Letters
Epidemiology Volume 26, Number 1, January 2015
be used in place of the relative risks in
Equations 1 and 2 in the setting of case
control studies with rare outcomes.
RJE = (Observed RR11 / Expected RR11)
(1)
Expected RR11 = RR10 + RR 01 1 (2)
Equation 3 shows that the RJE is a simple
function of the AP. The AP is the proportion
of the joint risk in the doubly exposed group
because of additive interaction, whereas the
RJE captures how many more times greater
or lower the observed joint effect is com-
pared with the joint effect that is expected
in the absence of interaction. Given that the
RJE compares observed and expected val-
ues similar to common epidemiologic and
statistical measures (eg, standardized mor-
tality ratio, chi-squared test), the RJE may
be more intuitive than the other measures
of additive interaction.
RJE = 1 / (1 AP)
(RR11 RR10 RR 01 + 1) more exposure is preventive, for appropri-
= 1 / 1
RR11 ate estimates and interpretation, the refer-
(3)
For 2 risk factors, an estimate of the RJE
that is approximately 1.0 indicates absence
of additive interaction. An RJE > 1.0 sug-
gests positive additive interaction. An RJE
< 1.0 indicates negative additive interaction.
The estimated RJE is mathematically non-
negative. The observed and expected RR11
and in turn the RJEcan be estimated from
a log binomial or modified Poisson regres-
sion model.4 eAppendix 1 (http://links.lww.
com/EDE/A846) provides simple STATA6
code to estimate the RJE and the corre-
sponding bootstrapped confidence interval
using a modified Poisson model for hypo-
thetical cohort data included in eAppendix
2 (http://links.lww.com/EDE/A846).
The upper and lower panels of the
Table present the RJE and other measures
in the setting of positive and negative
additive interaction between 2 risk fac-
tors, respectively, for the data shown in
eAppendix 2 (http://links.lww.com/EDE/
A846). eAppendix 3 (http://links.lww.
com/EDE/A846) provides the formulas
e4 | www.epidem.com 2014 Lippincott Williams & Wilkins
TABLE. Measures of Additive Interaction
REFERENCES
Based on 10,000 Bootstrapped Samples 1. Rothman KJ, Greenland S, Walker AM.
Concepts of interaction. Am J Epidemiol.
Measures (Bootstrapped 1980;112:467470.
of Additive Bootstrapped 95% Confidence 2. Greenland S. Interactions in epidemiol-
ogy: relevance, identification, and estimation.
Interaction Estimatesa Interval)
Epidemiology. 2009;20:1417.
3. Rothman KJ. Modern epidemiology. 1st ed.
Positive additive interaction Boston, MA: Little, Brown; 1986.
RJE 1.96 (1.1 to 2.5) 4. Spiegelman and Hertzmark Easy SAS
RERI 1.8 (1.4 to 2.6) calculations for risk or prevalence ratios and dif-
ferences. Am J Epidemiol 2005; 162:199200.
AP 0.5 (0.3 to 0.6) 5. Knol MJ, VanderWeele TJ, Groenwold RH,
S Index 1.4 (1.01 to 1.9) Klungel OH, Rovers MM, Grobbee DE.
Negative additive interaction Estimating measures of interaction on an ad-
ditive scale for preventive exposures. Eur J
RJE 0.66 (0.46 to 0.87)
Epidemiol. 2011;26:433438.
RERI 1.2 (2.4 to 0.03) 6. StataCorp. Stata Statistical Software: Release
AP 0.51 (0.98 to 0.04) 12. College Station, TX: StataCorp LP; 2011.
S Index 0.39 (0.2 to 0.6)
a
Bias-corrected mean of the bootstrapped samples.
