Document dDYDN9JBy4oRakELM5411e1Gb

DEPARTMENT OF HEALTH & HUMAN SERVICES CflO-G3>-?l-a* 01S-S-AS Public Health Service Centers for Disease Control Atlanta GA 30333 FTS 236-4111 December 7, 1982 David E. Weil, Ph.D. Project Manager, Environmental Criteria and Assessment Office (MD-52) U. S. Environmental Protection Agency Research Triangle Park, North Carolina 27711 Dear David: Enclosed is the comments on the DuPont Analysis of the NHANES II blood lead data which was done by Dr. James Pirkle of my staff. I understand that Dr. Annest is separately and Specifically addressing the Lucas paper on overascertainment. We will send separately a more detailed analysis as it relates to Hammond and Bornchein concerns about the reference laboratories. That concern has no bearing on the blind quality control that was in place during the analysis of the majority of the NHANES II data, which is discussed . on page 11 of the Pirkle comments. Would you please acknowledge rec'eipt of this material. See you on the 18th. Sincerely yours, Enclosure Vernon N. Houk, M.D. Acting Director Center for Environmental Health N33897 TEH 0533288 Subject: Comments on the Dupont Analysis of the NHANES XI Blood Lead Data From; James L. Pickle M.D. Ph.D. Clinical Chemistry Division Center for Environmental Health Centers for Disease Control Date: December 1, 1982 N33897.01 TEH 0533289 DUP050034556 *' INTRODUCTION 2 Recent comments from Dupont (I) on the NHANES II data warrant careful analysis. Dupont asserts that conclusions drawn on the NHANES II data are "suspect" because (1) demographic characteristics of subject sample groups changed during the study and (2) analytical error leads to overestimation of the percentage of the population with blood lead levels of 30 or more. In addition, DuPont claimed there were "several potential difficulties" in the blind quality control data that "could impact on the reported interpretations with NHANES II trends," The points concerning accounting for demographic variables and the "potential difficulties" in the blind quality control are summarized and discussed below. The question of analytical error leading to an overes timation of children with lead toxicity is addressed here and by Annest (2). SUMMARY (1) After accounting for the variability of demographic variables and incorporating the complex survey design, a statistically significant downward trend of 31.12 (blacks and whites together) is still present in the NHANES II lead values over the period of the survey. This drop is noted in blacks (27.12), whites (31.82), and in different age groupings of whites: .5-5 yrs (32.1%), 6-17 yrs. (31.3%) and 18-74 yrs, (35.4%). (Blacks were not broken into age groupings due to small sample sizes.) (2) Changes in leaded paint consumption is an unlikely explanation for the downward trend in blood lead levels since the trend is also present in adults. (3) The best available data on dietary lead intake indicate that lead in the diet certainly was not declining over the survey period and may have been rising* This makes changes in dietary lead intake also an unlikely explanation of the downward trend. (4) After accounting for the variability of demographic variables and incorporating the complex survey design, lead in gasoline is highly statistically significant (p < ,0001) and alone can account for the magnitude of the downward trend in lead values seen with time. Over the survey period, changes in gasoline lead account for the following decreases in blood lead levels: blacks and whites - 34.4%, blacks - 25.6%, whites - 35.3%, whites .5-5 yrs old - 34.7%, whites 6-17 yrs old - 34.9% and whites 18-74 yrs old -- 34.0%. (5) Multiple regression analysis of the NHANES II lead data must incorporate the complex survey design. Otherwise, all of the F statistics in the regression will be incorrect. The DuPont analysis did not incorporate the complex survey design;. TEH 0533290 DUP050034557  { 3 (6) The contention that the downward trend can he explained by variability in the sampling; of urban areas (cities with greater than 1 million population) has previously been shown to be incorrect (3). In fact, in subpopulations with no urban dwellers ( i. e. small cities with less than 1 million population and rural areas), the trend is still present. Furthermore, after accounting for urbanization, the trend is still present in the total population (see item (1) above), (7) The NHANES II data either correctly