Measuring Medical Care Prices

In document Three essays in health economics (Page 78-85)

Disease Treatment Prices

3.1 Measuring Medical Care Prices

3.1.1 Difficulties in Price Measurement

That medical care should be treated as an input into health has been known for some time. Thirty years ago, Michael Grossman (1972) introduced his seminal theoretical

model that focused on medical care as the key health production input. Forty years ago, Anne Scitovsky (1964) had argued for the creation of national disease accounts.

Despite having the conceptual framework for relating medical care to health for some time, success at empirically incorporating either of these ideas into national accounts has been limited. Currently, none of the national accounts relate medical care inputs to health outputs. However, newly available data and renewed interest has led to recent progress in the methods used to implement this conceptual framework.

The difficulties in applying these concepts are fundamental. The empirical relationship between medical care services and health outcomes is complicated by the difficulty in measuring initial health, health outcomes, medical care input prices, and the productive relationship between medical care and health outcomes. Trans-action prices for medical care are difficult to collect and may differ dramatically from listed charges because of insurance negotiation, bad debt, and the provision of free or discounted care. Medical care inputs and health outputs are both extremely het-erogeneous and accounting for this heterogeneity can be difficult. Moreover, even if health outcomes and medical care input prices could be perfectly measured, a productivity analysis requires that inputs are associated with outputs, which in this case relates medical care services to health outcomes. However, none of the govern-ment statistics such as the National Health Accounts, the Producer’s Price Index (PPI) and the Medical Care component of the Consumer Price Index (mCPI) relate the input services they consider to the health outputs they produce. Consequently, these statistics cannot enable even a cursory productivity evaluation.

The major innovation of the last decade has been the method in which to empirically implement these concepts. The literature is fast coming to the consensus that a disease-specific empirical implementation of medical care productivity evalua-tion is the proper approach. For example, Cutler et al (1996) examines heart attack treatments, Berndt et. al (1997) examines depression treatment, Evans (2005)

ex-amines AIDS treatment and other studies have examined psychoses (Duggan, 2005) and automobile injury treatments (Doyle, 2005). This cursory survey of the liter-ature suggests that any future empirical assessment of medical care productivity likely begins by attributing medical care services to the specific diseases they are used to treat. An examination of the productivity of medical care as a whole should account for the spending by all insurance types including the privately insured, the publicly insured, and the uninsured, and should also include the spending from all regions of the economy. The direction of this research suggests that steps should be taken to organize the construction of ”disease accounts” from a nationally rep-resentative dataset in order to attribute the medical care spending in the United States to disease treatments, thereby enabling future research related to the topic of medical care productivity.

The disease specific nature of medical care productivity arises because of the heterogeneity of disease characteristics. Diseases vary dramatically in their dura-tion, treatment protocols and health consequences. For instance, diabetes treatment is a very complex chronic condition that may involve a lifetime of hospital visits, insulin shots, and other drug regimens. In contrast, upper respiratory infections are acute conditions that can be very simple to treat and may involve a trip to the doc-tor’s office and a week’s worth of antibiotics. The health response to treatment can also vary dramatically across diseases. Diseases may be preventable with vaccine, curable, treated indefinitely but never cured, or have no available effective treat-ment. The differences in these outcomes may change over time with technological advancement. Related to the potential outcomes of the treatment are the health consequences of the disease itself. Some diseases may increase mortality risks but are not physically debilitating, such as high cholesterol. Other conditions may be physically debilitating but do not change the risk of death, such as arthritis. Other conditions, such as diabetes, can be both debilitating and mortal.

Although attributing observed health outcomes to medical care productivity requires the specification of a structural relationship between medical care services and health outcomes, the ”disease accounts”, presented here, provide a cursory pro-ductivity assessment by facilitating a ”back of the envelope” calculation of regional health outcome price levels and their changes over time in the aggregate.

3.1.2 Methodology of Treatment Price Measurement

Disease treatments are defined as the medical care service bundles consumed by in-dividuals for the purpose of treating disease type d. Disease d is determined by the three-digit ICD-9 code associated with the medical care event. Disease treatment prices are defined as regional price levels over the period 1996-2003 using the MEPS.

The medical care services used in disease treatment are referred to as Entry-Level-Items (ELIs). Input service type heterogeneity is controlled using a variant of the Country Product Dummy variable (CPD) model described in detail by Summers (1973). The CPD model is a hedonic pricing model that has been used extensively to estimate cross-country purchasing power parities. This paper follows Kokoski, Cardiff and Moulton (1994) in implementing a CPD model to construct price lev-els across geographic regions in the United States for medical care services. The services are organized into a hierarchy of two aggregation levels, where the most aggregated service level is the disease treatment. As an example, the treatment of depression is a commodity aggregate of the pharmaceutical products Fluoxetine HCl and Paroxetine HCl, Psychiatrist office visits, Psychologist office visits, and other ELI-level services used in the treatment of ICD-9 code 311 Depressive Disorders.

