| Methodology Report #6:
List Sample Design of the 1996 Medical Expenditure Panel Survey Insurance Component by John Paul Sommers, Ph.D., Agency for Health Care Policy and Research Select for more information on AbstractThe Medical Expenditure Panel Survey (MEPS) is the third in a series of nationally representative surveys of medical care use and expenditures sponsored by the Agency for Health Care Policy and Research (AHCPR). MEPS comprises four component surveys. The Insurance Component (IC) collects employment-related health insurance information, such as premiums and types of plans offered. Respondent characteristics–such as size of business, employee characteristics, and type of industry–also are collected. This report outlines the process used to allocate and select the MEPS IC list sample, including goals, development of allocation schemes, and selection methods. The list sample is collected from samples developed from three lists that, together, cover almost 100 percent of the employers in the United States. Select for information on The Medical Expenditure Panel Survey (MEPS).
The 1996 Medical Expenditure Panel Survey (MEPS) Insurance Component (IC) is a survey of employers, the self-employed with no employees (SENEs), unions, and insurance companies. The MEPS IC, which is sponsored by the Agency for Health Care Policy and Research (AHCPR) and conducted by the U.S. Bureau of the Census, is designed to collect employment-related health insurance information, such as premiums and types of plans offered. Respondent characteristics—such as size of business, employee characteristics, and type of industry—also are collected. Although the 1996 MEPS IC was first administered in 1997, data collected were for the entire calendar year 1996; the survey name refers to the data year, not the time of collection. The MEPS IC sample has two parts:
The IC household sample is defined by the sample design of the MEPS HC (Cohen, 1997) and has persons as sample units. Data are collected from the employers and other insurance providers of the household respondents from the MEPS HC. The employers and other providers are proxy respondents for supplemental information on health insurance offered to the household respondent through the employer or provider. Hence, the probabilities of selection and the corresponding weights for these employers are the same as those of the household sample members and come from the MEPS HC design. The data collected are linked to other information for household sample cases. The IC list sample is a random sample of establishments selected especially for the IC, and its selection is independent of the HC design. The two IC samples (household and list) are combined for collection purposes. The data collected for both samples are almost identical. The MEPS IC list sample provides a picture of State and national employer health insurance. Enrollments, premiums, employee contributions, types of plans offered, deductibles, coverages, and other characteristics are shown for the Nation and 40 individual States. Estimates for characteristics of employer-related health insurance for small establishments are of particular interest. The MEPS IC sample is a primary government source of information on employer-related health insurance in the United States. It should prove to be a key data provider for this crucial portion of health care. This report outlines the process used to allocate and select the MEPS IC list sample, including goals, development of allocation schemes, and selection methods. The list sample is so named because it is collected from samples developed from three lists that, together, cover almost 100 percent of the employers in the United States. These are as follows: The Standard Statistical Establishment List (SSEL) including agricultural units. The SSEL is a list of private-sector establishments that is developed and maintained by the Census Bureau. Although the Census Bureau obtains information on agricultural units from the same sources used to produce the SSEL, such units are normally not in scope for most surveys. However, agricultural employers are in scope for the MEPS IC. The SSEL is continually updated; the list used for sample selection was that of establishments in business as of the end of 1995. The 1992 Census of Governments (COG), a list containing the Federal, State, and local governments within the United States that is maintained by the Census Governments Division.