estimate or underestimate the number of persons with lead toxicity. Inasmuch as the original "toxicity" criteria of 30 ug/dl was based on blood lead data which included measurement error, the only pertinent question is Whether the measurement error in the NHANES II data is more or less than the data on which the 30 ug/dl cutoff was established. If the NHANES II measurement error were higher, then there would be potential for overestimating the number of persons with lead toxicity. Actually, the measurement error is equal to or less than the measurement error from standard or reference laboratories and thus the data either correctly estimate or underestimate the number of persons with lead toxicity. (8) The blind quality control data show no significant downward trend in lead values due to laboratory measurement. TEH 0533291 DUP050034558 4 DISCUSSION ACCOUNTING FOR THE DOWNWARD TREND IN THE NHANES II BLOOD LEAD DATA Demographic Variables Age, sex, race and urbanization are variables known to effect blood lead levels. Ideally, standard stepwise regression techniques could be used on the 9937 lead values to control for the influence of these variables. However, the NHANES II data was collected utilizing a stratified sampling design which has to be incorporated into the regression analysis for proper estimation of errors. The regression program SURREGR (4) permits multiple regression analysis incorporating the effects of the survey design. In the current analysis, the dependent variable was the natural log of blood lead. Originally, the lead data was analyzed with SURREGR in subgroups broken down by age, sex, race, urbanization, income level, season, and geographic location. The downward trend was consistently greater than 30% and statistically significant in all subgroups except urban areas with more than one million population. Normally, there would be plenty of degrees of freedom to include all these variables plus their interactions in a stepwise linear regression approach. However, incorporating the complex survey design reduces the degrees of freedom to 32. Thus some selection of variables is necessary. The following method was used to select the most significant variables to enter into SURREGR: (1) indicator variables were constructed representing: -- sex -- age (.5-5 yrs, 6-17 yrs, 18-74 yrs) -- urbanization (1 million or more, less than 1 million, rural) -- SMSA (whether the person lived in or out of an SHSA) -- all the first order interactions of the above variables (2) A model using these variables was fit using stepwise linear regression (5) with an F-to-enter of 2.7 and F-to-remove of 2.7 (6), In this manner, variables with even marginal significance would not be excluded' but clearly insignificant variables could be dropped. This process did not incorporate the complex survey design but consistently included more variables than found to be significant following subsequent analysis with SURREGR. (3) The remaining variables were then used with SURREGR. Depending on the regression, the variable "time" or the variable "gas" was entered in both a linear and quadratic term. A manual backward elimination was performed with SURREGR, deleting the least significant variable in each regression run until all variables were significant at approximately the .05 level. This elimination procedure had very little effect on the coefficients of "time" or "gas". TEH 0533292 DUP050034559 Regression on TIME 5 When time and time-squared were entered as variables along with the demographic variables mentioned above, time-squared was consistently significant at the p <.0001 level. The regression results by race and selected age groups are given in Appendix A, After accounting for the demographic variables, the downward trend over time was still present. The amount of the downward trend is given by race and selected age groups in Table Table 1 Regression Results for the "Time" Variable After Controlling Age, Sex, Urbanization and SMSA Blood Lead at Start of Survey Blacks & Whites Blacks Whites Whites .5-5 yrs Whites 6-17 yrs Whites 18-74 yrs 14,45 ug/dl 16.23 14.24 14.26 14.79 13.38 Blood Lead at End of Survey 9.96 11.84 9.71 9.69 10.17 8.64 Difference 4.50 4.39 4.53 4.57 4.62 4.74 Percent Decrease 31.1 27.1 31.8 32.1 31.3 35.4 Lead in Paint In attempting to account for this downward trend, the different major sources of lead exposure must be examined. Lead in paint is a well known source of lead exposure