Blackwell Science, LtdOxford, UKADDAddiction0965-2140© 2005 Society for the Study of Addiction

100
Original Article
Breath carbon monoxide and cigarette smoking
Martin A. Javors
et al.

METHODS AND TECHNIQUES

Cut-off levels for breath carbon monoxide as a marker

for cigarette smoking

Martin A. Javors1,2, John P. Hatch1,3 & Richard J. Lamb1,2

Departments of Psychiatry,1 Pharmacology2 and Orthodontics,3 University of Texas HSC, San Antonio, Texas, USA

Correspondence to: ABSTRACT

Martin Javors
Department of Psychiatry
Aims Current clinical studies often use a breath carbon monoxide (BCO) cut-
The University of Texas HSC
7703 Floyd Curl Drive off level of 8 parts per million (p.p.m.) or higher to identify smoking. In this
San Antonio study, the cut-off level of BCO as an indicator of smoking over the past 24 hours
Texas 78229 was re-examined.
USA
Design BCO and self-reported smoking were obtained each weekday for up to
Tel: 210 567 5532
Fax: 210 567 3759 14 weeks in 213 subjects paid to deliver reduced BCO values. Analysis of
E-mail: javors@uthscsa.edu 12 386 paired values for reported smoking and BCO were analyzed.
Findings The 25% quartile, median and 75% quartile values for BCO were 1,
Submitted 17 November 2003;
1 and 2 p.p.m. on non-smoking days and 2, 5 and 12 p.p.m. on smoking days,
initial review completed 4 July 2004;
final version accepted 2 September 2004 respectively. Receiver-operating characteristic (ROC) analysis indicated that
BCO provided high diagnostic accuracy to distinguish between smoking and
non-smoking days [area under the curve (AUC) = 0.853, P < 0.0001]. The
METHODS AND highest combined sensitivity and specificity was observed at a BCO cut-off level
TECHNIQUES of 3 p.p.m. (sensitivity = 71.5%; specificity = 84.8%). At a BCO cut-off of
8 p.p.m. sensitivity and specificity were 40.6% and 98.2%, respectively, indicat-
ing that many smokers would be falsely classified as abstinent. Finally, the per-
centage of true tests (positive and negative) was highest at a BCO cut-off of
2 p.p.m. (80.2%).
Conclusions BCO cut-off levels well below 8 p.p.m and as low as 2–3 p.p.m.
may be more useful when it is important to maximize identification of smoking
abstinence with a high degree of certainty.

KEYWORDS Breath carbon monoxide/ breath CO, cigarette, cut-off,

smoking.

INTRODUCTION these studies. BCO levels can be measured accurately and

Confirmation of reported smoking levels with biochemi-
cal tests such as breath carbon monoxide (BCO) or sali-
vary cotinine is an important feature of research studies
for the development of new approaches to facilitate smok-
ing cessation (Colletti et al. 1982). Cotinine levels in
serum, saliva or urine provide the most sensitive and spe-
cific marker for determining recent smoking status. How-
ever, there are no convenient on-site tests for cotinine
that are sufficiently sensitive to low levels of smoking to
provide a satisfactory marker for studies in which imme-
diate assessment of smoking is desired (Middleton & Mor-
ice 2000). As a result, BCO levels are better suited for
immediate results can be obtained using commercially
available, hand-held analyzers. BCO levels have a half-life
of 3–6 hours, depending on the level of exercise and envi-
ronmental CO exposure of the research subject (Benowitz
et al. 2002). Previous studies have shown that smoking
within the past 24 hours generally results in BCO levels
that are elevated above the normal physiological range,
although this can depend on amount and recency of
smoking (Benowitz et al. 2002).
BCO levels are used frequently as a biological marker
to confirm abstinence in smoking cessation studies,
including studies that place monetary contingencies on
participants’ ability to achieve and maintain abstinence

