Association of Recall Rates with

Downloaded from ajronline.org by 94.247.203.6 on 04/11/24 from IP address 94.247.203.6. Copyright ARRS. For personal use only; all rights reserved

Sensitivity and Positive Predictive

Values of Screening Mammography
Bonnie C. Yankaskas 1 OBJECTIVE. The performance of screening mammography is measured mainly by its sensi-
Rebecca J. Cleveland 2 tivity, positive predictive value, and cancer detection rate. Recall rates are also suggested as a sur-
Michael J. Schell 3 rogate measure. The main objective of this study was to measure the effect on sensitivity and
Robert Kozar 1 positive predictive value as recall rates increase in the community practice of mammography.
MATERIALS AND METHODS. Mammography and pathology data are linked in the
Carolina Mammography Registry, a population-based registry of screening mammography.
Our mammography database is created from prospectively collected data from mammogra-
phy facilities; the data include information on the woman and the imaging studies. Our pa-
thology database is created from prospectively collected breast pathology data received from
pathology sites and the Central Cancer Registry. Women in the registry who were 40 years old
and older and who underwent screening mammography between January 1994 and June 1998
were included. “Recall rate” was defined as the percentage of screening studies for which fur-
ther workup was recommended by the radiologist.
RESULTS. The study included 215,665 screening mammograms. The mean age of the women
was 56 years. The recall rates of the average practice ranged from 1.9% to 13.4%. Sensitivity rose
from a mean of 65% in the lowest recall rates to 80.2% at the highest level of recall rates. The pos-
itive predictive value of screening decreased from 7.2% in the lowest level of recall to 3.3% in the
highest. As recall rates increased, sensitivity increased very little beyond a recall rate of 4.8%, and
positive predictive value began decreasing significantly at a recall rate of 5.9%.
CONCLUSION. Practices with recall rates between 4.9% and 5.5% achieve the best
trade-off of sensitivity and positive predictive value.

B reast cancer is the most commonly

diagnosed malignancy in women
in the United States. Although the
mortality rates are declining because of early
detection and treatment at a more curable stage,
182,800 women in the United States are likely
Received January 24, 2001; accepted after revision to be diagnosed with breast cancer in 2001 [1].
March 19, 2001. Screening mammography has been shown to
Supported by grant NCI U01-CA-70040 from the National be an effective means for detecting early-stage
Cancer Institute. breast cancer in women and has resulted in a
1
Department of Radiology, CB 7515, Mason Farm Rd., decrease in breast cancer mortality [2, 3].
University of North Carolina at Chapel Hill, Chapel Hill, The performance of mammography is often
NC 27599-7515. Address correspondence to
B. C. Yankaskas. determined by examining accuracy indexes,
2
Department of Epidemiology, CB 7400, McGavran-
including sensitivity, specificity, positive pre-
Greenberg Hall, University of North Carolina at Chapel Hill, dictive value (PPV), and cancer detection
Chapel Hill, NC 27599-7400. rates. Additionally, recall rates are often calcu-
3
Lineberger Comprehensive Cancer Center, CB 7295, lated by mammography practices and are used
University of North Carolina at Chapel Hill, Chapel Hill, as a surrogate measure of practice perfor-
NC 27599-7295.
mance. Several studies have been conducted
AJR 2001;177:543–549
showing how these indexes relate to age, fam-
0361–803X/01/1773–543 ily, personal history of breast cancer, history of
© American Roentgen Ray Society biopsy, breast parenchymal density, the num-
ber of reviewers, and the number of images
[4–12]. Although guidelines suggest that set-
ting a recall rate at less than 10% will translate
into maximizing the trade-off between sensi-
tivity and PPV [13, 14], no study to date has
shown how recall rates affect sensitivity and
PPV. We performed this analysis on prospec-
tively collected screening assessment data
from a broad representation of mammography
facilities in community practice.
The objective of this study was to estimate the
association of recall rate with PPV and sensitiv-
ity of screening mammography among a wide
range of community-based mammography facil-
ities that are linked to a population-based tumor
registry for outcome data. Knowledge of how re-
call rates relate to both sensitivity and PPV, with
information of other factors that are indepen-
dently associated with these rates, should be a
useful guide for practices to interpret their own
recall rates as surrogate measures of the perfor-
mance of screening mammography.