for children, However, the downward trend also occurred in teens and adults which makes decreased leaded paint consumption an unlikely explanation of the trend. Lead in the Diet The Food and Drug Administration (FDA) has measured lead in its Total Diet ("Market Basket") Studies since 1973 (7). The survey includes 120 individual food items representative of the 2-week diet of a teenage male for the regions of the country in which it is collected. The results from 1976 through 1980 are given in Table 2, TEH 0533293 DUP050034S60 table 2 Lead in the Diet (micrograms/day) teenage Male 1976 1977 1978 1979 1980 71.1 79.3 95.1 81.7 82.9 6 No downward trend is present, suggesting that diet is also an unlikely explanation of the decrease in blood lead levels seen over the course of the NHANES II survey. Lead in gasoline the only other widespread source of lead exposure (other than occupational) is from leaded gasoline, the amount of lead used in gasoline production across the nation is available from the quarterly refiner's reports to EPA. During a 3-month period, the NHANES II caravans on-the-average sampled 4 different sites in the USA and over a 6-month period they sampled on-the-average 8 sites. Three gas variables were tested to cover both the 3-month and 6-month breakdown of gasoline lead data: (1) Gasl * amount of lead used in gasoline for each of the quarters of the year during the NHANES II survey (2) Gas2 * amount of lead used in gasoline for the January-June and ' July-Dscember 6-raonth periods during the NHANES II survey (3) Gas3 amount of lead used in gasoline for the April-September and 0ctober-March 6-month periods during the NHANES II survey The regression procedure included the demographic variables itemized in Appendix A and was performed separately for each of these gas variables. A "gas-squared" term was also permitted to enter each model to account for a curvilinear relation. The regression results by race and selected age groups are given in Appendix A. The amount of downward trend attributable to gasoline lead is tabulated by race and selected age groups in Table 3 and the gas variable with the generally most conservative results (Gasl) is shown in Figure 1. AFTER ACCOUNTING FOR DEMOGRAPHIC VARIABLES, THE GASOLINE LEAD VARIABLE (GAS1, GAS2 OR GAS3) WAS ALWAYS HIGHLY STATISTICALLY SIGNIFICANT (p .0001) AND ALONE COULD ACCOUNT FOR THE DOWNWARD TREND IN LEAD VALUES OVER TIME SEEN IN THE NHANES II SURVEY. TEH 0533294 DUP050034561 PERCENT DECREASE IN BLOOD LEAD VALUES EXPLAINED B Y LEAD NHANES II BLOOD LEAD DATA AFTER ACCOUNTING FOR AGE, SEX, URBANIZATION, SMSA FIGURE 1 DUP050034562 8 Table 3 Regression Results for the Three Gas Variables After Accounting for Age, Sex, Urbanisation and SMSA Blood lead during initial period of NHAKOES survey* Blacks & Whites Gasl Gas2 Gas3 Blacks Gasl Gas2 Gas3 Whites Gasl Gas2 Gas3 Whites .5-5 yrs Gasl Gas2 Gas3 Whites 6-17 yrs Gasl Gas2 Gas3 Whites 18-74 yrs Gasl Gas2 Gas3 14.72 14.94 14.80 16.18 16.72 16.31 14.53 14.76 14.68 14.44 14.65 14.60 15.03 15.33 15.22 13.36 13.75 13.65 Blood lead during final period of WHAWES survey* 9.67 9.43 9.96 12.05 11.66 11.84 9.40 9.19 9.47 9.43 9.27 9.49 9.79 9.63 9.93 8.82 8.34 8.69 Difference Percent decrease 5.06 5.50 4,97 4.13 5.07 4,39 5.13 5.58 5.21 5.01 5.38 5.10 5.25 5.70 5.29 4.54 5.41 4.97 34,4 36.9 33,6 25.6 30.3 25.5 35.3 37.8 35,5 34.7 36,7 35.0 34.9 37.2 34.8 34.0 39.3 36.4 *see text for definition of 3-month initial period of Gasl and 6-month initial period of Gas2 and Gas3. Importance of Incorporating the Complex Sample Design into the Regression The complex sample design must be considered in the regression analysis Or the F statistics will be incorrect. The DuPont analysis did not incorporate the complex sample design. TEH 0533296 DUP050034563 9 Explaining the Downward Trend by Variability in Sampling of the Urban Areas (Cities with more than one million population) :' This contention has previously been shown to be incorrect. In fact, in subpopulations with no "urban dwellers" (i.e. small cities with less than one million population or rural areas), the trend is still present. Furthermore, after accounting for urbanization as a demographic variable the trend is still present in the total population (see analysis above). IN SUMMARY, after accounting for all these demographic variables and considering alternative etiologies, lead in gasoline remains very significantly related to blood lead levels