© 2005 Society for the Study of Addiction doi:10.1111/j.1360-0443.2004.00957.x Addiction, 100, 159–167
160 Martin A. Javors et al.
(Rand et al. 1989; Gilbert et al. 1999; Shoptaw et al.
2002). In our own studies placing contingencies upon
BCO levels (Lamb et al. 2004a; Lamb et al. 2004b), we
noticed that many smokers could maintain low levels of
smoking while producing BCO levels below those used
typically as indicators of smoking in treatment outcome
studies [e.g. 8–10 parts per million (p.p.m.)]. As there is
little guidance in the literature for determination of an
optimal BCO cut-off that can distinguish between smok-
ing days and non-smoking days, we undertook a detailed
examination of data collected during the course of these
contingency management studies in order to determine
the characteristics of various cut-off criteria. Current
smokers who were paid for producing reduced BCO levels
were followed for about 3 months while 24-hour smok-
ing reports and daily BCO levels were monitored. The pur-
pose of the present paper is to examine the relationship
between reported smoking and BCO levels and to deter-
mine optimal cut-off levels for detection of smoking under
conditions designed to maximize the distinction between
days of low-rate smoking and abstinence.
METHODS
Research subjects
Subjects participated in any of three contingency man-
agement studies designed to reduce BCO. There was a
total of 213 smokers who reported smoking at least 20
cigarettes per day and had a BCO level ≥ 15 p.p.m. at
study intake. In two of the three studies, participants
were attempting to stop smoking. In the third study, par-
ticipants had no defined plans to quit smoking in the next
three months.
For all analyses except the correlations, data from all
213 subjects were used. Among those subjects, we col-
lected BCO levels on 12 649 days, reported smoking on
12 450 days and paired data for BCO levels and reported
smoking on 12 386 days. This group of 213 subjects
included 161 white, 35 Hispanic, nine African American
and eight Asian, Native American and other subjects.
The mean age was 38.8 years (range: 19–68 years).
There were 119 females and 94 males.
Using the Fagerström Test for Nicotine Dependence, it
was determined that, prior to entry into the study, 6% of
subjects smoked £10 cigarettes per day, 42% smoked 11–
20 cigarettes per day, 46% smoked 21–30 cigarettes per
day and 6% smoked ≥31 cigarettes per day. The mean of
BCO levels at entry into the studies was 24.4 p.p.m.
For the between-subject and within-subject correla-
tion analyses, 191 of the total 213 subjects’ data were
selected. Those subjects had 10 or more smoking days,
which provided a measure of validity and consistency for
© 2005 Society for the Study of Addiction
the correlations. This group of 191 subjects included 145
white, 32 Hispanic, nine African-American and five
Asian, Native American and other subjects. The mean
age was 38.4 years (range: 19–67 years). There were
107 females and 84 males.
Participant study procedures
Each weekday that the University was open, participants
delivered a breath sample at a time that was fixed for the
individual, between 11 a.m. and 5 p.m. For each partici-
pant’s first 10 visits, there were no contingencies placed
on BCO levels except for the first day for participants in
the two studies for individuals attempting to quit smok-
ing. In these two studies, participants received an extra
monetary payment if their BCO was <4 p.p.m. After these
initial 10 visits, monetary contingencies were designed to
reduce BCO level for the remaining 60 visits.
BCO measures
A Vitalograph CO monitor (Vitalograph Inc. Lenexa, KS,
USA) was used to measure BCO samples. Subjects were
required to take a deep breath, hold it for 20 seconds and
then to expire over 20 seconds into the disposable mouth-
piece of the CO monitor. The peak BCO reading (the mon-
itor was watched until the CO reading declined) was
taken as the subject’s BCO level. BCO levels and BCO cut-
off levels were expressed as integers in parts per million
(p.p.m.). The cut-off level for a BCO test indicates that a
test at or above the cut-off level would be a positive test for
a smoking day (presumed to have smoked). A BCO level
below the cut-off would be a negative test (presumed not
to have smoked). In this study, various BCO cut-off levels
were used to evaluate cut-off levels for their accuracy in
identifying smoking and non-smoking days. For example,
a BCO level of = 3 p.p.m. would be a positive test at a BCO
cut-off of 3 p.p.m. The repeatability of BCO determina-
tions was assessed by collecting two samples per day on 3
separate days. The intraclass correlation coefficients,
assuming a single fixed determination, were 0.994
(n = 22), 0.976 (n = 17) and 0.935 (n = 22) for the
3 days. These results indicate the high reliability of single
determinations as used in this study.
Reported smoking
Subjects were asked to keep track of the exact number of
cigarettes that they smoked within the past 24 hours.
Each day the number of times that a subject visited the
university clinic was reported. Smoking days were
defined as those days when reported smoking was at least
one cigarette smoked within the 24 hours before the time
of a given BCO level test. In 15 subjects, on a total of 45 of
Addiction, 100, 159–167

