THE NEISS SAMPLE (DESIGN AND IMPLEMENTATION) 1997 to Present. Prepared for public release by:

Transcription

1 THE NEISS SAMPLE (DESIGN AND IMPLEMENTATION) 1997 to Present Prepared for public release by: Tom Scroeder Kimberly Ault Division of Hazard and Injury Data Systems U.S. Consumer Product Safety Commission June 2001

2 TABLE OF CONTENTS INTRODUCTION 1 NEISS SAMPLE CHARACTERISTICS 1 JANUARY DECEMBER JANUARY 1999 TO PRESENT 2 TREND ANALYSIS OF NEISS DATA PRIOR TO HOSPITAL SELECTION PROCEDURE AND HOSPTIAL PARTICIPATION 4 HOSPITAL WEIGHTS, 1997 TO PRESENT 6 BASIC HOSPITAL WEIGHTS 6 ADJUSTMENTS FOR NON-RESPONSE 6 ADJUSTMENTS FOR HOSPITAL MERGERS 6 ADJUSTMENTS FOR CHANGES IN SAMPLING FRAME 8 FINAL NEISS WEIGHTS 10 NATIONAL ESTIMATES OF PRODUCT-RELATED INJURIES FROM NEISS 11 BIAS IN A RATIO ADJUSTED NEISS ESTIMATE 12 SAMPLING ERRORS ASSOCIATED WITH NEISS ESTIMATE 13 CALCULATING VARIANCES OF NEISS ESTIMATES 13 GENERALIZED SAMPLING ERRORS 14 APPENDIX 1 27 APPENDIX 2 29 i

3 REFERENCES 33 ii

4 LIST OF TABLES Table Page Table 1: NEISS Sample Caracteristics, January 1, 1997 troug December 31, Table 2: Ratio Adjustments to NEISS Weigts, January 1, 1999 troug December 17 31, 1999 Table 3: NEISS Sample Caracteristics, January 1, 2000 to present 18 Table 4: Ratio Adjustments to NEISS Weigts, January 1, 2000 troug December 19 31, 2000 Table 5: Ratio Adjustments to NEISS Weigts, January 1, 2001 troug December 20 31, 2001 Table 6: National Estimates of te Total Number of Product Related Injuries, 21 January September 30, 1997 Table 7: Number of Hospitals in te NEISS Sample and Number Participating by 22 Stratum for eac mont starting January 1997 Table 8: NEISS Weigts by Year, Mont, and Stratum from January 1997 to 24 December 2000 Table 9: Generalized Relative Sampling Errors for NEISS for Estimates of Various 26 Size iii

5 INTRODUCTION Te Consumer Product Safety operates a data system known as te National Electronic Injury Surveillance System (NEISS). Te NEISS is a probability sample of ospital emergency departments in te United States and its territories. Data collected from te NEISS sample is weigted based on te sample design to produce national estimates of te number of consumer products-related injuries treated in ospital emergency rooms. Additionally, te NEISS data provide a source for follow-up investigations of product-related injuries. Tis report documents te canges in NEISS sample design since January For documentation of te NEISS sample prior to 1997, see te report Te NEISS Sample Design (Design and Implementation) from 1979 to NEISS SAMPLE CHARACTERISTICS January December 1998 During 1996, under contract wit Westat, Inc., te NEISS sample was updated to reflect canges in te universe of ospitals (Marker & Lo, 1996). A new sampling frame was constructed based on te most currently available (1995) listing of ospitals and emergency room visits purcased from te SMG Marketing Group of Cicago, IL. Te sampling frame included ospitals wit 6+ beds aving an emergency department; excluded were psyciatric and penal institutions. Te updated sample contains five strata, four based on size (te total number of emergency room visits reported by te ospital) and one stratum consisting of cildren s ospitals. Te ospital size strata were constructed as sown in Table 1, wit a special stratum for cildren's ospitals. In selecting te sample, Westat used a resampling metod tat maximized te probability of retaining ospitals selected for te former 1991 sample. Te metod was an extension of te Keyfitz procedures for stratified simple random samples. Terefore, 76 of te previously selected ospitals were retained and 26 new ospitals were selected, for a total of 102 ospitals. Of te 76 ospitals retained, 70 were retained from te probability sample and six were retained from te cildren s sample. Data from tese six 1