Prevalence of
for and interpretation of the RERI, AP, Alzheimer Disease in
and S. In both the setting of positive and
negative additive interaction, the estimate US States
for the RJE is consistent with the values
To the Editors:
obtained from the RERI, AP, and S. Simi-
lar to other measures of additive interac-
tion between risk factors, when one or A lzheimer disease dementia (AD)
poses increasing challenges to
US states, which dedicate substantial
ence category must be recoded to be the Submitted 18 July 2014; accepted 22 July 2014.
group with the lowest risk.3,5 Disclosure: This work was funded by the Alzheim-
ers Association and National Institutes of
In summary, the potential greater Health/National Institute on Aging grant
intuitive appeal of the RJE combined with AG011101.
the example and straightforward code we J.W. is a consultant for the Alzheimers Asso-
ciation and the AlzRisk Project (www.alzrisk.
provide in this paper should encourage more org). She is also funded by Alzheimers Asso-
frequent assessment of additive interaction ciation grant NIRG-12-242395 and NIH grant
in the applied epidemiologic literature. R21ES020404. L.E.H. has no disclosures of
financial relationships. She is (or has been)
funded (Principal Investigator, Co-investigator
ACKNOWLEDGMENTS or Biostatistician) by NIH grants NR010211,
AG303544, AG011101, AG036650 and
We thank Omar de la Cruz, Tyler AG009966. P.A.S. reports no disclosures.
VanderWeele, Peter Samai, and Gregory D.A.E. has no disclosures of financial rela-
Wellenius for their helpful comments on tionships. He is funded (Principal Investigator
or Co-Investigator) by NIH grants AG11101,
an earlier version of this article. AG036650, AG09966, AG030146, AG10161,
AG021972, ES10902, NR009543, HL084209,
Hirut T. Gebrekristos and AG12505l.
Department of Epidemiology Supplemental digital content is avail-

able through direct URL citations in the
Tulane School of Public Health and Tropical
Medicine
HTML and PDF versions of this article
(www.epidem.com). This content is not
New Orleans, LA peer-reviewed or copy-edited; it is the sole
hgebrekr@tulane.edu responsibility of the authors.
Chanelle J. Howe
Correspondence: Jennifer Weuve, Rush Institute for
Healthy Aging, Rush University Medical Center,
Department of Epidemiology 1645W. Jackson Boulevard, Suite 675, Chicago,
Center for Population Health and Clinical IL 60612. E-mail: jennifer_weuve@rush.edu.
Epidemiology Copyright 2014 by Lippincott Williams & Wilkins
Brown University School of Public Health ISSN: 1044-3983/15/2601-0139
Providence, RI DOI: 10.1097/EDE.0000000000000199
Epidemiology Volume 26, Number 1, January 2015 Letters
FIGURE. Percentage change from
2010 to 2025 in the number of older
adults (65 years and older) with AD
dementia, by state.
resources to AD care. Chief among
these services is long-term nursing
home care, a service that many, possibly
most, persons with AD require during
the course of their illness.13 Federal and
state governments jointly fund, but indi-
vidual states manage, Medicaidthe
only public health insurance program
that provides coverage for this care.
However, surveillance systems are not
available for gauging the scale of these
challenges, and it would be extremely
difficult to devise mandatory reporting
requirements for AD.
Using an alternative approach,
we estimated the number of adults, age
65 years, with AD in each US state
and the District of Columbia (DC)
from 2010 to 2025. We computed these
estimates by applying the annual AD
incidence and AD mortality hazard
identified in a large, systematically
evaluated community to each states
population, accounting for each states
age structure, mortality patterns, and
other demographic characteristics. The
Supplementary material (eAppendix,
http://links.lww.com/EDE/A845) pro-
vides extensive detail on these com-
putations, outlined below (as well as
presenting further background and
sensitivity analyses).
