EPA/600/8-91/049A
                                                       Journal Supplement

               UNITED STATES ENVIRONMENTAL PROTECTION AGENCY
                  Environmental Criteria and Assessment Office (MD-52)
                      Research Triangle Park, North Carolina 27711


                  •       .  •          March 1992           •  . ,  •
Dear Requestor:

      The comment period for public review of the External Review Draft of
Air Quality Criteria for Oxides of Nitrogen  has been extended.  The revised
deadline for submitting comments on the document is March 30, 1992.   Because
of your specified interest in the document,  we have enclosed a copy of the
paper, Synthesis of Environmental  Evidence:  Nitrogen Dioxide Epidemiology
Studies, authored by Vic Hasselblad,  Dennis J. Kotchmar, and David M. Eddy.
The paper presents the information and analysis contained in Chapter 14 of
the document, Epidemiology Studies of Oxides of Nitrogen.  The paper has been
accepted for publication  by the Journal  of the  Air  and  Waste Management
Association.
                                          Sincerely yours,
                                          Lester D.  Grant,  Ph.D.
                                          Director,  Environmental  Criteria
                                            and Assessment  Office
Enclosure

-------

-------
                                                      EPA/600/8-91/049A
                                                      Journal Supplement
                         Synthesis of Environmental Evidence:
                        Nitrogen Dioxide Epidemiology Studies
                                   Vic Hasselblad
                                 Dennis J. Kotchmar
                                   David M. Eddy
     Vic Hasselblad, Ph.D., Research Associate Professor, Center for Health Policy
Research and Education, 125 Old Chemistry Building, Duke University, Durham, NC
27706.
     Dennis J. Kotchmar, M.D., Medical Research Officer, Environmental Criteria and
Assessment Office, U.S. Environmental Protection Agency, Research Triangle Park,
NC 27711.
     David M. Eddy, M.D., Ph.D., Professor, Center for Health Policy Research and
Education, Duke University, Durham, NC 27706.
                                                             Printed on Recycled Paper

-------
Disclaimer
     The views expressed in this paper are those of the authors and do not necessarily reflect
the views or policies of the U.S. Environmental Protection Agency.  The U.S. Government
has the right to retain a nonexclusive royalty-free license in and to any copyright covering
this article.
                                            2

-------
Acknowledgment




     The authors thank the members of the Oxides of Nitrogen Workshop Review Group for



their Suggestions that lead to this paper. The authors also thank James Ware, Lucas Neas,



and Jonathan Samet for their advice. Finally, the authors thank the journal's reviewers



whose suggestions added greatly to the paper.

-------
Practical Implications—
"Synthesis of Environmental Evidence:  Nitrogen Dioxide Epidemiology Studies"

     The assessment methodology discussed in the paper, meta-analysis, provides an

alternative approach for assessing environmental data bases.  Meta-analysis has the potential

to increase the ability to estimate a small but meaningful change in the risk of a health

outcome measure by analyzing the total evidence from all studies simultaneously.  The

specific example of the relationship of lower respiratory illness and nitrogen dioxide (NO2)

exposure is a case in point. The evidence of the individual studies was inconclusive. When

taken as a whole, the results of the meta-analysis suggest an increase of at least 20% in the

odds of respiratory illness in children exposed to an increase of 30 jug/m3  NO2.  This analysis

can be considered along with other evidence in assessing health effects of exposure to NO2.

-------
Abstract


     The use of meta-analysis is becoming more common in the medical literature, but it is


not common in the environmental literature.  Although meta-analysis cannot combine a group


of poorly executed, conflicting studies to get an unequivocal answer, there are certain


situations where it can be helpful.  The inability of studies to produce similar results may be


a function of the power of the studies rather than a reflection of their quality.  The literature


on the effects of nitrogen dioxide on the odds of respiratory illness in chiildren is such an


example.  Three quantitative methods for the synthesis of this evidence are presented.


Although the methods produce slighlty different results, the  conclusion from all three methods


is that the increase in the odds of respiratory illness in children exposed to a long-term

                   i*
increase of 30 jttg/nr (comparable to the increase resulting from exposure to a gas stove) is


about 20%.  This estimated increase is not sensitive to the method of analysis.

-------
Introduction



     The United States Environmental Protection Agency (U.S. EPA) is directed by the



Clean Air Act1 to promulgate standards that protect the public health from air pollutants and



are based on air quality criteria. Such criteria are to reflect the latest scientific information



useful in indicating the kind and extent of all identifiable public health effects that may be



expected from the presence of ambient air pollutants.  Air Quality Criteria Documents



(AQCDs) for these pollutants attempt to integrate and synthesize key information from



several disciplines to provide a coherent framework from which interpretation and judgments



can be made concerning the risk to human health.  Reducing the uncertainties inherent in



such information strengthens the conclusions that can be drawn.



     Over the past decade, quantitative approaches have been developed to synthesize



evidence from multiple studies.  Making use of such approaches in evaluating and



synthesizing epidemiologic evidence as part of AQCi) preparation requires that the methods



be able to handle the results of a variety of analyses, including multiple logistic regression



analyses, and provide combined estimates of the probability of a given type of health effect



occurring at a specified exposure level.   This eliminated many meta-analysis methods often



used with clinical trials, as well as the method of effect sizes. In preparing a revised AQCD



for nitrogen oxides, three methods were found to be useful.   These methods are described



below, and comparisons are then made between the results of the three methods as applied to



the evaluation and  synthesis of epidemiologic evidence concerning the effects of nitrogen



dioxide on respiratory disease symptoms in children.



      The purpose of this paper is to (1)  demonstrate that there are situations where the



synthesis of environmental evidence is feasible,  (2) describe some of the models used for this



synthesis, and (3) apply these models to a specific data set of interest.



                                            6

-------
Quantitative Methods for Synthesizing Evidence
     The three quantitative approaches employed in evaluating the nitrogen dioxide
health effects evidence are (1) the variance-weighted method, (2) random-effects models as
described by DerSimonian and Laird,2 and (3) the Confidence Profile Method as described by
Eddy3 and Eddy et al.4'5
                              Variance-Weighted Method
     One of the oldest methods for combining estimates of a parameter is the variance-
weighted method, which is described by Hald.6 Assume that there are n  studies, each giving
estimates, 6t, of a parameter 6, where i = 1,2,...,«.  The method assumes that each study

is independent of the other studies and that each study is estimating exactly  the same
parameter.  We shall refer to this as the fixed-effects model.  The minimum variance estimate
for any estimate, which is a linear combination of the 0/s, is:
           n
n
                                                                                  (1)
where wt =  1 / variance (^.  The statistic for testing the null hypothesis that all 0/s are
estimating the same parameter is
                                                                                  (2)
which is approximately distributed as a chi-square distribution with (n - 1) degrees of
freedom.