6 ospitals ad been collected since January 1995 to increase case findings for cildren. One ospital was no longer in business wen recruitment began in 1997, tus te in scope probability sample is comprised of 101 ospitals, instead of te initially planned 102. Tree ospitals ave unique weigts because te ospital merged wit anoter ospital after te updated sampling frame was cosen. Te specifics of te mergers are listed below: One ospital selected in te NEISS sample in stratum 2, merged wit anoter ospital not selected in te NEISS sample in stratum 2. Te ospital's administrative records are merged. Te merged ospital is considered in te statistical sample as one ospital beginning August 1, In te 1997 sample, te ospital as a weigt of Anoter ospital selected in te NEISS sample in stratum 2, merged wit anoter ospital not selected in te NEISS sample in stratum 2. Te merged ospital is considered in te statistical sample as one ospital beginning May In te 1997 sample, te ospital as a weigt of One ospital selected in te NEISS sample in stratum 3, merged wit anoter ospital not selected in te NEISS sample in stratum 3. Te merged ospital is considered in te statistical sample as one ospital beginning July In te 1997 sample, te ospital as a weigt of January 1999 to present Beginning in January 1999, NEISS weigts ave been adjusted annually to take into account canges in te most recent sampling frame of U.S. ospitals available. Tis new component of te basic NEISS weigt is te ratio of total emergency department visits as listed on te updated sampling frame to te total emergency department visits as estimated from te NEISS sample. Table 2 presents te total emergency department visits (ERVs) from te 1998 sampling frame, te estimated ERVs from te NEISS sample for 1998, and te computed ratio adjustments for January 1, 1999 December Te section labeled Adjustments for Canges in te Sampling Frame discusses te ratio adjustments in more detail. Te report Updated NEISS Weigts Using te 1998 SMG Hospital Frame written by Tom Scroeder (1999) gives a complete description of tis ratio adjustment procedure. 2

7 One ospital in te small stratum closed its emergency department on December 13, Tus as of December 14, 1999, te NEISS is considered to be a sample of 100 in-scope ospitals. Table 3 presents te caracteristics of te NEISS sample for tis period. Table 4 presents te total emergency department visits (ERVs) from te 1999 sampling frame, te estimated ERVs from te NEISS sample for 1999, and te computed ratio adjustments for January December Te report Updated NEISS Weigts Using te 1999 SMG Hospital Frame written by Tom Scroeder (2000) gives a complete description of tis ratio adjustment procedure. Table 5 presents te total emergency department visits (ERVs) from te 2000 sampling frame, te estimated ERVs from te NEISS sample for 2000, and te computed ratio adjustments for January December Te report Updated NEISS Weigts Using te 2000 SMG Hospital Frame written by Kim Ault and Tom Scroeder (2001) gives a complete description of tis ratio adjustment procedure. TREND ANALYSIS OF NEISS DATA PRIOR TO 1997 One of te advantages of a long running data series suc as te NEISS is te ability to track trends across time. Periodic updates to te sample, suc as te one described in tis report, can interfere wit suc analyses. One of te best ways to adjust for tese updates is to ave an overlap (bridge) during wic data are collected from bot te old and new samples. Between January and September of 1997, a nine-mont overlap study was implemented as part of te sample update. Selecting a new sample from te same frame increases te variances in measures across time. Wen te new sample is selected from a different frame (as is te case for te NEISS update) tere is a potential break in te time series. By comparing te estimates produced by te two samples and te two frames, it is possible to adjust old estimates (backcast) to be consistent wit te new sample. Suc an overlap can be expected to provide continuing useful information tat more tan compensates for te one time cost of te bridge sample. To backcast existing time series for te period from te previous frame update to te current update requires a two-step process. Te first step is to estimate te difference between estimates from te two samples for te overlap period. Te ratio of total number of injuries estimated from te new sample and new weigts, 3

8 divided by te estimate for te same period from te old sample and old weigts, sould be calculated. Tis is an estimate of cange due to fluctuations in te number and type of product-related injuries at existing emergency rooms, te addition of new emergency rooms to te frame, and sampling variation from one sample to anoter. Tis measure of cange could be computed separately by eac stratum. However, given te small sample sizes in some strata, te procedure cosen was to compute an overall measure of cange and apply it to all strata. Te second step is to adjust for tis cange. Since te previous NEISS sample design was selected based on 1985 ERV data and te sample frame (prior to 1997) was based on 1995 ERV data, tis cange represents fluctuations over a 10-year period. Wile te canges in number of emergency rooms and ERVs as probably been uneven over tis period, a reasonable approximation is to assume tat onetent of te cange as occurred eac year. Terefore, te adjustment procedure adopted was to add one-tent of te estimated cange to te 1986 estimates, twotents to te 1987 estimates, and so on up to nine-tents of te cange to te 1994 estimates. Te entire cange as been applied to te estimates. For additional information on analyzing NEISS data over time, see Trend Analysis of NEISS Data (February 2000). Overlap estimates were calculated for individual product codes as well as iger order groupings of product codes. Te overall estimate during te overlap period form te old sample was 9,258,592 wile te estimate from te new sample was 8,436,476. Tis represents an 8.9% decrease in te estimate from te new sample as compared to te estimate from te old sample. Te difference in estimates fluctuated by product code or product grouping wit some estimates increasing and oters decreasing wen comparing te estimates from te new sample to estimates from te old sample. Table 6 lists te estimates from te new and old sample at te overall level as well as at te product grouping level used in CPSC s annual report. HOSPITAL SELECTION PROCEDURE AND HOSPTIAL PARTICIPATION Westat, te contractor for te current NEISS sample redesign, provided CPSC wit te sampling frame, a set of primary ospital sample selections and procedures for selecting alternate ospitals to be used as substitutes for ospitals unwilling to participate. In eac recruitment process, CPSC recruiters make 4