2014 Lippincott Williams & Wilkins www.epidem.com | e5
We obtained AD incidence
and mortality data from the Chicago
Health and Aging Project (CHAP),46
a longitudinal, population-based study
of older adults (60% of whom were
black, 40% white). Each 3-year, in-
home data-collection cycle included
identical clinical evaluations for
AD dementia of a stratified, random
sample. Between 1997 and 2010, 402
cases of incident AD were identified
in 2577 evaluations among 1913 per-
sons who had been classified as free
of AD at the previous cycle. Criteria
for AD were those of the Work Group
of the National Institute of Neurologi-
cal and Communicative Disorders and
Stroke and the Alzheimers Disease
and Related Disorders Association
for probable AD,7 except that persons
who met these criteria and had another
condition impairing cognition were
retained. There were 990 deaths. From
these data, we calculated separate
annual incidence estimates for 432
groups defined by single year of age
(65100 years); sex, 2 race groups;
and 3 education groups.
For each state and DC, we
applied the incidence estimates com-
puted from CHAP data to the corre-
sponding state subpopulation jointly
defined by year of age (beginning with
age 65), sex, and calendar year. We
then estimated the AD prevalence pro-
portion in each state and DC in each
subpopulation jointly defined by year
of age, sex, race, and calendar year,
incorporating information on the AD
and mortality experience of the cor-
responding birth cohort in previous
years. Finally, we multiplied the pro-
portion of prevalent AD by the census
estimate of number of people in each
age, sex, and race group and summed
across groups to obtain total numbers
of people with AD.
States with larger older adult
populations had larger estimated AD
prevalences. Between 2010 and 2025,
all states, but not DC, are expected to
experience double-digit to triple-digit
percentage increases in AD prevalence
(range, 19% [Pennsylvania] to 116%
[Alaska]), with the largest increases
occurring in the West and South-
east (Figure1, e-Table 1, http://links.
lww.com/EDE/A845). Some states
with the largest predicted percentage
increasesCalifornia, Florida, and
Texasalready have large numbers of
adults with the condition. The burden
of AD on the AD-free population also
will grow: in 2025 older adults with
Letters Epidemiology Volume 26, Number 1, January 2015
AD are expected to comprise a larger
fraction of state populations than they
did in 2010 (eTable 2, http://links.lww.
com/EDE/A845).
Even within agesexrace
education strata, the experience of the
CHAP population might not general-
ize to state-specific populations or to
different points in time. Uncertainty in
state population projections contributes
additional uncertainty to our estimates.
Nonetheless, these limitations are minor
compared with the magnitude of the
estimated trajectories that portend a
substantial increase in the burden of AD
on state populations.
Jennifer Weuve
Liesi E. Hebert
Paul A. Scherr
Denis A. Evans
Rush Institute for Healthy Aging
Rush University Medical Center
Chicago, IL
jennifer_weuve@rush.edu
REFERENCES
1. Banaszak-Holl J, Fendrick AM, Foster NL,
et al. Predicting nursing home admission:
estimates from a 7-year follow-up of a
nationally representative sample of older
Americans. Alzheimer Dis Assoc Disord.
2004;18:8389.
2. Smith GE, Kokmen E, OBrien PC. Risk
factors for nursing home placement in a
population-based dementia cohort. J Am
Geriatr Soc. 2000;48:519525.
3. Arrighi HM, Neumann PJ, Lieberburg IM,
Townsend RJ. Lethality of Alzheimer dis-
ease and its impact on nursing home place-
ment. Alzheimer Dis Assoc Disord. 2010;24:
9095.
4. Hebert LE, Weuve J, Scherr PA, Evans DA.
Alzheimer disease in the United States
(20102050) estimated using the 2010 census.
Neurology. 2013;80:17781783.
5. Evans DA, Bennett DA, Wilson RS, et al.
Incidence of Alzheimer disease in a biracial
urban community: relation to apolipopro-
tein E allele status. Arch Neurol. 2003;60:
185189.
6. Bienias JL, Beckett LA, Bennett DA, Wilson
RS, Evans DA. Design of the Chicago Health
and Aging Project (CHAP). J Alzheimers Dis.
2003;5:349355.
7. McKhann G, Drachman D, Folstein M,
Katzman R, Price D, Stadlan EM. Clinical
diagnosis of Alzheimers disease: report
of the NINCDS-ADRDA Work Group
under the auspices of Department of
Health and Human Services Task Force
on Alzheimers Disease. Neurology.
