1 INTRODUCTION

The past two decades have seen a dramatic increase in the number of people with diabetes mellitus (DM) worldwide. In 2010, there were 150-220 million people with DM worldwide, and this number is expected to reach 300 million by 2025.1, 2 As of 2018, estimates showed that 34.2 million US residents had DM,3 and the prevalence of DM in the United States is expected to increase 120% by 2050.4 The increasing prevalence of DM over the next decades is alarming and calls for public health attention.

DM is a well-established and widely recognized risk factor for cardiovascular disease (CVD). However, studies suggest that the relationship may be reciprocal, that is, that CVD also increases the risk of developing DM.5, 6 The mechanisms underlying the latter association are not well understood. Several posited mechanisms for the association of CVD with DM include gluconeogenesis and glycogenolysis resulting from neurohormonal activation usually present in CVD, catecholamine-induced increase insulin resistance, shared inflammatory pathways in CVD and DM, and physical inactivity.6 None of these hypothesized explanations have been formally tested and several other mechanisms may be involved as well. For instance, medications used for secondary prevention of CVD might sometimes have unintended health consequences.7-9 Prior mechanistic, clinical, and epidemiologic research suggests that commonly used pharmacological agents such as beta-blockers, statins, and diuretics may cause metabolic derangements leading to increased risk for incident DM but findings have been inconclusive.7-9 In addition, studies have reported higher risk for incident DM after coronary artery bypass graft surgery or transplant.10, 11 Furthermore, the physical impairment associated with CVD may limit the capacity of people with CVD to engage in physical activity12 and promote weight gain,13 which are both associated with elevated risk of DM. The extent to which the association of CVD with incident DM can be attributed to weight gain or cardiovascular medications is not known.

In the current study, we analyzed data from the Hispanic Community Health Study/Study of Latinos (HCHS/SOL), using propensity score (PS) methods to test the hypothesis that CVD was associated with increased risk of incident DM. We further investigated whether this association was explained by weight gain or the use of beta-blockers, statins, and diuretics. This analytical framework may provide supportive evidence for the association of CVD with subsequent DM and elucidate potential underlying mechanisms.

2 METHODS 2.1 Study population

The HCHS/SOL is a population-based longitudinal study in the United States that enrolled participants between 2008 and 2011.14 Participants were self-identified Hispanic/Latino individuals aged 18-74 years at baseline, randomly selected from households in the Bronx, New York; San Diego, California; Chicago, Illinois; and Miami, Florida. A stratified, two-stage sampling method was used to select households. The design oversampled individuals aged 45 to 74 years. A total of 16 415 participants were enrolled in the study. Each participating institutional review board approved the study, and written informed consent was obtained from all participants.

We restricted our study population to individuals without DM at baseline (Visit 1) and who participated in Visit 2 data collection. Of the 16 415 participants at Visit 1, we excluded 4792 (29.2%) individuals who did not participate in the Visit 2 data collection. Of those who participated in Visit 2 data collection (N = 11 623), we excluded 2401 with diabetes and 8 with missing data on diabetes at Visit 1. We further excluded 149 individuals with chronic kidney disease (defined as estimated glomerular filtration rate <60 mL/min/1.73 m2), 162 individuals seropositive for hepatitis C virus or hepatitis B virus, 113 individuals who self-reported chronic obstructive pulmonary disease (diagnosed by a doctor), 147 individuals with history of gestational diabetes, and 635 individuals because of missing information on covariables at Visit 1 and diabetes at Visit 2, thus yielding 8008 eligible individuals (Figure 1). We restricted our main analysis to a subset of 3798 individuals: 1899 individuals who self-reported prior CVD at baseline, and 1899 individuals who did not report CVD but had similar probability or propensity to self-report CVD at baseline, as described later. This approach was taken to address differential loss to follow-up because 3163 (31.6%) participants free of self-reported CVD were lost to follow-up compared to 798 (25.4%) participants with self-reported CVD (P < 0.0001). By matching, we created a control group of individuals without self-reported CVD that are comparable to individuals with self-reported CVD with respect to observed covariates.

