of The New Approach

Top Risk Estimation And Spending... Data Source Relationship Between Risks And... Conceptual Approach To Resource... Setting The Budget Advantages Of The New... NOTES

 
While a randomized controlled trial would be required to show that outcomes were better or at least no worse with the ERV approach than with current practices, we believe that the ERV approach is promising because it is more systematic. It represents care standards where few or none now exist. Based as it is on model-derived risk estimates, it contributes to overcoming the shared uncertainty and the information asymmetry between the patient and the case manager regarding the actual risk of each adverse outcome faced by the patient. Case managers can better deal with heterogeneity of the home care population by at least loosely separating them into patients at high or low risk of suffering key adverse events. Basing payment on risk gives case managers an objective function for home care for each patient (reduction of their particular risk profile), as well as data for program evaluation (avoidance of adverse outcomes). Finally, with proper information on dose-response relationships, which the system would encourage gathering, the ERV approach would focus attention on marginal cost and its relationship to marginal benefit.

Sample cases. Exhibit 2 illustrates how the ERV budget targets would work for three actual patients of varying risks. It presents each patient’s monthly risk of each of four adverse outcomes, where that risk ranks that patient in percentage terms in relation to all other patients in the sample, and a suggested monthly budget associated with each risk and totaled across risks. The cases illustrate how target budgets based on effectiveness of home care, risk of adverse outcomes, and value of the outcome cause the budgets to vary. Case 1 is an individual whose risks are low for all four of the risks for which funds are budgeted in the example. The monthly budget target for this case is low—$149—although in actual practice additional funds targeted to additional risks not enumerated in this study might be added.

Finally, Case 3 is at high risk of all but functional decline, especially high for both death and nursing home admission. Her monthly budget is $3,419, among the highest produced by these four risks.

Greater flexibility. We envision that the budgets are targets only and may be adjusted up or down by the case manager; in actual practice, additional risks would be valued and estimated. But case managers should be encouraged by the totality of the information provided to focus their attention on specific risks faced by a given patient, and they should be encouraged to keep total spending within the cumulative budget for all patients. Ideally, they would design care plans intended to mitigate those risks, and they would begin trying to judge the expected payoff in risk reduction from specific kinds and amounts of care. Program managers and payers would be better equipped to evaluate care plans in relation to risks being mitigated, and program managers and policymakers would be able to measure improvement over time and among programs in risk reduction.

Better use of funds. As we considered how to implement a controlled trial of the new payment method, another weakness of current practice and potential strength of the new method became obvious. Current practice may not be well equipped to use the extra funds that would be transferred by the proposed system from low- to high-risk enrollees. That is, current payment methods limit the involvement of physicians and other highly skilled diagnosticians and treatment specialists in care planning for home care because average payment rates simply are not high enough. Under the system proposed here, diagnostic capabilities are an essential skill.

Can avoid unnecessary hospitalization. William Weissert and colleagues showed that nearly a third of hospitalizations from home care were among patients probably admitted to the hospital to die there, to be transferred to a nursing home, or merely to be evaluated.¹¹ Many of these actions could have been undertaken in the home, without hospitalization. Family counseling; better communication protocols between the aide, nurse, and physician; more training; better assessment tools; and better monitoring or other interventions might have avoided some hospitalizations. But judging the need for such interventions exceeds the skill and training of most of the personnel now most heavily involved in home care planning and delivery. If funds shifted by our methodology from low-risk to high-risk patients were made available for diagnosing and treating high-risk patients, funds to pay for physician evaluation and more involvement in care planning and monitoring might avoid some hospitalizations.

Home care interventions might include signing do-not-resuscitate orders, increasing pain medications, moving directly to a nursing home and bypassing the hospital, or simply accepting day-to-day variation in health status. Home care staff can also offer comfort, reassurance, observation, reporting, and personal assistance. They can reassure patient and caregiver that they are linked to potential services, advocate for them in a system that may not automatically respond to their unmet needs, and support the patient’s and family’s preferences in care choices. None of these contributions may alter the health status trajectory, but they may alter its consequences.

Care planning appears to be hampered by so-called targeting: selecting patients for home care but giving little guidance regarding how much care they should receive. Supervision and program performance could be enhanced if care were titrated to specific risks faced by a given patient, his or her potential to benefit, and the value of those benefits. Additional skills, especially physicians’ diagnostic skills, could be better deployed into the home care setting, and funds could be freed up from low-risk patients to support physician involvement in the care of high-risk patients. Finally, research into the relationship between doses of home care and outcomes would be strongly encouraged by the payment method proposed here, and home care could become a venue for improving outcomes and reducing costs rather than a problem for policymakers.

Editor's Notes

 
William Weissert is professor and chair of the Department of Health Management and Policy, University of Michigan School of Public Health. Michael Chernew and Richard Hirth are associate professors there.

This project was generously supported by a grant from the Robert Wood Johnson Foundation’s Home Care Research Initiative administered by the Center for Home Care Research, Visiting Nurse Service of New York.