Data from the 1994 National Employer Health Insurance Survey (NEHIS) were used as a basis for developing the design of the MEPS IC list sample (Marker, Bryant, Wallace, et al., 1996). NEHIS collected similar data from the same universe, and it was assumed that the variance structure for the MEPS IC would follow that of NEHIS. Therefore, the NEHIS variance results, sample size, and standard sample variance formulas were used to assist in the development of the IC list sample design. The IC also used the same strata structure as NEHIS. This structure seemed reasonable, and time constraints did not allow for the in depth analyses required to produce another structure. For the private sector, the structure is determined by the strata of establishments defined by the State in which the establishment is located, the size of the establishment, and the size of the firm that owns the establishment. An establishment is an economic unit at a single location where business is conducted or services or industrial operations are performed. A firm is a business entity consisting of one or more business establishments under common ownership or control. Exact categories of firm and establishment size groups used for private-sector stratification within each State are displayed in Tables 1 and 2. For governments, the strata are defined by the State in which the government is located and employment size (number of employees). The goals for the establishment sample are as follows:
For practical purposes, because of the number of variables involved, two of each type of variable were chosen for analysis. Return To Top
Note: An establishment is an economic unit at a single location where business is conducted or services or industrial operations are performed. A firm is a business entity consisting of one or more business establishments under common ownership or control. Source: Center for Financing, Access, and Cost Trends, Agency for Health Care Policy and Research: Medical Expenditure Panel Survey Insurance Component, 1996. Return To Top
Note: Private-sector list sample = 600. An establishment is an economic unit at a single location where business is conducted or services or industrial operations are performed. A firm is a business entity consisting of one or more business establishments under common ownership or control. Source: Center for Financing, Access, and Cost Trends, Agency for Health Care Policy and Research: Medical Expenditure Panel Survey Insurance Component, 1996. The formulas below were applied relative to the variance of a population estimate for a variable of interest based on a stratified sample (Cochran, 1977). Assuming:
Because the main concern is with the relative quality of two types of allocations, a useful ratio is that of the variances of two different allocations:
If Mj and Mk are the sample sizes for two different allocations, such that for all cells,
then
and
and
Using these formulas, sample sizes for the 1994 NEHIS, RSEs from the 1994 NEHIS, and proposed sample sizes for the 1997 MEPS IC, comparisons of various new allocation schemes can be produced and estimates made of their RSEs. Return To TopAllocation to SENEs The first decision was to allocate a national sample of 1,000 to the population of SENEs and the remaining sample of 29,000 to the private-sector and government employers. This was done by determining the total sample needs for the SENE sample. A brief assessment showed that this sample size appeared to give an average RSE of approximately 5 percent for survey estimates for a set of important variables. Allocation Among States The next step required the allocation of the sample among the States. The equations above were used, and it was assumed that the variance structure is the same within each State. In equation (1), the strata are assumed to be States and the values of unit variances for each State are assumed to be equal. These assumptions require the within-stratum variances and the relative sizes of the strata within States to be equal. The assumptions seemed reasonable after examining a set of within-stratum variances for several variables and the relative sizes of the strata for several States of various sizes. Thus, to compare allocations, only values of equation (1) where the values of sigma all equal to 1 must be compared. For the 1994 NEHIS, the sample allocation to each State was done relative to the values of the .3 power of the number of establishments (Marker, Bryant, Wallace, et al., 1996.) This allocation method was used to give adequate sample size to smaller States by increasing their relative size compared with the larger States. Using this method, 29,000 remaining units were allocated to the States, and an RSE for this survey was projected using equations (3), (4), (5), and (6) along with the RSE for the 1994 NEHIS. However, the allocation failed to meet the stated goals for two reasons: The national goal of an RSE of 1 percent or less could not be met. This happened because the sample was much smaller than the sample of 38,000 for the 1994 survey. The smaller IC sample, combined with the loss of efficiency for national estimates caused by the oversampling of all small States, contributed to failure to meet the goal.