9 repeated and intensive efforts to obtain te participation of every primary sample ospital selected. Only wen all approaces to obtain cooperation from a primary ospital fail do te recruiters turn to a replacement ospital. If te ospital being replaced is te primary selection, CPSC selects te alternate (replacement) ospital according to te procedures establised by Westat. If te first alternate ospital refuses to participate, te second alternate is selected, etc., until cooperation is obtained wit a replacement ospital. If te ospital to be replaced is not a primary selection, we use te opportunity to try again to recruit tat ospital wic is te primary sample selection. Tis as enabled us to obtain te cooperation of some primary ospitals, wic initially ad refused to participate. Over time, wenever ospitals remain in business but elect to drop out of te system, CPSC attempts to return to te original recruitment order: primary ospital, first alternate ospital, etc. Since te replacement ospitals ave te same probabilities of selection as te primary ospitals wit wic tey are associated, tey also ave te same statistical weigts. All ospitals on te system at te time of one of te updates of te NEISS are considered primary ospital selections as of tat time. 5

10 HOSPITAL WEIGHTS, 1997 to present Basic Hospital Weigts Te basic ospital weigts used by NEISS are equal to te inverse of te probability of selection for te ospitals in eac stratum. Te inverse of te probability of selection is simply te total number of ospitals on te 1995-sampling frame divided by te total number of ospitals in te 1997 NEISS sample calculated at te stratum level. Adjustments to tese basic weigts are made for non-response and ospital mergers. Annual estimates of injuries are derived by summing te montly estimates for all monts of te year. Adjustments for Non-Response Sown in Table 7 are te number of in-scope ospitals for te sample, togeter wit te number of ospitals tat actually participated in te system, by stratum, for eac year and mont since January Tis table may be used to compute te montly stratum non-response adjustment factor for te basic ospital weigts. Adjustments for Hospital Mergers Wen two ospitals merge and are in different size classes, te probability of selection of te merged ospitals is found by te formula for te union of two events: P = P + 1 P2 P1 P2 Were P i = probability of selection of ospital i; i = 1, 2 Taking into account any non-response adjustment, te basic merged weigt of ospital i is computed as: Wgt i = n N 1 1 r n n N 2 2 r n n N 1 1 r n 1 1 n N 2 2 r n 2 2 6

11 were: N 1(2) = n 1(2) = r 1(2) = n' 1(2) = Number of ospitals in te NEISS sampling frame for stratum 1(2) Number of ospitals selected for te NEISS sample for stratum 1(2) Number of ospitals participating in te NEISS sample for stratum 1(2)for te time period Number of in-scope ospitals in te NEISS sample for stratum 1(2) Wen two ospitals merge and are in te same size class, te situation is more complex because te sampling in a size class is done witout replacement. Te sample size in a particular size class was a fixed number n, and te total number of ospitals in te size class was N for te original sample and frame. If S denotes te sample, H 1 and H 2 te ospitals, te tree possibilities tat lead to te retention of te merged ospital in te sample are: A. H 1 S, H 2 S Hospital 1 is in te original sample, ospital 2 is not B. H 1 S, H 2 S Hospital 2 is in te original sample, ospital 1 is not C. H 1 S, H 2 S Bot ospitals are in te original sample Te probability of event A is: P(A) = P(H 1 is selected on te first draw and H 2 is not selected on any draw) 1 = n N N N 2 1 N N 3 N n N n + 2 N n = N n + 1 n N N n N 1 Te leading multiplier n accounts for te possibility tat H 1 may be selected on any of te n draws. Te probability of event B, P(B), is te same as P(A) from symmetry. Te probability of event C is: n P C) = 2 P(H 1 is selected n te first draw and H 2 n 1 1 = 2 2 N N 1 ( 2 is selected on te 2nd) Te probability of inclusion of te merged ospital is ten: n(2n n 1) P( A) + P( B) + P( C) = N( N 1) 2n n( n 1) = N N( N 1) 7

12 Taking into account any non-response adjustment, te basic merged weigt of ospital i is computed as: Wgt i = 2n N 1 n n ( n 1) r N ( N 1) Were: N = Number of ospitals in te NEISS sampling frame for stratum n = Number of ospitals selected from te NEISS sample for stratum r = Number of ospitals participating in te NEISS sample for stratum for te time period n' = Number of in-scope ospitals in te NEISS sample for stratum Adjustments for Canges in Sampling Frame Te ospital population does not remain static over time. Hospitals close, merge, and open as well as cange in te volume of emergency department visits. In order to stabilize te NEISS estimates over time witout taking a new NEISS sample and backcasting istorical estimates, a ratio adjustment to te basic NEISS weigt can occur. NEISS estimates te number of consumer product-related injuries treated in ospital emergency departments. A ratio adjustment takes advantage of knowledge about a igly correlated auxiliary variable, wic in tis case is te total number of emergency department visits. Te total number of emergency department visits can be obtained by purcasing a complete SMG ospital database on a yearly or biyearly basis. Te ratio adjustment applied to te basic NEISS weigt is te ratio of te known total number of emergency department visits in te population (from te frame) over te estimate of te total emergency department visits based on te sample of NEISS ospitals. For computing ratio adjustments, Westat as recommended combining te small and medium strata togeter and te large and very large strata togeter due to te relatively small number of NEISS ospitals in some of te larger strata. (Marker, et al, 1999) Witin eac combined stratum, te ratio-adjusted weigts, w* i, are computed as: 8