AJR:177, September 2001 543


Materials and Methods
The Carolina Mammography Registry was cre-
ated for the purpose of collecting population-based
data for the study of screening mammography prac-
tice and performance in the community. The Caro-
lina Mammography Registry is a member of the
national Breast Cancer Surveillance Consortium.
Full descriptions of the registry and the consortium
Downloaded from ajronline.org by 94.247.203.6 on 04/11/24 from IP address 94.247.203.6. Copyright ARRS. For personal use only; all rights reserved
have been published previously [15, 16]. The Caro-
lina Mammography Registry collects prospective
mammography data from mammography facilities
and links these data to a breast pathology database
that includes benign and malignant breast pathology
records collected in an ongoing process from the pa-
thology facilities themselves, from the North Caro-
lina Central Cancer Registry, and from hospital and
private pathology departments.
The study group included all screening mam-
mography examinations for women living in North
Carolina who underwent screening mammography
in any participating facility between January 1994
and June 1998. We identified all standard two-view
bilateral mammography examinations in the regis-
try, then selected those that were performed as
screening studies and excluded those that were diag-
nostic studies. Mammography was considered a
screening study if the patient was classified as as-
ymptomatic by the radiologist, the examination was
bilateral, and the examination was conducted at least
9 months after previous mammography.
The mammography database contains data col-
lected from all patients at the time their mammography
is performed, including demographic information, his-
tory of breast procedures, reason for visit, hormone
use, and family history of breast cancer (defined as a
first-degree relative with premenopausal breast can-
cer). At each visit, the technologist and radiologist
record information about the type of examination per-
formed (screening, diagnostic, sonography, or other di-
agnostic imaging modalities), comparison with
previous films, density of breast parenchyma, mam-
mographic impression using the American College of
Radiology Breast Imaging Reporting and Data System
(BI-RADS) [17] and recommendations for type of fol-
low-up study and follow-up time. Recommendations
for follow-up care may include additional mammogra-
phy at specified intervals (immediately, 6 months, or 1
year), additional mammographic images, breast imag-
ing using another modality, biopsy, or surgical consul-
tation. The number of mammograms obtained
annually at each facility can be calculated from data in
the Carolina Mammography Registry database.
Outcome data from the pathology database in-
clude date and type of pathology, method of bi-
opsy, date of biopsy, breast site, number of nodes
tested and number positive, estrogen and proges-
terone receptor status, and tumor grade and stage.
All data are reviewed for quality. Detected errors
are corrected before the data are entered into the
registry. This project has the approval of the inter-
nal review board of the Medical School of the Uni-
versity of North Carolina at Chapel Hill and holds
a certificate of confidentiality from the United
States Public Health Service.
544
Yankaskas et al.
“Recall rate” is defined as the proportion of
screening mammography examinations resulting
in a recommendation for further workup, includ-
ing recall for imaging studies and invasive proce-
dures if they were recommended on the basis of
the screening mammography findings. Women
who were recommended for recall were consid-
ered to have a positive screening mammogram. A
true-positive study was a positive mammographic
examination that showed invasive breast cancer or
ductal carcinoma in situ that was diagnosed within
12 months of the screening examination. A true-
negative study was a negative mammographic ex-
amination that did not show invasive breast cancer
or ductal carcinoma in situ diagnosed within 12
months of the examination. A false-negative study
was a negative mammographic examination that
resulted in a breast cancer diagnosis within 12
months of the study. A false-positive study was a
positive mammographic examination that resulted
in no diagnosis of cancer within 12 months.
Sensitivity was defined as the proportion of can-
cers that had positive mammographic findings
within 12 months before the cancer diagnosis date,
calculated as true-positive / (true-positive + false-
negative). The PPV for recall was defined as the pro-
portion of positive mammograms for which cancer
was diagnosed within 12 months of the screening,
calculated as true-positive / (true-positive + false-
positive). Practice volume was calculated as the
mean number of screening mammographic exami-
nations performed at the facility per year. For most
analyses, a missing response for family history of