-------
                       Combining Using Random-Effects Models
     Random-effects models have been used for many years, although they are sometimes
called two-stage or hierarchical models. Random-effects models arise when the parameter or
parameters of "mother nature" do not remain constant from study to study.  Instead, they
vary randomly and are in fact assumed to be random variables from some distribution. The
problem then becomes to estimate (or derive a posterior distribution for) some function of
these parameters.
     One such random-effects model is the so-called "normal-normal" model.  In this model
we assume that each estimate, 0,-, of the parameter  6 is sampled from a normal distribution
                          2                           2
with mean ^ and variance a..  For now, we assume that a. is known and that ^ is a random
value from another normal distribution with mean 0 and variance r2. The likelihood for the
estimates from n different studies is
     L « exp
E
z=
                                                                                   (3)
                                                        r\
The parameter 0 can be estimated, or a posterior for 6 and t can be calculated if prior
distributions (priors) are specified for 6 and r2  A flat prior [p(T2) =  1] will work for r2
but the natural noninformative prior 1/v,  leads to problems.  The integrals do not converge
for r2 near zero, and the intergrals do not converge as T2 approaches infinity unless n is at
least 4.  A safer prior for r2 is
 Usually, the actual variances of the estimates, oi ,  are not known and are replaced by their
 sample values.
                                           8

-------
       Often the study parameter that varies is an effect measure, such as an odds ratio. This


 parameter then has two sources of variation: (1) the variation resulting from -"mother nature"


 choosing different parameter values for each study and (2) the sampling variation from the


 study itself given a particular value from "mother nature".  DerSimoniati and Laird2 give


 formulas for partitioning the variation in a random-effects model without making any


 particular distributional assumptions about the parameters. Let &t  be the estimate of the


 parameter, as before, and let wt be the inverse of the sampling variance of dt. Define $w
 as
            n
                       n
                                                                                   (5)
 Estimate r2 by
= max "
                0, [*w- (n -  I)]/
                                    n
                                                                                  (6)
Now define v,- as l/(l/w,. + t2).  The estimate of 6 is


          n       .   n

                        i'
                                                                                  (7)
and the variance of 8  is approximately
          n

         £
         /=!
              1  /    v,.
                                                                                  (8)

-------
                               Confidence Profile Method

     The Confidence Profile Method is a very general method for combining virtually any

Mnd of evidence about various parameters, as long as those parameters can be described in a

model.  The Confidence Profile Method can be used in either a Bayesian mode, to estimate a

joint posterior distribution for the parameters of interest, or in a non-Bayesian mode, to

estimate a joint likelihood function for the parameters of interest.  The method will be

applied to the special case where each study is estimating the same endpoint.  The Confidence

Profile Method uses a model that consists of three elements: (1) basic parameters,

(2) functional parameters, and (3) likelihood functions relating evidence to basic or functional

parameters.  The description of these elements follows,

     Basic Parameters.  Basic parameters are those parameters that appear in  the model that
                                                                       i
are not functions of any other parameters.  For example, the respiratory disease rate in

children living in homes with electric stoves in a particular study could be a basic parameter.

For convenience, we will denote the basic parameters as 6 1,^2^ --^k'  ^ a Bayesian analysis

is to be used, then all basic parameters must have prior distributions,  Noninformative priors

for these parameters can be derived in a variety of ways,  One standard method is to use a

Jeffreys' prior.7

     Functional Parameters. Each functional parameter, 6-, is defined as a function of the

basic parameters 0j,02 ..... dk> and ^ previously defined functional parameters Ok+ii«»Qj-i:
Although the functions can be any mathematical expression, certain functions are very

common. For example, the multiple logistic model is often used:
                                          10

-------
                               =  1  /
1 + exp
k
-£''*

  (10)
 where the 0y s are logistic regression coefficients, the Xy& are known constants and
 81 represents the probability that the jth response is positive.
      Likelihood Functions.  Likelihood functions connect observed evidence to basic and
 functional parameters.  For example, the likelihood function for a multiple logistic regression
 problem could be:
                                              n
                                                                                 (11)
 where 6t is defined in equation (10) and where yt equals 1 if the response; is positive and
 equals zero otherwise.  For a more complete discussion of likelihood functions, see Barnard
 and Sprout.8
      The Model.  Once the basic parameters, functional parameters, and likelihoods are
 defined, the model has been formulated.  The general log-likelihood for the model (assuming
 independence of experiments) is:
(12)
                                     £ ln
                                s=\  i=l
where / indexes the observations of an individual study, s indexes the S different experiments,
ns is the number of observations in experiments S, and Yt is the observed data for the
/  study.  If priors have been defined, then the m dimensional posterior is; defined as:
                                          11

-------
                                   S  ns

                                   H  T
 k
 n
M
(13)
     Methods of Model Solution.  There are five different methods for solving the model


once it has been defined:  (1) maximum likelihood estimation, (2) maximum likelihood


methods for determining the posterior model, (3) exact solutions uncertain special cases,


(4) approximate solutions using moments, and (5) Monte Carlo simulation.  The choice


depends on the complexity of the model, the accuracy required, and the amount of


computational power and time available.  For additional discussion of these, see Eddy et al/


The Confidence Profile Method can be used for both fixed and random effects models.
An Application to Studies of Nitrogen Dioxide Effects on Respiratory Illness Symptoms


     Lower respiratory tract illness (LRI) is one of the major causes of childhood morbidity


in the United States,9 This is of public health importance because childhood respiratory


illness is extremely common and the potential for exposure to NO? is great.10  Lower


respiratory illness takes on added importance since recurrent childhood respiratory illness may


be a risk factor for later susceptibility to lung damage.10"1'3  Various studies of LRI have


reported rates ranging from about 20 to 30 illnesses per 100 children in the first year of


life.9'14'15 The rate of LRI in children is affected by several factors that include age,


immunologic status, prior viral infections, level of health, socioeconomic status, day care


attendance,1^1 environmental tobacco smoke (ETS), and exposure to  NO2 and other pollutants.
                                          12

-------
       Bacteria are not thought to be common causes of LRI in nonhospitalized infants in the




  United States.15 Seventy-five percent of the isolated microbes were one of four types:




  respiratory syncytial virus, parainfluenza virus Types 1 and 3, and M. pneumonia.9'14  Early




  insult from virus infection in the lower respiratory tract is an essential element of the




  development of chronic and persistent impairment.11"13  It is now recognized that infections,




  reactive airways, ETS, and other inhaled pollutants, are the most important risk factors in the




  development of chronic lung disease.17  Thus, factors such as NO2, which increase the risk




  for LRI, are important because of the associated public health concern and the potential for



  increase in the development of chronic lung disease.




      Epidemiological studies of the relationship between NO2 exposure and a health outcome




  such as LRI in children provide  the majority of the evidence for examination of such




 relationships.  Several factors arise in the interpretation of epidemiological studies of the




 health effects of NO2:  (1) measurement error in exposure, (2) misclassification of the health




 outcome, (3) selection bias, (4) adjustment for covariates, (5) publication bias, (6) internal



 consistency, and (7) plausibility of the effect based on other evidence.




      The effect of measurement  error on estimation has been studied by several authors,



 including Shy et al.,18 Gladen and  Rogan,19 StephansM and Carroll,20 Fuller,21 Schafer,22




 and Whittemore and Keller.23 In general, measurement error that is independent of the health




 outcome will result in estimated effects biased towards the null.   Whittemore and Keller23




 specifically consider the data of Melia et al.29 as described by Florey et al.34 and  show that a




 20% inisclassificatkMi, rate of the exposure category will result in an underestimate of the




logistic regression coefficient by as much as 50%.  StefansM and Carroll20 have shown that,



even without the independence of error related to outcome, the bias is towards the null in
                                           13

-------
Situations where the risks are not extremely clpse tP 0 or 1.  The use of the presence of a gas



stove as a surrogate for actual NQ2 exposure introduces misclassificatipn.



