Att 12_Papers and Presentations

SEARCH TECHNICAL REPORT
Estimation of Completeness of Case Ascertainment Using Capture-Recapture

Richard F. Hamman, MD DrPH1
Dana Dabelea, MD, PhD1
Angela D. Liese, PhD2
Lenna L. Liu, MD, MPH3
Jennifer W. Talton, MS4
Andrea Anderson, MS4
Mark A. Espeland, PhD4
Kristi Reynolds, PhD5
Scott Isom, MS4
Jasmin Divers, PHD4
July, 2015
based on prior reports from 2005, 2007 and 2013
1

2

3
4

5

University of Colorado Denver, Colorado School of Public Health, Department of
Epidemiology
Arnold School of Public Health, University of South Carolina, Department of Epidemiology
and Biostatistics
University of Washington School of Medicine, Department of Pediatrics
SEARCH Coordinating Center, Wake Forest University School of Medicine, Department of
Biostatistical Sciences
Kaiser Permanente Southern California

Background
The purpose of this technical report is to summarize SEARCH activities to estimate the
completeness of case ascertainment using the capture-recapture (C-R) method. The goal is to
estimate the total size of the population of youth with diabetes aged 0-19 in a population, when
the size of that population is not known. Case ascertainment through multiple sources provides a
count of the number of cases found, but the number not identified remains unknown and must be
estimated. The C-R method1 was developed from animal biology to estimate the size of rodent
populations, but it has been applied extensively to human disease situations 2-7. The unknown
total population size is estimated based on the number of cases found in more than one source
(e.g., duplicate records from multiple hospitals, health care offices, and other sources). The
approach is shown in Figure 1. Since two or more sources are required for C-R, it was not
possible to use it in the SEARCH sites primarily utilizing one data source. These include the
Kaiser Permanente Southern California site and participating Native American Tribes. The
Kaiser Permanente site uses information from multiple health databases (laboratory, pharmacy,
inpatient and outpatient encounters) and direct case reports from pediatric endocrinologists but
these sources are not independent. Native American tribes used a single source, the Indian
Health Service RPMS record system. This report further explores the use of two mode or
multiple-mode sources in a systematic way for all four geographic sites.
Figure 1. Estimation of the total (unknown) population size.

Total population
(unknown)

Sample 2
Sample 1
Number of duplicates

Table 1 summarizes the data display for a simple two source situation, where N can be estimated
algebraically as shown in the formula if it is assumed that the two sources identify cases
independently (this assumption cannot be verified without additional data).

Table 1: Summary of capture-recapture calculations
Source 1
Yes
No
Total

Yes
A
C
A+C

Source 2
No
B
X=?
?

Total
A+B
?
N=?

The addition of 1 to each cell counts prevents the estimated totals from taking nonsensical values
like 0 and infinity.

Assumptions
Traditional C-R methods, such as we have adopted, make the following assumptions1: Cases
are:
 From the same space and time. This means that geographic and temporal residence is the
same for all members of the population and can be determined similarly in all cases.


Identical with respect to how likely they are to be identified. This assumption means that
every case has the same probability of being identified by a given source, i.e., that some
cases are not inherently easier or less difficult to identify than others. This is rarely met in
health care studies.



Independently ascertained by separate modes. The assumption of independence of sources is
rarely met in disease ascertainment but can be dealt with using log-linear models with
interaction terms to estimate and model the source dependence when more than two modes of
ascertainment are involved. When only two ascertainment modes are available, the
assumption of independence cannot be assessed.



Matched between modes of ascertainment. The assumption of equal matching between
modes of ascertainment assumes that sufficient data are available on personal identifiers from
each source to be ‘certain’ that cases identified in multiple sources are, or are not, the same
person. This may vary across sites and within sites across sources, depending on the amount
of personal information provided by a source.



Cases have been validated. This assumes that cases truly have diabetes and that this can be
determined in each source.



Cases are from a closed population. This assumption means that cases in the total population
are not moving in or out of the population during the time interval.