and the most reasonable explanation of the downward trend in the NHANES II lead data. THE EFFECT OF ANALYTICAL ERROR ON ESTIMATING THE PERCENTAGE OF THE POPULATION WITH LEAD TOXICITY FROM THE NHANES II DMA ~ ................. "Lead toxicity" is a medical rather than a numeric determination. For example, suppose some unusual systematic bias pervaded all lead measurements causing them to be two times their true value. The same people who medically had "lead toxicity*11 would still have it, only the lead value with which they were associated would be different. This concept is important when considering the 30 ug/dl cutoff for defining "lead toxicity". Physicians examined and followed individuals and determined that those with lead values of 30 ug/dl or more medically had "lead toxicity". These blood lead values, on which the 30 ug/dl cutoff was based, came from several laboratories and all contained measurement error. If the measurement error was removed from this data, the cutoff of 30 ug/dl would decrease to a lower lead value so, for example, 23 or 27 ug/dl would be the cutoff value for "lead toxicity". This has not happened. No one has removed measurement error from the data on which the 30 ug/dl cutoff was based. Consequently, the laboratory determination of "lead toxicity" by blood lead levels is based on data which includes measurement error. The essential point in determining if the NHANES II lead data falsely classify persons "lead toxic" due to laboratory measurement error, is the comparison of the amount of measurement error in NHANES II data relative to the amount of 'measurement error in the studies on which the 30 ug/dl cutoff was based. If the NHANES II measurement error is greater than the measurement error of these other studies, then some persons would be falsely classified "lead toxic" due to this relatively greater measurement error (i.e. false positives). If the NHANES II measurement error is less than the measurement error of these studies on which the 30 ug/dl cutoff was based then the NHANES II data underestimates the number of persons who should be classified "lead toxic". TEH 0533297 DUP050034564 10 To make this comparison, the measurement error of the studies on which the 30 ug/dl cutoff was based needs to be known. This cannot be determined exactly since multiple studies using several laboratories influenced the setting of the 30 ug/dl cutoff. However, the NHANES II measurement error can be compared to the error from standard and reference laboratories which provide good estimates of what the error was in the studies of interest. The Centers for Disease Control (CDC) conducts a Blood Lead Proficiency Testing Program which has estimates of the coefficients of variation for standard end reference laboratories. The 1980 report (8) has data which are near the mean levels for the two NHANES II blind quality control pools (13.5 ug/dl and 25.5 ug/dl). This data was compiled from about 27 reference laboratories and about 150 standard laboratories. The coefficient of variation of lead measurements is known to increase as the magnitude of the lead value decreases. A comparison of these standard and reference labs with the NHANES II data is given below. Table 4 Comparison of Measurement Error in NHANES II Data with that of Standard and Reference Laboratories Standard Laboratory Reference Laboratory NHANES 11 Lead Data Lower Lead Level Mean (ug/dl) C.V. (%) not available 15.9 24.7 13.5 16.0 Higher Lead Level Mean (ug/dl) C.V. (%) 25,1 24.8 25.5 27.9 14.4 12.5 The important conclusion from Table 4 is that the NHANES II measurement error is at least as good as and probably better than standard or reference laboratories. Without any additional calculations, it can be concluded that RELATIVE TO THE STUDIES ON WHICH THE 30 UG/DL CUTOFF FOR "LEAD TOXICITY" WAS ESTABLISHED, THE NHANES II DATA.EITHER CORRECTLY ESTIMATE OR UNDERESTIMATE THE NUMBER OF PERSONS WITH "LEAD TOXICITY". It is altogether a separate issue as to whether measurement error correction procedures should be implemented on lead data from all studies. If so, it would certainly be incorrect to adjust the NHANES II data for measurement error and not adjust the data on which the 30 ug/dl cutoff was based. This would result in falsely assigning 30 ug/dl as the cutoff for "lead toxicity" in the NHANES II data when the actual measurement-error-free cutoff would be less than 30 ug/dl. In summary, after considering the effects