early breast cancer or personal history of breast can-
cer was classified as a negative response. The “can-
cer detection rate” was defined as the number of
cancers discovered through mammography per 1000
women screened. Sensitivity, PPV, cancer detection
rates, and recall rates were calculated for each prac-
tice.
Reduced monotonic regression analysis [18] was
used to model PPV and sensitivity rates as functions
of the recall rate for each of 31 radiology practices
included in the study. Reduced monotonic regres-
sion is a nonparametric approach based on the iso-
tonic regression theory. The reduced monotonic
regression method simplifies the isotonic fit, identi-
fying cut points from an isotonic regression fit, re-
flecting locations in the recall rate at which the trend
in the dependent variable (sensitivity or PPV) is
most manifest. To avoid identifying a cut point be-
tween two practices with nearly identical recall rates,
practices whose recall rates differed by less than 1%
of the range (i.e., by 0.12% in recall rate) were auto-
matically grouped together. For all analyses, the data
for each practice are weighted by the number of
mammograms interpreted.
Linear regression analysis was performed to ex-
amine the association of recall rates with sensitiv-
ity and with PPV and to adjust for relevant
covariates. In a second linear regression analysis,
the recall rates were put into the models according
to the groupings that resulted from the reduced
monotonic regression.
Results
Recall Rates and Screening Mammography Results
A total of 273,105 mammographic studies
were recorded in the Carolina Mammography
Registry for the period from January 1994
through June 1998. After the diagnostic mam-
mograms were excluded, 215,665 screening
mammograms obtained in 155,289 women
were included for study. The mean age of this
population was 56 years; two thirds of the
screening mammograms were obtained in
women younger than 60 years. These women
were 82.5% white, 16.2% black, and 1.3%
“other.” Of all screening mammograms,
13,552 led to a recommendation for further
workup—an overall recall rate of 6.3%. Recall
rates and associated sensitivity, PPV, and can-
cer detection rate are shown by demographics
and breast history in Table 1. The highest re-
call rate (13.6%) among all the subgroups was
seen in women who reported a breast problem
at screening; the lowest (2.4%) was in women
with breasts categorized as almost entirely fat.
The distribution of recommended follow-up
visits was 68%, further imaging follow-up;
16%, biopsy or surgical consultation; and
0.5%, clinical follow-up. Fifteen percent of the
recall follow-ups were missing the specific
recommendation for follow-up (not shown in
table). The 1-year cancer detection rate was
3.5 per 1000 mammograms obtained.
In this population, recall rates decreased with
increasing age, from 7.3% in women younger
than 50 years old to 4.9% in women 70 years
old and older. Women with a history of breast
cancer were recalled at a greater rate than
women whose mammograms were negative for
breast cancer: 7.4% and 6.3%, respectively. For
women who had a history of breast surgery or
breast procedures, the recall rate was 7.2%,
compared with 6.3% for women who did not
have this history. The recall rate decreased with
decreasing breast density, from nearly 7.0% in
the extremely dense and heterogeneously dense
groups to 2.4% in the almost entirely fat group.
Recall rates increased as elapsed time since pre-
vious mammogram increased: 5.2% in women
having recent previous mammography versus
8.7% in women whose previous mammography
was more than 36 months previous or who had
no previous mammography.
The largest difference was found between
women who reported symptoms at the time of
their screening mammogram and those who did
not report symptoms: 13.6% for women report-
ing symptoms, which was twice the 6.1% for
women not reporting symptoms. Recall rates
were greater in black women than in white and
AJR:177, September 2001
 Recall Rates of Screening Mammography other women. No difference was seen for family
TABLE 1 Mammography Performance by Demographic Characteristic
history of breast cancer. The characteristics of
Recall Cancer women giving rise to greater recall rates in this
No. of No. of Sensitivity PPV
Characteristic Rate Detection population included black race, age less than 50
Women Examinations (%) (%)
(%) Rate a years, extremely dense or heterogeneously dense
Age group (yr) breasts, a history of breast biopsy or surgery, re-
<40 13,596 14,454 7.3 60.6 1.9 1.4 port of a current breast problem, and no previous
40–49 49,595 61,378 7.3 69.1 2.7 2.0 mammogram or an elapsed time of more than 36
50–59 41,053 60,057 6.2 74.4 5.2 3.2 months from the previous mammogram.
The relationship of recall rates to practice vol-
Downloaded from ajronline.org by 94.247.203.6 on 04/11/24 from IP address 94.247.203.6. Copyright ARRS. For personal use only; all rights reserved