     Most studies pf respiratory disease and NQ^ exposures measured the important



covariates pf age, gender, soeio^econpmic level pf the parents, and parental smoking habits.



The estimated effect (regression coefficient of disease on NO^ exposure) will be an



overestimate when a missing soyariate is either positively or negatively correlated with both



the exposure variable and the health putepme. The estimated effect will be an underestimate



When a missing epvariate is positively correlated with the exposure variable and negatively



correlated with the health outcome, °r vice versa,  Ware et al.24 found that parents with



spme college education were more likely to report respiratory symptoms and were less likely



to use a gas stpve, leading  tP an underestimate of the health effect if education were left put




Of the analysis.



     Studies that exajnine NO2 relationships to respiratory illness, when reviewed



independently, produce somewhat mixed resylts.2^ The use of quantitative rnethods of



Synthesizing evidence presents the opportunity to examine the consistency between these



studies arid the strength of  the total data ba^e, Selected studies are discussed, followed by a



combined analysis.
                                     British Studies



     Results of British studies have, been reported by Melia et al.,26"31 Goldstein et al.,32'33



and Florey et al.34'3^ The initial study, reported by Melia et al,,2^ was based on a survey of



5,658 children  (excludes asthmatics, thus 100 less than the number reported), aged 6 to



11  years, with  sufficient information in 28 randomly selected areas of England and Scotland.



The Study included a self-administered, parent-completed questionnaire that obtained




                                           14                                        :

-------
  information on the presence of morning cough, day or night cough, colds going to chest,




  chest sounds of wheezing or whistling, and attacks of bronchitis.  The questionnaire was




  distributed in 1973 and asked about symptoms during the previous 12 months.  Colds going




  to chest accounted for the majority of the symptoms reported.  Information about cooking




  fuel (gas or electric), age, gender, and social class (manual or nonmanual labor) was




  obtained, but information on parental smoking was not. No measurements of NO2,  either




  indoors or outdoors, were given. The authors presented their results in the form of a




  contingency table with complete covariate information for nonasthmatic children under age




  eight.  The authors indicated that there was a trend for increased symptoms in homes with




 gas stoves,  but that the increase was only significant for girls in urban areas; however, they




 did not report odds ratios or other measures of increased risk.




      Our reanalysis of the authors'  data was performed using a multiple-logistic model.




 Because it had been suggested that gender had an effect on the relationship with "gas



 cooker," interaction terms for gender were included in the original model.  None of these




 proved to be significant,  and they were subsequently dropped from the model.  When




 separate terms for each gender were used for the effect of "gas cooker," an estimated odds




 ratio of 1.25 was obtained for boys  and an  odds ratio of 1.39 was obtained for girls, but the




 odds ratios were not significantly different. The combined odds ratio for both genders was




 1.31 (95% confidence limits of 1.16 to  1.48) and was statistically  significant from  1.00




 (p < .0001).  The other main effects of gender,  socio-economic status, and age were all



 statistically significant.




      Melia et al.28 report further results of the national survey covering two groups:




 (1) a new cohort of 4,827 boys and  girls, aged 5 to 10 years, from 27 randomly selected




areas who were examined in 1977 and (2) 2,408 children first examined in 1973 who were




                                          15

-------
followed-up for at least one year and whose parents reported the use of the same cooking fuel



for each year the child was studied. The 1977 study collected information on the number of



smokers in the homes. In the 1977 cross-sectional study, only the prevalence of day or night



cough in boys (p « 0.02) and colds going to chest in girls (p < 0.05) were found to be



significantly higher in children from homes where gas was used for cooking compared with



children from homes where electricity was used.  Grouping  responses according to the six



respiratory questions into one or more symptoms or diseases, or none, yielded a prevalence



higher in children from homes where gas was used for cooking than in those from homes



where electricity was used (p » 0.01 in boys, p = 0.07 in girls).  The authors examined the



effect of gender, social class, use of pilot lights, and number of smokers in the homes.



      Our reanalysis of the authors' data was performed applying a multiple-logistic model.



This model contained the same terms that were included in our analysis of Melia et al.26 As



in the previous analysis, none of the interaction terms proved to be significant, and they were



subsequently dropped from the  model.  The maximum likelihood estimate of the  odds ratio



was 1.24 (95% confidence limits of 1.09 to 1.42. This effect was statistically significant



(p < .0001). The other main effects of gender,  socio-economic status, and age were all



statistically significant.



      This study was followed by a study in 1978 of 808 schoolchildren,29 aged six to  seven



years, in Mddlesborough, an urban area in northern England. Respiratory illness was



defined in the same manner as in the previous  study.  Indoor NO2 measurements were



collected from 66% of the homes,  with the remaining 34% refusing to participate.  Nitrogen



dioxide was measured by Palmes tubes36 attached to walls in the kitchen areas and in



children's bedrooms.  In homes with gas stoves,  levels of NO2 in kitchens ranged from 10 to



596 fig/m3 (0.005 to 0.317 ppm [1 jug/m3 = 0.00053  ppm  at 25  °C, 760 mmHg]),  with a



                                           16

-------
  mean of 211 jig/m3; and levels in bedrooms ranged from 8 to 318 jug/m3,, with a mean of 56

      <•»
  /*g/m .  In homes with electric stoves, levels of NO2 in kitchens ranged from 11 to

           "2                       ^
  353 jtg/m , with a mean of 34 jttg/m , and in bedrooms NO2 levels ranged from 6 to

         *2                        ** '•
  70 /Kg/m  , with a mean of 26 /wg/m3.  Outdoor levels of NO2 were determined using


  diffusion tubes systematically located throughout the area, and the weekly average ranged


  from 26 to 45 jwg/m3.


      One analysis by the authors29 was restricted to those 103 children in homes where gas


  stoves were present and where bedroom NO2 exposure was  measured. A multiple logistic


 regression  model was fitted to the presence or absence of respiratory illness.  Measured NO2


 exposure was found to be associated with respiratory illness, independent of social class,  age,


 gender,  or the presence of a smoker in the house (p = 0.06). However, when social class


 was excluded from the regression, the association was weaker (p  = 0.11).  For the six- to


 seven-year-old children living in gas stove homes, there appeared to be an increase of


 respiratory illness with increasing levels of NO2 in their bedrooms (p  = 0.10), but no


 significant relationship was found between respiratory symptoms in those children'or their


 siblings  or parents and levels of NO2 in kitchens.