of measurement error, the NHANES II data either correctly estimate or underestimate the number of persons with lead toxicity. TEH 0533298 DUP050034565 11 QPALIIY CONTROL DATA The SHANES II lead samples were covered by two independent quality control systems: a bench and a blind quality control. The bench quality control started at the beginning of the survey and finished at the end. The blind quality control began approximately 12 months after the survey began and continued uninterrupted through the end of the survey. Approximately 902 of the drop in lead levels occurred while both bench and blind quality control were in effect. If the average of replicate values of either bench or blind quality control specimens fell outside of their esablished 95% confidence limits, the run was repeated. Also, if replicate absorbance values for any given specimen differed by more than 0,25 absorbance units or the difference between calculated concentrations for duplicates was greater than 7 ug/dl, the analysis was repeated. The bench quality control utilized a number of bovine pools over the course of the survey so a long term trend in lead values could not be statistically tested on this data. The blind quality control used two pools simultaneously for the entire period of blind quality control (a high pool with mean of 25.5 ug/dl and a low pool with mean of 13.5 ug/dl). In order to statistically test for a trend in the entire blind quality control, the high and the low pools were combined. First, the lead data from the low and high blind pools were checked to insure they were approximately normally distributed. Then the values in each pool were transformed using a z transformation; that is, (lead value -'lead pool mean)/lead pool standard deviation. Finally, the transformed data from each pool, having mean of zero and variance of one, were combined and linear regression of z on "time" was performed. The "time" variable was defined to correspond with the chronology of the collection of the SHANES II specimens. No significant linear trend was found from the regression (p.69, 11=544). DuPont separately analyzed the low pool and high pool and found that the high pool rose very slightly over time and the low pool declined very slightly over time and that there were statistically significant curvilinear terms in each pool when analyzed separately. We tested the entire blind quality control, and though there was no statistically significant linear trend with time, a Statistically significant (p=.Q001) quadratic term was found. This quadratic term accounts for a slight dome curvature in the blind quality control data but does not account for any decline in the blood, lead levels with time. On the contrary, including this quadratic term with the linear term predicts a very small net increase in lead levels of .09 ug/dl from the beginning to the end of the blind quality control data. If a first-order correction is applied to the data to compensate for the slight dome curvature, the average absolute amount of the correction is only about 1.8% of the lead value. The adjustment changes the Coefficients of the ''time" or "gas" variable by about 3%. For the sake of detail, all the lead results in this report include this slight adjustment. In summary, a small statistically significant curvature is present in the blind quality control data. However, it does not explain any of the downward trend seen in the NHANES blood lead data nor diminish the significance of the relationship between blood lead levels and gasoline lead. TEH 0533299 DUP050034566 12 CONCLUSIONS After considering the laboratory quality control and the effect of demographic variables, a statistically significant downward trend of about 31% remains in the NHANES II blood lead levels over the course of the survey. Lead in paint is an unlikely explanation of this trend since the trend is present in adults, the best available data indicate lead in the diet did not decrease over the survey period making it also an unlikely explanation of the trend. The only other widespread source of lead (other than occupational) is leaded gasoline. Multiple regression analysis accounting for demographic variables and incorporating the complex survey design indicates that lead in gasoline is significantly related to NHANES II blood lead levels and it alone can explain the entire magnitude of the downward trend seen in the NHANES II blood lead data. TEH 0533300 DUP050034567 