METHODS
Study design and patients. Prospective study of a cohort of
9658 patients newly diagnosed with HF during a period of 15 years (1
January 2004 to 31 December 2018). These patients are residents of a
community of 321,753 inhabitants in the south of Spain, served by the
Hospital Universitario Puerto Real (HUPR). The GAMIC cohort comprises
adults (>= 14 years) diagnosed with HF according to the
Framingham criteria (11). All patients have at least one valid
echocardiography (EcoC). Patients aged less than 14 years, those not
permanently resident in the community of reference and those without a
valid EcoC (386 patients, 4.0%) have been excluded. These 386 patients
were mostly women, older and sicker, dying before a valid EcoC could be
performed. We have BNP levels in 32% of the cohort, mainly in patients
with obesity and/or chronic pulmonary disease, and only with diagnostic
purposes without follow-up determinations. This study has been
undertaken with the approval of the Committee for Ethics and Clinical
Research, of the HUPR.Collection of data. The data collected provide information on:
Sociodemographic and clinical parameters, tests requested, previous
treatment, definitive diagnoses, treatment established, scheduled or
emergency outpatient visits, and hospitalizations. These data were
recorded not only at the time of the inclusion of the patients in the
study but also during the 15 years of monitoring. We have recorded data
corresponding the 12 months prior to the inclusion of those patients for
whom these data were available (data bases of the HUPR and of the Family
Doctor). We classified kidney function using the Modification of Diet in
Renal Disease equation for estimated glomerular filtration rate (eGFR)
based on serum creatinine determinations (12). Comorbidity was assessed
by the “modified” Charlson Index (13). The echocardiogram was analyzed
following the guidelines of the American Society for Echocardiography
(14). The limit for considering the LVEF ”normal” has been very
variable, between 40 and 50% (14, 15). The criterion that we have
chosen to define normal systolic function (LVEF > 50%) is
that customarily utilized in previous studies (15).Socioeconomic status (SES) was self-declared and characterized
using net annual household income (NAHIL) measured at inclusion and
revised yearly until death or censorship (16). NAHIL was categorized as
low (<16,000 euros; n=3909), mid-level (16,000–30000 euros ;
n=3377), high (>30000 euros;n=2372),or not reported (n=35)
(16). Participants in the last category were not included in our study
because of the small sample size (17). As there were no differences of
prognosis between patients with HF in the low vs. the NAHIL, the
prognostic cut-off for the annual household income level was established
using bootstrapping as the point at which the 12-month probability curve
for death exceeds the 97,5th confidence intervals of
the same curve for the highest NAHIL. A mean NAHIL of 30,000 euros was
associated with a 92-month mortality above the 97.5% bootstrap levels
for the highest NAHIL. Thus, a prognostic cut-off for the NAHIL of
30,000 euros was chosen (18). Educational level was defined as the
highest grade or year of school completed, divided into 3 categories, as
previously described (1, 2).
Using not only income but, the educational level, the marital status,
the living status dependency, number of households, occupational status,
total wealth and properties value, allows us to more comprehensively
capture the cumulative results of SES on cardiovascular health over the
life course in participants (2, 16, 17).
Outcomes. Patients were prospectively included from January 1,
2004, and censored at end of follow-up on December 31, 2018. Primary
outcomes included death (all cause and cardiovascular), hospitalizations
for HF, and visits for any cause. To confirm mortality and morbidity,
the histories of the patients (hospital or health center) were monitored
weekly during the period of study. Death was identified from national
health service and family practitioners’ databases and, deaths that
occurred in the emergency department or hospital. When the cause of
death was not clear, the physician certifying death was contacted. The
patients admitted with heart failure were identified by weekly review of
the 9th revision of the International Classification of Diseases
(ICD-9-CM). The codes of the ICD-9-CM included are those previously
utilized in other studies: 428, 402.01, 402.11, 425, 429.3, 514, 402.9,
404.01, 404.11, 404.90, 398, 416, and 429 (15). When the health status
of any patient was not known, they were contacted by telephone.
Estimation of propensity scores and matching. Because there
were significant differences in baseline characteristics among HF
patients by income level (Tables 1-3 suppl ), we used
propensity scores-matching to achieve balance (19). We estimated
propensity scores (PS) for high income level using a non-parsimonious
multivariable logistic regression model (20-22). In that model, all
baseline patient characteristics displayed in Tables 1 and 2, and
clinically plausible interactions were included as covariates (20-22).