13 w * i = w i * ERV i w * up, ' i erv up,i = w i * R * (Equation 1) were w i = NEISS basic weigt (adjusted for ospital mergers if necessary) w i = NEISS basic weigt (adjusted for ospital mergers if necessary: see discussion below) ERV up,* = Total ERVs on updated SMG file for combined stratum * Erv up,i = Number of ERVs from te updated SMG file for NEISS ospital i R * = Ratio adjustment for combined stratum * Before applying equation 1 to compute te ratio adjusted NEISS weigts, tree issues need to be resolved. Several NEISS ospitals ave been replaced since te redesigned sample was selected. It was decided tat te ERVs from te set of ospitals tat are currently on te NEISS at te time of te update would be used in te denominator of equation 1. Wit te possible exception of te tree ospitals wit merger weigts in NEISS, te two basic NEISS weigts, w i and w i, are equal. A NEISS ospital is given a merged weigt if tat ospital as merged wit anoter ospital and te NEISS coder cannot distinguis in wic ospital emergency department an injury was treated. However, it is possible on te updated frame tat te total ERVs from eac of te merged ospitals are listed separately, and for estimating total ERVs te ospitals would not be considered merged. If tis is te case ten w i equals te merged weigt and w i equals te NEISS basic weigt. Te following are ow eac of te merged ospitals will be treated: For one of te merged ospitals in stratum 2: Of te two ospitals tat merged, one eventually closed and one built a new facility. For computing ERVs, tis will be treated as two separate ospitals on te updated frame, one wit zero ERVs and te oter wit a specified number of ERVs. Tus w i w i. For te oter merged ospital in stratum 2: Bot of te merged ospitals ave separate ERVs listed on te updated frame and bot emergency rooms remain operational. For computing ERVs, tis will be treated as two separate ospitals on te updated frame. Tus w i w i. 9

14 For te merged ospital in stratum 3: Bot emergency rooms remain operational. However, te updated sampling frame lists ERVs for only one of te ospitals. Tis will be treated as a merged ospital in computing ERV totals. Tus w i = w i. A ratio adjustment in its true form would only sum te total ERVs on te updated frame for te ospitals tat were eligible for NEISS on te 1995 frame. However, summing in tis manner would not account for te ERVs from any of te new emergency departments tat opened in te interim and tus underestimate te total number of injuries. Troug NEISS, CPSC wants to estimate te total number of product-related injuries treated at all eligible U.S. ospitals. Tis would include any new emergency departments tat ave opened since Tus, te total ERVs from te new emergency departments are added into teir appropriate combined strata. Final NEISS Weigts Te final NEISS weigt calculated eac mont and used for national estimates can consist of te following parts: basic weigts, adjustments for non-response, adjustments for merged ospitals, and adjustments for canges in te sampling frame. Te final weigt (for all non-merged ospitals) can be written as: ( N * n' ) R NEISS wt = ( n * r ) (Equation 2) were: N = Number of ospitals in te 1995 sampling frame for stratum n = Number of ospitals selected for te NEISS sample for stratum n = Number of in-scope ospitals in te NEISS sample for stratum r = Number of NEISS ospitals participating in stratum for te given mont R = Ratio adjustment for combined stratum Table 8 sows te final NEISS weigts for 1997 to

15 NATIONAL ESTIMATES OF PRODUCT-RELATED INJURIES FROM NEISS National estimates for a given mont of NEISS are calculated using te following formula were: wgt i = x i = Estimate = wgt x i i i (Equation 3) Weigt of ospital i for te mont Number of cases for a specified product or type of injury reported by ospital i for te given mont Except for te unique weigts of merged ospitals, te weigts of te ospitals are te same witin a stratum and equation 3 can be written as: Estimate = m r = 1 i= 1 N n R n r * x i (Equation 4) were: m = N = n = n' = r = R * = x i = Number of strata in te NEISS sample during te given time period Number of ospitals in te NEISS sampling frame for stratum Number of ospitals selected for te NEISS sample for stratum Number of in-scope ospitals in te NEISS sample for stratum Number of NEISS ospitals participating for stratum for te given mont Ratio adjustment factor for stratum for te given mont Number of cases for a specified product or type of injury reported by ospital i in stratum for te given mont Note tat N /n, te reciprocal of te probability of selection of a ospital in stratum for te (current) sample, is te basic weigt associated wit eac ospital in stratum. Te factor n' /r is used to adjust eac ospital in a given stratum for te lack of participation of one or more ospitals in te stratum, wen necessary. If all ospitals in te stratum participate during te given mont, te non-response adjustment factor is one. Non-response adjustment factors can be computed for 11