13 Appendix A Regression Variables The variables for tbe regression model were constructed as follows: Characteristic male female age .5-5 yrs age 6-17 yrs age 18-74 yrs live in SMSA live outside SMSA live in city with one million or more pop. live in city with less than million pop. live in rural area Interaction terms: maleout " male X outsmsa smallout " small X outsmsa malesmal " male X small malerurl " male X rural malecbild * male X child maleteen " male X teen cbildsmal * child X small teensmal " teen X small childrural " child X rural teenrurl " teen X rural cbildout " child X outsmsa teenout " teen X outsmsa Variable value male*! male"0 ebild*l teen"0 cbild"0 teenml child"0 teen*0 outsmsa"0 outsmsa"! small^O rural*0 small"! rural"0 smal1*0 rural"! For the regression on blacks, whites or blacks + whites, all of the above terms were initially in the model. For regression on whites .5-5 yrs, whites 6-17 yrs or whites 18-74 yrs, tbe population was already divided into an age group so no age variables were used in the model. Regression Results on "Time11 Variable Time and time-squared were entered as variables together with tbe demographic variables listed above. Time was defined as the number of days from 2/20/76 to the sample collection date divided by 2800. This reduced the magnitude of the time variable to approximately that of tbe 0-1 indicator variables. Some TEH 0533301 DUP050034568 14 of tbe final models contain interaction terms without the corresponding main effects. All of tbe main effects were reintroduced into each of these final models and the regression rerun to insure that the omission of these main effects did not significantly alter tbe "time" coefficients or tbe percent decrease in lead calculated over time (Table 1). Tbe "time" coefficients cbanged less than 2% and tbe percent decrease in lead changed less than 0.5% (e.g. 35,5% to 35.0%). The final models resulted from tbe manual backward elimination procedure described on page 3. In each regression tbe denominator degrees of freedom Cdf) is 32. Results are given below for each race and selected age groups of wbites. Tbe time-squared variable is consistently significant. Group: Black & White Variable MALE MALECHILD CHILD MALETEEN TIMESO RURAL CHILDRURAL MALEODT SMALLOUT TEEN INTERCEPT F stat; 901.9 355.0 210.1 87,6 67.6 23.6 13.2 11.7 10.6 4.9 DF 1 1 1 1 1 1 1 1 1 1 Overall Model 256,2 10 Prob. 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0010 0.0017 0.0027 0.0335 0.0000 Coeff. 0.3420 -0.3361 0.3002 -0.1921 -1.3576 -0.1207 -0.0833 0.0631 -0.0965 -0.0387 2.5692 Variance 0.0001300 0.0003200 0.0004300 0.0004200 0.0272500 0.0006200 0.0005300. 0,0003400 0.0008800 0,0003000 0,0004700 Group: Black Variable MALE CHILD MALECHILD TIMES0 ma l e t e e n RURAL SMALLOUT OUTSMSA INTERCEPT Overall Model F Star. 211.0 120.8 52.9 51,7 23.2 10.5 6.7 4.3 DF 1 1 1 1 1 1 1 1 72,6 8 Prob. 0.0000 0.0000 0,0000 0.0000 0.0000 0.0028 0.0142 0.0456 0,0000 Coeff. 0.4035 0,4690 -0.4163 -1,1492 -0.2073 -0.2170 -0.2205 0.1616 2.6232 Variance 0.0007700 0.0018215 0.0032780. 0,0255330 0.0018554 0.0044866 0.0072169 0.0060313 0.0011671 TEH 0533302 DUP050034569 Group: White Variable MALE MALECHILD CHILD MALETEEN TIMESQ RURAL TEEM SMALL CHILDRURAL INTERCEPT F Stat; 1334.1 245.3 109.3 59.6 54.8 11.4 7.4 6.4 4.5 DF 1 1 1 1 1 1 1 l l Overall Model 146.7 9 Prob. 0.0000 0.0000 0.0000 0.0000 0.0000 0.0020 0.0105 0.0165 0.0422 0.0000 Coeff. 0.3619 -0.3229 0.2425 -0.1880 -1.3938 -0.1011 -0.0540 -0.0581 -0.0456 2.5664 Variance 0.0000980 0.0004300 0.0005400 0.0005900 0.0354706 0.0009000 0.0004000 0.0005300 0.0004600 0.0005300 15 Group: White .5-5 yrs. Variable MALE TIMESQ RURAL INTERCEPT F Stat. 1212.7 55.6 8.9 DF 1 1 1 Overall Model 392.7 3 Prob. 0,0000 0.0000 0.0055 0.0000 Coeff, 0.3169 -1.4079 -0.0683 2.5218 Variance 0.0000830 0.0356367 0.0005300 0.0004100' Group: White 6-17 vrs. Variable MALE TIMESQ RURAL SMALL INTERCEPT F Stat. 1358.2 49.1 12.6 7.7 DF 1 I 1 1 Overall Model 330.6 4 Prob* 0.0000 0.0000 0.0012 0.0092 0.0000 Coeff; 0.3426 -1.3654 -0.1043 -0.0614 2.5793 Variance 0.0000860 0.0379896 0.0008600 0.0004900 0.0004300 Group: White 18-74 yrs. Variable MALE TIMESQ RURAL INTERCEPT F Stat. 58.3 47.0 5,5 DF 1 1 1 Overall Model 35.3 3 Prob. 0,0000 0,0000 0.0258 0.0000 Coeff. 