Supporting Statement Part B

B. Collection of Information Employing Statistical Methods

1. Sample Selection and Universe

The total budgeted sample size of the list sample is approximately 44,600 sample units before non response and out-of-scope units and any additional units bought by States (in some past survey years, individual States have paid to increase the sample for their State in order to provide more accurate or detailed State or sub-State estimates). These units include business establishments which employ at least one person and governments. The sampling goal is to produce adequate estimates for 1) the private sector for all 50 States and the District of Columbia; 2) Census Division estimates for Governments; and 3) National estimates. This goal is discussed later in this section.

The list frame sample is derived from two lists:

16. The Census Bureau’s Business Register (BR), a list that contains private sector establishments in the United States which employ at least one person. The list is derived from tax records, and is continually updated to add births and remove those establishments that have closed. This list contains over 7,000,000 establishments and is thought to be very complete, with any under-coverage caused by cases of domestic workers not reported for tax purposes.

16. The Census of Governments, which is collected every five years with data updated in non-Census years using a sample survey. Currently, the most recently available Census of Governments is for the year 2002. A new Census will be available for the year 2007 which will be used for all three years of the survey. If by chance the 2007 Census is not available in time for the 2008 survey, an updated version of the 2002 Census will be used for that year. The Census of Governments contains over 90,000 units.

Together these two lists cover almost 100% of all places with at least one employee, excluding the Federal government.

With samples from these frames, one can study the insurance available through employers to working persons. One can also estimate total costs for the year, total enrollment and other employer health insurance statistics. One can also determine estimates for different parts of the population, such as different firm sizes or industries in order to assess the characteristics of employer provided health insurance for each subgroup.

In order to meet the goals of the survey, the list sample is developed in several steps. These are:

16. the sample is allocated to the public and private sectors

17. a small sample is set aside for certainty and birth units in the private sector

18. the remaining sample within the private sector is allocated by state and the government sample is allocated by Census Division

19. sample is allocated to strata within each sector.

Allocation to the Public and Private Sectors

The division of sample between the public and private sector is based upon past allocations. There are several precision goals for the survey. There are National and State private sector goals and National and Census Division goals for government.

Among the national error goals for the private sector for the survey are the following:

• a .005 relative standard error for national estimates of single and family premiums

• a .0150 relative standard error for national estimates of single and family employee contributions

• a .0075 relative standard error for national estimates of important proportions, such as the percent of employees enrolled in health insurance

Among the national error goals for the state and local government sector for the survey are the following:

• a .0075 relative standard error for national estimates of single and family premiums

• a .020 relative standard error for national estimates of single and family contributions

• a .010 relative standard error for national estimates of important proportions

State estimate goals for the private sector are that relative standard errors for state estimates have errors less than 6 times the similar national private sector goals. Census Division goals for the state and local government sector are that relative standard errors be less than 5 times the national state and local government goals.

These goals reflect the bias among users for higher quality private sector estimates as opposed to estimates for governments.

Given, these goals and the budget limitation for the list sample of 44,600 units, approximately 42,000 sample units were allocated to the private sector and 2,600 to governments. This allocation was based upon relative standard errors obtained from past surveys.

Within the private sector, the allocation is broken into 3 parts:

• a small number of approximately 100 large certainty units,

• a sample of 200 birth units which are added to the frame late in the process and do not contain all information necessary for complete use in the full within-state stratification scheme, and

• the remaining sample which is allocated to individual states.

State Allocation for the Private Sector

The optimal national allocation to states would be to allocate proportional to the size of each State. However, for most states this would result in far too small a sample to meet state estimation goals. This occurs because the 9 largest states under proportional allocation would receive over half the sample and many of the remaining states would have a sample of less than 100 establishments. From experience with past MEPS-IC surveys, it has been determined that a sample of approximately 500 establishments appears to give estimates that meet most state estimation goals using state stratification and allocation processes. To meet state precision goals, an equal size sample could be allocated to each state. An allocation of equal sample to each state would produce state estimates that meet state estimation goals but would be 50% worse than proportional allocation and would not produce adequate national estimates.

The compromise allocation starts by proportionally allocating 17,000 responding sample establishments among the states. The allocation is then supplemented for the 42 smallest states so that each of the 11 smallest states receive about 495 sample establishments and each of the next 31 largest states receive 535 sample units. The 9 largest states receive the allocation from the proportional allocation of the 17,000 units. This allocation has an error about 20% higher than if the entire sample were proportionally allocated. However, national estimation goals should still be met and the state goals should also be met, with the exception of the 11 smallest states where target quality could be missed by a small degree.

Once the desired responding sample size is determined, the sample allocations are increased to allow for expected nonresponse and out of scope establishments to arrive at the final sample size per state. These sizes are listed in Attachment F.