16 eac stratum of te NEISS since January 1997 by using Table 8. For a given mont and stratum, te non-response adjustment factor is obtained by dividing te number in te column labeled "SAMPLE" by te number in te column labeled "PARTICIPANTS." For example, te non-response adjustment factor for te small stratum for October 1999 is 48/47 or Wen te montly non-response adjustment factor for te given stratum is multiplied by te basic ospital weigt, te result is te adjusted basic NEISS weigt for te small stratum for October Te basic weigt for te small stratum in October 1999 was 3179/48 or (See Table 1). Terefore, te adjusted basic NEISS weigt is x 66.23, or Te last adjustment is for canges in te sampling frame, so te finally NEISS weigt for October 1999 is x or Bias in a Ratio Adjusted NEISS Estimate Ratio estimates in general are biased estimates. In practice, te bias is usually negligible and unimportant if te sample is of moderate size. As discussed in Cocran, te bias in eac stratum of a stratified sample as an upper bound less tan te standard error of te estimate multiplied by te coefficient of variation of te estimated total ERVs. Bias in NEISS Estimate <= SE( NEISS Estimate ) * C.V. (Estimated ERV ) were: C.V. (Estimated ERV ) <=.10 in te combined small and medium strata.09 in te combined large and very large strata.11 in e cildren s stratum Bias in a stratified sample is additive, but due to te relatively small upper bounds of te bias in terms of te standard error of te estimate for eac stratum, te bias in ratio adjusted NEISS estimates is considered negligible. 12

17 SAMPLING ERRORS ASSOCIATED WITH NEISS ESTIMATE Calculating Variances of NEISS Estimates Variances of NEISS estimates are calculated using te classical formula for te variance of a total from a stratified sample. Wit adjustments to NEISS weigts made for non-response, ospital mergers, and ratio adjustments, te classical variance formula doesn t fit NEISS exactly. Oter metods for approximating te variance are available suc as Taylor series approximation or various replication (Jackknife, Balanced Repeated Replication) metods. Taylor series approximations and a leave one out Jackknife approac available using SUDAAN software produce similar if not identical results to te classical variance formula used by Data Systems. Because of te similar results to oter metods of calculating variances, Westat recommended tat te classical variance formula sould be used in calculating variances for annual NEISS estimates. Because NEISS estimates are based on a sample of ospital emergency rooms rater tan on a census of all ospital emergency rooms, tey may differ somewat from te figures tat would ave been obtained if product-related injuries ad been obtained from all ospital emergency rooms in te U.S. Standard (or sampling) errors are measures of te sampling variability, tat is, of te variations in te estimates tat occur by cance because a sample rater tan te entire set of emergency rooms is surveyed. Measures of sampling variation are frequently expressed as coefficients of variation (c.v.'s). Te coefficients of variation are te standard errors divided by te estimates. Te c.v. is a measure of te proportionate error due to sampling and te standard error is a measure of te absolute error. Te square of te standard error is referred to as te sampling variance. Te variance of an estimate based on a sample can be calculated from te sample data, and tis as been done for NEISS. Te estimates of variances for NEISS take into account te probabilities of selection, stratification, and weigting. Te variance estimating formula currently used is: σ 2 x = m = 1 2 r r N n 1 i r i= 1 n r = 1 r 1 i= 1 m r ( ) ( ) 2 2 r x x = wgt x wgtx i i i 13

18 m = Number of strata in te sample for te time period r = te number of ospitals participating in stratum for te time period N = Number of ospitals in te NEISS sampling frame for stratum n = Number of ospitals selected for te (current) sample for stratum n' = Number of in-scope ospitals in te (current) sample or stratum x i = te number of injuries reported for te time period in te i-t ospital in stratum wgt i = te weigt of ospital i in stratum for te time period x r = i= 1 x r i and wgtx r = i= 1 wgt i x r i Te equation above applies to estimates of injuries during any period -- montly, quarterly, annually, etc. -- wit x i interpreted as te number of injuries during tat period. For periods greater tan one mont, te formula assumes tat te sample size is constant over te period. Wen tere ave been variations in sample size, r is defined as te number of ospitals reporting during all, or te majority of te monts in te period. A majority n (number of ospitals reporting a majority of te time) or fractional n (sum of number of ospital monts divided by te number of monts) can be used wit little difference in te results between te two metods. Tis formula actually sligtly overstates te true sampling variance, because it does not take into account te effect of te secondary stratification factor, geograpy. Test calculations indicate tat tere are only sligt differences between calculations using te formula above, and ones tat consider te geograpic substratification. Appendix 1 contains sample SAS code for calculating variances associated wit a particular NEISS estimate. Appendix 2 contains sample SAS /SUDAAN code for calculating variances associated wit a particular NEISS estimate. Generalized Sampling Errors "Generalized sampling errors" are also produced for NEISS estimates. Tese smooted values are derived from fitting a curve to all calculated sampling errors for a defined set. Generalized sampling errors are commonly used by U.S. Government statistical agencies suc as te Bureau of te Census, te Bureau of Labor Statistics, and te National Center for Healt Statistics to convey to te 14

