C LIN I CA L R ES E AR CH A RT IC LE

Treatment-Free Survival in Patients With Differentiated

Thyroid Cancer

Mousumi Banerjee,1,2 David Reyes-Gastelum,2,3 and Megan R. Haymart2,3

1
Department of Biostatistics, School of Public Health, University of Michigan, Ann Arbor, Michigan 48109;
2
Institute for Healthcare Policy and Innovation, University of Michigan, Ann Arbor, Michigan 48109; and
3
Division of Metabolism, Endocrinology and Diabetes, University of Michigan, Ann Arbor, Michigan 48109

Objective: Cancer recurrence is a primary concern for patients with differentiated thyroid cancer;
however, population-level data on recurrent or persistent disease do not currently exist. The
objective of this study was to determine treated recurrent or persistent thyroid cancer by using a
population-based registry, identify correlates of poor treatment-free survival, and define
prognostic groups for treatment-free survival.

Methods: In this population-based study, we evaluated treatment-free survival in 9273 patients

from the Surveillance, Epidemiology, and End Results Program–Medicare with a diagnosis of
differentiated thyroid cancer between 1998 and 2012. Treated recurrence was defined by
treatment of recurrent or persistent differentiated thyroid cancer with surgery, radioactive
iodine, or radiation therapy at $1 year after diagnosis. Multivariable analysis was performed
with Cox proportional hazards regression, survival trees, and random survival forests.

Results: In this cohort the median patient age at time of diagnosis was 69 years, and 75% of the
patients were female. Using survival tree analyses, we identified five distinct prognostic groups (P ,
0.001), with a prediction accuracy of 88.7%. The 5-year treatment-free survival rates of these prognostic
groups were 96%, 91%, 85%, 72%, and 52%, respectively, and the 10-year treatment-free survival
rates were 94%, 87%, 80%, 64%, and 39%. Based on survival forest analysis, the most important
factors for predicting treatment-free survival were stage, tumor size, and receipt of radioactive iodine.

Conclusion: In this population-based cohort, five prognostic groups for treatment-free survival
were identified. Understanding treatment-free survival has implications for the care and long-
term surveillance of patients with differentiated thyroid cancer. (J Clin Endocrinol Metab 103:
2720–2727, 2018)

or patients with differentiated thyroid cancer, death
F from thyroid cancer is rare, but recurrent or persis-
tent thyroid cancer remains common (1–3). Confirmed
or suspected thyroid cancer recurrence can be associated
with psychological distress and heightened cancer-related
worry (4, 5). Treatment of recurrent or persistent thyroid
cancer increases the risk of complications including vocal
fold paralysis, hypoparathyroidism, sialadenitis, and
lacrimal duct damage (6–9).
Treatment-free survival, which is survival without the
need for treatment of recurrent or persistent cancer, is an
important outcome for patients. Yet population-level
data on treatment-free survival do not currently exist.
Recurrence is not routinely captured by large national
cancer registries such as Surveillance, Epidemiology, and
End Results (SEER) or the National Cancer Database
(10). The applicability of recurrence data from previous
single- and multi-institution studies may be limited by
small sample size and selection bias (3, 11–13).
Understanding treatment-free survival at a population
level is key to improving patient care, tailoring long-term
surveillance, and, in some scenarios, reducing patient

ISSN Print 0021-972X ISSN Online 1945-7197 Abbreviations: AHR, adjusted hazard ratio; HMO, health maintenance organization;
Printed in USA SEER, Surveillance, Epidemiology, and End Results.
Copyright © 2018 Endocrine Society
Received 5 March 2018. Accepted 11 May 2018.
First Published Online 16 May 2018

2720 https://academic.oup.com/jcem J Clin Endocrinol Metab, July 2018, 103(7):2720–2727 doi: 10.1210/jc.2018-00511

Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515

by Kaohsiung Medical University user
on 07 July 2018
 doi: 10.1210/jc.2018-00511
worry. We hypothesized that treatment-free survival in
patients with differentiated thyroid cancer would vary
greatly, allowing us to identify distinct prognostic
groups. In this population-based study evaluating pa-
tients with differentiated thyroid cancer affiliated with
SEER-Medicare, we obtained details on additional
neck surgery, use of radioactive iodine, and use of
radiation at $1 years after diagnosis to determine treated
recurrence, identify correlates of poor treatment-free
survival, and define prognostic groups for treatment-
free survival.
Methods
Data source and study population
SEER-Medicare is a linkage of SEER data to Medicare
claims. SEER is the primary source for cancer statistics in the
United States, and it represents ~28% of the US population (14).
Medicare represents ~15% of the US population, including
most people $65 years old (15). The SEER program of cancer
registries includes clinical and demographic details in addition
to cause of death for patients with cancer. Medicare includes
claims for covered health services. The linkage of these two data
sets allows assessment of long-term cancer care in a population-
based cohort (16).
Data from 43,813 patients with a diagnosis of differentiated
thyroid cancer (papillary, follicular, Hürthle cell) between 1998
and 2012 were queried from SEER-Medicare. We evaluated
patient enrollment in Medicare Part A and B, non–health
maintenance organization (HMO), for the month before di-
agnosis, month of diagnosis, and 12 months after diagnosis. We
did not include patients enrolled in Medicare who were also
enrolled in HMOs because some of their care may not be
captured by Medicare. To increase the likelihood that treatment
of recurrent or persistent disease would be captured by the
registry, only patients enrolled for $11 of the 14 months
evaluated were included (N = 16,054). To avoid capturing
treatment of another cancer, we excluded patients who also
had a diagnosis of a nonthyroid malignancy (N = 11,336).
Because we were evaluating treatment of thyroid cancer re-
currence, only the patients who had an initial thyroid surgery
recorded in both SEER and Medicare were selected for final
analysis (N = 9273).
Institutional review board approval was not required Be-
cause this study involves research using publicly available data
and cannot be tracked to human subjects.
Measures
Patient age at diagnosis was analyzed as a continuous
variable. Patient race was categorized as white, black, and
“other.” Because of the small sample size, Asian, American
Indian/Alaska Native, Native Hawaiian, and Pacific Islander
were grouped into the “other” category. The percentages of
patients $25 years old with a high school diploma only and
household income were categorized based on household zip
code at the time of diagnosis, matched to the 2000 Census data
and the American Community Survey 2008 to 2012. Tumor
characteristics included stage, histology, and tumor size. Stage
was based on SEER stage, with localized including tumors
Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515
by Kaohsiung Medical University user
on 07 July 2018
https://academic.oup.com/jcem 2721
confined to the thyroid or extending into the capsule but not
beyond. Regional included direct extension into blood vessels,
nerves, muscles, thyroid cartilage, and so on, and tumors de-
scribed as “fixed to adjacent tissues.” It also included regional
lymph node involvement. Distant included distant lymph nodes
and extension to bone, mediastinal tissues, and so on (3).
Histology was restricted to International Classification of
Diseases for Oncology classification codes for papillary, fol-
licular, or Hürthle cell cancer (16). Tumor size was catego-
rized based on categories previously defined by American
Joint Committee on Cancer TNM staging: #1 cm, .1 cm
and #2 cm, .2 cm and #4 cm, and .4 cm (17). Initial
treatment characteristics included type of surgery, lymph node
resection, receipt of radioactive iodine, and receipt of radiation.
Surgery was categorized as lobectomy, which involves resection
of one lobe of the thyroid, or total thyroidectomy, which in-
cludes resection of the entire thyroid or near-total thyroidec-
tomy. Lymph node resection was classified as not resected or
resected and negative for lymph node metastases or resected and
positive for lymph node metastases. Receipt of radioactive
iodine and receipt of radiation were dichotomized into yes
or no.
Previous work found that delaying risk stratification until 8
to 12 months after initial treatment improves the predictive
value of risk stratification (12). Because additional treatment in
the first year after diagnosis may be secondary to incomplete
initial treatment and may vary based on treating physician, in
this study treated recurrence was defined as additional
treatment $1 year after diagnosis. This additional treatment
included surgery to resect lymph nodes in the neck, receipt of
radioactive iodine, or receipt of radiation. Treatment-free
survival was defined as the time interval from diagnosis to
treatment of recurrent or persistent disease or to time of cen-
soring. Disease-specific survival was defined as the time in-
terval from diagnosis to death from thyroid cancer or time
of censoring.
Statistical analysis
Survival tree analysis was used to construct distinct prog-
nostic groups, such that within each group patients have similar
treatment-free survival, but treatment-free survival differs be-
tween groups (17–19). In the tree paradigm, the covariate space
is partitioned recursively in a binary fashion. The partitioning
is intended to increase within-group homogeneity. For the
treatment-free survival endpoint (censored), within-group ho-
mogeneity was measured via deviance based on a proportional
hazards model (20). The two groups were recursively parti-
tioned (each group being split on the same or other variables),
creating a tree structure. At each step, to select the best split the
tree-growing paradigm examined every possible cutoff point for
each prognostic variable. This process was continued until the
groups reached a minimum size (,10 patients in each group).
Because the resulting tree was overgrown (thereby overfitting
the data), a subtree was chosen via cost complexity pruning
(20–22). The final tree contained “terminal” groups with
similar treatment-free survival. Five- and 10-year treatment-free
survival rates were calculated as summary measures for each