Flow chart of study population selection. CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; CVD, cardiovascular disease; HCHS/SOL, Hispanic Community Health Study/Study of Latinos

2.2 Exposure and outcome assessment

Self-reported CVD at baseline (Visit 1) was defined as the presence of self-reported coronary heart disease, cerebrovascular events, peripheral artery disease, or heart failure. Incident DM at 6-year follow-up (visit 2) was defined by any of the following American Diabetes Association criteria: fasting time >8 hours and fasting blood glucose of 126 mg/dL or greater; fasting time less than 8 hours and fasting glucose of 200 mg/dL or greater; post-oral glucose tolerance test glucose of 200 mg/dL or greater; hemoglobin A1C of 6.5% or greater; or use of antihyperglycemic medications.15

2.3 Covariables

At HCHS/SOL Visit 1, sociodemographic data including age, sex, and Hispanic/Latino background were self-reported. Additional self-reported information assessed at Visit 1 included smoking status, the average number of cigarettes smoked per day, alcohol use, family history of diabetes, and 24-hour dietary recalls. Blood pressure, height, weight, and waist circumference were ascertained on physical examination. Physical activity was assessed using Actical accelerometers and categorized as high, moderate, and low. Venous blood specimens were also collected at Visit 1 and analyzed to measure serum creatinine, blood glucose, hemoglobin A1C, total and high-density lipoprotein (HDL) cholesterol, and triglycerides. Alternate Healthy Eating Index 2010 (AHEI-2010), a measure of dietary quality, was calculated from the 24-hour dietary recalls.16 Height and weight were used to calculate body mass index as weight (in kilograms) divided by height (in meters) squared. Hypertension was defined as systolic blood pressure 140 mm Hg or greater, diastolic blood pressure 90 mm Hg or greater, or use of antihypertensive medications.17 Prediabetes was defined as fasting time >8 and fasting blood glucose in range 100-125 mg/dL, or post-oral glucose tolerance test glucose in range 140-199 mg/dL, or 5.7% ≤ A1C < 6.5%. The diagnosis of metabolic syndrome was made when three or more of the following factors were present: hypertension, triglyceride ≥150 mg/dL, low HDL (HDL cholesterol <40 mg/dL in men and <50 mg/dL in women), fasting blood glucose ≥100 mg/dL, and waist circumference ≥102 cm in men and ≥88 cm in women.18

2.4 Mediating variables

We examined whether the association between self-reported CVD and incident DM was explained by medication classes such as beta-blockers, statins, and diuretics as well as by weight gain. HCHS/SOL obtained information on medications taken in the past 4 weeks before the baseline examination. Participants were asked to bring in all prescribed or over-the-counter medications. The medication information was used to uniquely identify drug products based on their generic ingredients and then assigned to their medication classes. Data were also gathered on weight at both Visit 1 and Visit 2, and weight gain was defined as the difference in weight between the two visits.

2.5 Selection of the analytical sample

We matched HCHS/SOL participants based on their probability or propensity to self-report CVD at baseline. The PS is the conditional probability of being exposed (eg, CVD) given a vector of measured covariates and can be used to adjust for selection bias when assessing causal effects in observational studies.19 We estimated the PS for self-reported CVD for each participant using a multivariable logistic regression model, in which self-reported CVD was modeled using all baseline covariates in Table 1, as well as interaction effects. We then used the estimated PS to match participants without self-reported CVD to participants with self-reported CVD who had very similar PS using greedy algorithms (nearest best match).20 Under the greedy algorithms, matches with the highest digit on PS are the best ones and selected first, and then the next-best matches are selected in a sequential process until no further matches can be made.20 In our matching algorithm, we first attempted to match each individual without self-reported CVD with a self-reported CVD individual who had a similar PS to eight decimal places. Then we removed those matched pairs of individuals and repeated the process matching to seven, six, five, four, three, two, and one decimal places.