Private Sector Sample Selection

Once the allocations are complete for each state, samples are selected for the private sector within each state. The frame to be used for the private sector will be the most recent update of the Business Register (BR). The BR currently contains 7.0 million establishments. Before final sample allocation and selections, the universe in each state is stratified into cells. There are 14 stratification cells per state, plus a certainty stratum which contains establishments with projected enrollments of above 5000 employees and a birth stratum of recently added units on the frame. Sample for certainty and birth strata are allocated for the country as a whole and are not part of the state allocation process.

The table in Attachment G gives the national breakdown of the percent of private sector establishments in each of the 14 state strata and defines the strata boundaries.

Once these cells are created, the frame broken is into 51x14 strata. The breakdown is made by state to allow for the best sample within each state to assure quality state estimates. The beginning allocation to an individual cell within a state is:

where

N_si is the number of establishments in the ith stratum in the sth state,

is a unit variance for the sth state and the ith stratum determined using 2001-2002 MEPS-IC data and

n_si is the allocation to the ith stratum in the sth state.

The unit variance is an average of unit variances of a selection of important variables derived from past data. The same variance is used for similar cells across all states. These were derived by pooling data across all states for the same strata. This was done because an examination of state by state variances failed to provide significant evidence that the unit variances differed across states.

The allocation is not an optimal allocation for any one variable. There is interest in a wide array of estimates. Some are establishment population estimates, such as the percent of all establishments which offer health insurance. These types of variables require large samples of small establishments. Others are employee population estimates, such as total enrollments in health insurance and require the sample have relatively large numbers of establishments with many employees. The final allocation is the weighted allocation obtained by taking the weighted value of the optimal allocations for several variables and provides a compromise allocation. Trial allocations using this variance compare favorably with the optimal allocations for the individual variables. Results are monitored annually and efforts are continually made to improve strata definitions and allocation to strata.

An allocation to the strata based upon proportions of establishments in the states is given in Attachment G. This gives a reasonable approximation to the proportion of the private sector sample allocated in each stratum within a state.

These allocations are increased for potential non-response and out-of-business. The non-response and out-of-business rates are based upon experience gained in the 2004 through 2006 MEPS-IC. Allocations are increased within each cell to reflect these losses.

Once these allocations are done, they are checked. If a base allocation is too small to assure that variances can be calculated, that is, less than two, the cell sample is increased to assure a large enough sample. This process adds less than 10 units to the total sample. If the allocation is larger than the available establishments, the entire stratum will be selected with certainty. Remaining sample is allocated across the remaining strata in a state using the original allocation method applied to a smaller number of cells. After these checks are performed, each establishment in a cell is given the same chance of selection equal to

p_si = a_si/N_si where a_si is the final cell allocation within the state.

At this point, in order to create a more efficient sample and to reduce burden on large firms, where a single respondent may sometimes provide the information for more than one establishment owned by that firm, these probabilities are altered for some establishments.

The values of the p_si's for all establishments linked to the same firm on the BR are summed. This yields the number of establishments that are expected to be selected for that firm. For a small number of firms this expected value is too large and must be reduced to minimize burden. This is also good for the sample, since the insurance offered in establishments with a common owner is correlated.

To reduce this expected number of establishments, the probabilities of selection are simply reduced to a level where their sum is the desired new sample for that firm. To make up for this reduction in sample, the probability of selection for all other establishments in a cell which contains an establishment with a reduced probability of selection is increased by the amount necessary to have the sum of the probabilities of selection within the strata equal a_si.

Once these probabilities of selection are finalized, the allocated samples are selected using sequential sampling with a random start. To perform this selection, the file is sorted by state, strata, industry and employment size. This assures a good balance of establishments within strata.

Allocation and Selection of the Government Sample

The government sample will use the 2007 Census of Governments as the frame, if available. This contains approximately 90,000 governments. For this selection there will be only two strata per Census Division. There is a certainty stratum which includes all governments with over 5,000 employees. The certainty stratum for governments accounts for approximately half of all government employment.

The non-certainty governments stratum contains all other governments. A sample size of 180 governments is allocated to the non-certainty government stratum for each of the 9 Census Divisions.

To perform the selection, each non-certainty government is given a measure of size equal to the square root of its total employment. This increases the sample of smaller governments relative to their total employment. The selection probability for a single government is determined as the total final Census Division non-certainty state government allocation, times the government’s measure of size, divided by the sum of all measures of size within the Census Division

The non-certainty government sample within each Census Division is selected sequentially from a file sorted by state, type of government (county, city, school district, etc.), and employment using a random start.