12 386 test days, reported smoking included a partial
cigarette. In those cases, the number of cigarettes was
rounded-up to the next higher integer. Non-smoking
days were those days when reported smoking was zero
cigarettes smoked in the past 24 hours.
Statistical methods
The within-subjects correlation between number of ciga-
rettes smoked and BCO level was calculated using analy-
sis of covariance with subjects dummy-coded as a
categorical variable. This method removes between-
subject variability and assesses the within-subject rela-
tion between number of cigarettes smoked and BCO level.
The between-subjects correlation was computed by first
calculating for each subject the mean CO level and mean
number of cigarettes smoked per day taken over all avail-
able smoking days and then calculating the Pearson
product–moment correlation coefficient between pairs of
means, weighted by the number of smoking days. Data
for cigarettes smoked per day and BCO levels showed pos-
itive skew and kurtosis. After transformation to square
roots, both distributions approximated normality. There-
fore, these two variables were subjected to square root
transformation for statistical analysis. All tests were two-
tailed with an alpha level of P < 0.05.
Outcome definitions
Sensitivity was defined as the percentage of positive BCO
tests (a BCO level at or above a given cut-off) on smoking
days among all subjects, i.e. the percentage of all smok-
ing days for which there was a positive BCO test. Specific-
ity was defined as the percentage of negative BCO tests (a
BCO level below a given cut-off) for non-smoking days,
i.e. percentage of all non-smoking days for which there
was a negative BCO test. Sensitivity and specificity were
determined for integer BCO cut-off levels from 1 to
12 p.p.m. The area under the receiver-operating charac-
teristic (ROC) plot was estimated using the non-paramet-
ric method to determine diagnostic accuracy (Zweig &
Campbell 1993), i.e. the ability of the BCO test to distin-
guish between smoking and non-smoking days. The
term ‘accuracy’ is used to indicate the BCO cut-off that
provides the best balance of sensitivity and specificity.
Positive predictive value (PPV) was defined as the num-
ber of true positive BCO tests divided by the total number
of positive tests for a given BCO cut-off, expressed as a
percentage. PPV is also called the percentage of true pos-
itive tests. Negative predictive value (NPV) was defined
as the number of true negative BCO tests divided by the
total number of negative BCO tests, expressed as a per-
centage. NPV is also called the percentage of true nega-
tive tests.
© 2005 Society for the Study of Addiction
Breath carbon monoxide and cigarette smoking 161
RESULTS
The distribution of BCO levels for smoking days and non-
smoking days for all subjects is shown in Fig. 1. Median
BCO level was 1 p.p.m. for non-smoking days and
5 p.p.m. for smoking days. The 10th, 25th, 75th and
90th percentile BCO levels for non-smoking days were 1,
1, 2 and 3 p.p.m., respectively. Corresponding values for
smoking days were 1, 2, 12 and 19 p.p.m., respectively.
Mean ± standard error of the mean (SEM) of BCO levels
was 1.64 ± 0.04 p.p.m. for non-smoking days and
8.26 ± 0.09 p.p.m. for all smoking days combined
(P < 0.0001). Mean ± SEM for cigarettes smoked in the
past 24 hours for the smoking days was 9.36 ± 0.067
cigarettes.
At a BCO cut-off level of 3 p.p.m., 2792 of 3292 non-
smoking days had negative tests (specificity = 84.8%),
whereas 6500 of 9094 smoking days had positive tests
(sensitivity = 71.5%). This contrasts sharply with a BCO
cut-off level of 8 p.p.m., where 3232 of 3292 non-
smoking days had negative tests (specificity = 98.2%),
but only 3688 of 9094 smoking days had positive tests
(sensitivity = 40.6%).
Figure 2a shows the mean and 95% confidence limits
of BCO levels at integer levels of number of cigarettes
smoked for 12 386 smoking and non-smoking test days.
For all smoking days on which 10 or more cigarettes were
smoked, the data were combined. These results show that
BCO levels were directly proportional to the number of
n = 9094
20
BCO Levels (ppm)
16
12
8
n = 3292
4
0
Non-Smoking Days Smoking Days
Smoking Status
Figure 1 Breath carbon monoxide (BCO) levels on smoking days
and non-smoking days. Smoking days were defined as those days
when reported smoking was at least one cigarette smoked within
the 24 hours previous to the time of the BCO level. Non-smoking
days were those days when reported smoking was zero cigarettes in
the past 24 hours. BCO levels were measured with a Vitalograph
BCO monitor. The box and error bars represent the median (line in
the box), 25th and 75th percentiles (top and bottom of the boxes),
and 10th and 90th percentiles (error bars). N = number of samples
included in the analysis. Data from 213 subjects enrolled in the study
were included in this figure
Addiction, 100, 159–167
162 Martin A. Javors et al.

(a) (a)
Between Subject Correlation
15 45
40 n = 191

Breath CO Level (ppm)

r = 0.617
12 35 p < 0.001
BCO Level

30
9
(ppm)

25
20
6
15
10
3
5
0
0
0 10 20 30 40 50
0 1 2 3 4 5 6 7 8 9 ≥ 10 Reported Smoking
Cigarettes Smoked (Cigarettes during previous 24 h)

(b)
(b)
18 Histogram of Within Subject Correlations
45
Cigarettes Smoked

(Number of Subjects at each level)

(mean ± 95% CL)

15 40

12 35 Total number of correlations = 191

30
Frequency

9
25
6 20

3 15

10
0
5
0
0 1 2 3 4 5 6 7 8 9 ≥ 10 –0.75 –0.55 –0.35 –0.15 0.05 0.25 0.45 0.65 0.85 1.05

BCO Level (ppm) Level of Within Subject Correlation (r value)