TABLE 1. Covariate balance check at baseline - before and after propensity score matching among HCHS/SOL participants without diabetes at baseline (Visit 1) Baseline characteristics Before matching (n = 8008) After matching (n = 3798) Self-reported CVD Self-reported CVD Yes (n = 1956) No (n = 6052) Standardized difference (%) Yes (n = 1899) No (n = 1899) Standardized difference (%) Age, years, mean (SD) 52.3 (10.4) 42.7 (13.1) 81.7 52.0 (10.3) 51.9 (10.7) 0.6 Sex (male), n (%) 607 (31.0) 233 (38.6) −15.9 593 (31.2) 697 (36.7) −11.6 Hispanic/Latino Background, n (%) Dominican 0 243 (12.4) 478 (7.9) 15.0 231 (12.2) 232 (12.2) −0.2 Central American 1 192 (9.8) 689 (11.4) −5.1 190 (10.0) 208 (10.9) −3.1 Cuban 2 279 (14.3) 897 (14.8) −1.6 274 (14.4) 259 (13.6) 2.3 Mexican 3 661 (33.8) 2670 (44.1) −21.3 654 (34.5) 670 (35.3) −1.8 Puerto Rican 4 396 (20.2) 650 (10.7) 26.5 367 (19.3) 326 (17.2) 5.6 South American 5 142 (7.3) 485 (8.0) −2.8 141 (7.4) 157 (8.3) −3.1 More than one/Other heritage 6 43 (2.2) 183 (3.0) −5.2 42 (2.2) 47 (2.5) −1.7 Cigarette use, n (%) Never 1195 (61.1) 4059 (67.07) −12.5 1140 (60.0) 1163 (61.24) 2.5 Former 399 (20.4) 988 (16.33) 10.5 425 (22.4) 384 (20.22) −5.3 Current 362 (18.5) 1005 (16.61) 5.0 334 (17.6) 352 (18.54) 2.5 Cigarette pack years, mean (SD) 6.2 (13.7) 3.8 (10.6) 19.5 6.1 (13.5) 6.5 (15.1) −3.2 Alcohol use, n (%) Never 375 (19.17) 1223 (20.21) −2.6 370 (19.5) 328 (17.3) 5.7 Former 689 (35.22) 1786 (29.51) 12.2 660 (34.7) 666 (35.1) −0.7 Current 892 (45.60) 3043 (50.28) −9.4 869 (45.8) 905 (47.6) −3.8 Physical activity level, n (%) Low 655 (10.8) 178 (9.1) −5.7 178 (9.4) 185 (9.7) −1.3 Moderate 2807 (46.4) 834 (42.6) −7.5 811 (42.7) 861 (45.4) −5.3 High 2590 (42.8) 944 (48.3) 11.0 910 (47.9) 853 (44.9) 6.0 AHEI-2010, mean (SD) 49.7 (7.5) 48.9 (7.5) 10.8 49.7 (7.6) 50.1 (7.5) −4.4 Family history of diabetes, n (%) 910 (46.5) 2382 (39.4) 14.5 881 (46.4) 851 (44.8) 3.2 Metabolic syndrome, n (%) 696 (35.6) 1547 (25.6) 23.6 692 (36.4) 681 (35.9) 1.2 Hypertension, n (%) 652 (33.3) 1000 (16.5) 39.6 613 (32.3) 614 (32.3) −0.1 High total cholesterol, n (%) 876 (44.8) 2318 (38.3) 13.2 847 (44.6) 828 (43.6) 2.0 Prediabetes, n (%) 1188 (60.7) 2773 (45.8) 30.2 1146 (60.4) 1162 (61.2) −1.7 BMI categories, n (%) Underweight (BMI < 18.5) 14 (0.7) 48 (0.8) −0.9 13 (0.7) 17 (0.9) −2.4 Normal (18.5 ≤ BMI < 25) 333 (17.0) 1318 (21.8) −12.0 323 (17.0) 400 (21.1) −10.3 Overweight (25 ≤ BMI < 30) 737 (37.7) 2479 (40.9) −6.7 726 (38.2) 674 (35.5) 5.7 Obese I (30 ≤ BMI < 35) 523 (26.8) 1474 (24.4) 5.5 511 (26.9) 525 (27.6) −1.6 Obese II (35 ≤ BMI < 40) 251 (12.8) 486 (8.0) 15.8 235 (12.4) 207 (10.9) 4.6 Obese III (BMI ≥ 40) 98 (5.0) 247 (4.1) 4.5 91 (4.8) 76 (4.0) 3.8 Hemoglobin A1C, mean (SD) 5.6 (0.4) 5.5 (0.3) 32.7 5.6 (0.4) 5.6 (0.4) 4.0 Note: Standardized difference is the mean difference divided by the pooled SD, expressed as percentage. A standardized difference > 10% is suggestive of a meaningful covariate imbalance. Abbreviations: AHEI 2010, Alternate Healthy Eating Index; BMI, body mass index; CVD, cardiovascular disease; HCHS/SOL, Hispanic Community Health Study/Study of Latinos. 2.6 Assessment of baseline covariate balance