NOTES

Top Risk Estimation And Spending... Data Source Relationship Between Risks And... Conceptual Approach To Resource... Setting The Budget Advantages Of The New... NOTES

 

1. S.L. Hughes et al., "Impact of Home Care on Hospital Days: A Meta Analysis," Health Services Research 32, no. 4 (1997): 415–432; W.G. Weissert and S.C. Hedrick, "Lessons Learned from Research on Effects of Community-Based Long-Term Care," Journal of the American Geriatrics Society 42, no. 3 (1994): 348–353; W.G. Weissert, C.M. Cready, and J.E. Pawelak, "The Past and Future of Home and Community Based Long-Term Care," Milbank Quarterly 66, no. 2 (1988): 309–388; S.C. Hedrick, T.D. Koepsell, and T. Inui, "Meta-Analysis of Home-Care Effects on Mortality and Nursing-Home Placement," Medical Care (November 1989): 1015–1026; and P. Kemper, R. Applebaum, and M. Harrigan, "Community Care Demonstrations: What Have We Learned?" Health Care Financing Review (Summer 1987): 87–100.
2. We would like to have examined a richer set of outcomes, more specific components of health status and condition-specific outcomes, and a variety of intermediate outcomes; data limitations prevented this. As use of the Outcome and Assessment Information Set (OASIS), recently mandated for collection of data on Medicare home health patients, becomes more widespread, expanding our approach would become more feasible. "Medicare and Medicaid Programs: Comprehensive Assessment and Use of the OASIS as Part of the Conditions of Participation for Home Health Agencies," Federal Register 64, no. 15 (25 January 1999): 3764–3784.
3. N. McCall et al., "Managed Medicaid Cost Savings: The Arizona Experience," Health Affairs (Spring II 1994): 234–245.
4. More details regarding risk estimation and results are available from the authors, <weissert{at}umich.edu>.
5. The hazard model we estimate is a logistic regression in which the unit of observation is a patient/month. The number of observations in the analysis data set for any given client is equal to the number of months that client was in the community. Such a model yields similar results as a standard duration model with time-varying covariates. R.B. D’Agostino et al., "Relation of Pooled Logistic Regression to Time Dependent Cox Regression Analysis: The Framingham Heart Study," Statistics in Medicine (December 1990): 1501–1515. However, the hazard model is much easier to work with and to compute predicted risks. The predicted risks were the fitted values for the logistic regression (exp (xß)/(1 + exp (xß)). The model fits the data very well. Adjustments were made to the standard errors to reflect multiple observations on the same individual. More detail is available from the authors.
6. All patients in our sample had received some home care. Thus, the risks that we estimate are posttreatment, not pretreatment. As our model illustrates, the appropriate risks to measure are pretreatment risks. Throughout the analysis we adjust the estimated risk by our effectiveness assumptions to approximate pretreatment risks. The extent to which the relationship between the risk we present and spending varies from the relationship between true pretreatment risk and spending depends on the extent to which effectiveness varies within risk strata.
7. Ideally, we would relate risk to provision of specific services. But given the large numbers of potential services, this is beyond the scope of this paper.
8. We use the term "expected benefit" to denote average benefit over a large group of patients. Because outcomes are stochastic, benefits will not equal costs in every case.
9. Effectiveness calculations assume independence in effectiveness over time. In our model, increases in effectiveness shift the cumulative incidence function to the right, effectively delaying the expected time a patient will suffer any given outcome.
10. Our sample was all patient months in which hospitalization, death, or nursing home admission did not occur. This is because actual home care spending in months with one of these events could be lower than average simply because the patients were not in home care for the entire month.
11. W.G. Weissert et al., "Toward a Strategy for Reducing Potentially Avoidable Hospital Admissions among Home Care Clients," Medical Care Research and Review (December 1997): 439–455.

                      What's this?

This article has been cited by other articles:

				N. J. Holland and D. E. Northrop Young Adults With Multiple Sclerosis: Management in the Home Home Health Care Management Practice, April 1, 2006; 18(3): 186 - 195. [Abstract] [PDF]

				V. L. Greene Prospective Budgeting for Home Care: Making Titration Work J Aging Health, August 1, 2005; 17(4): 399 - 424. [Abstract] [PDF]

				T. H. Wagner and S. Chen An Economic Evaluation of Inpatient Residential Treatment Programs in the Department of Veterans Affairs Med Care Res Rev, April 1, 2005; 62(2): 187 - 204. [Abstract] [PDF]

				W. G. Weissert, R. A. Hirth, M. E. Chernew, S. Diwan, and J. Kim Case Management: Effects of Improved Risk and Value Information Gerontologist, December 1, 2003; 43(6): 797 - 805. [Abstract] [Full Text] [PDF]

				J. A. Capitman Guest Editorial: Verve not ERVs for Care Planners Gerontologist, December 1, 2003; 43(6): 806 - 807. [Full Text] [PDF]

				E. A. Miller and W. G. Weissert Strategies for Integrating Medicare and Medicaid: Design Features and Incentives Med Care Res Rev, June 1, 2003; 60(2): 123 - 157. [Abstract] [PDF]

				W. Weissert, M. Chernew, and R. Hirth Titrating Versus Targeting Home Care Services to Frail Elderly Clients:: An Application of Agency Theory and Cost-Benefit Analysis to Home Care Policy J Aging Health, February 1, 2003; 15(1): 99 - 123. [Abstract] [PDF]