Figure 2 (a) Mean BCO levels at various levels of cigarettes
smoked; (b) mean cigarettes smoked at various BCO levels. In both
panels, symbols represent the mean ± 95% confidence limits (CL) of
BCO levels for each number of cigarettes smoked during the past 24
hours (a) or number of cigarettes smoked in the past 24 hours at
each level of BCO (b). Number of reports for each level of smoking:
0 = 3292; 1 = 639; 2 = 761; 3 = 566; 4 = 584; 5 = 655; 6 = 547;
7 = 359; 8 = 410; 9 = 140; 10 = 4433. Number of BCO readings for
each level: 0 = 698; 1 = 2436; 2 = 2252; 3 = 1398; 4 = 499; 5 = 486;
6 = 426; 7 = 443; 8 = 409; 9 = 342; 10 = 2997. Data from 213 sub-
jects enrolled in the study were included in this figure
cigarettes smoked over the past 24 hours. Smoking days
during which three cigarettes were reported resulted in a
mean BCO level between 3 and 4, while smoking days
during which eight or nine cigarettes were smoked pro-
duced BCO levels at a mean of approximately 8 p.p.m., the
often-recommended cut-off.
Figure 2b shows the mean and 95% confidence limits
of cigarettes smoked at various BCO levels for 12 386
smoking and non-smoking test days. For all days on
which BCO levels were 10 p.p.m. or greater, the data were
combined. These results show that number of cigarettes
smoked over the past 24 hours was directly proportional
to detected BCO levels, with mean smoking levels ranging
from 0.61 ± 0.13 cigarettes for days on which BCO level
was zero to 15.9 ± 0.3 cigarettes for days on which BCO
Figure 3 Correlations between reported smoking and BCO levels.
Mean number of cigarettes smoked is plotted as a function of mean
BCO level for 191 subjects. All 191 subjects had at least 10 smoking
days (cigarettes ≥ 1). The range was 10–70 days. The median was
55 days, the mean was 47.5 days; and the SD was 20.6 days. Non-
smoking days were not included in this analysis. (a) Between-subject
correlation. Each symbol represents data from one subject, i.e. aver-
age reported number of cigarettes smoked per day versus average
value of BCO level for all smoking days. The between-subject cor-
relation was plotted using Prism 4 software from Graph Pad. (b) His-
togram distribution of within-subject correlations between number
of cigarettes smoked and BCO values. The histogram was plotted
using Prism 4 software from Graph Pad. Each correlation value
shown on the x-axis represents the mid-point of a 0.1 point range of
r-values. For example, the level labeled 0.25 represents r-values rang-
ing from 0.20 to 0.30
levels were 10 p.p.m. or greater. On test days when BCO
levels were 3 p.p.m., a mean of five cigarettes smoked dur-
ing the past 24 h was reported. On test days when BCO
levels were 8 p.p.m., a mean of 11.9 cigarettes smoked
was reported.
Figure 3a shows the between- and Fig. 3b the within-
subject correlation analyses for reported smoking versus
BCO levels. For these analyses, we utilized data of 191
subjects. All subjects had at least 10 days on which they
reported smoking at least one cigarette (10 smoking
days). Data from non-smoking days were not included in

© 2005 Society for the Study of Addiction Addiction, 100, 159–167

 Breath carbon monoxide and cigarette smoking 163

this analysis. The between-subjects Pearson correlation
coefficient was 0.617 (P < 0.0001) for mean BCO versus
mean cigarettes smoked, supporting a significant linear
relationship (Fig. 3a).
Figure 3b shows a histogram of the distribution of r-
values for within-subject correlations between reported
smoking and BCO levels. The same data set that was
used for the analysis shown in Fig. 3a was used for this
analysis, except that each 24-hour smoking value
reported was used to calculate a within-subject corre-
lation coefficient. The overall r-value for within-subject
correlations of 191 subjects as determined by analysis
of covariance (ANCOVA) was 0.726 (Bland & Altman
1995; Altman & Bland 1996). Over 75% of the sub-
jects had statistically significant correlation coeffi-
cients greater than 0.50.
Figure 4a and Table 1 show results of the sensitivity
and specificity analyses for various BCO cut-off criteria.
Sensitivity is the percentage of smoking days with a BCO
level at or above a given BCO cut-off and specificity is the
percentage of non-smoking days with a BCO level below a
given BCO cut-off. As the BCO cut-off increased in value
from 1 to 12 p.p.m., sensitivity decreased and specificity
increased. It is noteworthy that the sensitivity and speci-
ficity curves in Fig. 4a intersect at a BCO level between 2
and 3 p.p.m. This intersection indicates that the highest
combined sensitivity and specificity was observed at a
BCO cut-off level less than 3 p.p.m.
A receiver-operating characteristic (ROC) plot was
also used to assess the performance (diagnostic accuracy)
of BCO across the full range of potential BCO cutoff values
(Fig. 4b). An ROC curve is derived from sensitivity and

(b)
1.0
(a) 1
0.8 2
100
3
Sensitivity

80 Specificity 0.6
Percent

4
Sensitivity
60
0.4
40

20 0.2
AUC = 0.853
P < 0.0001
0
0.0
0 3 6 9 12 15 0.0 0.2 0.4 0.6 0.8 1.0

BCO Cutoff (ppm) 1-Specificity

(c)
(d)
100
Percent of True Tests

100
80 PPV (% true pos)
80
Percent

(+ and –)

NPV (% true neg)

60 60

40 40

20 20

0 0

0 3 6 9 12 15 0 3 6 9 12 15

BCO Cutoff (ppm) BCO Cutoff (ppm)

Figure 4 Additional analyses of BCO cut-off criteria. (a) Sensitivity and specificity were plotted at BCO cut-off levels from 1 to 12 p.p.m.
Sensitivity is the percentage of positive BCO tests at a specified cut-off on all smoking days. Specificity is the percentage of negative BCO
tests at a specified cut-off for all non-smoking days. (b) For ROC analysis, 1-specificity (x-axis) was plotted against sensitivity at BCO cut-off
levels from 1 p.p.m. to 12 p.p.m. AUC was 0.853 (P < 0.0001).The numbers placed along the ROC curve indicate BCO cut-off levels. (c) BCO
cut-off level (x-axis) was plotted against positive predictive value (PPV; percentage of true positive tests) and negative predictive value (NPV;
percentage of true negative tests). A true positive test is defined as a BCO level above a specified cut-off in a subject who reported smoking
at least one cigarette in the past 24 hours. A true negative test is defined as a BCO level below a specific cut-off for a subject who reported
zero cigarettes in the past 24 hours. (d) Accuracy of classification at various BCO cut-off levels is shown by plotting BCO cut-off against per-
centage of true detections. The percentage of true detections was calculated by adding the number of true positive tests plus the number
of true negative tests, dividing by the total number of tests and multiplying by 100 to obtain a percentage. Data from 213 subjects enrolled
in the study were included in this figure

© 2005 Society for the Study of Addiction Addiction, 100, 159–167

164 Martin A. Javors et al.

Table 1 Sensitivity and specificity of various BCO cut-off levels.