0.1490 -1.5944 . -0.0670 2.5452 Variance 0.0003800 0.0540572 0.0008200 0.0009100 TEH 0533303 DUP050034570 16 Regression on the Lead in Gasoline Variables Eacb of tbe gas variables is in units of 1000 tons of lead divided by 100 to reduce tbe magnitude of-the variable to approximately that of tbe 0-1 indicator variables. As with the time regressions, some of the final models contain interaction terms without tbe corresponding main effects. All of the main effects were reintroduced.to these final models and tbe regression rerun to insure that their omission did not significantly alter the "gas" coefficients or tbe percent decrease in blood lead levels accounted for by lead in gasoline. Tbe "gas" coefficients changes less than 3.5% and the percent decrease in blood lead levels accounted for by gasoline lead changed less than 1.1% (e.g. 34.4% to 33.3%). Tbe final models resulted from a manual backward elimination procedure (decribed on page 3) performed separately for each of tbe three gas variables. In eacb regression the denominator degrees of freedom (df) is 32. Results are given below for eacb race and selected age groups of whites. Tbe "gas" variable (Gasl, Gas2, or Gas3) is consistently significant. lead in Gasoline Variable: Gael Group: Black & White Variable MALE MALECHILD CHILD MALETEEN GAS1 MALERURL SMALL0UT TEENOUT MALEOUT INTERCEPT Overall Model F Stat; 928.2 327.6 200.4 128.5 89.6 13.8 5,6 4.4 4.0 DF 1 1 1 l 1 1 1 1 1 256.2 9 Prob, 0.0000 0.0000 0.0000 0.0000 0.0000 0.0008 0.0247 0.0434 0.0541 0.0000 Coeffi 0.37584 -0.33918 0.27565 -0.19898 1.42693 -0.08638 -0.05091 -0.07768 0.05167 1.78655 Variance 0.0001500 0.0003500 0.0003800 0,0003100 0.0227163 0.0005400 0,0004700 0.0013639 0.0006700 0.0044702 Group: Black Variable MALE CHILD GAS1 MALECHILD MALETEEN TEEN TEENOUT MALEOUT INTERCEPT Overall Model F Stati 196.3 145.3 91.2 68.2 18.5 6.4 3.0 2.5 DF i i i i i i i i 110,2 8 Prob; 0,0000 0.0000 0.0000 0,0000 0.0002 0.0166 0.0919 0.1288 0.0000 Coeffi 0.39318 0.48829 0.99997 -0.44610 -0.23137 0.07328 -0.17989 0.08737 2.06451 Variance 0.0007900 0.0016415 0.0109703 0.0029167 0.0028991 0.0008400 0.0107213 0.0030057 0.0008300 TEH 0533304 DUP050034571 / 17 Group: White Variable MALE MALECHILD CHILD GAS1 MALETEEN RURAL SMALL TEEN MALEO'OT INTERCEPT Overall Model F Stat. 674.9 250.3 126.9 104.2 59.1 22.3 13.3 7.7 6.1 DF 1 1 1 1 1 1 1 1 1 160.6 9 Prob. 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0009 0.0090 0.0193 0.0000 Coeff 0.34352 -0.32582 0.23235 1.47503 -0.18393 -0.13970 -0.10110 -0.05611 0.04796 1.83529 Variance 0.0001700 0.0004200 0.0004300 0.0208812 0.0005700 0.0008800 0.0007700 0.0004100 0.0003800 0.0038429 Group: White .5-5 yrs Variable MALE GAS1 RURAL SMALL MALEOUT INTERCEPT F Stat. 570.9 92.3 23.8 14.1 5.9 DF l 1 1 1 1 Overall Model 257.0 5 Prob; 0.0000 0.0000 0.0000 0.0007 0.0209 0.0000 Coeff. 0.29737 1.44410 -0.14286 -0.10304 0.05095 1,83879 Variance 0.0001500 0.0225929 0.0008600 0.0007500 0.0004400 0.0041133 Group: White 6-17 yrs. Variable MALE GAS1 RURAL SMALL INTERCEPT F Stat. 1174.3 114,6 20.6 14.9 DF 1 1 1 1 Overall Model 300.8 4 Prob; 0,0000 0.0000 0.0001 0.0005 0.0000 Coeff; 0.34271 1.45531 -0.12087 -0.09452 1.84415 Variance 0,0001000 0.0187422 0.0007100 0.0006000 0.0037919 Group: White 18-74 yrs. Variable MALE GAS1 RURAL INTERCEPT F Stat. 66.6 31.8 5.1 DF 1 1 1 Overall Model 35.5 3 Prob. 0.0000 0.0000 0,0312 0.0000 Coeff. 0.15381 1.40769 -0.07161 1.79792 Variance 0.0003600 0.0622089 0.0010101 0.0111279 TEH 0533305 DUP050034572 18 Lead In Gasoline Variable; GAS2 Group: Blacks & Whites Variable MALE MALECHILD CHILD GAS 2 MALETEEN RURAL MALEOUT CHILDRURAL TEEN SMALL SMALLOUT INTERCEPT F Stat; 909.4 377.9 234.0 100.6 95.8 23.8 15.1 14.9 5.4 3.8 3.6 OF 1 1 1 1 1 1 1 1 1 1 1 Overall Model 200.0 11 Prob. 0.0000 0.0000 0.0000 0,0000 0.0000 0.0000 0.0005 0.0005 0.0273 0.0598 0.0657 0.0000 Coeff; 0.34076 -0.33764 0.30608 0.77418 -0.19386 -0.14514 0.06695 -0.09014 -0.03967 -0.05578 -0.06046 1.80786 Variance 0.0001300 0.0003000 0.0004000 0.0059587 0.0003900 0.0008900 0.0003000 0.0005500 0.0002900 0.0008200 0.0010061 0,0048407 Black * Variable MALE CHILD MALECHILD GAS2SQ MALETEEN MALERURL MALEOUT SMALLOUT INTERCEPT Overall Model F Stat; 166.0 136.7 56,8 36,2 21.9 6.8 5.2 4.2 DF 1 1 1 1 vI 1 1 1 46.7 8 Prob. 0.0000 0.0000 0.0000 0,0000 0.0001 0.0134 0.0291 0,0479 0.0000 Coeff; 0.38893 0.47271 -0.42610 0.39912 -0.20361 -0.17387 0.15268 -0.10112 2.20442 Variance 0.0009100 0.0016347 0.0031954 0.0044039 0.0018978 0.0044112 0.0044635 0.0024159 0.0026534 White Variable MALE MALECHILD CHILD GAS 2 MALETEEN F- Stat; 748.7 262,3 129.9 83.6 64.0 DF 1 1 1 1 1 