60–69 28,096 43,544 5.8 77.0 8.2 4.8

ume is seen in Table 2. Recall rates decreased
≥70 22,949 35,232 4.9 83.9 12.7 6.2
with increasing practice volume: 8.3% in the
Race
practices with mean volumes of fewer than 200
Black 26,163 33,779 7.2 74.8 4.8 3.4 mammographic examinations per month com-
White 120,986 171,907 6.2 76.8 5.8 3.6 pared with 5.6% in practices with mean monthly
Other 2,259 2,792 6.2 75.0 3.5 2.1 volumes equal to or greater than 500.
Unknown 5,881 7,187
History of breast cancer Recall Rate and Sensitivity
Self Sensitivity was inversely related to recall rates
Yes 1,387 2,022 7.4 71.4 10.1 7.5 for age; that is, as age increased, the recall rate
No 153,902 213,643 6.3 76.3 5.6 3.5 decreased and sensitivity increased. This same
inverse relationship (decrease in recall rate and
Family
increase in sensitivity) was seen for a decrease in
Yes 10,977 18,650 6.6 74.8 7.0 4.6
breast density, for a personal history of breast
No 99,921 142,708 6.7 75.4 4.9 3.3
cancer, for a history of breast surgery, and for the
Unknown 44,391 54,307 presence of breast symptoms. Mammograms of
History of breast biopsy or women with almost entirely fat breasts showed a
surgery high sensitivity (91.9%) and the lowest recall
Yes 34,522 52,545 7.2 70.0 6.0 4.3 rate (2.4%). Sensitivity was directly related to
No 120,767 163,120 6.0 79.2 5.5 3.3 time since the last mammogram, increasing with
Parenchymal density increasing recall rates as the time increased.
Extremely dense 10,875 14,394 6.9 58.0 4.1 2.8 Screening mammograms with the shortest time
Heterogeneously dense 58,165 82,783 7.0 68.3 4.4 3.1 since the previous mammogram had a recall rate
Scattered fibroglandular 71,952 99,290 5.9 84.0 6.6 3.9 of 5.2% and a sensitivity of 73%, and screening
densities mammograms with the longest time or with no
Almost entirely fat 9,526 13,349 2.4 91.9 10.8 2.5 previous mammogram had a recall rate of 8.7%
Unknown 4,771 5,849 and a sensitivity of 80.7%.
With respect to the volume of mammogra-
Current breast problem
phy by practice, we found that sensitivity did
Yes 3,619 4,873 13.6 64.9 7.5 10.3
not show an association with recall rates.
No 151,670 210,792 6.1 77.1 5.5 3.4
Sensitivity seemed to remain in the mid 70%
Time since last mammogram range, although the recall rate decreased with
Within 24 months 71,962 125,809 5.2 73.3 6.4 3.3 increasing mean practice volume (Table 2).
24–36 months 23,165 27,635 6.0 78.3 5.7 3.4
>36 months 60,162 62,107 8.7 80.7 4.7 4.0 Recall Rate and PPV
Total 155,289 215,665 6.3 76.2 5.6 3.5 PPV was inversely related to recall rate for
Note.—PPV = positive predictive value. age; that is, as age increased, the recall rate for
a Number detected per 1000 mammographic examinations performed. age decreased and the PPV increased. This same
pattern was seen for an increase in breast density
TABLE 2 Mammography Performance by Monthly Volume of Practice and an increase in the time elapsed since previ-
ous mammography. For age, the recall rate de-
Cancer
Mean Practice No. of No. of Recall Rate Sensitivity PPV creased from 7.3% for the youngest women to
Detection
Volume a Women Examinations (%) (%) (%) 4.9% for the oldest women, and PPV rose from
Rate b
1.9% to 12.7%. For breast density, as the recall
<200 30,344 37,132 8.3 77.3 4.4 3.7 rate rose from 2.4% in the almost-entirely-fat
200–499 36,539 46,605 6.7 73.8 4.9 3.3 category to 6.8% in the extremely dense cate-
≥500 88,406 131,928 5.6 76.7 6.4 3.6 gory, the PPV decreased from 10.8% to 4.1%.
Note.—PPV = positive predictive value. PPV was directly related to the recall rate for his-
a Mean number of screening mammographic examinations. tory of breast cancer, history of breast procedure,
b Number detected per 1000 mammographic examinations performed.
and presence of a current breast problem. No as-

AJR:177, September 2001 545

 Yankaskas et al.

for the 31 practices ranged from 1.9% to 13.4%.

TABLE 3 Mammography Performance by Practice
These data were fit using an isotonic regression
Cancer technique for evaluating the relationship be-
No. of No. of Recall Rate Sensitivity PPV
Practice Detection tween practice-specific recall rate and sensitiv-
Women Examinations (%) (%) (%)
Rate a ity. A graphic representation of the results is
1 95 106 1.9 0.0 0.0 0.0 presented in Figure 1. The line running through
2 1,097 1,563 2.1 50 6.1 1.3 the data points shows the isotonic regression fit
Downloaded from ajronline.org by 94.247.203.6 on 04/11/24 from IP address 94.247.203.6. Copyright ARRS. For personal use only; all rights reserved