Sampling of Health Insurance Plans

In private sector and non-certainty government cases where an employer offers several health insurance plans, a procedure for scientifically selecting a sample of plans will be implemented. The first step is to identify these cases. During telephone prescreening, establishments will be ask to provide the number of health insurance plans they offer their employees. If there are four plans or less, all are taken with certainty. For the few establishments that offer more than four plans, the names of the plans are collected along with estimates of enrollments. The CATI software will be used to randomly select a subset of plans for collection. The respondent will receive a preprinted plan collection supplement for the selected plans.

All plans are collected for large certainty governments.

Response Rates

The following table presents the approximate sizes of each of the universes from which sample data are being collected, the approximate expected sample size including out-of-scope, the expected in-scope sample size, and the expected number of responses and the response rates. The rates are based upon experience from the 2002 to 2004 MEPS-IC data collection.

Projected Response Rates 2008-2009 MEPS-IC

Response Group	Universe	Sample	In-Scope	Respondents	Rate
List Sample- Private Sector	7.0 x 106	42,019	39,693	30,836	.777
List Sample- Government	90,000	2,582	2,582	2,426	.940
Total		44,601	42,275	33,262	.787

2. Procedures for Data Collection

Data Collection

Data collection for the MEPS-IC takes place in three phases: prescreening interview, questionnaire mailout, and nonresponse follow-up. The prescreening interview is conducted by telephone. Its goal is to obtain the name and title of an appropriate person in each establishment to whom a MEPS-IC questionnaire will be mailed. Interviewers also verify addresses and identify businesses that no longer exist, have closed, have merged, etc. Establishments from certain large firms and governments are not prescreened because they are known to offer health insurance. Due to their size and importance to all Census establishment surveys, the Census Bureau maintains up-to-date contacts with these employers.

For establishments which do not offer health insurance, a brief set of questions about establishment characteristics is administered at the end of the prescreening interview to close out the case. The MEPS-IC questionnaires are mailed to those establishments, which during the prescreening phase, were:

1. not contacted,

2. refused to cooperate,

3. were contacted and acknowledged that they did provide health insurance,

4. were from large firms or governments specified at the start of collection for mail-only, or

5. had no known phone number.

Establishments which do not respond to the initial MEPS-IC mail questionnaire are mailed a nonresponse follow-up package. Those establishments which fail to respond to the second mailing are contacted for a telephone follow-up using computer-assisted interviewing.

Data for the largest governments and private sector firms, reporting for multiple establishments, are collected using specialized staff and forms. This is done to make the collection process flexible, simple, and as little burden for these important respondents as possible. Sometimes multiple telephone contacts and person visits are used to collect these data. For some collections, abstraction from company records or plan brochures is used if the firm insists on such methods.

Samples of survey materials for the MEPS-IC are included in Attachments H-R. Included are all letters, questionnaires and the Computer Assisted Telephone Interviewing (CATI) text used for the prescreener:

6. Letter for original mailing to a private sector list sample unit, Attachment H

7. Letter for original mailing to a government sample unit, Attachment I

8. Letter for follow-up mailing to a private sector list sample unit, Attachment J

9. Letter for follow-up mailing to a government sample unit, Attachment K

10. Letter for follow-up fax to a private sector list sample unit, Attachment L

11. Letter for follow-up fax to a government sample unit, Attachment M

12. Thank you letter, Attachment N

13. Questionnaire 20(D) – Definitions, Attachment O

14. Text of CATI prescreener, Attachment P

15. Establishment-level questionnaires, Attachment Q

16. Plan-level questionnaires, Attachment R

Materials displayed are the most recent version of the respective item. Some minor changes will be made in the final materials to reflect current needs. For example, signatures on letters are updated to reflect current management at AHRQ and Census.

Weighting

Each of the list samples will be weighted separately. Beginning with the inverse probability of selection as a base weight, the data will be adjusted for non-response. Logistic regression is used to determine variables which relate to response. Among other characteristics, the number of plans, size of establishment, size of firm or government, industry, and whether the respondent received a mail survey must be considered. Adjustments are made so that the sum of respondents in a cell equals the beginning weight for the cells. It would be desirable to adjust for non-response in each state, but cell size may preclude that adjustment in all but the largest states.

Once adjustments for non-response are made, data will be post stratified to a set of new control totals. For the private sector this would be new frame counts from the most recent BR for cells determined by state, industry, size of firm and size of establishment. For governments, post stratification will be done to counts provided by the Census of Governments Division. Cells will be determined by size of government and state.

Estimation and Accuracy

Estimation is done using sampling weights and variances are calculated using standardized software available, such as SUDAAN or the random group method. Both account for the specialized sampling methods used for the survey.