19 public information about te precision of teir estimates. One reason for teir use is to reduce te cost of computing and presenting te large number of variances tat would be required if eac required a separate computation. Anoter reason is to produce more stable estimates of variances. Sampling errors, wen estimated from sample data, ave variances of teir own; fitting a curve frequently reduces tese errors. To fit curves for generalized sampling errors, te relationsip between te size of te estimate x k for te k-t product group and te variance of tat estimate, σ 2 xk is expressed by te formula: σ xk x k = C. V. xk 1 = a + ( b * LN( x k )) were C.V. is te coefficient of variation, "a" and "b" are estimated by an iterative procedure. Te iterative procedure produces estimates of "a" and "b" wic minimize te expression: k x k C. V. x k 1 a + ( b* LN( x Recent approximate generalized standard errors and coefficients of variation for annual estimates of NEISS are presented in Table 9. k )) 2 Te standard errors derived from NEISS can be used in te following way: Te sample estimate and its standard error enable one to construct confidence intervals, ranges tat would include te average results of all possible samples wit a known probability. For example, if all possible samples were selected and surveyed, and an estimate and its standard error calculated from eac, ten: 1. Approximately 95 percent of te time te interval from two standard errors below te estimate to two standard errors above te estimate would include te average result of all possible samples. 2. About 90 percent of similar confidence intervals using 1.6 standard errors would include te average results of all samples. 3. About 68 percent of confidence intervals using one standard error would include te average results of all possible samples. 15

20 January 1, 1997 troug December 31, 1998 Stratum Number of Range of Total Total Out of In Scope Hospitals in ERVs Sample Scope /1 Sample Universe ,830 3, ,831-28,150 1, ,151-41, , Cildren's Various Total 5, / Out of scope ospitals included ospitals no longer in existence, ospitals witout emergency departments, or ospital emergency departments no longer in business. a) Hospitals no longer in existence: Table 1 NEISS Sample Caracteristics One ospital selected in te small stratum was no longer in business as a ospital wen te recruitment began in January

21 Table 2 Ratio Adjustments to NEISS Weigts January 1, 1999 troug December 31, 1999 Stratum 1998 ERVs from Estimated 1998 ERVs from Emergency Rooms ERVs from Ratio Emergency Rooms Total ERVs New to 1998 NEISS Adjustment on 1995 Frame Frame Sample 1 24,864, ,250 24,992,905 26,587, ,919, ,767 24,037,811 22,104, ,385, ,385,639 24,429, ,027,969 77,000 24,104,969 25,480, Cildren's 1,967,015 94,582 2,061,597 2,355, Total 98,164, ,599 98,582, ,957,582 17

22 Table 3 NEISS Sample Caracteristics January 1, 2000 to present Stratum Number of Range of Total Total Out of Scope In Scope Hospitals in ERVs Sample /1 Sample Universe ,830 3, ,831-28,150 1, ,151-41, , Cildren's Various Total 5, / Out of scope ospitals included ospitals no longer in existence, ospitals witout emergency departments, or ospital emergency departments no longer in business. a) Hospitals no longer in existence: One ospital selected in te small stratum was no longer in business as a ospital wen te recruitment began in January One ospital in te small stratum closed its emergency department on December 13,

23 Table 4 Ratio Adjustments to NEISS Weigts January 1, 2000 troug December 31, 2000 Stratum 1999 ERVs from Estimated 1999 ERVs from Emergency Rooms ERVs from Ratio Emergency Rooms Total ERVs New to 1999 NEISS Adjustment on 1995 Frame Frame Sample 1 24,895, ,975 25,073,528 26,076, ,759, ,898 24,029,274 21,927, ,253,312 38,516 23,291,828 24,265, ,441,652 77,000 23,518,652 25,787, Cildren's 1,984, ,222 2,253,890 2,381, Total 97,334, ,611 98,167, ,437,879 19

24 Table 5 Ratio Adjustments to NEISS Weigts January 1, 2001 troug December 31, 2001 Stratum 2000 ERVs from Estimated 2000 ERVs from Emergency Rooms ERVs from Ratio Emergency Rooms Total ERVs New to 2000 NEISS Adjustment on 1995 Frame Frame Sample 1 24,890, ,436 25,182,483 26,484, ,810, ,161 24,138,701 22,397, ,411, ,949 23,549,111 26,982, ,573, ,811 23,817,834 24,756, Cildren's 2,006, ,380 2,259,298 2,297, Total 97,691,690 1,255,737 98,947, ,919,163 20

25 Table 6 National Estimates of Te Total Number Of Product Related Injuries January 1 September 30, NEISS Hospital Sample 101 NEISS Hospital Sample Product Group Cases Estimate CV* Cases Estimate CV* Percent Cange Overall 225,692 9,258, ,649 8,436, Nursery Equip. 1,852 68, ,045 64, Toys 2, , , , Sports & Rec. 79,868 3,331, ,914 3,072, Home Entertain. 1,945 81, ,851 76, Personal Use 10, , , , Packaging 6, , , , Yard & Garden 4, , , , Home Worksop 5, , , , Home Maintain. 2,375 98, ,107 88, House Appliance. 2, , ,474 98, Heat/Cool/Vent 2, , ,249 89, Housewares 14, , , , Home Furnis 37,698 1,521, ,507 1,387, Home Struct 63,478 2,561, ,094 2,280, Miscellaneous 4, , , , * CV= coefficient of variation. Te coefficient of variation is te standard deviation of te estimate expressed as a proportion of te estimate. 21