terminal group. Although this method ensures that left and right
terminal groups from the same parent are significantly different
in terms of treatment-free survival, it is possible that terminal
groups from distinct parents may have similar treatment-free
survival. Therefore, further amalgamation of terminal groups
 2722 Banerjee et al Treatment-Free Survival in Thyroid Cancer
with similar treatment-free survival was performed. For amal-
gamation, we first ordered the terminal groups based on hazard
ratio of a terminal group relative to the leftmost terminal group
of the final tree. The monotone ordering was then coded as a single
ordered covariate, and the survival tree algorithm was used again
to form the final prognostic groups (22).
Next, we performed random survival forest analyses of our
data (17, 23, 24). A random survival forest is an ensemble of
unpruned survival trees, induced from bootstrap samples of the
data, with random feature selection used in the tree induction
process. At each step of the splitting process, instead of eval-
uating all allowable splits on all variables, as is done when
growing a single tree, a subset of the covariates is drawn at
random. We grew 1000 trees in the forest with the treatment-
free survival endpoint. Predictions for each patient were ob-
tained by averaging the predictions across all trees in the forest.
Furthermore, a byproduct of the forest is a ranking of variables
in terms of their relative importance in the forest.
We also performed Cox proportional hazards regression
analyses on this patient cohort with the same variables used in
the tree and forest analyses.
Finally, to address potential confounding by indication with
respect to radioactive iodine use, we performed a propensity
scores–based analysis. The propensity scores were estimated as
the conditional probability of receiving radioactive iodine based
on a multivariable logistic regression model of radioactive io-
dine use (binary), with age, sex, race, education, income, SEER
stage at diagnosis, histology, tumor size, initial surgery, lymph
nodes resected, and receipt of radiation used as covariates.
These scores were rank-ordered and grouped into quartiles.
Subsequently, the propensity score quartiles were used in
the Cox model, tree, and random forest analyses as an
additional covariate.
Prediction accuracy for the Cox model, tree, and random
forest were assessed with a generalized version of the C-statistic
for censored outcomes (25). To safeguard against getting overly
optimistic results by using the same data to build and test the
models, we used out-of-bag predictions for obtaining the ac-
curacy estimates (17, 25).
All analyses were performed in R (R Foundation for Sta-
tistical Computing) and SAS (SAS Institute Inc.) (26, 27).
Specifically, for growing the survival tree and random survival
forest, we used rpart and randomForestSRC packages in R,
respectively (24, 28–30).
Results
Median follow-up was 6.4 years. A total of 1332
(14.4%) of the patients underwent treatment of recurrent
or persistent disease at $1 year after diagnosis. Some
patients received more than one treatment and more
than one type of treatment. At $1 year after initial
treatment, a total of 301 (3.3%) patients received ad-
ditional neck surgery, 978 (10.6%) received treatment
with radioactive iodine, and 435 (4.7%) received
radiation treatment.
Table 1 shows the descriptive data and results from the
Cox proportional hazards regression analysis relating
patient, tumor, and initial treatment characteristics to
Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515
by Kaohsiung Medical University user
on 07 July 2018
J Clin Endocrinol Metab, July 2018, 103(7):2720–2727
treatment-free survival. The median patient age at time of
diagnosis was 69 years (range 21 to 98), and 75% of the
patients were female. Male sex [adjusted hazard ratio
(AHR) 1.35; 95% CI, 1.20 to 1.52], regional (AHR 2.13;
95% CI, 1.82 to 2.49) and distant disease (AHR
4.43; 95% CI, 3.68 to 5.33), Hürthle cell cancer (AHR
1.36; 95% CI, 1.09 to 1.69), and larger tumor size
(.1 cm and #2 cm AHR 1.36; 95% CI, 1.14 to 1.62;
.2 cm and #4 cm AHR 1.69; 95% CI, 1.43 to 2.01;
.4 cm AHR 2.11; 95% CI, 1.74 to 2.55) were signifi-
cantly associated with worse treatment-free survival.
Treatment characteristics including resection of positive
lymph nodes (AHR 1.38; 95% CI, 1.14 to 1.67), receipt
of radioactive iodine (AHR 1.47; 95% CI, 1.29 to 1.68),
and receipt of radiation (AHR 1.68; 95% CI, 1.33 to
2.12) also were significantly associated with worse
treatment-free survival. Results from the propensity
score–based analyses were similar to those of the original
analyses; therefore, we report only the latter. The gen-
eralized C-statistic for the Cox model was 72.4%.
Figure 1 demonstrates the treatment-free survival tree
with the 9273 patients from SEER-Medicare. At each
level of the tree, we show the best splitter. Circles denote
terminal groups in the tree. Within each terminal group,
the first percentage denotes the 5-year treatment-free
survival and the second percentage denotes the 10-year
treatment-free survival. For each terminal group we also
report the crude number of patients with treated re-
current or persistent disease and the total number of
patients in the terminal group. The tree was initially split
by stage (localized vs regional and distant). The localized
group was subsequently split by tumor size (#1 cm
and .1 cm). The group with regional or distant stage was
further split into regional vs distant. Based on the
amalgamation method described earlier, we ultimately
defined five prognostic groups for treatment-free sur-
vival (P , 0.001). At lowest risk for treated recurrence is
group 1 (blue), which includes patients with localized
stage, tumor size #1 cm, and no receipt of radioactive
iodine. Group 2 (orange) includes patients with local-
ized stage, tumor size #1 cm, and receipt of radioactive
iodine along with patients with localized stage, tumor
size .1 cm, and female sex. Group 3 (green) includes
patients with localized stage, tumor size .1 cm, and
male sex in addition to patients with regional stage,
tumor size #2 cm, lymph nodes not resected or lymph
nodes resected and negative. Group 4 (yellow) includes
patients with regional stage, tumor size #2 cm, and
lymph nodes resected and positive along with patients
with regional stage and tumor size .2 cm. Group 5 (red)
has the highest risk of treated recurrence and includes all
patients with distant disease. The generalized C-statistic
for the tree was 88.7%.
 doi: 10.1210/jc.2018-00511 https://academic.oup.com/jcem 2723