The following table shows estimated relative standard errors for some key results that were produced using the list sample for the 2004 MEPS-IC. This survey contained a similar sample allocation to that being proposed for future surveys. The precisions are very adequate given the MEPS-IC goals, discussed earlier, for the quality of various national and sub national estimates for the private sector government estimates.

Sample Relative Standard Errors

Domain Variable	National	Largest 9 States	Middle 31 States	Smallest 11 States	National Governments
Average Family Contribution	1.44%	7.04%	7.60%	7.90%	1.36%
Average Family Premium	0.52%	2.19%	2.99%	3.11%	0.61%
Average Single Contribution	1.41%	6.32%	8.00%	8.31%	1.99%
Average Single Premium	0.59%	2.58%	2.76%	2.87%	0.52%
Percent Employed Where Insurance Offered	0.34%	1.27%	2.19%	2.28%	NA
Percent Enrolled Where Insurance Offered	0.74%	3.45%	4.57%	4.75%	1.17%
Percent of All Employees Enrolled	0.79%	3.62%	5.02%	5.22%	NA
Percent of Establishments That Offer Health Insurance	0.64%	3.00%	3.87%	4.02%	NA

NA – Estimates not made because the number of governments that do not offer health insurance is negligible and not of interest.

Certain key variables, such as premiums, contributions and enrollments are imputed when item non response occurs and values are missing. Selection of donors is accomplished using a nearest neighbor type hot deck process which chooses the best donor given a set of matching variables and their order of importance. Variables used to match to determine donors are chosen for their correlation with the variable to be imputed, with special care taken to select variables which are also correlated with non response.

Actual values are derived from donors in various ways. In some cases ratios are derived from donors and applied to other values the recipient has reported. Great care is taken to maintain consistency of relationships within the data by using ratios and other means that apply data from the donor to data which have been reported by the recipient. The data are also processed in a specific order to assure that variables important to the imputation of variables later in the process are imputed first. For instance, if one must impute total single premium and total family premiums for the same health plan, the values are imputed in order so that the imputed family premium will depend on the single premium that has already been imputed. This is necessary because these values are highly correlated.

3. Methods to Maximize Response Rates

To achieve maximum response rates the following methods are to be used:

1) Perform a screening phone call to identify the best contacts, number of insurance plans and to complete simple cases, such as establishments which offer no insurance.

2) Interviewers schedule calls at convenient times for respondents.

3) Supervisors regularly attempt to convert refusals.

4) Mail a self-administered questionnaire.

5) Mail a second questionnaire to those not responding in adequate time.

6) After an adequate interval, do a telephone follow-up to either remind the respondent or to collect the data.

7) For questionnaires returned by mail which fail key edits, perform a callback to verify data.

A further method to improve response, developed to overcome problems encountered in the 1994 NEHIS and the 1996 MEPS-IC, is a special handling group. This group consists of highly trained analysts, interviewers and statisticians. Their purpose is to collect data from large firms with high burdens. These respondents are given a primary focus early in the data collection process. Each large firm is assigned a collection coordinator. The special handling group can use any reasonable means to accommodate respondents. They can use personal visits and arrange to collect information in formats easiest for the respondents. There are also special collection forms that allow these firms to list plans only once that may be repeated across multiple establishments. Improved methods of collection for large firms allowed the Census Bureau to significantly improve response rates after the initial 1996 MEPS-IC survey. Another positive effect of these efforts has been a lowering of burden for the larger firms and large governments.

4. Tests of Procedures

This is a renewal of a current survey. As part of general data collection activities, data collection results, interviews, comments made by respondents and estimates are monitored. New data items added to collection are pretested under a separate Census Bureau testing clearance. Because of this monitoring and collection experience, no special pretest is required at this time for the general survey questionnaire.

5. Persons Responsible for Statistical Aspects of the Design

The following persons participated in either the design or review of the MEPS-IC sample design.

Dr. John Sommers

Mathematical Statistician

Center for Financing, Access and Cost Trends

Agency for Healthcare Research and Quality

Rockville, MD 20850

301-427-1474

Michael Kornbau

Chief, Program Research and Methods Branch

Bureau of the Census

Washington, DC 20233

301-763-6779

David Kashihara

Mathematical Statistician

Center for Financing, Access and Cost Trends

Agency for Healthcare Research and Quality

Rockville, MD 20850

301-427-1474

Dr. Steven B. Cohen

Center for Financing, Access and Cost Trends

Agency for Healthcare Research and Quality

Rockville, MD 20850

301-427-1406

Anne Kearney

Program Research and Methods Branch

Bureau of the Census

Washington, DC 20233

301-763-6780

Maribel Aponte

Program Research and Methods Branch

Bureau of the Census

Washington, DC 20233

301-763-6777

11