Methods
In each of the geographic sites (Colorado, Ohio, South Carolina, and Washington) cases were
identified from multiple sources (CO: 13+; Ohio: 20; SC: 41; WA: 26). A “source” was defined
as any location where cases were reported. Sources were then aggregated. First, all individual
small practices were usually grouped into ‘practices’, but these were initially maintained
distinctly from larger pediatric endocrine practices, HMOs providing larger numbers of cases,
etc. Individual hospitals were also maintained separately. Matching across sources was done on a
regular basis as cases were reported to identify potential duplicate records. Initial computerized
listings were generated, sorted and compared using available personal health identifiers (PHI),
and then manual matching was completed. The amount of PHI available to conduct the
matching across sites differed by site, with some sites unable to identify names at the first receipt
of data. Once matching was accomplished across sources, the sources were further grouped into
‘modes’ of ascertainment. After exploratory analyses, all provider sources were aggregated, as
were all hospital system records and modes were defined for all sites as ‘provider’ and ‘hospital’.
Several sources were large health care systems that included both ambulatory and inpatient
facilities (e.g., Children’s Hospital, Seattle). In these cases, manual review of records categorized
youth by whether they had been cared for in either one or both portions of the system to allow
better classification of the mode of ascertainment. As of January 2015, the 2 modes of
ascertainment in SEARCH will be defined as “inpatient” and “outpatient”. A SEARCH study
participant is considered to be inpatient if the participant was ascertained during a hospital visit
that included at least one overnight stay. Cases ascertained during a visit that did not involve an
overnight stay will be classified as outpatient. For example, a participant identified during an
emergency room visit will be classified as inpatient if the visit went from 6:00 PM to 6:00 AM
the next day, or outpatient if the visit started at 6:00 AM and ended at 6:00 PM the same day.
The inpatient mode will include most of the sources that were initially classified as ‘hospital’.
Similarly, the outpatient mode will consist of the majority of sources previously included in the
‘provider/other’ source. The list of sources and corresponding mode of ascertainment is provided
in the appendix.
Once two modes were identified and their duplicates noted, log linear models8,24,25 were fitted to
the data to estimate the total (unknown) population. These estimates were computed separately
for prevalent 2001, prevalent 2009 and incident 2002 to 2009 youth. The models were fitted
using all the data that was available in each subset adjusting for relevant covariates including and
site. Multiple mode interaction models were evaluated systematically for each of the four
geographic sites. For models with more than two modes, an estimate of the ‘best’ model was
based on identifying the minimum value (best fit) of an information criterion statistic16 defined
as:

Where:
G2: Likelihood ratio statistic (-2 logarithm of the ratio of the likelihood of the fitted model to the
likelihood of the saturated model);

df: Number of degrees of freedom for the comparison of any fitted model with the saturated
model;
c: A constant that varies with the method use to estimate the information criterion. For the
minimum Akaike information criterion (AIC), c = 2.
The percent completeness of ascertainment for any group was estimated as the number of
observed cases divided by the total number estimated from C-R. Estimates of the ascertainment
rates pooled across clinical sites were produced from a global log-linear model 8, which allowed
for separate intra-site performance. The rates were estimated using maximum likelihood and the
standard errors were estimated using the delta method 9.

Results
Four different ascertainment modes were initially defined for the analysis. In the “provider”
mode, practices were split into “endocrine” and “other”, and “hospitals” were divided into
“hospitals only” and “integrated practices”. There were too few cases in some locations in each
of the four modes to successfully use this approach, so “hospitals” was used as a single mode,
and there were two practice modes (endocrine and other) to allow a three mode model. Three
mode models were explored allowing all possible 3-way interactions, and the ‘best’ model was
chosen with the lowest IC value. The overall estimate of completeness of ascertainment was then
compared to the 2 mode model (from which no IC value can be calculated). As shown in Figure
2 below, there was wide variability in the 3 mode estimates across centers, ranging (for
prevalence) from 44 to 99% complete. This 3 mode estimate can be compared to the range of the
2 mode estimates from 89 to ~100% across sites. For incidence estimates, the 3 mode models
ranged from 15 to 98%, whereas the 2 mode model was much more consistent – from 86% to ~
100%. Results of the different models within site also showed substantial heterogeneity. Across
sites, there were several different patterns of interactions between sites – that is, there were not
consistent types of modes that interacted across centers. In Ohio, the 2 and 3- mode approaches
gave almost identical results, since all estimates were > 97%. The widest changes within a single
site occurred in Washington, where for prevalence, the 2 mode estimate was 94% and the 3 mode
was 49%; for incidence it was 86% for 2 mode and 15% for the “best” 3 mode model.