Table 1. Association of Patient, Tumor, and Initial Treatment Characteristics With Treatment-Free Survival
Characteristic N (%) AHR (95% CI)
Patient characteristic
Age, y
Median (range) 69 (21–98) 1.00 (0.99–1.00)
Sex
Female 6938 (74.8) 1 (Reference)
Male 2335 (25.2) 1.35 (1.20–1.52)
Race
White 7613 (82.1) 1 (Reference)
Black 680 (7.3) 1.06 (0.85–1.34)
Other 980 (10.6) 1.07 (0.90–1.27)
Percentage with high school education only
,20% 2195 (23.7) 1 (Reference)
20%–29.9% 3168 (34.2) 0.95 (0.82–1.10)
$30% 3554 (38.3) 0.92 (0.79–1.08)
Unknown 356 (3.8)
Household income
,$35,000 1627 (17.6) 1.03 (0.86–1.23)
$35,000–$59,000 4188 (45.2) 1.03 (0.90–1.18)
$$60,000 3098 (33.4) 1 (Reference)
Unknown 360 (3.9)
Tumor characteristic
Stage
Localized 6920 (74.6) 1 (Reference)
Regional 1902 (20.5) 2.13 (1.82–2.49)
Distant 451 (4.9) 4.43 (3.68–5.33)
Histology
Follicular 763 (8.2) 1.19 (0.99–1.44)
Hürthle cell 518 (5.6) 1.36 (1.09–1.69)
Papillary 7992 (86.2) 1 (Reference)
Tumor size
#1 cm 3580 (38.6) 1 (Reference)
.1 cm and #2 cm 2082 (22.5) 1.36 (1.14–1.62)
.2 cm and #4 cm 1972 (21.3) 1.69 (1.43–2.01)
.4 cm 1213 (13.1) 2.11 (1.74–2.55)
Unknown 426 (4.6) 2.00 (1.56–2.57)
Initial treatment characteristic
Surgery
Lobectomy 1922 (20.7) 0.88 (0.73–1.05)
Total thyroidectomy 7351 (79.3) 1 (Reference)
Lymph node resection
Not resected 6047 (65.2) 1.01 (0.87–1.18)
Resected and negative 2065 (22.3) 1 (Reference)
Resected and positive 1161 (12.5) 1.38 (1.14–1.67)
Radioactive iodine
No 5569 (60.1) 1 (Reference)
Yes 3704 (39.9) 1.47 (1.29–1.68)
Radiation
No 8918 (96.2) 1 (Reference)
Yes 355 (3.8) 1.68 (1.33–2.12)
Abbreviation: AHR, adjusted hazard ratio.