      Since no exposure-response estimates were given by the authors, a multiple-logistic


 model was  fitted to the data with a linear slope for NO2 and separate intercepts for boys and


 girls. Nitrogen dioxide levels for the groups were estimated by fitting a lognormal


 distribution to the NO2 data,  which was reported by intervals, and the average exposures


 within each interval were estimated.37  The estimated logistic regression coefficient for NO2


 (in /*g/m3) was 0.015 with a  standard error of 0.007.  This result is not directly comparable


with the  previous two analyses since it gives the increase  in the logarithm of the odds of


respiratory illness per unit increase in NO2 exposure. Since  most studies of gas stove


                                           17

-------
exposure (both in the United States and the United Kingdom) show an approximate increase



of 30 jtg/m3 in (he NO2 levels, the slope was multiplied by 30 to get the increase due to gas



StOY<5 exposure, and then converted to an odds ratio by exponentiation.  All of this assumed



that the logarithm of the odds ratio was linear in N02 exposure.  The result was an odds ratio



of 1.53, with 95% confidence limits of 1,04 to 2,24.



     The study was repeated January through March of 1980 by Melia et al.30  This time,



five- and, six-year-old children were sampled from the same neighborhood as the previous



study, but only families with gas stoves were recruited.  Environmental measurements were



madi? and covariate data were collected in a manner similar to the previous study.



Measurement^ of NO2 were available from 54% of the homes. The unadjusted rates of one



or more symptoms by gender and exppsure level were analyzed by the authors,  and they



concluded that;".,, no relation was found between the prevalence of respiratory illness and



levels of NQ2,"  A reanalysis of the data was  made using a multiple-logistic model similar to



the one used for the previous study.  The model included a linear slope for NO2 and separate



intercepts for boys and girls. Nitrogen dioxide levels for the groups were estimated by fitting



a lognprmal distribution to the grouped bedroom NO2 data. The estimated logistic regression



coefficient for NO2 (in jig/m3) was 0.004 with a standard error of 0.005. As for Melia



et al.,29 the regression coefficient was converted, to an odds ratio for an increase of 30 /ig/m3



in NO2 assuming that the logarithm of the odds ratio was linear in NO2 exposure.  This gave



an odds ratio  of 1.11, with. 95% confidence limits of .83 and  1.49.



     Melia et al,31 investigated the association between gas cooking in the home and



respiratory illness in a study of 390 infants bom between 1975 and 1978, When a child



reached one year of age, the child's mother was interviewed by a trained field worker who



completed a questionnaire.  The mother was asked whether the child usually experienced




                                          18

-------
  morning cough, day or night cough, wheeze, or colds going to chest, and whether the child




  had experienced bronchitis, asthma, or pneumonia during the past 12 months.  No relation




  was found between the type of fuel used for cooking in the home and the prevalence of




  respiratory symptoms and diseases recalled by the mother after allowing for the effects of




  gender, social class, and parental smoking.  The authors gave prevalence rates of children




  having at least one symptom by gas stove use and gender.  The combined odds ratio for




  presence of symptoms by gas stove use was 0.63, with 95% confidence limits of 0.36  to



  1.10.









                             United States Six-Cities Studies




      Several authors24'35"46 have reported on a series of studies conducted in six U.S. cities.




 The six cities were selected  to represent a range of air quality based on their historic levels of




 outdoor pollution and included Watertown, MA; Kingston and Harriman, TN; southeast



 St. Louis, MO; Steubenville, OH; Portage, WI; and Topeka, KS. Approximately




 1,500  grade-school children were enrolled in each community and were followed for several




 years.  Families reported the number of persons living in their homes and their smoking




 habits, parental occupations and educational backgrounds, and fuels used for cooking and




 heating.  Outdoor pollution was measured at fixed sites in the communities and at selected




 households. Indoor pollution, including NO2, was measured in several rooms of selected




 households. Results of monitoring in Portage, WI, verify that the presence of a gas stove



 contributes dramatically to indoor NO2 levels.  The results clearly show the effect of a  gas




 stove on not only the indoor concentrations but also on the personal exposure of the




individual. The study42 was conducted very  carefully with excellent quality control.  It gave
                                          19

-------
an average estimate of 29 jig/in3 increase in exposure resulting from the use of gas stoves in



cities studied in the United States.



     Ware et al.24 reported results from the six-cities studies based on 8,120 children, aged



6 to 10 years, who were followed from 1974 to 1979.  An initial report on a subset of the



data was given by Speizer et al.39 Health endpoints were measured by a standard respiratory



questionnaire that was completed by parents of the children.  The authors used log-linear



models to estimate the effect of gas stoves versus electric stoves on the rates of serious



respiratory illness before age two. Directly standardized rates of reported illnesses and



symptoms did not show any consistent pattern of increased risk for children from homes with



gas stoves.  Logistic-regression analyses controlling for age, gender,  citys and maternal



smoMng level gave estimated odds ratios fof the effect of gas stoves ranging from 0.93 to



1.07 for bronchitis, cough> wheeze, LRI index, and illness for the past year.  The index for



LRI was defined as the presence of either bronchitis, respiratory illness that kept the child



home 3 days or more, or persistent cough for 3 months of the past year.  None of these odds



ratios were statistically different from 1.   Only two odds ratios approached statistical



significance: (1) history of bronchitis (odds ratio =  0.86, 95% confidence interval 0.74 to



1.00) and (2) respiratory illness before age two (odds ratio =  1.13, 95% confidence



interval 0.99 to 1.28). When the odds ratio for respiratory illness before age two was



adjusted for parental education, the odds ratio was 1.11, with 95% confidence limits of



0.97 to 1.27 (p = 0.14).  Thus, the study suggests an increase in respiratory illness of about



11%, although the increase was not statistically significant at the 0.05 level.  The endpoint in



the Ware et al.24 study most similar to that of the Melia studies was the LRI index.  The



authors gave the unadjusted rates,  and from those an estimated odds ratio of 1.08, with 95%
                                           20

-------
  confidence limits of 0.97 to 1.19, were calculated. Although this rate was not adjusted for




  other covariates, the effect of those adjustments on other endpoints was minimal.




      Neas et al.45'46 studied a cohort of 6,273 children from the same six cities. This cohort




  included children that were part of the Dockery et al.44 analysis but was restricted to white




  children 7 to 11 years of age with complete covariate information and at least one valid




  indoor measurement of both NO2 and respirable particles.  This resulted in 1,286 children




 being included in the analysis. Methods for measuring indoor pollutants were described by




 Spengler et al.38  Indoor pollutants were measured in each child's home for two weeks




 during the heating season and two weeks during the cooling season.  Nitrogen dioxide was




 measured by Palmes tubes at three locations in each home.




      The analysis of the Neas et al.45'46 study was based on the third symptom questionnaire




 that was completed by parents following the indoor measurements. The questionnaire




 reported symptoms during  the previous year, including shortness of breath, chronic wheeze,



 chronic cough, chronic phlegm,  and bronchitis.  The authors used a multiple-logistic model,




 which had separate-city intercepts, indicator variables for gender and age, parental history of




 chronic obstructive pulmonary disease and asthma, parental education, and single-parent




 family status.  The sampling strategy minimized the association between NO2 and passive-




 smoking exposure. The increases in symptoms were estimated for an additional 31 jtg/m3




 NO2 exposure.  This corresponded to the average difference in NO2 concentrations monitored




 in homes with a gas stove with a pilot light, based on exposure information from the study.




 Table I shows the odds ratios for the five separate symptoms associated with the increase in



 NO2 exposure.