We compared the balance of all baseline covariates in Table 1 between individuals with and without self-reported CVD before and after PS matching using the standardized differences. A standardized difference greater than 10% is suggestive of a meaningful covariate imbalance.21 We also used the standardized difference to compare the mean PS score before and after matching, which directly quantifies the bias in the mean PS across the two groups (individuals with and without self-reported CVD), expressed as a percentage of the pooled SD. Before matching, the mean PS for people without self-reported CVD (n = 6052) was 0.193 (SD = 0.142) and in those with self-reported CVD (n = 1956) was 0.328 (SD = 0.163), with an associated standardized difference of 87.8% (t test P value, <0.0001). After matching, the mean PS for individuals without self-reported CVD (n = 1899) was 0.318 (SD = 0.154) and for those with self-reported CVD (n = 1899) was 0.319 (SD = 0.155), which yields a standardized difference of 0.7% (t test P-value 0.81).

2.7 Statistical analysis

A generalized estimating equation (GEE) was used to calculate the odds ratio (OR) and corresponding 95% confidence intervals (CIs) for the association between self-reported CVD and incident DM. This method allowed accounting for the lack of independence induced by PS matching. To assess the robustness of our findings, we analyzed data from the unmatched sample (N = 8008) by performing weighted logistic regression adjustment to estimate the association of self-reported CVD with incident DM; these analyses adjusted for covariates used in the logistic regression model for PS. Because individuals were matched regardless of their sampling units, analyses for the matched data were not weighted. We also evaluated the association between self-reported CVD and incident DM among males and females. The interaction between sex and self-reported CVD was assessed. Finally, we examined the associations between CVD subtypes (heart failure and myocardial infarction) and incident DM.

We further examined whether the association between self-reported CVD and incident DM was explained by either weight gain or cardiovascular medication classes such as beta-blockers, statins, and diuretics. For each mediating factor, we decomposed the total effect into two different pathways: (a) the effect of self-reported CVD on DM through the mediating pathway (ie, the natural indirect effect) and (b) the effect of self-reported CVD on DM that is not through the mediating pathway (ie, the natural direct effect). For each mediator, we determined the proportion mediated by calculating the ratio of the log indirect effect and the log total effect. The mediation analysis was conducted using semiparametric methods.22, 23 For each medication class, we estimated the total effect by regressing DM on self-reported CVD using GEE model accounting for matching and follow-up time between Visit 1 and Visit 2. To estimate the natural direct effect of self-reported CVD, we regressed DM on self-reported CVD using GEE model in a subsample of individuals who were not on that medication (individuals who were not on beta-blockers, statins, and diuretics were 3486, 3437, and 3377, respectively). Restricting the analysis to medication free individuals ensured that we isolated the effect of self-reported CVD that is not mediated by the use of that medication (ie, direct effect). Finally, the natural indirect effect was determined by subtracting the direct effect from the total effect.22, 23 All analyses were conducted using SAS version 9.4 (SAS Institute, Inc., Cary, NC).