Figure 2 illustrates the relative importance of each of
the patient, tumor, and treatment variables for predicting
treatment-free survival based on the random forest anal-
ysis. Stage was most important in predicting treatment-
free survival, followed by tumor size and receipt of
radioactive iodine.
Table 2 lists the five prognostic groups and their re-
spective 5-year and 10-year treatment-free survival along
with their 5-year and 10-year disease-specific survival
Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515
by Kaohsiung Medical University user
on 07 July 2018
rates. The 5-year treatment-free survival rates of groups 1
to 5 were 96%, 91%, 85%, 72%, and 52%, respectively,
and the 10-year treatment-free survival rates were 94%,
87%, 80%, 64%, and 39% (P , 0.001). The 5-year
disease-specific survival rates were 100%, 99%, 99%,
95%, 76%, respectively, and the 10-year disease-specific
survival rates were 99%, 98%, 96%, 89%, and 59%.
Although treatment-free survival decreases incrementally
with progression of the prognostic groups, the patients in
 2724 Banerjee et al Treatment-Free Survival in Thyroid Cancer
Figure 1. Treatment-free survival tree. 5yr, 5-year treatment-free survival; 10yr, 10-year treatment-free survival; N, total number of patients in
the terminal group; R, treated recurrent or persistent disease; RAI, radioactive iodine.
prognostic groups 1 through 4 have a high disease-
specific survival. In contrast, the patients with distant
metastases (group 5, red), have both poor treatment-free
survival and poor disease-specific survival.
Discussion
This study determined treated recurrent or persistent
disease by using data from a population-based registry,
identified correlates of poor treatment-free survival, and
defined prognostic groups for treatment-free survival in
patients with a diagnosis of differentiated thyroid cancer.
Figure 2. Relative importance of each patient, tumor, or treatment
characteristic to treatment-free survival based on the random
survival forest analysis. RAI, radioactive iodine.
Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515
by Kaohsiung Medical University user
on 07 July 2018
J Clin Endocrinol Metab, July 2018, 103(7):2720–2727
Five distinct prognostic groups for treatment-free sur-
vival were identified, with stage, tumor size, and receipt
of radioactive iodine emerging as important correlates of
treatment-free survival.
In the past, studies on thyroid cancer recurrence were
performed at the single- or multi-institution level (3,
11–13, 31–35). In addition to defining risk groups,
previous smaller studies found a relationship between
positive lymph nodes, male sex, and likelihood of re-
currence (31, 32, 36–38). Although age is known to
strongly correlate with risk of death from differentiated
thyroid cancer, there are conflicting data on the re-
lationship between age and risk of recurrence (19, 31, 33,
34). Strengths of previous single- and multi-institution
studies include their ability to include tumor character-
istics and biochemical markers not readily available in
large national cancer registries and the relative homo-
geneity of their treatment protocols (3, 11–13). Limita-
tions include small sample sizes, selection bias, and
potential lack of applicability to the population at large.
Recurrent and persistent disease are not routinely
captured by large national cancer registries such as SEER
and the National Cancer Database (10). Our study uses
Medicare claim data linked to a population-based cancer
registry (SEER) to identify treatment of recurrent or
persistent thyroid cancer. The results of our population-
based study using SEER-Medicare data complement the
findings from previous single- and multi-institution studies
but also differ secondary to our use of a population-based
sample and our focus on recurrence necessitating
an intervention.
In addition to using of claim data to identify treatment
of thyroid cancer recurrence in a population-based co-
hort, our study also uses a method to naturally identify
 doi: 10.1210/jc.2018-00511
Table 2. Prognostic Groups for Treatment-Free Survival and Corresponding Disease-Specific Survival
Group Description
1 Localized stage, tumor #1 cm, no radioactive
iodine
2 Localized stage, tumor #1 cm, + radioactive iodine
or
Localized stage, tumor .1 cm, female
3 Localized stage, tumor .1 cm, male
or
Regional stage, tumor #2 cm, lymph nodes not
resected or resected and negative
4 Regional stage, tumor #2 cm, lymph nodes
resected and positive
or
Regional stage, tumor .2 cm
5 Distant stage
risk groups. Although our initial assessment was with
Cox proportional hazards regression, which allowed us
to define correlates of poor treatment-free survival,
limitations of Cox proportional hazards regression are
the need to determine interaction a priori and the in-
ability to group patients into well-characterized risk
groups in a natural way. The survival tree analysis
partitions the cohort into risk groups that are most
homogeneous within and most heterogeneous between
with regard to the defined outcome (i.e., treatment-free
survival in our context). Terminal groups characterized
by different branches can then be regrouped based on
similar outcomes, via a post hoc amalgamation process.
In this study we identified five distinct prognostic groups
for treatment-free survival. Previous recurrence risk