     All of these odds ratios are consistent with the size of effect seen in the other analyses




of the six-city data and the  analyses of the British studies.  The authors defined a combined




                                          21

-------
Table I.   Odds ratios and 95% confidence intervals for the effect of an additional 31 jig/m3
          nitrogen dioxide on the symptom prevalence.
Symptom
Shortness of breath
Chronic wheeze
Chronic cough
Chronic phlegm
Bronchitis
Combined symptoms score
Odds Ratio
1.27
1.19
1.21
1.29
1.05
1.47
95% Confidence Interval
0.92 to 1.73
0.87 to 1.61
0.86 to 1.71
0.93 to 1..79
0.71 to L56
1.17 to L86
Source: Neas et al.45


symptom, which was the presence of one or more of the symptoms just reported, and an

analysis of this combined indicator of respiratory symptoms gave an estimated odds ratio of

1.47, with 95% a confidence interval of 1.17 to 1.86.  When split by gender, the odds ratio

was higher in girls, and when split by smoking and nonsmoking homes, it was higher in

smoking homes.
                                    Tayside Study

     Ogston et al.47 studied infant mortality and morbidity in the Tayside region of northern

Scotland.  The subjects were 1,565 infants born to mothers who were living in Tayside in

1980.  Episodes of respiratory illness were recorded during the first year of life. The

information was supplemented by observations made by a health visitor and scrutinized by a

pediatrician who checked diagnostic criteria and validity.  One health endpoint assessed was

defined as the presence of any respiratory illness during the year. This endpoint was

analyzed by the authors using a multiple-logistic regression model that included  terms for

parental smoking, age of mother, and presence of a gas stove.  The estimated odds ratio for



                                          22

-------
 the presence of a gas stove was 1.14, with 95% confidence limits of 0.86 to 1.50. Only the




 coefficient for parental smoking was statistically significant (p < 0.01).








                                       Iowa Study




      Ekwo et al.48 surveyed 1,355 children 6 to 12 years of age for respiratory symptoms




 and lung function in the Iowa City School District. Parents of the school children completed




 a questionnaire that was a modification of the questionnaire developed by the American




 Thoracic Society.49  Eight different measures of respiratory ilhiess were reported by the




 authors, but only the endpoint of chest congestion and phlegm with colds was similar to the




 endpoints used in the British studies and the six-city studies. Information on parental




 smoking was obtained  and used as a covariate in the analysis.  The result of the analysis,




 which was based on 1,138 children, was an odds ratio of 1.10 for gas stove use.  The 95%




 confidence limits of 0.79 and 1.53 were derived from the authors'  data. No NO2



 concentrations, either inside or outside the homes, were reported.








                                     Dutch Studies




     In the Netherlands, Houthuijs et al.,50 Brunekreef et al.,51 and Dijkstra et al.52 studied



 the effects of indoor factors on respiratory health in children. The population consisted of




 6- to 9-year-old children from 10 primary schools in five nonindustrial communities in the




 southeast  region of the Netherlands.  Personal exposure to NO2 and homes concentrations




 were measured.  An important NO2 emission and exposure source in  these homes are




 geysers, which are unvented gas-fired hot  water sources at the water tap.  Exposure to




 tobacco smoke was assessed with a questionnaire that also reported symptom information.



Pulmonary function was measured at school.  The study used Palmes  tubes to measure a




                                          23

-------
single weekly average personal NO2 exposure. Potential high peak exposures from the



geysers may not be well characterized by the weekly average personal exposure



measurements.  In January and February of 1985, the homes of 593 children who had not



moved in the last four years were measured for one week for NO2.  Personal exposure was



also estimated from time budgets and room monitoring.



     Three measures of health (cough, wheeze, and asthma) were obtained from the



questionnaire, which was a modified form of the World Health Organization questionnaire.53



Asthma was defined as attacks of shortness of breath with wheezing in the last year.  The



presence of any of the three symptoms was used as a combination variable, and a logistic-



regression model was used to fit the combination variable.  Exposure was estimated by fitting



a lognorm'al distribution to the exposure data,  which was reported in intervals; and the mean



exposure values for each group were estimated by a maximum likelihood technique.37 The



estimated logistic-regression coefficient was —0.002, corresponding to an odds ratio of .94



for an increase of 30 /tg/m3 in NO2, with 95% confidence limits of 0.66 to 1.33.  This



assumed a linear relationship between  the logarithm of the odds ratio and the NO2 exposure.



The rates were not adjusted for covariates such as parental smoking and age of the child.








                                     Ohio Study



     Keller et al.54 and Mitchell et al.55 originally conducted a 12-month study of respiratory



illness and pulmonary function in families in Columbus, OH, prior to 1978.  The study



measured NO2 exposure by both the Jacobs-Hochheiser and continuous-chemiluminescence



methods.  The electric stove users averaged 38 jug/m3 NO2 exposure, whereas the gas stove



users averaged 94 jtg/m3.  Thus, the estimated average difference between gas and electric



stove use was 58 /zg/m3.  The paper did not report which rooms were measured in order to



                                         24

-------
  get these averages.  In a second related study,56 580 persons drawn from households that




  participated in the earlier study were examined to confirm the reports and to determine the




  frequency distribution of reported symptoms among parents and children in gas or electric




  stove homes.  A nurse-epidemiologist examined selected persons who were reported ill.




  Unfortunately, these rates were not adjusted for other covariates. The percentage of children




  having lower respiratory symptoms in homes with a gas stove was 53.2% (n = 267) and




  50.7% (n = 286) in homes with electric stoves.  Although the difference is not statistically




  significant, these rates give an estimated odds ratio of 1.10, with 95%  confidence limits of



  0.74 to 1.54.








 Synthesis of the Evidence




      In order to combine the studies just described, several assumptions were necessary.




 First, although each  study used a slightly different health outcome as an endpoint, we



 assumed  that the endpoints are similar enough to warrant their combination. Second, the




 exposure levels were different in each study. An increase of 30  pg/m3 was used as a




 standard increase, and all studies were used  to estimate the effect of an increase of 30 /ig/m3,




 even if they had a different exposure range.  Third we assumed that each study controlled  for




 key covariates, or that those covariates were properly adjusted  for or are of minimal




 significance. The omission of covariates such as parental education  almost certainly biases



 the results towards the null, and for this reason we retained some studies, which arguably



 could have been excluded.




     The studies described used different indicators to study health endpoints. The symptoms



describing LRI evaluated in the studies varied but are, in general, reasonable indicators  of




LRI.  They include colds going to chest, chronic wheeze and cough, bronchitis, chest cough




                                           25

-------
with phlegm, episodes of respiratory illness, and various respiratory indexes, which are



combinations of more than one of these symptoms.  These symptpms are comparable to



indicators of LRI in children that were used in other studies. In order to compare these



studies on respiratory effects of NO2, a cqmmqn endpqint was defined, and then each study



was compared with this standard endpoint.  The endpoint was the presence of reported LRI



symptoms in children age  12 or younger.