Figure 2. Estimates of case ascertainment completeness for prevalence (2001) and incidence
(2002) using a 2 mode (blue bar) and 3 mode (red bar) capture-recapture model, by geographic
site.
Prevalence 2001

2 mode
3 mode

Estimated % complete

100
80
60
40
20
0
Colorado

South Carolina

Washington

Incidence 2002

Ohio

2 mode
3 mode

Estimated % complete

100
80
60
40
20
0
Colorado

South Carolina

Washington

Ohio

Based on the heterogeneity of results from the 3 mode (interaction) models, which appeared to
be due largely to site specific differences in patterns of care, reporting, location of duplicate
cases, and statistical variability, it was decided that the 2 mode model provided more consistent
results with better face validity. For example, the 3 mode model suggested that Washington
missed over 1500 cases in 2002, more than 3 times the number actually identified. Another
rationale for choosing the 2 mode estimate comes from the consistency of the incidence rates by
site. This consistency is shown for total incidence (all types) in 2002 in Table 2. If the 3 mode
model were correct, it would suggest that rates in Washington and South Carolina would be
substantially lower than actually observed.

Table 2. Total incidence rates of diabetes (all types) by geographic sites, 2002.
Center

Youth with
DM

Population
Denominator (Personyears)

Incidence Rates
(per
100,000/year)

95% CI
(per
100,000/year)

Ohio
South Carolina
Washington
Colorado

355
539
509
655

1,097,960
2,170,362
1,927,958
2,553,884

32.3
24.8
26.4
25.7

29.1-35.9
22.9-27.0
24.2-28.8
23.8-27.8

The incidence rate in Ohio was ~ 25% higher than rates in the other three geographic sites and
Ohio also had the highest estimated completeness. However, rates for South Carolina,
Washington and Colorado were quite similar (24.8-26.4) while estimates of completeness ranged
from 86% (Washington) to 97% (Colorado) (Figure 3). It cannot be ruled out that higher rates in
Ohio were due in part to slightly higher estimated completeness, however, over this narrow
range, it did not appear to influence rates in the other three sites.
Figure 3. Incidence rates (per 100,000/yr.) by estimated completeness from capture-recapture (2
mode) by site, SEARCH 2002 incidence

For these capture-recapture analyses, we therefore chose a 2 mode model to estimate
completeness by site and overall. Table 3 shows the results using C-R for the four geographic
sites using a two mode ascertainment model (‘hospital’ vs. ‘provider sources’ combined) for
prevalent and incident cases, by site and age group.

Table 3. Summary of percent completeness of ascertainment by capture-recapture analysis by
year, site and age-group.
SEARCH 2001-2002

Year
Site
2001
Colorado (1)
Prevalent Washington
Ohio
South Carolina (1)
All sites (2)

0-4
95.3
95.9
100.0
94.4
96.3

Colorado (3)
Washington
Ohio
South Carolina (3)
All sites (2)

99.0
94.4
--(4)
97.2
97.5

2002
Incident

Age group at diagnosis
5-9
10 – 14 15 – 19
91.6
92.0
83.9
92.5
89.8
87.3
100.0
99.9
99.4
96.6
99.6
94.9
94.1
93.3
90.9
98.8
87.3
99.9
98.3
95.1

93.7
83.7
100.0
96.2
92.8

91.9
79.6
97.9
86.0
87.9

Total

95% CI
LL
UL

88.8
89.3
99.8
97.0
92.2

91.1

93.3

96.9
85.9
99.7
95.5
93.8

91.9

95.6

(1) Prevalence sub-areas of state (2) Weighted average using observed cases at each site as weight (3) Entire state
(4) Too few cases to estimate

The C-R analyses suggest that over all four sites, both prevalent and incident cases are at least
91% ascertained. Ascertainment appeared somewhat lower in the older than younger age groups,
reflecting clinic experience at the difficulty of identifying and recruiting older youth. Analyses
will be updated periodically and reported in relevant manuscripts.