26 TABLE 7 Number of Hospitals in te NEISS Sample and Number of Participating Hospitals By Stratum For Eac Mont Starting Wit January 1997 Small Stratum Medium Stratum Large Stratum Verge Large Cildren Stratum Total Year Mont Sample Participants Sample Participants Sample Participants Sample Participants Sample Participants Total Sample Total Participants

27 TABLE 7 Number of Hospitals in te NEISS Sample and Number of Participating Hospitals By Stratum For Eac Mont Starting Wit January 1997 Small Stratum Medium Stratum Large Stratum Verge Large Cildren Stratum Total Year Mont Sample Participants Sample Participants Sample Participants Sample Participants Sample Participants Total Sample Total Participants

28 Table 8 NEISS Weigts by Year, Mont, and Stratum From January 1997 to Present Year Mont Small Stratum Medium Stratum Large Stratum Very Large Stratum Cildren's Stratum 1997 JANUARY FEBRUARY MARCH APRIL MAY JUNE JULY AUGUST SEPTEMBER OCTOBER NOVEMBER DECEMBER JANUARY FEBRUARY MARCH APRIL MAY JUNE JULY AUGUST SEPTEMBER OCTOBER NOVEMBER DECEMBER JANUARY FEBRUARY MARCH APRIL MAY JUNE JULY AUGUST SEPTEMBER OCTOBER NOVEMBER DECEMBER JANUARY FEBRUARY

29 Table 8 NEISS Weigts by Year, Mont, and Stratum From January 1997 to Present Year Mont Small Stratum Medium Stratum Large Stratum Very Large Stratum Cildren's Stratum 2000 MARCH APRIL MAY JUNE JULY AUGUST SEPTEMBER OCTOBER NOVEMBER DECEMBER

30 Table 9 Generalized Relative Sampling Errors for NEISS for Estimates of Various Size Estimated Number of Injuries Approximate Standard Error Generalized Sampling Error Approximate 68% Confidence Interval Approximate 95% Confidence Interval 1, , ,720 5, ,250 5,750 3,500 6,500 10,000 1, ,700 11,300 7,400 12,600 25,000 2, ,250 27,750 19,500 30,500 50,000 4, ,500 54,500 41,000 59,000 75,000 6, ,250 81,750 61,500 88, ,000 8, , ,000 84, , ,000 10, , , , , ,000 12, , , , , ,000 14, , , , , ,000 16, , , , , ,000 21, , , , , ,000 28, , , , , ,000 35, , , , , ,000 42, , , , , ,000 49, , , , ,000 1,500,000 90, ,410,000 1,590,000 1,320,000 1,680,000 26

31 APPENDIX 1 SAS CODE FOR CALCULATING VARIANCES %include 'g:\users\epds\rdwrite\formats\stratum.fmt'; /**************************************/ /* STRATUM FORMATS */ /* 1990 Estimates - $strt90. */ /* Estimates - $strt91. */ /* 1997-present Estimates - $strt97. */ /**************************************/; DATA IN ERROR(KEEP = HID WT STRATUM DUMMY); SET g.neiss; id = put(id,$subosp.); stratum = put(id,$strt97.); dummy = 1; WT = WT/10000; *if wt as a format of 7.4 delete tis line; if wt<1 ten delete; if stratum in ('C','S','M','L','V') ten output in; else output error; PROC PRINT DATA = ERROR; TITLE 'Possible Error! Hospital wit no stratum'; PROC SORT DATA = IN; BY HID DUMMY; /**************************************************************/ /* SUM THE TOTAL WEIGHT AND TOTAL NUMBER OF CASES BY HOSPITAL */ /**************************************************************/ ; DATA TOTAL (KEEP = HID TOT_WT TOT_CNT DUMMY STRATUM); SET IN; BY HID DUMMY; TOT_WT + WT; TOT_CNT + 1; IF LAST.DUMMY THEN DO; OUTPUT; TOT_WT = 0; TOT_CNT = 0; END; RETAIN TOT_WT 0 TOT_CNT 0; PROC SORT DATA = TOTAL; BY DUMMY STRATUM; DATA VARPROD (KEEP = ESTIMATE COUNT CV VAR); SET TOTAL; 27

32 BY DUMMY STRATUM; COUNT + TOT_CNT; TWCNT + TOT_WT; SUMSQ = (TOT_WT)**2; TSUMSQ + SUMSQ; IF LAST.STRATUM THEN DO; IF STRATUM ='S' THEN N = 47; ELSE IF STRATUM = 'M' THEN N = 13; ELSE IF STRATUM = 'L' THEN N = 8; ELSE IF STRATUM = 'V' THEN N = 23; ELSE IF STRATUM = 'C' THEN N = 8; STR_VAR = (TSUMSQ-(TWCNT**2)/N)*N/(N-1); VAR+STR_VAR; TSUMSQ = 0; ESTIMATE+TWCNT; TWCNT=0; END; IF LAST.DUMMY THEN DO; SD = SQRT(VAR); CV = SD/ESTIMATE; OUTPUT VARPROD; ESTIMATE = 0; VAR = 0; COUNT = 0; END; RUN; PROC PRINT DATA = VARPROD; TITLE 'VARIANCE ESTIMATES FOR NEISS'; VAR ESTIMATE COUNT CV VAR; FORMAT ESTIMATE 7.0 CV 7.4 VAR 12.; SUM ESTIMATE COUNT VAR; RUN; 28