As a result, neither quality estimates for 40 States nor a quality national estimate could be made with the 1994 NEHIS sample design and the current smaller sample size. To remedy this problem, it was necessary to produce a minimum sample in 40 States and to generate a more efficient allocation for national estimates. Examination of equation (1) in the case of equal values of shows that the best allocation for national estimates allocates sample proportional to the size of Ni for each State. Thus, an allocation was needed that was closer to proportional than the 1994 NEHIS was. At the same time, the allocation required a minimum sample within at least 40 States. After a brief analysis, this number was set at 600. This number would not produce State estimates that would meet the quality goal of 5 percent or less RSE for estimates of total enrollment. However, the number was set at 600 for two reasons:
Allocation Within States Government and Private Sectors The next step was to divide each State’s allocation between the public and private sector. In the 1994 NEHIS, this was done in two steps:
Because of this allocation, the RSEs for many government statistics were less than those for the private sector, although the government sample was smaller than that of the private sector. This fact and an examination of equation (1) indicate that the unit variances and design effects for governments are smaller than those for the private sector. This was the reason the 1994 government sample was allocated as only 2/3 the proportion of government employment. However, this allocation to governments appears to be still too large, and it would be better to further lessen the government allocation. To do this, assume that in equation (1) there are two strata—governments and the private sector. Let the number of governments and private-sector establishments be Ng and Np. Likewise, let mg and mg be the means for a variable of interest for governments and private-sector establishments, and RSEg and RSEp be the relative standard errors for the estimates of the means of the variables for the two groups. Then the variance V of the estimate of the total over the entire population can be written as
For allocation between the government and private sector within a State, a fixed sample size is divided between the two strata of the population. Under these conditions, one can show that
This relationship can be developed using the Neyman optimal allocation of the sample when there are two strata (Cochran, 1977). For the NEHIS government allocation, all values within this equation are known. The size of RSEs for governments for many variables, the relationship of the 1994 NEHIS allocations to government and the private sector within States, and the population values all are available. When these known values were entered into equation (8), the right side of the equation for the 1994 allocation appeared to be too large. Therefore, the value of the RSE for governments must be raised, while the value for the private sector must be lowered. To do this, the allocation to governments within each State is decreased. Different reductions are needed to achieve an optimal design for each of the different variables. Thus, a conservative method was developed that would reduce the government allocation but would not be extreme. The method is as follows:
As an example, suppose there are two States, each with 50,000 total employees, half of whom are government employees. Assume the sample allocation is 120, or 60 per State. The government allocation is .5(2/3)120, or 40. Any government with more than 1,250 employees is selected with certainty. Assume each State has 6 governments with more than 1,250 employees and the total employment of the 6 governments in each State is 20,000. Remove the 12 governments and their employees. This leaves 60,000 total employees, of whom 10,000 are government employees. The remaining sample is 108, or 54 per State. Government gets (10,000)/(60,000)(2/3)108 = 12. Check this and find nothing large enough to be a certainty—that is, no government in the remainder larger than 10,000/12 = 867. Now allocate the remaining 54 sample units in each State. Because the remaining government in each State is 1/6th of the employment, government gets 2/3 of that proportion of the remaining sample, or 6. The final allocation in each State under this assumed scenario is 48 private sector, 6 certainty governments, and 6 noncertainty governments. If the 1994 NEHIS method had been used, the government allocation would have been 20 per State. This method considers the disproportionate amount represented by large local governments. It essentially applies optimal allocation of the sample to the remainder of the universe after these very large governments are removed. It produced a sample for governments within States that was approximately 80 percent of the size produced using the 1994 NEHIS methods. Table 3 shows the breakdown of the samples within States. Return To TopPrivate-Sector Sample In allocation of private-sector establishments within States, two types of variables are considered:
The two variables tend to call for two opposite types of allocation methods. With the first, allocation is proportional to the number of establishments in a stratum. With the second, allocation involves increasing the sample according to the number of employees in establishments within a stratum. The first allocates large percentages of the sample to strata with small establishments. The second allocates larger percentages of the sample to large establishments. (Strata definitions are given in Table 1.) The probability of allocation for the 1994 NEHIS sample was proportional to size, with the size measure for an establishment being the square root of the establishment’s employment size (Marker, Bryant, Wallace, et al., 1996). This was a compromise between the two extremes above. The method worked reasonably well and gave good error results. To determine the allocation for the MEPS IC, a compromise allocation was developed using unit variances estimated from 1994 NEHIS data. This analysis was performed to further improve State estimates to meet the State quality goals that could not be met using the 1994 NEHIS methodology. To perform the analyses, two different sets of unit variances were used, one to represent each type of variable. There are 18 possible strata, 6 establishment sizes crossed with 3 firm sizes. These are listed in Table 1. Note that only 14 strata have establishments in them because of restrictions placed on establishment sizes by the size of the total firm. Unit variances for a given establishment size were assumed to be equal for all firm sizes. Thus, there are 6 unit variances for each variable type. For a variable of the first type, which creates an allocation relative to total numbers of establishments, these values will be equal; for variables of the second type, these variances will increase as establishment sizes increase. The relative values of these variances are as follows:
After a preliminary analysis, six methods of allocation were attempted: 1. Optimal for variable type 1, or allocation proportional to the number of establishments in a stratum. 2. Optimal for variable type 2, or Neyman allocation (Cochran, 1977). 3. Allocation proportional to the total employment size in each stratum. 4. Neyman allocation using variances for each stratum, which are an average of the stratum variances for the two types of variables to be estimated. 5. Allocation equal to the average of the allocations determined by the two optimal allocation schemes presented above. 6. Allocation proportional to the sum of the square roots of the employment sizes of all establishments in the stratum. (This is the method from the 1994 NEHIS.) The sizes of the variances obtained from each method relative to method 1 are shown below. The values were calculated using equations (1) and (2).
Method 4 was selected for the MEPS IC because variable type 2 was considered to be slightly more important than type 1 in this survey. However, both methods 4 and 5 were good, and both were robust when tested under changes in unit variance. Table 2 gives the percent allocation to each stratum in the private sector for a typical State using this method. Final Allocations At this point in the process, all samples were of the desired sample size. Samples must be inflated to account for potential losses due to nonresponse and out-of- scope units. To develop these numbers for the private sector, losses were calculated by strata from previous Census Bureau SSELs for time periods from the time the frame was developed to the time of the survey. The loss rates are shown in Table 5. The general tendency is for rates to decrease as both the firm size and the establishment size increase. Losses also vary from year to year based on economic conditions. The assumption made was that conditions would be stable with no recession. No losses were assumed for governments because governments that close are small boards and commissions with few, if any, employees. The Census Bureau estimates that total losses in employment during a year would amount to much less than .5 percent of government employment. To avoid these types of governments that are of little interest analytically, two steps were taken: • All governments with less than one full-time equivalent employee were removed. This removed a large number of small boards and the like, but less than .01 percent of the total employment. • The sample was selected using employment as a measure of size. These two steps resulted in a sample in which there was little chance of selecting any governments that would be lost because of closure. Thus, no inflation was made for losses because of closure in the government sample. A 90-percent response rate was assumed for governments. A similar rate had been achieved in the 1994 NEHIS, and this was believed to be reasonable. There was no experience as a guide to estimating losses in the SENE group except a supposition that it would be a large number. A one-third loss rate because of closure was chosen for this group. Sample increases were also made to account for losses because of nonresponse. Tables 4 and 5 contain these values for private-sector strata, governments, and SENEs. The results for the private sector and government were based on results from the 1994 NEHIS, which were slightly lower than the rates assumed. The rates from NEHIS were increased slightly because of the smaller questionnaire being used for MEPS and because the Census Bureau was the collection agent. The rate chosen for SENEs was similar to that assumed for small businesses. The sample selection for the three frames was done independently. All selections employed stratified sequential sampling. No clustering of units was done because the data collection was primarily by mail and telephone. To allocate the SENEs, only counts by income class were available. Because only those in higher income classes were likely to have insurance through their work, the SENEs were divided into five strata by total annual income and allocations were made. The allocations relative to the number of persons in each stratum increased with income to allow for the fact that it was very unlikely that any of the SENEs with lower incomes had insurance through their jobs. Assuming a binomial distribution of persons with insurance within each stratum means that the unit variances increase as the incomes within the strata increase. Under these assumptions, the resulting optimal allocation (Cochran, 1977) to the strata places a high percentage of the sample in the strata with higher income persons.