3 2,172 2,862 3.6 62.5 4.8 1.7 to the data and provides an estimate of how sen-
4 18,186 28,053 3.8 63.7 7.3 2.8 sitivity increases with increasing recall rate. Fig-
ure 2 shows the analogous results for the
5 17,127 21,415 4.4 67.6 7.3 3.2
relationship of PPV as a function of recall rate,
6 812 814 4.9 100 2.5 1.2
where PPV declines with increasing recall rate.
7 897 909 5.2 0.0 0.0 0.0
The reduced monotonic regression shows that
8 4,743 6,230 5.2 76.5 8.0 4.2 the trends for both sensitivity and PPV are sta-
9 1,310 1,310 5.5 77.8 9.7 5.3 tistically significant. (Reduced monotonic re-
10 627 634 5.8 50.0 16.2 9.5 gression models the outcome variable as a
11 7,867 11,284 5.9 65.2 4.5 2.7 constant within groups defined by cut points.)
12 33,485 47,338 5.9 83.9 5.6 3.3 For sensitivity, a single cut point was identified
13 1,460 1,743 6.2 75.0 2.8 1.7 at a recall rate of 4.6% ( p < 0.0001). A cut point
for the relationship between recall rate and PPV
14 4,163 4,368 6.4 72.2 4.7 3.0
was identified at a recall rate of 8.8% ( p <
15 921 926 6.4 100 6.8 4.3
0.001). A second suggestive, although statisti-
16 4,282 5,205 6.5 72.7 7.1 4.6
cally not significant, cut point for PPV ( p =
17 3,133 4,497 7.0 77.8 6.6 4.7 0.13) was obtained at a recall rate of 5.7%. For
18 7,404 8,100 7.2 78.0 5.5 4.0 practices with recall rates between 1.9 and
19 19,608 35,122 7.2 82.7 6.6 4.8 4.4%, the average sensitivity was 65%; for the
20 1,984 2,149 7.5 80.0 5.0 3.7 remaining practices with recall rates greater
21 1,100 1,417 7.6 80.0 3.7 2.8 than 4.4%, the average sensitivity was 80% (Ta-
22 906 1,330 8.3 90.0 8.1 6.8 ble 4). The average PPV was 7.2% for practices
with recall rates of 1.9–5.5%, 5.9% for practices
23 695 782 8.6 75.0 4.5 3.8
with recall rates of 5.8–8.7%, and 3.3% for
24 833 848 8.7 80.0 5.4 4.7
practices with recall rates of 8.9% or greater.
25 1,963 2,692 8.9 83.3 2.1 1.9
Linear regression analysis was first per-
26 5,199 8,360 8.9 78.6 3.0 2.6 formed for recall as a continuous variable con-
27 2,374 2,945 9.5 84.6 3.9 3.7 trolling for the covariates of age, race, family
28 898 907 9.8 0.0 0.0 0.0 and personal history of breast cancer, history of
29 4,124 4,955 9.8 75.0 4.9 4.8 breast surgery or biopsy, presence of symptoms,
30 4,496 5,448 12.0 87.5 3.2 3.9 and breast density. The analysis showed that
31 1,328 1,353 13.4 75.0 3.3 4.4 practices with greater recall rates have greater
sensitivity ( p = 0.003, r 2 = 0.64). When the re-
Note.—Reported values of zero indicate no cancers occurred for that practice during the study period. PPV = positive pre-
dictive value. duced monotonic regression cut point of 4.6%
a Number detected per 1000 mammographic examinations performed. was used for the recall rate, the fit improved to
r 2 = 0.68. Linear regression analysis likewise
sociation of the recall rate with the PPV was 1000. The cancer detection rates are displayed showed that PPV decreased as recall rate in-
seen for racial group or family history of breast in the last column of Table 1. The largest differ- creased ( p = 0.0002), with an overall r 2 of 0.56.
cancer, although the PPV was greater in women ence in recall rate and associated cancer detec- When the practices are split into groups based
with a family history of breast cancer (Table 1). tion was seen for the presence of symptoms at on the reduced monotonic regression cut points,
Looking at practice volume, we found an in- screening. Recall rates rose from 6.1% in the model fit improved to 0.66.
verse relationship of PPV to recall rate for mean women who did not have symptoms to 13.6%
practice volume, with PPV increasing from 4.4% in symptomatic women, with the associated
in practices with the lowest volume to 6.4% in cancer detection rate rising from 3.4 to 10.3 per Discussion
practices with the largest volume, and recall rates 1000. No association between recall rate and The main purpose of this study was to ex-
decreased from 8.3% to 5.6% (Table 2). cancer detection was seen with respect to race, amine the association of recall rates with
family history of breast cancer, or breast density. sensitivity and PPV. We found no published
Recall Rate and Cancer Detection Rate study that focuses on recall rates alone. A
The overall cancer detection rate (cancers Recall Rate, Sensitivity, and PPV search using the terms “recall rates” and
seen on screening mammograms as a percent- The individual data for each practice in the “performance” turned up studies that men-
age of the total screened population) was 3.5 per study are displayed in Table 3. The recall rates tioned recall rates in the pursuit of measuring

546 AJR:177, September 2001

 Recall Rates of Screening Mammography

rates are in the range of what has been re-

ported previously in this literature.
Comparing our results directly with those in
the literature is difficult because of the variation
in the definition of recall (often, a definition is
lacking). In previous studies, “recall” has been
defined on the basis of clinical judgment [29];
Downloaded from ajronline.org by 94.247.203.6 on 04/11/24 from IP address 94.247.203.6. Copyright ARRS. For personal use only; all rights reserved