classification systems have fewer risk groups (3, 31, 35).
Although some of the previously defined risk groups
overlap with ours, unlike our study previous risk strat-
ification systems do not separate patients with distant
metastases into a separate “very high” risk group (3, 31,
35, 39). However, our large cohort size and risk strati-
fication method (i.e., tree analysis) allowed us to identify
patients at very high risk of treated recurrent or persistent
disease (group 5). We also evaluated the role of initial
treatment and likelihood of treated recurrence. Receipt of
radioactive iodine and radiation correlated significantly
with worse treatment-free survival based on Cox pro-
portional hazards regression, and after stage and tumor
size, receipt of radioactive iodine was the third most
important correlate of treated recurrent or persistent
disease in the survival forest analysis. We used propensity
score analysis, which can reduce confounding by in-
dication due to measured variables. However, the per-
sistent relationship between receipt of radioactive iodine
and worse treatment-free survival after the addition of
propensity score analysis is unlikely to be secondary to a
Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515
by Kaohsiung Medical University user
on 07 July 2018
https://academic.oup.com/jcem 2725
Treatment-Free Survival Disease-Specific Survival
5-y (95% CI) 10-y (95% CI) 5-y (95% CI) 10-y (95% CI)
0.96 (0.95–0.97) 0.94 (0.93–0.95) 1.00 (0.99–1.00) 0.99 (0.99–1.00)
0.91 (0.90–0.092) 0.87 (0.86–0.89) 0.99 (0.98–0.99) 0.98 (0.97–0.98)
0.85 (0.83–0.87) 0.80 (0.78–0.83) 0.99 (0.98–1.00) 0.96 (0.95–0.98)
0.72 (0.70–0.75) 0.64 (0.60–0.67) 0.95 (0.94–0.96) 0.89 (0.86–0.91)
0.52 (0.47–0.58) 0.39 (0.33–0.46) 0.76 (0.72–0.80) 0.59 (0.53–0.66)
causal relationship but instead may occur because receipt
of radioactive iodine correlates with additional un-
measured variables that may also be associated with
greater likelihood of treated recurrence. For example, it is
plausible that a higher postoperative thyroglobulin, a
variable not captured by this data set, would be asso-
ciated with both greater likelihood of receiving radio-
active iodine and greater likelihood of recurrence.
Interestingly, although treatment-free survival de-
creased sequentially within each recurrence risk group,
disease-specific survival remained excellent with the
exception of group 5, which included patients with
distant metastases. The fact that there is not a one-to-one
correlation between recurrent or persistent disease and
mortality is an expected finding for physicians who treat
patients with differentiated thyroid cancer, but the ex-
planation for this discordance remains unclear. The
marked difference between treatment-free survival and
disease-specific survival in the majority of the prognostic
groups could be secondary to the efficacy of treatments
for recurrence, our ability to detect low-volume indolent
recurrence that may not affect survival, or pathogenic
differences in cancers, with some cancers at greater risk
for progression that leads to death.
As in other studies using claim data, limitations of our
study include risk for coding errors and reporting bias.
However, to reduce the risk of reporting bias, we re-
stricted the cohort to patients who were enrolled in
Medicare Part A and B, non-HMO (~80% or more),
during the first year after their thyroid cancer diagnosis.
An additional limitation is the fact that adequacy of
initial treatment, response to initial treatment, bio-
chemical evidence of recurrence, and structural re-
currence not necessitating an intervention cannot be
captured through the use of Medicare data. SEER-
Medicare data can be used to identify treated recurrent
 2726 Banerjee et al Treatment-Free Survival in Thyroid Cancer
or persistent disease, but recurrences that remain un-
treated will be missed (40). However, defining recurrence
as additional treatment at $1 year after diagnosis reduces
the likelihood of identifying correlates of incomplete
initial therapy and focuses instead on the recurrent or
persistent disease that is most clinically relevant. With use
of this data set, we also lacked the ability to reliably
capture use of tyrosine kinase inhibitors, thus potentially
missing some patients with progressive recurrent or
persistent disease. In addition, when using SEER data to
capture initial treatment, such as initial treatment with
radioactive iodine, there is a risk of underascertainment if
adjuvant therapy occurred outside the hospital setting.
Finally, although this cohort includes patients as young
as 21 years, the median patient age at time of diagnosis
was 69 years. SEER-Medicare represents a largely older
patient cohort, and this bias may reduce the applicability
of the findings to younger patients. Previous work in
other cancer types has found that compared with
younger patients, patients $70 years old may be less
likely to have their recurrence treated (40). However, the
potential limitations of using SEER-Medicare data are
outweighed by the fact that this is a source of population-
based recurrence data in the United States.
Recurrent or persistent disease is a primary concern
for many patients with cancer but especially for patients
with differentiated thyroid cancer because death from
thyroid cancer is rare, but recurrent or persistent disease