BCO cut-off1 (p.p.m) Sensitivity2 Specificity3 1-Specificity4 Sensitivity + specificity5

0 1.000 0.000 1.000 1.000

1 0.985 0.171 0.829 1.156
2 0.874 0.603 0.397 1.477
3 0.715 0.848 0.152 1.563
4 0.596 0.945 0.055 1.541
5 0.546 0.959 0.041 1.505
6 0.497 0.970 0.030 1.467
7 0.452 0.976 0.024 1.428
8 0.406 0.982 0.018 1.388
9 0.362 0.985 0.015 1.347
10 0.325 0.987 0.013 1.312

1
The cut-off level for a BCO test indicates that a test at or above the cut-off level would be a positive test for a smoking day (presumed to have smoked). A
BCO level below the cut-off would be a negative test (presumed to have not smoked). In this study, various BCO cut-off levels were used to evaluate cut-
off levels for their accuracy in identifying smoking and non-smoking days. For example, BCO level of = 3 p.p.m. would be a positive test at a BCO cut-off
of 3 p.p.m.
2
Sensitivity was defined as the percent of positive BCO tests (a BCO level at or above a given cut-off) on smoking days among all subjects, i.e. the percentage
of all smoking days for which there were a positive BCO test.
3
Specificity was defined as the percentage of negative BCO tests (a BCO level below a given cut-off) on non-smoking days, i.e. percentage of all non-smoking
days for which there was a negative BCO test.
4
1-Specificity was calculated for the ROC plots. This term indicates the fraction of false positive tests at the various BCO cut-off levels, i.e. the proportion
of non-smoking days identified as smoking days.
5
Sensitivity + specificity was calculated to identify the BCO cut-off level that would have the highest combined sensitivity plus specificity.

specificity values by plotting 1-specificity against sensitiv-
ity. The area under the ROC curve is used as a measure of
the ability of a marker such as BCO to discriminate
between smoking and non-smoking days. The results of
the ROC analysis indicate that BCO levels were a highly
significant predictor of smoking versus non-smoking
days [area under the curve (AUC) = 0.853, P < 0.0001].
BCO cut-off levels corresponding to break-points along
the ROC curve indicate that cut-off levels of 2, 3 and
4 p.p.m. contributed significantly to the maximization of
AUC.
Figure 4c shows the impact of BCO cut-off criteria on
positive and negative predictive values. As BCO cut-off
criteria increase from 1 to 12 p.p.m., there is an increase
in positive predictive value (PPV). That is, as the cut-off
increases, a higher percentage of smoking days are cor-
rectly identified and we can be more certain that a posi-
tive test indicates a true smoking day. At the same time,
the increase in PPV is accompanied by a decrease in
negative predictive value (NPV). As the BCO cut-off
increases, it becomes less likely that non-smoking days
will be correctly identified and we are less certain that a
negative test indicates a true non-smoking day.
Figure 4d shows an additional analysis to support the
notion that a low BCO cut-off at the level of 2 or 3 p.p.m.
provides the highest accuracy to distinguish between
smoking and non-smoking days. For this analysis, BCO
cut-off levels were plotted against the combined percent-
age of true positive and true negative BCO tests, with the
dependent variables calculated by dividing the sum of
true positive and true negative tests (see Table 1) by the
total number of tests and expressing this as a percentage.
As can be seen on the graph, the highest percentage of
true tests, which was 80.2%, was observed at a BCO cut-
off of 2 p.p.m.
DISCUSSION
The novel finding of this study was that a BCO cut-off level
of 2 or 3 p.p.m. provided the highest accuracy for assess-
ment of abstinence versus any cigarette smoking within
the past 24 hours under conditions where smokers were
being paid to produce low BCO readings. This conclusion
is supported by several approaches to evaluating the data
from a large sample of smoking and non-smoking days in
subjects whose smoking was suppressed from normal lev-
els due to the operation of contingency management
interventions. First, sensitivity and specificity curves
crossed at a cut-off value of 3 p.p.m. (Fig. 4a), indicating
that this value yields the maximum combined sensitivity
and specificity. Secondly, the BCO cut-off levels that con-
tributed most to the area under the ROC curve in Fig. 4b
were 2, 3 and 4 p.p.m. (Fig. 4b). Thirdly, the highest per-
centage of true positive and negative tests combined
occurred at a BCO cut-off level of 2 p.p.m. (Fig. 4d).
The study also showed that there is a strong and sta-
tistically significant relationship between level of reported
smoking during the past 24 hours and BCO level. This
relationship was seen in a between-subjects analysis that