need for further workup (including clinical

breast examination, sonography, or cytology)
[26]; recommendation for further imaging, in-
cluding repeated mammography, for technical
reasons [11]; recommendation for workup for
clinical or radiographic suspicion of malignancy
[32]; and other factors. Some studies restrict re-
call to further imaging only; other studies may
include recommendations for biopsy. Obviously,
the definition affects the recall rate and thereby
its relationship with PPV and sensitivity. Our es-
timates are internally consistent and were based
Fig. 1.—Line graph shows isotonic regression of recall association with sensitivity according to mammography practice. on recall defined as the proportion of screening
mammography examinations resulting in a rec-
ommendation for further workup, including re-
call for imaging studies and invasive procedures
if they were recommended on the basis of the
screening mammography findings. We did not
include cases recalled for repeated mammogra-
phy requested for technical reasons.
Recall rates in the non–United States litera-
ture tend to be reported separately for prevalent
and incident screening programs [13]. That
method of reporting occurs less often in the
United States, where both are usually com-
bined. We were able to analyze our data by time
elapsed since the previous screening examina-
tion, which approximates the prevalence and in-
cidence screening data. In the Canadian
screening programs, the recall rates were 9.5%
for initial screening and 4.6% for subsequent
screening [24]. Our recall rates were 8.7% for
an interval longer than 36 months (including no
Fig. 2.—Graph shows isotonic regression of recall association with positive predictive value according to mam- previous screening) and 5.2% for an interval of
mography practice. 24 months or less. More than 60% of the
women had previous screening within 3 years
of the present screening. Our evaluation of the
TABLE 4 Sensitivity and Positive Predictive Value of Mammography by Recall Rate
effect of elapsed time between mammograms
No. of No. of No. of No. of shows that decreasing the time between subse-
Recall Rate (%) Sensitivity (%) PPV (%)
Examinations Practices Recalls Cancers quent mammograms reduces recall and im-
1.9–4.4 53,999 5 2,165 238 65.1 (59.1–71.1) 7.2 (6.1–8.3) proves other performance measures. Hunt et al.
4.8–5.5 9,263 4 482 44 77.3 (64.9–89.6) 7.1 (4.8–9.4)
[33] found similar results in their study, indicat-
ing a significant 30% reduction in recall for an-
5.8–8.7 125,743 15 8,232 607 79.7 (76.5–82.8) 5.9 (5.4–6.4)
nual mammography versus biennial, a finding
8.9–13.4 26,660 7 2,673 111 80.2 (72.7–87.6) 3.3 (2.6 –4.0)
that was consistent across age groups.
Note.—Numbers in parentheses are 95% confidence interval. PPV = positive predictive value. Our study was performed on data prospec-
tively collected from 31 community practices
performance [8, 19–25], recall rates related addition, a few specific reports related recall for mammography in North Carolina. The
to the question of double interpretation [10, rates to history of a previous benign biopsy women in this screened population represent
26], and recall rates obtained from single- [8] and the presence of implants or a history approximately 25% of the women more than
view versus two-view screening [27–30]. In of breast cancer [31]. Our findings of recall 40 years old in North Carolina. The women in