1984;34:939944.
e6 | www.epidem.com 2014 Lippincott Williams & Wilkins
Limitations of
Indicators of HIV Case
Finding
To the Editors:
T he Centers for Disease Control and
Prevention recommends routine
human immunodeficiency virus (HIV)
screening in health care settings.1 To evalu-
ate HIV case finding, which is the process
of identifying HIV-infected persons who
have not been diagnosed, programs use 2
outcome measuresmedian CD4 count at
diagnosis and proportion of late diagno-
ses.24 It is expected that an improvement
in case finding would result in an increas-
ing median CD4 count at diagnosis and a
decreasing proportion of late diagnoses.
However, despite continuing efforts, the
2 measures remain stable. A recent review
article reported a minimal rise of 1.5 cells/
mm3/year in the CD4 count at entry into
care and a negligible change in the propor-
tion of late diagnoses in 11 high-income
countries from 1992 to 2011.5
When 2 commonly used indica-
tors remain stable over a long period in
so many countries despite continuing
efforts to expand HIV testing, we should
reexamine how well the indicators mea-
sure case finding. Here, we use hypo-
thetical data to show the limitations of
these 2 indicators. (Detailed methods are
presented in the eAppendix, http://links.
lww.com/EDE/A844).
As shown in the Table, the HIV
case detection rate, stratified by dura-
tion of infection, is identical across the
3 communities. However, Community
A, with an emerging epidemic, has the
highest median CD4 count at diagnosis
(418 cells/mm3), lowest proportion of
late diagnoses (33.0%) and lowest crude
Submitted 06 August 2014; accepted 25 August 2014.
Supplemental digital content is avail-
 but must be estimated based on the distri-
able through direct URL citations in the
HTML and PDF versions of this article
(www.epidem.com). This content is not
peer-reviewed or copy-edited; it is the sole
responsibility of the author.
Copyright 2014 by Lippincott Williams & Wilkins
ISSN: 1044-3983/15/2601-0141
DOI: 10.1097/EDE.0000000000000202
case detection rate (15.8%); Commu-
nity C, with a declining epidemic, has
the lowest median CD4 count at diagno-
sis (331 cells/mm3), highest proportion
of late diagnoses (39.7%) and highest
crude case detection rate (16.6%). All 3
communities have an identical adjusted
case detection rate (16.2%) and an iden-
tical case detection rate among new
infections (25.0%).
Using hypothetical data, we dem-
onstrate the limitations of median CD4
count at diagnosis and proportion of
late diagnoses. Despite an identical
case detection rate stratified by dura-
tion of infection, an emerging epidemic
would see a higher median CD4 count
at diagnosis and a lower proportion of
late diagnoses, and a declining epidemic
would see a lower median CD4 count at
diagnosis and a higher proportion of late
diagnoses. Real-world data also show
the same phenomenon, eg, Eastern Euro-
pean countries with emerging epidemics
had a lower proportion of late diagnoses
than Western European countries.6
Median CD4 count at diagnosis and
proportion of late diagnoses are measur-
ing the distribution of duration of infection
among persons who are newly diagnosed,
not case finding. Since a change in case
finding usually happens across the board,
the change would have little effect on
the 2 measures. An improvement in case
finding would increase not only the case
detection rate among early infections but
also late infections, leaving the distribu-
tion of duration of infection among the
newly diagnosed unchanged.7
We propose 2 new measures for
HIV case finding: adjusted case detection
rate and case detection rate among new
infections. These 2 case detection rates
directly measure HIV case finding and are
not affected by HIV incidence or distribu-
tion of duration of infection. The 2 new
indicators cannot be directly measured,
bution of duration of infection among the
undiagnosed and newly diagnosed, which
is not available using existing methods.
With recent advances in estimating HIV
incidence, we may soon be able to pro-
duce these 2 estimates, or at least the case