© 2005 Society for the Study of Addiction Addiction, 100, 159–167


included both smoking and non-smoking days (Fig. 3a)
and in a within-subjects analysis with only smoking days
considered (Fig. 3b). The positive relationship between
cigarettes smoked and BCO level is expected and has been
reported previously by others (Fortmann et al. 1984;
Biglan et al. 1985). However, the strength of correlations
obtained in specific studies will probably vary depending
on circumstances of data collection. For example, it is
likely that certain features of the present study acted to
reduce within-subject correlations. First, cash incentives
for low BCO levels motivated participants to reduce their
smoking, and may also have reduced the range of ciga-
rettes smoked for some subjects, a feature of data sets that
can result in low correlations. Secondly, the contingen-
cies may have influenced patterns of smoking in a man-
ner that impacted correlations. Thus, it is possible that
some participants smoked the majority of their daily cig-
arettes immediately after the BCO level was taken so that
they would have low BCO levels at the next day’s testing,
as was observed previously (Gariti et al. 2002). The half-
life of BCO is about 4 hours (Benowitz et al. 2002), which
results in the relatively rapid disappearance of CO from
the breath. Many of the participants were probably aware
of this fact. Higher correlations might be found if smoking
and CO levels were determined over shorter (e.g. 6 hours)
segments of time. Alternatively, the time from last ciga-
rette could be factored into the correlation, as has previ-
ously been done (Colletti et al. 1982).
The choice of a BCO cut-off level depends on the inten-
tions of the clinic or the research study in which BCO lev-
els are used as a smoking marker. For situations in which
it is most important to identify smokers with a high
degree of certainty (i.e. maximize sensitivity), such as in a
public health screening project, a higher cut-off such as
8 p.p.m. should be used. The higher cut-off would result
in a higher PPV (% true positives) for the BCO test, but a
lower NPV (% true negatives) (Fig. 4c). For example, our
data show that at a cut-off of 8 p.p.m., 98.3% of the pos-
itive tests occurred on days for which smoking was
reported. In many cases, however, clinicians and
researchers are trying to verify whether or not a person
has refrained from smoking. A BCO cut-off criterion of
8 p.p.m. has been adopted widely within the research
community as the standard cut-off for abstinence verifi-
cation (Fortmann et al. 1984; Benowitz et al. 2002;
Shoptaw et al. 2002). When a cut-off of 8 p.p.m. is used,
however, our data show that only 39.2% of tests classified
as negative occurred on non-smoking days, while 70.8%
of tests classified as negative occurred on days when the
participant reported smoking. Thus, many participants
who reported smoking within the past 24 hours would be
classified as non-smokers under an 8 p.p.m. criterion.
Shoptaw et al. (2002) reported a similar specificity (93%)
in their study of methadone-maintained tobacco smokers
© 2005 Society for the Study of Addiction
Breath carbon monoxide and cigarette smoking 165
and concluded that BCO levels using a cut-off of 8 p.p.m.
resulted in lack of identification of many who were
actively smoking. These data suggest, therefore, that
reports using the 8 p.p.m. BCO cut-off criterion as verifi-
cation of smoking abstinence could be significantly exag-
gerating true abstinence rates.
In contrast, when a BCO cut-off of 3 p.p.m. is
employed, detection of smoking is somewhat less likely
(71% of all smoking days were associated with a positive
test) but accurate detection of abstinence is significantly
enhanced, with 84% of all non-smoking days yielding a
negative test. The 3 p.p.m. cut-off provided the highest
accuracy when sensitivity and specificity were simply
added together (Table 1), while a BCO cut-off of 2 p.p.m.
produced the highest accuracy when the percentage of
correctly identified smoking and non-smoking days com-
bined were determined (Fig. 4d). Figure 4a and b suggests
that the highest accuracy was obtained at a cut-off
between 2 and 3 p.p.m. Use of these very low BCO cut-off
values may be especially important in clinical or research
situations that employ incentives for smoking abstinence,
where it is important to minimize false negative
readings in order to ensure accurate implementation of
contingencies
Once-daily measurement of BCO does not provide a
precise index of recent smoking, no matter which cut-off
criterion is selected. For example, BCO level was zero in
the present study on 135 (1.5%) and was at or below
3 p.p.m. on 2594 (28.5%) of 9094 reported smoking
days. However, the number of cigarettes smoked on
smoking days at low BCO levels was low. For example, the
mean number of cigarettes smoked on smoking days
when BCO level was zero was 3.13 ± 3.0 (SD) cigarettes
per day. For smoking days on which BCO level was 0, 1
and 2 p.p.m., mean cigarettes smoked was 4.52 ± 3.9
(SD). In contrast, for smoking days on which BCO level
was 7 or less p.p.m., mean cigarettes smoked was
6.56 ± 5.2 (SD).
As discussed previously, amount and recency of smok-
ing as well as individual differences in CO elimination
kinetics and exercise levels can exert an important influ-
ence on the association between smoking and BCO. Some
recent studies have tried to obviate these factors by col-
lecting CO measures three times per day (Roll et al. 1996;
Roll & Higgins 2000; Tidey et al. 2002). However, these
studies used BCO cut-off levels of £11 p.p.m. that would
classify incorrectly some smoking as non-smoking. For
situations for which it is critical to detect any smoking,
both multiple daily measurements and a very low BCO
cut-off level would need to be employed.
A potential limitation of the present study is that all
subjects were paid during portions of their participation