Limitations
These analyses have a number of limitations, and while the C-R method is often touted as
the best way to estimate completeness of ascertainment 2,10, several authors have identified
significant problems with the method 1,11-23. In the context of the current US healthcare system
and HIPAA regulations, several of the limitations of the method were encountered. These
include: a) possible incomplete matching across sources due to restrictions on access to names
for matching in some states (thus violating the assumption that cases can be matched in all
sources); b) uncertainty about the residence location of some cases (thus violating the
assumption that cases were from the study area); and c) design of the ascertainment system for
efficiency (thus avoiding sources of likely duplicate cases). Each of these problems is known to
inflate the estimated number of total cases in the C-R analysis, leading to an underestimate of the
percent completeness. In addition, given the multiple sources of information used to identify
cases, it was possible to arbitrarily combine these sources into two modes in many alternate ways
instead of the one chosen: ‘hospital’ vs. ‘provider’. If this was done on the identical dataset, it
was possible to drive the estimates of completeness from 72.7% to 86.5% completeness (in
South Carolina as an example). An example of another problem came from Colorado.
Preliminary analyses conducted in December 2003 suggested that prevalent cases were 87%

complete, and that there were approximately 1246 estimated cases if all cases had been
identified. By December of 2004, Colorado had identified a total of 1366 prevalent cases;
however, the C-R estimate dropped to 81.5% complete. Addition of duplicates changed this to
88.8% complete. We, therefore, believe that the C-R estimates shown in Table 2 are a ‘lower
bound’ on the completeness of ascertainment in these four sites. While some redesign of the
case ascertainment system might provide better estimates of completeness, inherent limitations
of access to records in all sources with incomplete personal identifiers make the use of C-R in
the US difficult. Nonetheless, given the large geographic areas covered, and the multiple
providers and hospitals contacted and used during case ascertainment, SEARCH achieved at
least 90% ascertainment of prevalent and incident cases across all four geographic sites.
Systematic evaluation of models allowing interaction terms between 3 ascertainment modes did
not improve estimates of ascertainment completeness, and were inconsistent with the observed
rate consistency. Given the limitations noted above, the 2-mode model will continue to be used.
In January of 2013, capture-recapture estimates were updated to reflect on-going case
ascertainment and inclusion of later registered cases. In order to estimate completeness by
race/ethnicity and type of diabetes, the approach taken for fitting the log-linear model was
revised. The current approach relies on adjusted models instead of the stratified models that were
used previously. As the number of stratification variables increased, the cell counts observed in
some cases were too small, which prevented the maximum likelihood estimation routines from
converging. The adjusted models do not suffer from this limitation since they use all the
available data24-25.

Conclusions and use of results in SEARCH
Capture-recapture methods in the four geographic sites resulted in an overall estimate of
completeness of at least 90% for both prevalence and incidence. No estimates are possible in the
California and Native American sites. Given the closed nature of these data systems and the
comparable methods used to identify cases in these health systems, it seems likely (though
untested) that ascertainment rates were at least as good, if not better, than in the geographic sites.
It is likely, given the limitations of the use of C-R methods as implemented in SEARCH, the
estimates of completeness of ascertainment are a lower bound on the actual completeness.

References
1. Hook EB, Regal RR. Capture-recapture methods in epidemiology: methods and
limitations. Epidemiol Rev. 1995;17:243-264.
2. LaPorte RE, McCarty DJ, Tull ES. Counting birds, bees, and NCDs. Lancet.
1992;339:494-495.
3. Cormack RM, Chang YF, Smith GS. Estimating deaths from industrial injury by capturerecapture: a cautionary tale. Int J Epidemiol. 2000;29:1053-1059.