Our PS model discriminated well between patients in the high NAHIL and
those in the low/middle NAHIL. The model was fit to data during all
steps of the regression analyses (Hosmer and Lemeshow goodness-of-fit
test X 2 = 11.34; P = 0.10, andc -statistic = 0.83). We then used the PS to match each patient in
the high NAHIL to another patient in the low/middle NAHIL, who had a
similar PS. Thus, matching 2372 of those HF patients in the high NAHIL
to another 4744 HF patients in the low/middle NAHIL (Table
4 suppl ). Similarly, 1338 of those patients with HFrEF in the high
NAHIL were matched to another 2676 HFrEF patients in the low/middle
NAHIL (Table 1 ), and 1034 of those patients with HFpEF in the
high NAHIL were matched to another 2068 HFpEF patients in the low/middle
NAHIL (Table 2 ). We used a greedy matching algorithm, which
first look for matches to five decimal places. The efficacy of PS models
is best assessed by estimating post-match absolute standardized
differences between baseline covariates. We therefore calculated pre-
and post-match absolute standardized differences and presented those
findings as Love plots (23). Before matching, the mean PS for HF
patients in the low/middle income group was 0.17362 while, in the high
income HF group was 0.22186, in the high income HFrEF group was 0.22218
and, in the high income HFpEF group was 0.22237, which yielded a
standardized difference of 30.7%, 29.9% and 31.0%, respectively,
t-test P-value < 0.0001 in all cases. After matching, the mean
PS for HF patients in the low/middle income group was 0.21443 while, in
the high income HF group was 0.21449 (standardized difference of 3.0%
and t-test P = 0.995), in the high income HFrEF group was 0.21449
(standardized difference of 3.1% and t-test P = 0.996), and in the high
income HFpEF group was 0.21439 (standardized difference of 3.0% and
t-test P = 0.994).
Analysis of recurrent hospitalizations. Cumulative incidence of
hospitalizations and of 30-day readmissions over time were calculated
for the NAHIL groups using the Ghosh and Lin non-parametric analysis
(24). The rate ratio of HF hospitalizations per 100 patient-years of
follow-up (95%CI and p-value) was calculated based on the Poisson
distribution (25, 26). We assessed for over-dispersion by consulting the
deviance statistic of the Poisson model, and conducted supplementary
analyses using negative binomial regression when the deviance statistic
exceeded one (26). The negative binomial model was used to modify such
an estimated rate ratio by recognizing the heterogeneity (different
frailties) of patients with respect to their risks of recurrent
hospitalizations (27). Recurrent HF hospitalizations were also analyzed
using the Anderson-Gill approach with robust variance estimator that
allows for heterogeneity in hospitalization rates between patients (28).
Statistical analysis. All statistical analyses were done with
SPSS v 17.0 (SPSS Inc, Chicago, IL). No losses of patients initially
included in the study were recorded. We used chi-square tests and
independent sample t tests, as appropriate, for descriptive
analysis to compare baseline characteristics between pre-match patients
by income. For descriptive analysis of post-match cohorts, McNemar tests
and paired-sample t tests were used as appropriate. Kaplan-Meier
survival analyses and matched Cox proportional hazards models were used
to estimate the association between SES and prognosis. We confirm the
assumption of proportional hazards by a visual examination of the log
(minus log) curves. We conducted formal sensitivity analyses to describe
the weight of our evidence, by quantifying the degree of hidden bias
that, would need to be present to invalidate our main conclusions (23).
To evaluate the independent risk associated with each of the variables,
we utilized multivariate analysis. We identified the most likely
predictive variables within each category by backward steps selection,
with the variables with probabilistic value > 0.01 being
eliminated from the model. The variables with predictive significance
were combined and pre-established covariates were added, in the event of
not having been considered in the model. These covariates have been
included by virtue of the theoretical likelihood of their association
with the prognosis, or by previous studies (2, 7, 15), or by being
considered clinically important for the prediction of morbidity and
mortality. In analyzing hospitalizations for heart failure and visits, a
sandwich variant estimator was applied in the calculation of the 95%
confidence intervals to account for multiple hospitalizations or visits
by the same patient (24, 27).
Given that income varies overtime, and in order to avoid an optimistic
estimation of the results on high income level, performing an analysis
with intention to treat, we have considered that patients’ NAHIL at
inclusion has been the same up to their death, independently of whether
NAHIL changed during follow-up. A secondary analysis, incorporated
time-varying estimates of high NAHIL and assigned exposure status at the
time of an outcome event based on our high NAHIL-exposure algorithm. As
comorbidity has been associated with SES, we performed stratified models
on patients who, at the time of inclusion, did or did not present
comorbidities (Tables 1 and 2 ).
Heterogeneity of effects in pre-specified subgroups was examined by
testing for treatment-covariate interaction with the Cox proportional
hazards regression model, using p < .05 (29).