for meeting reduced BCO criteria. This resulted in rela-
tively low overall levels of smoking among participants,
Addiction, 100, 159–167
166 Martin A. Javors et al.
and in assessment of smoking–BCO correlations within
lower ranges of smoking than has previously been the
case. While this could be viewed as a strength that
heightens relevance to abstinence detection, it is also pos-
sible that the contingencies resulted in misrepresentation
(under-reporting) of smoking, which in turn would influ-
ence the sensitivity–specificity analyses and BCO cut-off
recommendations. In fact, however, under-reporting of
true smoking levels would tend to result in higher rather
than lower recommended BCO cut-offs. Thus, any under-
reporting would have a conservative effect on recommen-
dations. Additionally, participants were fully aware that
contingencies were based only on BCO levels and inde-
pendent of reported smoking, a factor that should have
limited under-reporting.
BCO levels can be elevated in certain medical condi-
tions including chronic obstructive pulmonary disease
(Montuschi et al. 2001), bronchiectasis (Horvath et al.
1998) and cystic fibrosis (Antuni et al. 2000), as well as
by marijuana smoking, and these factors may need to be
considered in individual cases. Another factor that may
need to be taken into consideration when setting BCO
cut-off criteria is baseline CO levels for non-smokers. The
issue of optimal cut-offs has been addressed in previous
studies that utilized general populations of smokers who
were not trying to quit as well as non-smokers. Jarvis
et al. (1987) suggested a BCO cut-off of 10 p.p.m. to min-
imize the number of misclassifications of smokers and
non-smokers. The mean BCO level for non-smokers in the
Jarvis study (Jarvis et al. 1987) was 5.6 p.p.m. versus
1.6 p.p.m. in our study. Similarly, Sato et al. (2003)
reported an optimal BCO cut-off of 10 p.p.m. with a back-
ground level for never smokers of 6.1 p.p.m. Finally, Ruth
& Neaton (1991) suggest an optimal cut-off of 10 p.p.m.
when non-smokers had a BCO level of 7.3 p.p.m. How-
ever, Woodward & Tunstall-Pedoe (1992) found an opti-
mal BCO cut-off of 6 p.p.m. when non-smokers had a
median BCO level of 2 p.p.m. These results suggest that if
BCO data were corrected for background levels, the dis-
crepancy between studies would lessen.
Cummings & Richard (1988) used the original data
from the Jarvis study (Jarvis et al. 1987) to demonstrate
that the optimal BCO cut-off value is also influenced by
smoking prevalence; the higher the prevalence of smok-
ing, the lower the optimal BCO cut-off value. At 50%
smoking prevalence, the optimal cut-off was 9 p.p.m. in
that study. In our study, the prevalence of smoking was
high in that 73.4% of days tested were smoking days and
all subjects were smokers. Thus, our cut-off recom-
mendations may be influenced by this high smoking
prevalence.
Nevertheless, findings consistent with those of the
present study were reported by Gariti et al. (2002), who
studied 240 male and female adults at 9, 26 and
© 2005 Society for the Study of Addiction
52 weeks after the completion of a smoking cessation
treatment. They evaluated BCO levels to confirm self
report of smoking using a BCO cut-off of 10 p.p.m. based
on Jarvis et al. (1987). Their findings suggest that sensi-
tivity was sacrificed for specificity, i.e. the rate of false neg-
ative tests was much higher than the rate of false positive
tests. The use of a BCO cut-off of 10 p.p.m. enabled the
correct classification of non-smokers in over 98% of cases
at all three time-points, but the correct classification of
smokers was lower, ranging from 62% to 89%.
In summary, our findings confirm the value of BCO as
an indicator of smoking. We conclude from our results
that BCO cut-off levels well below 8 p.p.m and as low as
2–3 p.p.m. may be more useful when it is important to
maximize identification of non-smokers or to distinguish
abstainers from light smokers. Our results indicate that
when BCO is measured once daily, cut-off levels of 2 or
3 p.p.m. provide the highest percentage of true identifica-
tion of both smoking and abstinence outcomes and there-
fore the highest accuracy for assessment of whether or
not someone has smoked within the past 24 hours.
Acknowledgements
The authors appreciate the editing efforts of Dr Thomas S.
King and gratefully acknowledge the excellent technical
assistance of Mr Floyd Jones and Ms Wendi Stewart-
Rodriguez. The authors also wish to express gratitude
to Dr Maxine Stitzer for exceptional editorial assistance.
This study was possible through a grant from NIDA
(DA 13304).
References
Altman, D. G. & Bland, J. M. (1996) Statistics notes: presenta-
tion of numerical data. BMJ, 312, 572.
Antuni, J. D., Kharitonov, S. A., Hughes, D., Hodson, M. E. &
Barnes, P. J. (2000) Increase in exhaled carbon monoxide
during exacerbations of cystic fibrosis. Thorax, 55, 138–
142.
Benowitz, N. L., Jacob, P., Ahijevych, K., Jarvis, M. F., Hall, S.,
LeHouezec, J., Hansson, A., Lichtenstein, E., Henningfield, J.,
Tsoh, J., Hurt, R. D. & Velicer, W. (2002) Biochemical verifi-
cation of tobacco use and cessation. Nicotine and Tobacco
Research, 4, 149–159.
Biglan, A., Gallison, C., Ary, D. & Thompson, R. (1985) Expired
air carbon monoxide and salive thiocyanate: relationships to
self-reports of marijuana and cigarette smoking. Addictive
Behaviors, 10, 137–144.
Bland, J. M. & Altman, D. G. (1995) Calculating correlation coef-
ficients with repeated observations: Part 2—correlation
between subjects. BMJ, 310, 633.
Colletti, G., Supnick, J. A. & Abueg, F. R. (1982) Assessment of
the relationship between self-reported smoking rate and eco-
lyzer measurement. Addictive Behaviors, 7, 183–188.
Cummings, S. R. & Richard, R. J. (1988) Optimum cutoff points
for biochemical validation of smoking status. American Journal
of Public Health, 78, 574–575.
Addiction, 100, 159–167