4. Corrao G, Bagnardi V, Vittadini G, Favilli S. Capture-recapture methods to size alcohol
related problems in a population. J Epidemiol Community Health. 2000;54:603-610.
5. Ballivet S, Salmi LR, Dubourdieu D. Capture-recapture method to determine the best
design of a surveillance system. Application to a thyroid cancer registry. Eur J
Epidemiol. 2000;16:147-153.
6. Tilling K, Sterne JA, Wolfe CD. Estimation of the incidence of stroke using a capturerecapture model including covariates. Int J Epidemiol. 2001;30:1351-1359.
7. Verstraeten T, Baughman AL, Cadwell B, Zanardi L, Haber P, Chen RT. Enhancing
vaccine safety surveillance: a capture-recapture analysis of intussusception after rotavirus
vaccination. Am J Epidemiol. 2001;154:1006-1012.
8. Bishop YMM, Fienberg SE, Holland PW. Chapter 6. Discrete multivariate analysis.
Cambridge. MA: MIT Press; 1975.
9. Espeland MA. A general class of models for discrete multivariate data. Communications
in Statistics: Simulation and Computation. 1986;15:405-424.
10. LaPorte RE, McCarty D, Bruno G, Tajima N, Baba S. Counting diabetes in the next
millennium. Application of capture-recapture technology. Diab Care. 1993;16:528-534.
11. Hook EB, Regal RR. The value of capture-recapture methods even for apparent
exhaustive surveys. The need for adjustment for source of ascertainment intersection in
attempted complete prevalence studies. Am J Epidemiol. 1992;135:1060-1067.
12. Papoz L, Balkau B, Lellouch J. Case counting in epidemiology: limitations of methods
based on multiple data sources. Int J Epidemiol. 1996;25:474-478.
13. Nanan DJ, White F. Capture-recapture: reconnaissance of a demographic technique in
epidemiology. Chronic Dis Can. 1997;18:144-148.
14. Domingo-Salvany A, Hartnoll RL, Maguire A et al. Analytical considerations in the use
of capture-recapture to estimate prevalence: case studies of the estimation of opiate use in
the metropolitan area of Barcelona, Spain. Am J Epidemiol. 1998;148:732-740.
15. Hook EB, Regal RR. Recommendations for presentation and evaluation of capturerecapture estimates in epidemiology. J Clin Epidemiol. 1999;52:917-926.
16. Hook EB, Regal RR. Accuracy of alternative approaches to capture-recapture estimates
of disease frequency: internal validity analysis of data from five sources. Am J Epidemiol.
2000;152:771-779.
17. Ismail AA, Beeching NJ, Gill GV, Bellis MA. How many data sources are needed to
determine diabetes prevalence by capture-recapture? Int J Epidemiol. 2000;29:536-541.
18. Hook EB, Regal RR. On the need for a 16th and 17th recommendations for capturerecapture analysis. J Clin Epidemiol. 2000;53:1275-1277.
19. Chao A, Tsay PK, Lin SH, Shau WY, Chao DY. The applications of capture-recapture
models to epidemiological data. Stat Med. 2001;20:3123-3157.
20. Tilling K. Capture-recapture methods--useful or misleading? Int J Epidemiol.
2001;30:12-14.

21. Laska EM, Meisner M, Wanderling J, Siegel C. Estimating population size and
duplication rates when records cannot be linked. Stat Med. 2003;22:3403-3417.
22. Carle, F., Gesuita, R., and Gregorio, F. Effect of overlapping among sources on the
validity of capture-recapture estimates of prevalence of diabetes. 39th Annual Meeting
EDEG . 2004.
23. Verlato G, Muggeo M. Capture-recapture method in the epidemiology of type 2 diabetes:
a contribution from the Verona Diabetes Study. Diab Care. 2000;23:759-764.
24. Chao A, Tsay PK, Lin SH et al.: The applications of capture-recapture models to
epidemiological data. Statist Med 2001; 20:3123-3157.
25. Cormack RM. Loglinear models for capture-recapture. Biometrics 1989; 45:395–413.

Supplementary Table 1: Sources included in each mode of ascertainment in South Carolina
South Carolina

Old classification

Anderson Area Medical Center

Hospital

Anmed Child Health Center

Hospital

Beaufort Hospital

Hospital

Carolinas Hospital - Florence

Hospital

Greenville Hospital System

Hospital

Greenville Memorial Hospital

Hospital

Lexington Medical Center

Hospital

Mauldin Medical Center

Hospital

McLeod Hospital

Hospital

McLeod Regional Medical Center

Hospital

Orangeburg Hospital

Hospital

Palmetto Bapstist Medical Center Easley

Hospital

Palmetto Baptist Medical Center Columbia

Hospital

Palmetto Health Alliance/RMH

Hospital

Palmetto Health Baptist

Hospital

Palmetto Health Easley

Hospital

Palmetto Health Richland

Hospital

Roper St Francis Hospital

Hospital

Spartanburg Regional Healthcare System

Hospital

Spartanburg Regional Medical Center

Hospital

The Regional Medical Center of Orangeburg and Calhoun Counties

Hospital

Amrhein

Other

Broome

Other

Carolina Diabetes and Kidney Center

Other

Coulter

Other

GHS Pediatric Endocrinology

Other

Heinze

Other

Hoffman

Other

Jackson

Other

Jocelyn Myers

Other

Laurel Endocrine-Brennan

Other

McLeod Pediatric Subspecialists

Other

McLeod/Woodberry

Other

Mendes

Other

MUSC

Other

Parker

Other

Raine

Other

New classification

Schwartz

Other

USC Pediatric Endocrinology

Other

Willi

Other

Benedict College

Other

Black River Community Health Care

Other

Brooks Health Center

Other

C.S.R.A. Renal Services

Other

Care-South Carolina

Other

Carolina Health Greenwood

Other

Carolina Peds

Other

Catawba Longhouse

Other

Children and Family HealthCare Center (USC College of Nursing)