Within each stratum, the units were sorted by State, industry, and income. Each unit within a stratum was assigned an equal probability, and selection was done using a systematic selection process. Governments not selected with certainty were stratified by State. Each government was given a measure of size based on the square root of its employment size. The governments were sorted by type (city, county, school district, and other) and employment and were selected systematically within each State. Each private-sector establishment within a State was first assigned a probability of selection. This probability was equal to the sample number allocated to the establishment’s stratum within its State divided by the number of establishments in the stratum within the State. For each large firm, the probabilities of selection of the firm’s establishments were summed to obtain an expected number of establishments. To reduce the burden on central office respondents for firms with multiple sampled establishments, the probability of selection of each establishment within a firm was reduced by a constant. After these reductions, the expected sample within each stratum was recalculated determine the shortfall caused by these reductions in probabilities of selection. To account for the shortfall, probability of selection of each establishment in a stratum that had not been reduced was increased slightly the total probabilities of selection within the stratum again equaled the allocation for the stratum within the State. Using the new probabilities of selection, establishments were sorted by State, stratum, industry, and size, and sequential selection was accomplished. SummaryThe MEPS IC sample consists of two parts—the household and list sample components. Similar data on employer health insurance are collected for both samples. The purpose of the household sample is to provide data that can be attached to information collected from respondents in the MEPS HC. The combination of information collected from households and employers for this household sample allows analysts to study household health expenditures and how these expenditures are related to the choices of health insurance offered to the households. The MEPS IC list sample is designed to provide quality estimates for a variety of employer-related health insurance characteristics. The sample is designed to provide a picture of State and national employer health insurance. Among the characteristics shown are enrollments, premiums, employee contributions, types of plans offered, deductibles, and coverages. Estimates made for the Nation should have excellent precision, with RSEs of about 1 percent for characteristics for the entire country and for the entire private sector. In addition, estimates for characteristics of employer-related health insurance within the 40 largest States can be made with good precision. Estimates for characteristics of employer-related health insurance for small establishments will be of particular interest. Oversampling of these establishments, along with annual surveys, should permit analysts to track changes and effects of government policy for this sensitive portion of the population—for the Nation as well as for the 40 largest States. The MEPS IC sample is a primary government source of information on employer-related health insurance in the United States. It should prove to be a key data provider for this crucial portion of health care. Return To TopCochran WG. Sampling techniques. New York: John Wiley and Sons; 1977. Cohen S. Sample design of the 1996 Medical Expenditure Panel Survey Household Component. Rockville (MD): Agency for Health Care Policy and Research; 1997. MEPS Methodology Report No. 2. AHCPR Pub. No. 97-0027. Kish L. Survey sampling. New York: John Wiley and Sons; 1965. Marker D, Bryant E, Wallace L, Yansaneh I. National Employer Health Insurance Survey (NEHIS): draft final methodology report; volume I: statistical methodology. Rockville (MD): Westat, Inc.; 1996.
Source: Center for Financing, Access, and Cost Trends, Agency for Health Care Policy and Research: Medical Expenditure Panel Survey Insurance Component, 1996.
Note: An establishment is an economic unit at a single location where business is conducted or services or industrial operations are performed. A firm is a business entity consisting of one or more business establishments under common ownership or control. N/A is not applicable. Source: Center for Financing, Access, and Cost Trends, Agency for Health Care Policy and Research: Medical Expenditure Panel Survey Insurance Component, 1996.
Note: Projected response rates, government: .90. Projected losses from governments going out of business: 0. Projected response rates, self-employed with no employees (SENEs): .75. Projected losses from SENEs going out of business: .33. An establishment is an economic unit at a single location where business is conducted or services or industrial operations are performed. A firm is a business entity consisting of one or more business establishments under common ownership or control. N/A is not applicable. Source: Center for Financing, Access, and Cost Trends, Agency for Health Care Policy and Research: Medical Expenditure Panel Survey Insurance Component, 1996.
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