Fortmann, S. P., Rogers, T., Vranizan, K., Haskell, W. L.,
Solomon, D. S. & Farquhar, J. W. (1984) Indirect measures of
cigarette use: expired-air carbon monoxide versus plasma
thiocyanate. Preventive Medicine, 13, 127–135.
Gariti, P., Alterman, A. I., Ehrman, R., Mulvaney, F. D. & O’Brien,
C. P. (2002) Detecting smoking following smoking cessation
treatment. Drug and Alcohol Dependence, 65, 191–196.
Gilbert, D. G., Crauthers, D. M., Mooney, D. K., McClernon, F. J.
& Jensen, R. A. (1999) Effects of monetary contingencies on
smoking relapse: influences of trait depression, personality,
and habitual nicotine intake. Experimental and Clinical Psy-
chopharmacology, 7, 174–181.
Horvath, I., Loukides, S., Wodehouse, T., Kharitonov, S. A., Cole,
P. J. & Barnes, P. J. (1998) Increased levels of exhaled carbon
monoxide in bronchiectasis: a new marker of oxidative stress.
Thorax, 53, 867–870.
Jarvis, M. J., Tunstall-Pedoe, H., Feyerabend, C., Vesey, C. &
Saloojee, Y. (1987) Comparison of tests used to distinguish
smokers from nonsmokers. American Journal of Public Health,
77, 1435–1438.
Lamb, R. J., Kirby, K. C., Morral, A. R., Galbicka, G. & Iguchi, M.
Y. (2004a) Improving contingency management programs
for addiction. Addictive Behaviors, 29, 507–523.
Lamb, R. J., Morral, A. R., Kirby, K. C., Iguchi, M. Y. & Galbicka,
G. (2004b) Shaping smoking cesssation using percentile
schedules. Drug and Alcohol Dependence, 76, 247–259.
Middleton, E. T. & Morice, A. H. (2000) Breath carbon monoxide
as an indication of smoking habit. Chest, 117, 758–763.
Montuschi, P., Kharitonov, S. A. & Barnes, P. J. (2001) Exhaled
carbon monoxide and nitric oxide in COPD. Chest, 120, 496–
501.
Rand, C. S., Stitzer, M. L., Bigelow, G. E. & Mead, A. M. (1989)
The effects of contingent payment and frequent workplace
© 2005 Society for the Study of Addiction
Breath carbon monoxide and cigarette smoking 167
monitoring on smoking abstinence. Addictive Behaviors, 14,
121–128.
Roll, J. M. & Higgins, S. T. (2000) A within-subject comparison
of three different schedules of reinforcement of drug absti-
nence using cigarette smoking as an exemplar. Drug and Alco-
hol Dependence, 58, 103–109.
Roll, J. M., Higgins, S. T. & Badger, G. J. (1996) An experimental
comparison of three different schedules of reinforcement of
drug abstinence using cigarette smoking as an exemplar. Jour-
nal of Applied Behavior Analysis, 29, 495–505.
Ruth, K. J. & Neaton, J. D. (1991) Evaluation of two biological
markers of tobacco exposure. Preventive Medicine, 20, 574–
589.
Sato, S., Nishimura, K., Koyama, H., Tsukino, M., Oga, T.,
Hajiro, T. & Mishima, M. (2003) Optimal cutoff level of breath
carbon monoxide for assessing smoking status in patients
with asthma and COPD. Chest, 124, 1749–1754.
Shoptaw, S., Rotheram-Fuller, E., Yang, X., Frosch, D., Nahom,
D., Jarvik, M. E., Rawson, R. A. & Ling, W. (2002) Smoking
cessation in methadone maintenance. Addiction, 97, 1317–
1328.
Tidey, J. W., O’Neill, S. C. & Higgins, S. T. (2002) Contingent
monetary reinforcement of smoking reductions, with and
without transdermal nicotine, in outpatients with schizophre-
nia. Experimental and Clinical Psychopharmacology, 10, 241–
247.
Woodward, M. & Tunstall-Pedoe, J. (1992) An iterative tech-
nique for identifying smoking deceivers with application to
the Scottish heart health study. Preventive Medicine, 21, 88–
97.
Zweig, M. H. & Campbell, G. (1993) Receiver-operating charac-
teristic (ROC) plots: a fundamental evaluation tool in clinical
medicine. Clinical Chemistry, 39, 561–577.
Addiction, 100, 159–167