     An attempt was made to include as many studies as possible.  The requirements for



inclusion were (1) the health endpoint measured must be reasonably close to the standard



endpoint; (2) exposure differences must, exist, and some estimate of exposure (either direct or



indirect) must be  available; and (3) an odds ratio for a specified exposure must have been



calculated, or data presented so that it can be calculated. These studies are summarized in




Table H.



     The approximate likelihoods for each study are shown in Figure  1.  Each curve can be



treated as a likelihood function or pQsteriQr-'probability distribution. If treated as a likelihood



function, then 95% confidence limits for the odds ratio can be calculated as those two points



on the horizontal  axis between which 95% of the area under the curve is contained.  If



treated as a posterior-probability distribution, then the area under the curve between any  two



points is the probability that the odds ratio lies, between those two points.   Note that all



11 likelihoods show some overlap. A ehi-square goodness-of-fit test of the homogeneity of



the 11 studies gives a chi-square of 18,75 with 10 degrees of freedom (p — 0.0436),



suggesting some lack of homogeneity in the 11 studies.



     The studies  were combined using four methods:  (1) the variance-weighted method,



assuming a fixed-effects model; (2) the Confidence Profile Method, assuming a fixed-effects,
                                          2:6

-------
 I
 .g

 S
 o
 i1
I
 s
i
.3

,8

-------
Figure 1.  Meta-analysis of epiderniologic studies of 30 jug/nr  nitrogen dioxide exposure
           increase on respiratory illness in children $12 years  old.
model; (3) the DerSimonian and Laird method, assuming a random-effects model; and (4) the

Confidence Profile Method, assuming a random-effects model.  Results of the use of these

models in synthesizing the NO2 evidence are presented in Table III for four subsets of the

studies.  The first includes all 11 studies; the second excludes the two studies on children less

than one year of age; the third excludes the younger children and those studies that  did not

measure NO2 directly; and the fourth excludes the younger children and those studies that

measured NO2 directly.


                                           28                                        ;

-------
  Table m. Summary of synthesis of studies on respiratory illness effects of nitrogen dioxide.

                                                 Model-Method
                                     Fixed
                                            Random
        Studies
  Variance-      Confidence
  Weighted        Profile
  Method6       Method5
                                                            DerSimonian    Confidence
                                                                and          Profile
   Laird2
Method'
   All 11 studies
   Children aged
   5-12 years,
   9 studies
    1.18           1.18
(1.11, 1.25)a   (1.11,  1.25)
    1.19           1.19
(1.12, 1.27)    (1.12, 1.27)
                                                                1.18           1.18
                                                            (1.08,  1.30)    (1.08, 1.29)
    1.19.           1.20
(1.09, 1.30)    (1.10, 1.31)
Measured NO2,
children aged
5-12 years,
4 studies
Surrogate NO2
estimate based
on presence of
gas stove,
children aged
5-12 years,
5 studies
1.27 1.27
(1.09, 1.47) (1.09, 1.47)
1.18 1.18
(1.10, 1.26) (1.10, 1.26)

1.27 1.25
(1.02, 1.58) (0.99, 1.58)
1.18 1.18
(1.07, 1.29) (1.09, 1.28)

 °95 % confidence limits given in parentheses.



     The results from all analyses are reasonably similar.  The variance-weighted method and

 the Confidence Profile Method have identical answers because the log normal approximation

 for the likelihood function was used in the calculation of the solution by the Confidence

 Profile Method. In general, the results should be nearly identical for reasonable sample

 sizes. The DerSimonian and Laird2 method and Confidence Profile Method5 for the analysis

of a random-effects model gave similar but not identical results.
                                          29

-------
     The analysis of the nine studies with children 5 to 12 years old was done separately


because the other two studies were of infants. The exclusion of these two studies made little


difference in the results.


     All studies that used the presence of a gas stove as a surrogate for NO2 exposure


obviously suffer from measurement error.  In general, measurement error will decrease the


estimated effect.  When the four studies of children over age 5 years with measured NO2


levels were combined, the estimated o
-------
  may be greatly improved by analyzing the total evidence from all studies; simultaneously, that




  is, by conducting a meta-analysis. A potential limitation of a meta-analysis is that the studies




  available for use may represent a spectrum of quality.  Well executed studies may be mixed




  with studies containing flaws-studies with missing data or confused definitions and outcome




  measures. On the other hand, such problems may be slight compared to problems with




  alternative traditional reviews.  Meta-analysis fills a need by assisting in the reconciliation of




  conflicting research results.  While some physical sciences may allow the; identical replication




  of experiments, many fields such as environmental science allow only the repetition of studies




  that introduce variation and produce uncertainty.  Meta-analysis is one way of dealing with



  uncertainty.




      All meta-analysis methods previously discussed assume that each piece of evidence




  (study) is  independent of the others.  Under the fixed-effects model, the evidence is assumed




 to pertain  to a common parameter. Under the random-effects model,  this assumption is




 relaxed to allow for a distribution for the parameter of interest. The computations for both




 models are relatively straightforward and can be made on a personal computer.  The method'



 of calculation has less impact on the conclusion than does the choice of model.




      In Air Quality Criteria for Oxides of Nitrogen, prepared 10 years ago by U.S. EPA,58




 a group of studies examining the relationship  between respiratory illness  and exposure in the




 home to gas combustion products from cooking fuel were evaluated.  At that time, those



 studies inferred the presence of NO2 by the presence of gas combustion emission sources.




The evidence from individual studies of the effect of NO2 on respiratory illness  was




somewhat mixed.  Since then, new studies have been conducted, and earlier ones updated,



that provide data on NO2 concentrations and estimates of exposure.
                                          31

-------
     The studies of respiratory illness in children exposed to increased levels of NQ2


provides an excellent example of the application of meta-analysis.  Taken by themselves,
                                                                          t
most of the 11 studies were reported as not being statistically significant at the 0.05 level


based on analyses performed by the original authors.  The studies differed in design and


sample size, and this likely contributed to the lack of significance of some of the studies.


However, use of the meta-analysis methods described above indicates that, taken  as a whole,


the collective evidence from the evaluated studies strongly suggests an increase of at least

                                                                              /j
20% in the odds of respiratory illness in children exposed to an increase of 30  jug/nr NO2


for extended periods of time.


     The choices of model (fixed or random) and computational method make little


difference in the estimates in this particular example.  In particular, the estimates do not


depend strongly on the assumption that each study is estimating the same parameter.  Thus,


any lack of homogeneity is not a major concern.  The choice of the computational method


(e.g.,  DerSimonian and Laird2 versus the Confidence Profile Method5) also makes little


difference in the estimates when restricted to the particular problem described in  this paper.


The Confidence Profile Method can be applied to a much broader class of problems,


however.

     There is always the concern that the studies described are not the complete  list of


studies, but contain primarily the positive studies, since these are the studies most likely to be


published. This is referred to as "publication bias".  There are two reasons not to be


concerned with publication bias in this particular situation.  First, prospective


 epidemiological studies are very expensive and require the work of many individuals.  The


 studies are usually described to the scientific community before the results are  even known.


 Second, most of the studies cited were reported as negative studies by the authors themselves,


                                           32

-------
 indicating that there was no difficulty in publishing negative results.  In spite of this, it is of



 interest to contemplate an undiscovered study with results so negative that, when combined



 with the other studies, produces a confidence interval for the odds ratio Ihat includes the



 value 1.  If we assume that the hypothetical study is  the size of the Ware et al.24 study, then



 its odds ratio for increased respiratory symptoms as the result of a 30 jug/m3 exposure would



 have to be 0.766.