is common. The results of this study have implications for
patients and their providers. Understanding population-
based data on treatment-free survival will allow im-
proved decision making at time of initial treatment.
Patients with characteristics associated with both ex-
cellent treatment-free survival and disease-specific sur-
vival may be candidates for less intensive treatment,
which has a lower risk of treatment-related complica-
tions. Similarly, patients with characteristics associated
with both excellent treatment-free survival and disease-
specific survival may be able to have their long-term
surveillance transferred from specialists to primary
care doctors sooner than those with characteristics as-
sociated with poor treatment-free survival. Because some
patients with poor treatment-free survival still have ex-
cellent disease-specific survival, information on the
prognostic groups for treatment-free survival and their
corresponding disease-specific survival could be used to
help reduce worry in many patients with recurrence. For
patients with cancer, fear of recurrence can be associated
with psychological distress, and treatment of recurrence
can be associated with increased morbidity (4–8). Un-
derstanding the risk of recurrent or persistent disease,
correlates of poor treatment-free survival, and prognostic
groups for treatment-free survival is key to improving
Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515
by Kaohsiung Medical University user
on 07 July 2018
J Clin Endocrinol Metab, July 2018, 103(7):2720–2727
treatment decision making, tailoring the intensity of
long-term surveillance, and in some cases reassuring
worried patients.
Acknowledgments
The authors thank Ms. Brittany Gay, who assisted with creating
the tables and editing the manuscript.
Financial Support: M.R.H. is supported by grants from the
National Institutes of Health (R01 CA201198) and the Agency
for Healthcare Research and Quality (R01 HS024512-01A1).
Correspondence and Reprint Requests: Megan R. Haymart,
MD, Division of Metabolism, Endocrinology, and Diabetes, and
Hematology/Oncology, University of Michigan Health System,
North Campus Research Complex, 2800 Plymouth Road
Building 16, Room 408E, Ann Arbor, Michigan 48109. E-mail:
meganhay@med.umich.edu.
Disclosure Summary: The authors have nothing to
disclose.
References
1. Davies L, Welch HG. Increasing incidence of thyroid cancer in the
United States, 1973–2002. JAMA. 2006;295(18):2164–2167.
2. Lim H, Devesa SS, Sosa JA, Check D, Kitahara CM. Trends in
thyroid cancer incidence and mortality in the United States,
1974–2013. JAMA. 2017;317(13):1338–1348.
3. Tuttle RM, Tala H, Shah J, Leboeuf R, Ghossein R, Gonen M,
Brokhin M, Omry G, Fagin JA, Shaha A. Estimating risk of re-
currence in differentiated thyroid cancer after total thyroidectomy
and radioactive iodine remnant ablation: using response to therapy
variables to modify the initial risk estimates predicted by the new
American Thyroid Association staging system. Thyroid. 2010;
20(12):1341–1349.
4. Bresner L, Banach R, Rodin G, Thabane L, Ezzat S, Sawka AM.
Cancer-related worry in Canadian thyroid cancer survivors. J Clin
Endocrinol Metab. 2015;100(3):977–985.
5. Misra S, Meiyappan S, Heus L, Freeman J, Rotstein L, Brierley JD,
Tsang RW, Rodin G, Ezzat S, Goldstein DP, Sawka AM. Patients’
experiences following local-regional recurrence of thyroid cancer:
a qualitative study. J Surg Oncol. 2013;108(1):47–51.
6. Lamartina L, Borget I, Mirghani H, Al Ghuzlan A, Berdelou A,
Bidault F, Deandreis D, Baudin E, Travagli JP, Schlumberger M,
Hartl DM, Leboulleux S. Surgery for neck recurrence of differ-
entiated thyroid cancer: outcomes and risk factors. J Clin Endo-
crinol Metab. 2017;102(3):1020–1031.
7. Starmer H, Noureldine SI, Ozgursoy OB, Tufano RP. Voice out-
comes following reoperative central neck dissection for recurrent/
persistent thyroid cancer. Laryngoscope. 2015;125(11):2621–2625.
8. Mendoza A, Shaffer B, Karakla D, Mason ME, Elkins D, Goffman
TE. Quality of life with well-differentiated thyroid cancer: treat-
ment toxicities and their reduction. Thyroid. 2004;14(2):133–140.
9. da Fonseca FL, Yamanaka PK, Kato JM, Matayoshi S. Lacrimal
system obstruction after radioiodine therapy in differentiated
thyroid carcinomas: a prospective comparative study. Thyroid.
2016;26(12):1761–1767.
10. In H, Bilimoria KY, Stewart AK, Wroblewski KE, Posner MC,
Talamonti MS, Winchester DP. Cancer recurrence: an important
but missing variable in national cancer registries. Ann Surg Oncol.
2014;21(5):1520–1529.
11. Vaisman F, Momesso D, Bulzico DA, Pessoa CH, Dias F, Corbo R,
Vaisman M, Tuttle RM. Spontaneous remission in thyroid cancer
patients after biochemical incomplete response to initial therapy.
Clin Endocrinol (Oxf). 2012;77(1):132–138.
 doi: 10.1210/jc.2018-00511
12. Castagna MG, Maino F, Cipri C, Belardini V, Theodoropoulou A,
Cevenini G, Pacini F. Delayed risk stratification, to include the
response to initial treatment (surgery and radioiodine ablation), has
better outcome predictivity in differentiated thyroid cancer pa-
tients. Eur J Endocrinol. 2011;165(3):441–446.
13. Jonklaas J, Sarlis NJ, Litofsky D, Ain KB, Bigos ST, Brierley JD,
Cooper DS, Haugen BR, Ladenson PW, Magner J, Robbins J, Ross
DS, Skarulis M, Maxon HR, Sherman SI. Outcomes of patients