AJR:177, September 2001 547


our population are similar to the population
distribution for North Carolina in their age and
racial distribution except for underrepresenta-
tion of older black women, in keeping with the
lower screening rates for these women in gen-
eral, not just in North Carolina [34, 35]. As
others have reported, we found that recall rates
Downloaded from ajronline.org by 94.247.203.6 on 04/11/24 from IP address 94.247.203.6. Copyright ARRS. For personal use only; all rights reserved
were higher for women who were younger;
were black; had a history of breast biopsy,
breast surgery, or breast cancer; or had higher
breast density [7, 9, 36]. We also found that re-
call rates were greater for the presence of re-
ported symptoms at screening and longer time
elapsed since the previous screening. We
found recall rates for women with and those
without a family history of breast cancer to be
virtually the same.
In our data, for the most part, recall rates strat-
ified within the covariates of interest showed an
association with sensitivity. As age increased, re-
call rates decreased and sensitivity increased.
Recall rates increased and sensitivity decreased
with increasing breast density when there was a
history of breast (or other) cancer, previous
breast surgery, or biopsy; or when symptoms
were reported at screening. Only with an in-
crease in time elapsed since the last screening
did recall rates and sensitivity both increase.
For the association of recall rates with PPV,
we found that women in the oldest age group
(≥ 70 years) had a recall rate of 4.9% with a
PPV of 12.7%, and women 60–69 years old
had a recall rate of 5.8% with a PPV of 8.2%.
Welch and Fisher [22] studied recall from
Medicare claims data in women 65 and older
and reported a recall rate of 8.5% and a PPV of
8% for women having further testing within 8
months of screening. We would have expected
this PPV to be higher because it was based on
women who were actually followed up for fur-
ther workup, whereas ours is based on those
for whom further workup was only recom-
mended. Our findings for the relationship be-
tween recall and history of biopsy or surgery
are the same as those reported by Slanetz et al.
[8], although those authors reported recall only
for women with a history of excisional biopsy.
The recall rates reported by Slanetz et al. were
6% in women who did not have a history of bi-
opsy and 7% for women who did have.
The association of recall rates with PPV in
our study showed that as history of breast bi-
opsy, surgery, or cancer, and the presence of
symptoms at screening increased, an in-
crease in recall rates was associated with an
increase in PPV. As age decreased, breast
density increased, time elapsed since previ-
ous mammography increased, recall rates in-
548
Yankaskas et al.
creased, and PPV decreased. Black women
had greater recall rates and lower PPVs than
white women. The percentage of women
with longer than 36 months from the previ-
ous screening or with no previous screening
was also greater for black women than for
white women. This trend of recall and PPV
with increasing elapsed time since screening
may explain the different findings in black
and white women.
No reports exist of the association of recall
with both PPV and sensitivity to compare with
our results. When grouping practices by vol-
ume, we found that an increase in mean vol-
ume of mammograms was associated with a
decrease in recall rate for screening volumes of
more than 200 per month with no difference in
sensitivity and an increase in PPV.
The key question we posed in this study
was whether a point existed at which sensitiv-
ity reached a plateau as the recall rate in-
creased, while PPV continued to decline. We
did not find previous reports addressing this
question. Our regression analyses showed that
practices with recall rates of 4.4% or less had
lower sensitivity than the remaining practices,
which had recall rates of 4.8% or greater.
However, no statistically significant increase in
sensitivity was noted among the practices once
a recall rate of 4.8% was reached. Although we
must be cautious not to overinterpret the loca-
tion of the cut point, both Figure 1 and Table 4
clearly show that, for these data, the sensitivity
increased very little, if at all, beyond a recall
rate of 4.8%. Conversely, a decline in PPV was
observed with increasing recall rates for prac-
tices with recall rates of 8.9% or greater, and to
a lesser extent for practices with recall rates of
5.9–8.7% (Fig. 2). Combining these findings,
we conclude that practices with recall rates be-
tween 4.9% and 5.5% achieve the best trade-
off of sensitivity and PPV.
References
1. Greenlee RT, Murray T, Bolden S, Wingo PA.
Cancer Statistics, 2000. CA Cancer J Clin 2000;
50:7–33
2. Tabar L, Vitak B, Chen HH, et al. The Swedish
two-county trial twenty years later. Radiol Clin
North Am 2000;38:625–651
3. Kerlikowske K, Grady D, Rubin SM, Sandrock
C, Ernster VL. Efficacy of screening mammogra-
phy: a meta-analysis. JAMA 1995;273:149–154
4. Kerlikowske K, Grady D, Barclay J, Sickles EA,
Eaton A, Ernster V. Positive predictive value of
screening mammography by age and family his-
tory of breast cancer. JAMA 1993;270:2444–2450
5. Fracheboud J, De Koning HJ, Beemsterboer PMM,
et al. Nation-wide breast cancer screening in the
Netherlands: results of initial and subsequent screen-
ing 1990-1995. Int J Cancer 1998;75:694–698
6. Kollias J, Sibbering DM, Blamey RW, et al.
Screening women aged less than 50 years with a
family history of breast cancer. Eur J Cancer
1998;34:878–883
7. Kerlikowske K, Grady D, Barclay J, Sickles EA,
Ernster V. Effect of age, breast density, and fam-
ily history on the sensitivity of first screening
mammography. JAMA 1996;276:33–38
8. Slanetz PJ, Giardino AA, McCarthy KA, et al.
Previous breast biopsy for benign disease rarely
complicates or alters interpretation on screening
mammography. AJR 1998;170:1539–1541
9. Stomper PC, D’Souza DJ, DiNitto PA,
Arredondo MA. Analysis of parenchymal density
on mammograms in 1353 women 25–79 years
old. AJR 1996;167:1261–1265
10. Warren RML, Duffy SW. Comparison of single
reading with double reading of mammograms,
and change in effectiveness with experience. Br J
Radiol 1995;68:958–962
11. Seguret F, Daures JP, Guizard AV, et al. Herault
breast screening programme: results after 30
months of a mobile French schedule. Eur J Can-
cer Prev 1995;4:299–305
12. Blanks RG, Wallis RM, Given-Wilson RM. Ob-
server variability in cancer detection during rou-
tine repeat (incident) mammographic screening
in a study of two versus one view mammography.
J Med Screen 1999;6:152–158
13. Perry NM, Kirkpatrick A. European guidelines
for quality assurance in mammography screen-
ing. In: de Wolf CJM, Perry NM, eds. European
guidelines for quality assurance in mammogra-
phy screening, 2nd ed. Luxembourg: Office for
Official Publications of the European Communi-
ties, 1996:I-1–I-15
14. Quality Determinants of Mammography Guide-
line Panel. Quality determinants of mammogra-
phy. Rockville, MD: United States Department of
Health and Human Services, Public Health Ser-
vice, Agency for Health Care Policy and Re-
search, 1994:78–86
15. Yankaskas BC, Jones MB, Aldrich TE. The Caro-
lina Mammography Registry: a population-based
mammography and cancer surveillance project. J
Registry Manage 1996;23:175–178
16. Ballard-Barbash R, Taplin SH, Yankaskas BC, et
al. Breast Cancer Surveillance Consortium: a na-
tional mammography screening and outcomes
database. AJR 1997;169:1001–1008
17. American College of Radiology. Breast imaging re-
porting and data system (BI-RADS), 3rd ed. Res-
ton, VA: American College of Radiology, 1998
18. Schell MJ, Singh B. The reduced monotonic regres-
sion method. J Am Statist Assoc 1997;92:128–135
19. Ng EH, Ng FC, Tan PH, et al. Results of interme-
diate measures from a population-based, random-
ized trial of mammographic screening prevalence
and detection of breast carcinoma among Asian
women. Cancer 1998;82:1521–1528
20. Giordano L, Giorgi D, Fasolo G, Segnan N, Del
Turco MR. Breast cancer screening: characteris-
tics and results of the Italian programmes in the
Italian Group for Planning and Evaluating Breast
Cancer Screening Programmes (GISMa). Tumori
AJR:177, September 2001