      Although there may be reasons to weight certain studies or groups of studies more



 heavily than  others, the final conclusion has to be that there is an increase in the odds of



 respiratory illness of children, especially those of elementary-school age.  The estimates are



 generally  centered about an odds ratio of 1.2, with 95%  confidence limits of 1.1 to  1.3,



although the  studies using measured NO2 give a slightly  higher estimate of the odds ratio.



This kind  of  synthesis may be possible for other areas of environmental aissessment where



multiple studies of a given health endpoint are available.
                                          33

-------
References

 1.  U.S. Code, "Clean Air Act, section 108, air quality criteria and control techniques,
     section 109, national ambient air quality standards," U.S,  C. 42: sections 74Q8-74Q9
     (1991).

 2.  R. DerSimonian and N. Laird, "Meta-analysis in clinical trials," Controlled Clin.  Trials
     7: 177 (1986).

 3.  D,M. Eddy, "The confidence profile method: a Bayesian method for assessing health
     technologies," Oper. Res. 37: 210 (1989).

 4.  D.M. Eddy, V. Hasselblad and R. Shachter, "An introduction to a Bayesian method for
     meta-analysis: the confidence profile method," Med. Decision Making 10: 15 (1990). •

 5.  D.M. Eddy, V. Hasselblad and R. Shachter, "A Bayesian method for synthesizing
     evidence: the confidence profile method," Int. J. Technol. Assess. Health Care
     6: 31 (1990).

 6.  A. Hald, Statistical theory mth engineering applications, John Wiley and Sons, Inc.,
     New York, 1962, pp. 243-244.

 7.  J.O. Berger, Statistical decision theory and Bayesian analysis: second edition,^,
     Springer-Verlag, New York, 1985.

 8.  G.A. Barnard and D.A. Sprout,  ""Likelihood"," in Encyclopedia of statistical sciences,
     v. 4: icing the tails to limit theorems, S. Kotz and N.L. Johnson, eds., John Wiley and
     Sons, New York, 1982, pp. 639-644.

 9.  K.M. McConnochie,  C.B. Hall and W,H. Barker,  "Lower respiratory tract illness in
     the first two years of life: epidemiologic patterns and costs in a suburban pediatric
     practice," Am. J. Public Health 78: 34 (1988).

 10.  J.M. Samet, I.B. Tager and F.E. Speizer, "The relationship between respiratory illness
     in childhood and chronic air-flow obstruction in adulthood," Am. Rev. Respir. Dis.
     127: 508 (1983).

 11.  W.P. Glezen, "Antecedents of chronic and recurrent lung disease: childhood respiratory
     trouble," Am. Rev. Respir. Dis. 140:  873 (1989).

 12.  D.R. Gold, I.E. Tager, S.T. Weiss et al., "Acute lower respiratory illness in childhood
     as a predictor of lung function and chronic respiratory symptoms," Am.  Rev. Respir.
     Dis. 140: 877 (1989).
                                           34

-------
 13.   B. Burrows, J.W. Bloom, G.A. Traver et al., "The course and prognosis of different
      forms of chronic airways obstruction in a sample from the general population," N. Engl
      J. Med.  317: 1309 (1987).

 14.   W.P. Glezen and F.W. Denny, "Epidemiology of acute lower respiratory disease in
      children," N. Engl. J. Med. 288: 498 (1973).

 15.   A.L. Wright, L.M. Taussig, C.G.  Ray et al., "The Tucson children's respiratory study:
      n. lower respiratory tract illness in the first year of life," Am. J. Epidemiol.
      129: 1232 (1989).

 16.   R.M. Chanock, K. Mclntosh, B.R. Murphy et al., "Respiratory syncytial virus," in
      Viral infections of humans: epidemiology and control, A.S. Evans, ed., Plenum
      Publishing Corporation, New York, 1989, pp. 525-544.

 17.   F.W. Denny and W.A. Clyde, "Acute lower respiratory  tract infections in
      nonhospitalized children,"  J. Pediatr.  (St. Louis) 108: 635  (1986).

 18.   C.M. Shy, D.G. Kleinbaum and H. Morgenstern, "The effect of misclassification of
      exposure status in epidemiological studies of air pollution health effects," Butt. N. Y.
     Acad. Med.  54: 1155 (1978).

 19.   B. Gladen and WJ. Rogan, "Misclassification and the design of environmental studies,"
     Am. J. Epidemiol 109: 607 (1979).

20.  L.A. Stefanski and R.J. Carroll, "Covariate measurement error in logistic regression,"
     Ann. Stat. 13: 1335 (1985).

21.  W.A. Fuller, "A single explanatory variable," in Measurement error models, John  •
     Wiley & Sons,  New York, 1987, pp.  1-9, 13-20.

22.  D.W. Schafer,  "Covariate measurement error in generalized linear models," Biometrika
     74: 385 (1987).

23.  A.S. Whittemore and J.B.  Keller,  "Approximations for regression with covariate
     measurement error," JASA J. Am.  Stat. Assoc. 83: 1057  (1988).

24.  J.H. Ware, D.W. Dockery, A.I. Spiro et al., "Passive smoking, gas cooking, and
     respiratory health of children living in six cities," Am. Rev. Respir. Dis.
     129: 366 (1984).

25.  J.M. Samet, M.C. Marbury and J.D. Spengler,  "Health effects and sources of indoor
     air pollution. Part I," Am.  Rev. Respir. Dis. 136: 1486 (1987).

26.  R.J.W. Melia, C.V. Florey, D.G. Altaian et al., "Association between gas cooking and
     respiratory disease in children," Br. Med.  J. 2: 149 (1977).
                                          35

-------
27.  RJ.W. Melia, C.V. Florey, S.C. Darby et al., "Differences in NO2 levels in kitchens
     with gas or electric cookers," Amos. Environ.  12: 1379 (1978).

28.  RJ.W. Melia, C.V. Florey and S. Chinn, "The relation between respiratory illness in
     primary schoolchildren and the use of gas for cooking: I. - results from a national
     survey," Int. J. Epidemiol 8: 333 (1979).

29.  RJ.W. Melia, C.V. Florey, S. Chinn et al-., "The relation between indoor air pollution
     from nitrogen dioxide and respiratory illness in primary schoolchildren," Clin. Respir.
     Physiol. 16: 7P (1980).
          *
30.  RJ.W. MeUa, C.V. Florey, R.W. Morris et al.* "Childhood respiratory illness and the
     home environment: n. association between respiratory illness and nitrogen dioxide,
     temperature and relative humidity," Int. J. Epidemiol. 11: 164 (1982).

31.  J. Melia, C. Florey, Y. Sittampalam et al., "The relation between respiratory illness in
     infants and gas cooking in the UK: a preliminary report," in Air quality VLth world
     congress: [proceedings of the International Union of Air Pollution Prevention
     Associations}, SEPIC (APPA), Paris, France, 1983, pp. 263-269.