with differentiated thyroid carcinoma following initial therapy.
Thyroid. 2006;16(12):1229–1242.
14. Surveillance, Epidemiology, and End Results Program (SEER).
Available at: www.seer.cancer.gov. Accessed 29 January 2018.
15. Medicare.gov: the official U.S. government site for Medicare.
Available at: www.medicare.gov. Accessed 1 February 2018.
16. National Cancer Institute, Division of Cancer Control and Pop-
ulation Sciences. SEER-Medicare database. 2017. Available at:
healthcaredelivery.cancer.gov/seermedicare/overview/. Accessed 20
November 2017.
17. Banerjee M, Noone AM. Tree-based methods for survival data. In:
Biswas A, Datta S, Fine J, Segal M, eds. Statistical Advances in
Biomedical Sciences: State of the Art and Future Directions. New
York, NY: Wiley; 2008:265–285.
18. Banerjee M, George J, Song EY, Roy A, Hryniuk W. Tree-based
model for breast cancer prognostication. J Clin Oncol. 2004;
22(13):2567–2575.
19. Banerjee M, Muenz DG, Chang JT, Papaleontiou M, Haymart
MR. Tree-based model for thyroid cancer prognostication. J Clin
Endocrinol Metab. 2014;99(10):3737–3745.
20. LeBlanc M, Crowley J. Relative risk trees for censored survival
data. Biometrics. 1992;48(2):411–425.
21. Breiman L, Frieman JH, Olshen RA, Stone CJ. Classification and
Regression Trees. Belmont, CA: Wadsworth; 1984.
22. LeBlanc M. Tree-based methods for prognostication stratification.
In: Crowley J, ed. Handbook of Statistics in Clinical Oncology.
New York, NY: Marcel Drekker, Inc.; 2001:457–472.
23. Breiman L. Random forests. Mach Learn. 2001;45(1):5–32.
24. Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random
survival forests. Ann Appl Stat. 2008;2(3):841–860.
25. May M, Royston P, Egger M, Justice AC, Sterne JA; ART Cohort
Collaboration. Development and validation of a prognostic model
for survival time data: application to prognosis of HIV positive
patients treated with antiretroviral therapy. Stat Med. 2004;23(15):
2375–2398.
26. SAS Institute Inc. SAS version 9.4. Cary, NC: SAS Institute Inc.;
2002–2012.
27. R Core Team. R: A Language and Environment for Statistical
Computing. Vienna, Austria: R Foundation for Statistical Com-
puting; 2017.
28. Therneau T, Atkinson B, Ripley B. rpart: Recursive partitioning
and regression trees. R package [computer program]. Version 4.1-
11; 2017.
Downloaded from https://academic.oup.com/jcem/article-abstract/103/7/2720/4996515
by Kaohsiung Medical University user
on 07 July 2018
https://academic.oup.com/jcem 2727
29. Ishwaran H, Kogalur UB. Random forests for survival, regression,
and classification (RF-SRC). R package [computer program].
Version 2.5.1; 2017.
30. Ishwaran H, Kogalur UB. Random survival forests for R. R News.
2007;7(2):25–31.
31. Kim KM, Park JB, Bae KS, Kim CB, Kang DR, Kang SJ. Clinical
prognostic index for recurrence of papillary thyroid carcinoma
including intraoperative findings. Endocr J. 2013;60(3):291–297.
32. Ito Y, Higashiyama T, Takamura Y, Miya A, Kobayashi K,
Matsuzuka F, Kuma K, Miyauchi A. Risk factors for recurrence to
the lymph node in papillary thyroid carcinoma patients without
preoperatively detectable lateral node metastasis: validity of pro-
phylactic modified radical neck dissection. World J Surg. 2007;
31(11):2085–2091.
33. Adam MA, Thomas S, Hyslop T, Scheri RP, Roman SA, Sosa JA.
Exploring the relationship between patient age and cancer-specific
survival in papillary thyroid cancer: rethinking current staging
systems. J Clin Oncol. 2016;34(36):4415–4420.
34. Pitoia F, Jerkovich F, Smulever A, Brenta G, Bueno F, Cross G.
Should age at diagnosis be included as an additional variable in the
risk of recurrence classification system in patients with differen-
tiated thyroid cancer. Eur Thyroid J. 2017;6(3):160–166.
35. Haugen BR, Alexander EK, Bible KC, Doherty GM, Mandel SJ,
Nikiforov YE, Pacini F, Randolph GW, Sawka AM, Schlumberger
M, Schuff KG, Sherman SI, Sosa JA, Steward DL, Tuttle RM,
Wartofsky L. 2015 American Thyroid Association management
guidelines for adult patients with thyroid nodules and differentiated
thyroid cancer: the American Thyroid Association Guidelines Task
Force on Thyroid Nodules and Differentiated Thyroid Cancer.
Thyroid. 2016;26:1–133.
36. Kim Y, Roh JL, Gong G, Cho KJ, Choi SH, Nam SY, Kim SY. Risk
factors for lateral neck recurrence of N0/N1a papillary thyroid
cancer. Ann Surg Oncol. 2017;24(12):3609–3616.
37. Kim SY, Kim BW, Pyo JY, Hong SW, Chang HS, Park CS.
Macrometastasis in papillary thyroid cancer patients is associated
with higher recurrence in lateral neck nodes. World J Surg. 2018;
42(1):123–129.
38. Kluijfhout WP, Drake FT, Pasternak JD, Beninato T, Vriens MR,
Shen WT, Gosnell JE, Liu C, Suh I, Duh QY. Incidental positive
lymph nodes in patients with papillary thyroid cancer is in-
dependently associated with recurrent disease. J Surg Oncol. 2017;
116(3):275–280.
39. Pitoia F, Bueno F, Urciuoli C, Abelleira E, Cross G, Tuttle RM.
Outcomes of patients with differentiated thyroid cancer risk-
stratified according to the American Thyroid Association and
Latin American Thyroid Society risk of recurrence classification
systems. Thyroid. 2013;23:1401–1407.
40. Warren JL, Mariotto A, Melbert D, Schrag D, Doria-Rose P,
Penson D, Yabroff KR. Sensitivity of Medicare claims to identify
cancer recurrence in elderly colorectal and breast cancer patients.
Med Care. 2016;54(8):e47–e54.