Other

CSRA Renal Services

Other

Debbie Yoman

Other

Diabetes Education Center in Lancaster

Other

Doctors Care (statewide)

Other

Eau Claire

Other

Eau Claire Cooperative Health Center

Other

Eau Claire Cooperative Health Centers

Other

Family Health Care Center--Orangeburg

Other

Family Health Centers, Inc.

Other

Family Practice Center-Palmetto Health

Other

Franklin Coulter

Other

Grand Strand Ped

Other

Grand Strand Pediatrics

Other

Lexington Pediatrics

Other

Longcreek Family Practice

Other

Orangeburg Hospital Diabetes Educator

Other

Orangeburg Hospital-Diabetes Educator

Other

Pediatric Associates, P.A.

Other

Pediatric Associates, PA

Other

Richland Community Health Care Association

Other

SandHills Pediatrics-Wessinger

Other

Sea Island Pediatrics P.A.

Other

Self Report

Other

The Pediatric Clinic

Other

Undefined

Other

USC Central Billing

Other

USC Department of Family & Preventive Medicine

Other

USC OB/GYN clinic (1801 Sunset)

Other

USC-Central Billing-Dr. Bryant

Other

Yoman

Other

Supplementary Table 2: Sources included in each mode of ascertainment in Ohio
Ohio

Old classification

FHH

Hospital

StLuke

Hospital

UniversityHosp

Hospital

Christ

Hospital

Mercy

Hospital

MRH

Hospital

Jewish

Hospital

CCHMC

Hospital

StElizabeth

Hospital

GoodSam/ Bethesda

Hospital

McCullough

Hospital

EndoAdult

Other

EndoPeds

Other

PrimaryMDs

Other

CDEs

Other

Universities

Other

Other

Other

CintiHealthDept

Other

Anthem

Other

Aetna

Other

KYMedicaid

Other

BCMH

Other

CareSource

Other

New classification

Supplementary Table 3: Sources included in each mode of ascertainment in Colorado

Colorado
St Mary's in Grand Junction
Exempla Hospitals
The Children's Hospital/The Children's Hospital Colorado

Old classification
Hospital
Hospital
Hospital

Centura Hospitals

Hospital

Boulder Community Hospital

Hospital

Pueblo, CO Hospitals/Metro Community Hospital

Hospital

Barbara Davis Center
Pediatric Endocrine Associates
San Luis Valley/Valley Wide Health System
Western Ped. in Grand Junction
Salud Family Health Centers
Denver Health
Kaiser Permanente
Providers/San Luis Valley Case Reports

Other
Other
Other
Other
Other
Other
Other
Other

New classification

Supplementary Table 4: Sources included in each mode of ascertainment in Washington
Washington

Old classification

Boldt

Hospital

CHRMC

Hospital

CHRMC inpatient

Hospital

Harborview Medical Center

Hospital

Madigan Medical Center

Hospital

Mary Bridge

Hospital

Mary Bridge inpatient

Hospital

Providence St. Pete’s

Hospital

Seattle Children's inpatient

Hospital

Swedish Medical Center

Hospital

UW Medical Center

Hospital

Valley Medical Center

Hospital

Virginia Mason

Hospital
Hospital
Hospital
Hospital
Hospital
Hospital

CHCKC

Other

CHRMC Endo Clinic

Other

CHRMC outpatient

Other

Diabetes Care Center

Other

Dr McGowen

Other

Green

Other

Joslin

Other

Mauseth

Other

MB outpatient

Other

Minor & James clinic

Other

N Sea Pub Health

Other

Neighborcare

Other

Ped Asso Olympia

Other

PSNHC

Other

SeaMar

Other

Seattle Children's outpatient

Other

Summit View Clinic

Other

Swedish Joslin

Other

UW Physicians Network

Other

ADA

Other

New classification

Camp Leo

Other

GHC

Other

Newspaper Advertisements

Other

Other

Other

SKWIDDS

Other
Other
Other
Other
Other