32.  B.D. Goldstein, RJ.W. Melia, S. Chinn et al., "The relation between respiratory
     illness in primary schoolchildren and the use of gas for cooking: II - factors affecting
     nitrogen dioxide levels in the home," Int. J. Epidemiol.  8: 339 (1979).

33.  B.D. Goldstein, RJ.W. Melia and C.V. Florey, "Indoor nitrogen oxides," Bull. N. Y.
     Acad. Med. 57: 873 (1981).

34.  C.V. Florey, RJ.W. Melia, S. Chinn et al., "The relation between respiratory illness
     in primary  schoolchildren and the use of gas for cooking: III - nitrogen dioxide,
     respiratory  illness and lung infection," Int. J. Epidemiol. 8: 347 (1979).

35.  C.V. Florey, R. Melia, B.  Goldstein et al., "Stickstoffdioxid im Innenraum -
     Bpidemiologische Auswirkungen im Vereinigten Koenigreich [The epidemiology of
     indoor nitrogen dioxide in the U. K.]," in Luftqualitaet in Innenraeumen [Indoor air
     quality], K. Aurand, B. Seifert and J. Wegner, eds., Gustav Fischer Verlag, Stuttgart,
     Federal Republic of Germany,  1982, pp. 209-218.

36.  E.D. Palmes, A.F. Gunnison, J.  DiMattio et al., "Personal sampler for nitrogen
     dioxide," Am. Ind. Hyg. Assoc.^J. 37: 570 (1976).

37.  V. Hasselblad, A.G. Stead and W. Galke, "Analysis of coarsely grouped data from the
     lognormal distribution," JASA J. Am. Stat. Assoc. 75: 111 (1980).

38.  J.D. Spengler, B.G. Ferris, Jr.,'IXW. Dockery et al., "Sulfur dioxide and nitrogen
     dioxide levels inside and outside homes and the implications on health effects research,"
     Environ. Scl Technol 13:  1276 (1979).
                                          36

-------
 39.  F.E. Speizer, B.J. Ferris, Y.M.M. Bishop et al., "Respiratory disease rates and
      pulmonary function in children associated with NO2 exposure," Am. Rev. Respir. Dis.
      121: 3 (1980).

 40.  E.G. Ferris, Jr., D.W. Dockery, J.H. Ware et al., "The six-city study:  examples of
      problems in analysis of the data," Environ. Health Perspect. 52; 115 (1983).

 41.  C.S. Berkey, J.H. Ware, D.W. Dockery et al., "Indoor air pollution and pulmonary
      function growth in preadolescent children," Am. J. Epidemiol 123: 250 (1986).

 42.  J.J. Quackenboss, J.D. Spengler, M.S. Kanarek et al.,  "Personal exposure to nitrogen
      dioxide: relationship to indoor/outdoor air quality and activity patterns," Environ. Sci.
      Technol 20: 775 (1986).

 43.  D.W. Dockery, F.E. Speizer, E.G. Ferris, Jr. et al., "Cumulative and reversible effects
      of lifetime smoking on simple tests of lung function in adults," Am. Rev. Respir. Dis.
      137: 286 (1988).

 44.  D.W. Dockery, J.D. Spengler, L.M. Neas et al.,  "An epidemiologic study of
      respiratory health status and indicators of indoor air pollution from combustion
      sources," in Combustion processes and the quality of the indoor environment:
      transactions of an international specialty conference, J.P. Harper, ed., Air & Waste
      Management Association, Pittsburgh, PA, 1989, pp. 262-271.

 45.   L.M. Neas, J.H. Ware, D.W. Dockery et al., "The association of indoor nitrogen
      dioxide levels with respiratory symptoms  and pulmonary function in. children," in
      Indoor air '90: precedings of the 5th international conference on indoor air quality and
      climate, volume 1, human health, comfort and performance, International Conference on
      Indoor Air Quality and Climate, Inc., Ottawa, ON, Canada, 1990,  pp. 381-386.

 46.   L.M. Neas, D.W. Dockery, J.H.  Ware et al., "Association of indoor nitrogen dioxide
      with respiratory symptoms and pulmonary function in children," Am. J. Epidemiol.
      134: 204 (1991).

 47.   S.A. Ogston, C.V. Florey and C.H.M. Walker, "The Tayside infant morbidity and
      mortality study:  effect on health of using gas for cooking," Br. Med. J.
     290: 957 (1985).

48.  E.E. Ekwo, M.M. Weinberger, P.A. Lachenbruch et al., "Relationship of parental
     smoking and gas cooking to respiratory disease in children," Chest 84: 662 (1983).

49.  B.G. Ferris, "Epidemiology standardization project," Am. Rev. Respir. Dis.
     118: 1 (1978).
                                          37

-------
50.  D. Houthuijs, B. Remijn, B. Brunekreef et al., "Exposure to nitrogen dioxide and
     tobacco smoke and respiratory health of children," in Indoor air '87: proceedings of the
     4th international conference on indoor air quality and climate,  v. 1, volatile organic
     compounds, combustion gases, particles and fibres, microbiological agents, B. Seifert,
     H. Esdom, M. Fischer et al., eds., Institute for Water, Soil and Air Hygiene, Berlin,
     Federal Republic of Germany, 1987, pp. 463-467.

51.  B. Brunekreef, P. Fischer, D. Houthuijs et al., "Health effects of indoor NO2
     pollution," in Indoor air '87: proceedings  of the 4th international conference on indoor
     air quality and climate,  v. 1, volatile organic compounds, combustion gases, particles
     and fibres, microbiological agents, B. Seifert,  H.  Esdorn, M. Fischer et al., eds.
     Institute for Water, Soil and Air Hygiene,  Berlin, Federal Republic of Germany, 1987,
     pp. 304-308.

52.  L. Dijkstra, D. Houthuijs, B. Brunekreef et al., "Respiratory health effects of the
     indoor environment in a population of Dutch children," Am. Rev. Respir. Dis.
     142: 1172 (1990).

53.  C.V. Florey and S.R. Leeder, Methods for cohort studies of chronic airflow limitation,
     World Health Organization, WHO regional publications, European series no. 12,
     Geneva, Switzerland, 1982.

54.  M.D. Keller, R.R. Lanese, RJ. Mitchell et al., "Respiratory illness in households using
     gas and electricity for cooking:  I. survey of incidence," Environ. Res. 19: 495 (1979).

55.  DJ. Mitchell, R.B. Ireland, A- Hume et al., "Epidemiologic notes  and reports: carbon
     monoxide poisoning - Mississippi," Wkly. Morb. Mortal Rep. 23:  1 (1974).

56.  M.D. Keller, R.R. Lanese, R.I. Mitchell et al., "Respiratory illness in households using
 ;   gas and electricity for cooking:  n. symptoms and  objective findings," Environ. Res.
     19: 504 (1979).

57.  C. Mann, "Meta-analysis in the breech," Science (Washington,  DC) 249: 476 (1990).

58.  U.S. Environmental Protection Agency,  Air quality criteria for oxides of nitrogen,
     EPA-600/8-82-026, Office of Health and Environmental Assessment, Environmental
     Criteria and Assessment Office, Research Triangle Park, 1982.
                                          38

-------