Prob. 0.0000 0.0000 0.0000 0,0000 0.0000 . Coeff; 0.34426 -0.32447 0.22909 0.79861 -0.18931 Variance 0.0001600 0.0004000 0.0004000 0.0076249 0.0005600 TEH 0533306 DUP050034573 RURAL SMALL TEEN MAlEOUT INTERCEPT Overall Model 19.5 8.7 8.2 6.6 165.4 1 1 1 1 9 0.0001 0.0059 0.0073 0.0148 0.0000 -0.12810 -0.07655 -0.05520 0.04425 1.77408 0.0008400 0.0006700 0.0003700 0.0002900 0.0063072 19 Croup: White .5-5 yrs. Variable MALE GAS 2 RURAL SMALL MALEOUT INTERCEPT F Stat, 621.6 71.9 20.3 9.0 5,7 t)F 1 1 1 1 1 Overall Model 247.8 5 Prob. 0.0000 0.0000 0.0001 0,0052 0.0230 0,0000 Coeff; 0.29721 0.77009 -0.13155 -0.07902 0.04678 1.78925 Variance 0.0001400 0.0082465 0.0008500 0.0006900 0.0003800 0.0068279 \ Group: Wbite 6-17 yrs Variable MALE GAS2 RURAL SMALL INTERCEPT F Stat. 1343.7 77.0 16.2 8.6 OF 1 1 l 1 Overall Model 315.5 4 Prob, 0.0000 0.0000 0.0003 0.0061 0.0000 Coeff. 0.34215 0.78245 -0.11078 -0,07120 1.79346 Variance 0.0000870 0.0079501 0.0007600 0.0005900 0.0067350 Group: White 18-74 yrs Variable GAS 2 MALE RURAL in t er c ept 1P Stat; 66.5 62.3 5.7 DF 1 1 1 Overall Model 45.5 3 Prob. 0.0000 0,0000 0.0235 0.0000 Coeff; 0,84111 0.14837 -0.07527 1.68533 Variance 0.0106363 0,0003500 0.0010012 0.0074795 TEH 0533307 DUP050034574 Lead in Gasoline variable: GAS3 Black & White Variable HALE MALECH1LD CHILD GAS 3 MALETEEN RURAL CHILDRURAL SHALL MALEOUT TEEN INTERCEPT F Stat'. 880.3 336, 3 212.8 102.8 85.3 23.2 12.8 12.7 5.3 4.3 DF 1 1 1 1 1 1 1 1 1 1 Overall Model 185.9 10 Prob. 0.0000 0.0000 0.0000 0,0000 0.0000 0.0000 0.0011 0,0012 0.0279 0.0462 0,0000 Coeff:. 0,34937 -0.33742 0.30832 0.68839 -0.18984 -0.14733 -0.08462 -0.09803 0.04429 -0.03654 1.88479 . Variance 0.0001400 0.0003400 0.0004500 0.0046114 0. 0004200 0.0009300 0.0005600 0.0007600 0.0003700 0.0003100 0.0038981 20 Group: Black Variable MALE CHILD GAS3 MALECHILD MALETEEN INTERCEPT F Stat'. 204.2 144.3 103,5 64.6 16.9 DF 1 1 1 1 1 Overall Model 167.6 5 Prob. 0.0000 0.0000 0.0000 0.0000 0.0003 0.0000 Coeff; 0.39472 0,47893 0.49571 -0.47215 -0.19477 2.08256 Variance 0.0007600 0,0015895 0.0023736 0.0028230 0.0022499 0.0013784 Group: White Variable HALE MALECHILD CHILD GAS3 MALETEEN RURAL SMALL TEEN MALEOUT INTERCEPT F Stat. 641,2 238.6 121.2 101,0 58.1 20.9 14.8 6.8 5,4 DF 1 1 1 1 1 1 1 1 1 Overall Model 160.7 9 Prob. 0.0000 0,0000 0.0000 0.0000 0,0000 0.0001 0.0005 0.0136 0,0263 0.0000 Coeff. 0.34266 -0.32357 0.23069 0.73855 -0.18459 -0.13380 -0.10004 -0.05338 0.04778 1.82802 Variance 0.0001800 0.0004400 0.0004400 0.0054024 0.0005900 0.0008600 0.0006800 0.0004200 0.0004200 0.0042439 TEH 0533308 DUP050034575 . --r > 21 Group: White .5-3 yrs* Variable MALE GAS 3 RURAL SMALL MALEOUT INTERCEPT F Stat. 554.3 91,9 22.4 15.8 6.0 DF 1 1 1 1 1 Overall Model 250,3 5 Prob. 0,0000 0.0000 0.0000 0.0004 0.0200 0.0000 Coeff. 0.29635 0.72462 -0.13704 -0.10193 0.05124 1.83083 Variance 0.0001600 0.0057151 0.0008400 0.0006600 0.0004400 0.0044237 Group: White 6-17 yrs. Variable MALE GAS 3 RURAL SMALL INTERCEPT F Stat. 1151,6 105.1 16.6 14,1 DF 1 1 1 1 Overall Model 280.2 4 Prob; 0.0000 0.0000 0.0003 0.0007 0.0000 Coeff; 0.34204 0.71918 -0.11581 -0.09344 1,84978 Variance 0,0001000 0.0049208 0.0008100 0.0006200 0.0042993 Group: White 18-74 yrs. Variable MALE GAS 3 RURAL INTERCEPT F Stat; 65.4 32.8 3.9 DF 1 1 1 Overall Model 36.9 3 Prob, 0.0000 0.0000 0.0578 0.0000 Coeff; 0.15290 0.76154 -0.06269 1.74183 Variance 0.0003600 0.0177065 0.0010148 0.0124679 TEH 0533309 DUP050034576 sf 22 REFERENCES 1) Comments bv E.I. du Font de Nemours and Co. in response to EPA's proposed rule regarding lead in gasoline. October 7, 1982. 2) Annest J.L., et al., manuscript in preparation 3) Pirkle J.L., Sampling of tbe Urban Population and tbe Downward Trend in tbe NHANES II Blood Lead Levels, August 24, 1982 (EPA docket) 4) Bolt M.M., Revised by Sbab B.V., SURBEGR: Standard Errors of Regression Coefficients from Sample Survey Data. Research Triangle Institute, Research Triangle Park, North Carolina April 1982. 5) BMDP Statistical Software, Stepwise Regression Program P2R, 1981. 6) Neter J, Wasserman W, Applied Linear Statistical Models, p. 382, 1974. 7) Jelinek C.V., "Levels of Lead in the U.S. Food Supply" presented at the Analytical Methodology for Lead ip Foods Symposium, Assoc, of Official Analytical Chemists Meeting, Oct. 19, 1981. 8) Dudley M.T., Boone D.J., Blood Lead Analysis 1980, Laboratory Improvement Program Office, Centers for Disease Control. TEH 0533310 DUP050034577