Implementation proceeds as follows. The price of an event i that belongs to ELI v is estimated using the following hedonic regression:

ln Piv = α0+

In equation 1, βkv and γjv are estimated coefficients, Piv is the total expenditures paid for the event i, and Rki is a region-year dummy variable that is one if the event occurred in region-year k, and zero if it did not. Cji is the jth characteristic of event i. The characteristics included in the hedonic regressions vary by ELI but often include disease and insurance type indicators as well as demographic variables such as age and gender. The price of the vth service in region-year k relative to the reference area is exp( ˆβkv).

In order to construct disease price levels for region-year k, ln Pkd, the ag-gregation procedure employs a bilateral Tornquist comparison of ELI items. Using the following procedure, the Tornquist bilateral price comparison of region-year j to region-year k for ELI v uses the following weighted geometric average of the V ELI-level prices that are used to treat condition d.

ln Pjkd = (1/2)

V

X

v∈d

(wdvj+ wvkd ) ln(Pjv/Pkv) (3.2)

In Equation 2, wvjd and wvkd represents ELI v’s fraction of total disease d expenditure in regions j and k, respectively. If there are K region-years, then this procedure results in a KxK matrix of bilateral prices. These bilateral prices are used to con-struct a multi-lateral price index for each disease-level region year using the Elteto, Koves, and Szulc (EKS) method described by Dreschler (1973), and implemented in Caves, Christensen, and Diewert (1982). The multi-lateral price for a region-year, ln Pkd, is determined by taking the weighted share of the K bilateral prices such that:

where sdj is the share of spending in region-year j on total disease d spending.

This procedures is conducted for twelve of the most costly disease types found in a representative sample of the United States population.

3.1.3 Observed Care versus Defined Care

The method used to define the bundle of services used to treat a disease implicitly defines the treatment protocol using the observed service shares. However, the observed services may be endogenous, because the bundle may depend on the relative prices of the services in the bundle. For example, a physician may prescribe an unnecessary treatment if the relative price of that treatment is high and the patient is insured against the full cost of the procedure.

In order to control for this issue one could use pre-defined protocols as defined by medical science to define the relative quantities of care used in the determina-tion of a treatment. However, many legitimate and highly substitutable protocol options are available for the treatment of a single disease, and the protocol used in practice varies for a multitude of reasons usually not observed in the MEPS data.

For instance, the treatment of depression may involve combining drugs with visits to psychiatrists and psychologists. The observed combination of drugs and visits depends on whether the patient responds better to the physician sessions or drugs, and potentially the aversion of the patient to taking drugs. The dispersal of med-ical knowledge also plays an important role in determining the protocol for many conditions.

Many of these unobserved changes are orthogonal to price, and we would like to incorporate them in our price index. We proceed with the assumption that these aspects of treatment are more important components of treatment than are the potential ”churning” and ”moral hazard” aspects of care introduced using the strict-protocol method.

3.1.4 Methodology of Health Outcome Measurement

There are two types of health outcome metrics used to measure output. The health measures are the age-adjusted mortality rate and the probability that an individual

has a disease-specific negative health outcome. Age-adjustment mortality rates for disease type j are calculated using the direct method which is described with the following equation:

Mj =

n

X

i=1

rji· (pi/P ) (3.4)

The following list defines notation:

• Mj is the age-adjusted mortality rate of disease j per 100,000 people.

• rji is the mortality rate of disease j per 100,000 people in age-group i.

• pi is the fraction of the population in age-group i.

• n is the total number of age-groups

• P =Pni=1pi represent the total population, which is the sum of all age groups i = 1, ..., n.

For the mortality statistics reported here, the age distribution of the popu-lation used is that of the 2000 census. Eleven age-groups are used for the mortality calculations that include less than 1, 1-4, 5-14, 15-24, 25-34, 35-44, 45-54, 55-64, 65-74, 75-84, and 85 and over.

The probability that an individual has a disease-specific negative health out-come is calculated for changes in observed health indicators. For instance, the observed indicator may be an indicator for whether mobility decreased during the period. The observed health indicators, I, are modelled as being directly related to a latent health variable, h. The relationship between latent health and the health indicator may either be increasing or decreasing in health. To fix ideas consider indicators that are decreasing in health. If the indicator variable takes on only two values, such as whether health deteriorated, then the indicator function is defined as follows:

• if h< h then I=1

• if h>= h then I=0

This binomial function describes the relationship between latent health and health outcomes. Latent health is modelled as a function of demographic character-istics such as age, sex, income, region, insurance type, race, and most importantly, year. In order to be consistent with the other measures of health, latent health depends on age as measured by eleven dummy variables representing the age-groups listed above. The probability that a health outcome occurs depends on patient characteristics, and the total expenditure spent on treatment. The probability that an observed health outcome occurs for an indicator function is defined using the following equation:

In this equation, demographic characteristics denoted by the variable Z and time de-noted by Y determine health. This equation depends on eight year dummy variables, Y , and l demographic variables, Z, that include the eleven age dummy variables, and six insurance type dummy variables. If υ = (h − ) is distributed standard nor-mal then Φ(·) is the cumulative nornor-mal distribution function, and this relationship can be estimated as a probit using standard techniques. The predicted probability of the health outcome for a specific individual evaluated at different years is the health measure used in the analysis.

In document Three essays in health economics (Page 78-85)