1996;82:31–37
21. Robinson JI, Crane CEB, King JM, Scarce DI, Hoff-
mann CEJ. The South Australian Breast X-ray Ser-
vice: results from a statewide mammographic
screening programme. Br J Cancer 1996;73:837–842
22. Welch HG, Fisher ES. Diagnostic testing follow-
ing screening mammography in the elderly. J
Natl Cancer Inst 1998;90:1389–1392
23. Kerlikowske K, Carney PA, Geller B, et al. Perfor-
Downloaded from ajronline.org by 94.247.203.6 on 04/11/24 from IP address 94.247.203.6. Copyright ARRS. For personal use only; all rights reserved
mance of screening mammography among women
with and without a first-degree relative with breast
cancer. Ann Intern Med 2000;133:855–863
24. Paquette D, Snider J, Bouchard F, et al. Performance
of screening mammography in organized programs
in Canada in 1996. CMAJ 2000;163:1133–1138
25. Libstug AR, Moravan V, Aitken SE. Results from
the Ontario breast screening program, 1990–
1995. J Med Screen 1998;5:73–80
26. Brown J, Bryan S, Warren R. Mammography
screening: an incremental cost effectiveness anal-
ysis of double versus single reading of mammo-
AJR:177, September 2001
Recall Rates of Screening Mammography
grams. BMJ 1996;312:809–812
27. Wald NJ, Murphy P, Major P, Parkes C,
Townsend J, Frost C. UKCCCR multicentre ran-
domised controlled trial of one and two view
mammography in breast cancer screening. BMJ
1995;311:1189–1193
28. Warren RML, Duggy SW, Bashir S. The value of
the second view in screening mammography. Br J
Radiol 1996;69:105–108
29. Bryan S, Brown J, Warren R. Mammography screen-
ing: an incremental cost-effectiveness analysis of
two view versus one view procedures in London. J
Epidemiol Community Health 1995;49:70–78
30. Blanks RG, Wallis MG, Moss SM. A comparison
of cancer detection rates achieved by breast can-
cer screening programmes by number of readers,
for one and two view mammography: results
from the UK NHS breast screening programme. J
Med Screen 1998;5:195–201
31. Kopans DB, Moore RH, McCarthy KA, et al.
Should women with implants or a history of treat-
ment for breast cancer be excluded from mammog-
raphy screening programs? AJR 1997;168:29–31
32. De Placido S, Nuzzo F, Perrone F, et al. The first
breast cancer screening program in southern It-
aly: preliminary results from three municipalities
of the Naples province. Tumori 1995;81:7–12
33. Hunt KA, Rosen EL, Sickles EA. Outcome anal-
ysis for women undergoing annual versus bien-
nial screening mammography: a review of 24,211
examinations. AJR 1999;173:285–289
34. Burns RB, McCarthy EP, Freund KM, et al. Black
women receive less mammography even with simi-
lar use of primary care. Ann Intern Med 1996;125:
173–182
35. Caplan LS, Wells BL, Haynes S. Breast cancer screen-
ing among older racial/ethnic minorities and whites:
barriers to early detection. J Gerontol 1992;47:101–110
36. Mandelson MT, Oestreicher N, Porter PL, et al.
Breast density as a predictor of mammographic de-
tection: comparison of interval- and screen-detected
cancers. J Natl Cancer Inst 2000;92:1081–1087
549