33 APPENDIX 2 SUDAAN CODE FOR CALCULATING VARIANCES Example to calculate estimates and coefficient of variations of NEISS estimates by gender DATA ALL; INPUT HOSPITAL $CHAR3. STRAT $CHAR1.; CARDS; 1 M 2 S 3 S 4 S 5 V 6 S 7 M 8 S 9 S 10 V 11 S 12 S 13 S 14 S 15 S 16 M 17 S 18 V 19 L 20 S 21 S 22 S 23 S 24 S 25 L 26 V 27 M 28 S 29 S 30 M 31 V 32 C 33 M 34 S 35 S 36 S 29

34 37 S 38 S 39 S 40 S 41 S 42 S 43 S 44 S 45 S 46 M 47 V 48 V 49 S 50 S 51 V 52 L 53 V 54 S 55 L 56 M 57 V 58 V 59 V 60 V 61 M 62 M 63 L 64 L 65 L 66 C 67 L 68 L 69 V 70 V 71 C 72 C 73 C 74 C 75 C 76 C 77 S 78 S 79 S 80 V 81 M 82 S 83 V 84 M 30

35 85 V 86 S 87 S 88 S 89 S 90 V 91 S 92 V 93 M 94 S 95 S 96 V 97 V 98 S 99 S 100 V 101 S ; PROC SORT DATA=DATASET; BY STRAT PSU; PROC SORT DATA=ALL; BY STRAT PSU; /* CREATE AT LEAST ONE RECORD FOR EVERY HOSPITAL. */ /* DATASET MUST CONTAIN VARIABLES NAMED PSU AND STRAT. */ DATA ALL; MERGE DATASET (IN=A) ALL (IN=B); BY PSU STRAT; IF ^A AND B THEN IN_STUDY = 0; IF A THEN IN_STUDY = 1; OUTPUT; PROC SORT; BY STRAT PSU; DATA ALL; FORMAT STRATUM 1.; IF STRAT = 'S' THEN STRATUM = 1; IF STRAT = 'M' THEN STRATUM = 2; IF STRAT = 'L' THEN STRATUM = 3; IF STRAT = 'V' THEN STRATUM = 4; IF STRAT = 'C' THEN STRATUM = 5; IF STRAT = ' ' THEN STRATUM =.; PROC SORT; BY STRATUM PSU; /***************************************************/ /* DATA SET IS NOW SORTED BY STRATUM PSU */ /* READY FOR USE IN SUDAAN */ /***************************************************/; 31

36 /* SUDAAN ESTIMATES OF TOTALS */; DATA _NULL_; TITLE1 'VARIANCE ESTIMATES FOR GENDER'; TITLE2 'SUDAAN: DESIGN = WR'; PROC DESCRIPT DATA = "ALL" FILETYPE = SAS DESIGN = WR; SUBPOPN IN_STUDY = 1; NEST STRATUM PSU; WEIGHT WT; SUBGROUP DISP; LEVELS 2; VAR NEISS; TABLES SEX; PRINT NSUM TOTAL SETOTAL / TOTALFMT=F10. SETOTALFMT = F10. STYLE = NCHS; OUTPUT NSUM TOTAL SETOTAL / TOTALFMT=F10. SETOTALFMT = F10. FILENAME = SUDAAN1; DATA SUDAAN1; SET SUDAAN1; FORMAT CV 7.4; CV = SETOTAL/TOTAL; PROC PRINT; VAR SEX TOTAL NSUM CV SETOTAL; RUN; 32

37 REFERENCES Ault, Kimberly & Scroeder, Tom, Updated NEISS Weigts Using 2000 SMG Hospital Frame, U.S. Consumer Product Safety Commission, May 12, Cocran, W. G. Sampling Tecniques, 3 rd ed., Capter 6, Jon Wiley, New York, Kessler, Eileen & Scroeder, Tom, National Electronic Injury Surveillance System (NEISS) Estimated Generalized Relative Sampling Errors, U.S. Consumer Product Safety Commission, September Marker, David & Lo, Annie, Update of te NEISS Sampling Frame and Sample, Westat, October 11, Marker, David, Lo, Annie, Brick, Mike, & Davis, Bill, Westat, Inc. Comparisons of National Estimates from Different Samples and Different Sampling Frames of te National Electronic Injury Surveillance System (NEISS), Westat, January 25, Scroeder, Tom, National Electronic Injury Surveillance System (NEISS) Improving Estimated Generalized Relative Sampling Errors, U.S. Consumer Product Safety Commission, April Scroeder, Tom, Updated NEISS Weigts Using 1998 SMG Hospital Frame, memorandum, U.S. Consumer Product Safety Commission, February 12, Scroeder, Tom, Updated NEISS Weigts Using 1999 SMG Hospital Frame, memorandum, U.S. Consumer Product Safety Commission, January 19, Scroeder, Tom, Trend Analysis of NEISS Data, U.S. Consumer Product Safety Commission, February