3 RESULTS

The mean (+SD) age of the 3798 PS-matched individuals was 52 (SD = 10.5) years (median: 53; range: 18-74), and 34.0% were males. Table 1 compares baseline characteristics by CVD status before and after PS matching. Before matching, individuals with self-reported CVD were more likely to be female and older as compared to individuals without self-reported CVD. Individuals with self-reported CVD were more likely to have DM risk factors including prediabetes, metabolic syndrome, obesity, family history of DM, high cholesterol, and cigarette smoking at baseline. They more often had hypertension at baseline. Of the 1956 individuals with self-reported CVD and 6052 individuals without self-reported CVD at baseline, 296 (15.1%) and 581 (9.6%) developed incident DM respectively at 6-year follow-up (P < 0.0001).

After matching, individuals with and without self-reported CVD were very similar with regards to baseline covariates (Table 1). The standardized difference for the mean PS was 0.7% in absolute value, thus demonstrating an excellent balance in measured covariates across the two groups. The distributions of the PS score for individuals with and without self-reported CVD matched were nearly identical (Figure 2). In the matched cohort, two-thirds of beta-blocker users and about three in five statin and diuretics users self-reported CVD (Table 2). Significantly higher proportions of people who were on beta-blockers, statins, or diuretics at baseline developed DM at 6-year follow-up (beta-blockers and statins, P < 0.0001; diuretics, P = 0.0043) (Table 2).

Propensity scores for participants with and without cardiovascular disease. CVD, cardiovascular disease

TABLE 2. Distribution of self-reported cardiovascular disease and incident diabetes by medication use at baseline (Visit 1) in the matched cohort (n = 3798) Self-reported CVD at Visit 1 (baseline) Incident diabetes at Visit 2 (Follow-up) Medications Yes No P value Yes No P -value Beta-blocker use, n (%) <0.0001 <0.0001 Yes 216 (69.2) 96 (30.8) 69 (22.1) 243 (77.9) No 1645 (48.3) 1762 (51.7) 451 (13.2) 2956 (86.8) Statin use, n (%) <0.0001 <0.0001 Yes 224 (62.1) 137 (37.9) 75 (20.8) 286 (79.2) No 1637 (48.7) 1721 (51.2) 445 (13.3) 2913 (86.7) Diuretics use, n (%) 0.0118 0.0043 Yes 235 (55.8) 186 (44.2) 78 (18.5) 343 (81.5) No 1626 (49.3) 1672 (50.7) 442 (13.4) 2856 (86.6) Note: Seventy-nine matched individuals had missing information on medication use. Row percentages are reported. Abbreviation: CVD, cardiovascular disease.

At 6-year follow-up, 290 (15.3%) of the 1899 individuals with self-reported CVD and 242 (12.7%) of the 1899 individuals without self-reported CVD in the propensity-matched cohort developed incident DM (Figure 1). Compared to individuals without self-reported CVD, individuals with self-reported CVD had a 24% increased risk for incident DM (OR = 1.24, 95% CI = 1.01, 1.51) after adjusting for baseline hemoglobin A1C (Table 3). When we adjusted for baseline covariables in the full (unmatched) cohort (n = 8008), we observed similar but nonsignificant effect of CVD on incident DM (OR = 1.22, 95% CI = 0.93, 1.58). We found no evidence of interaction between self-reported CVD and sex (P = 0.4273). The data showed that individuals with myocardial infarction and heart failure had increased odds for incident DM, but the associations were not significant (myocardial infarction: OR = 1.45, 95% CI = 0.87, 2.41; heart failure: OR = 1.33, 95% CI = 0.71, 2.48).

TABLE 3. Odds ratios and 95% confidence intervals of the association between cardiovascular disease and incident diabetes Propensity-matched cohort (n = 3798) Overall cohort (n = 8008) Model 1a Model 2a Model 1b Model 2b Model 3b CVD All 1.23 (1.03, 1.48) 1.24 (1.01, 1.51) 2.15 (1.71, 2.69) 1.30 (1.01, 1.69) 1.22 (0.93, 1.58) Male 1.32 (0.97, 1.79) 1.36 (0.98, 1.89) 2.45 (1.70, 3.55) 1.54 (1.03, 2.29) 1.49 (0.98, 2.28) Female 1.20 (0.96, 1.51) 1.16 (0.91, 1.49) 1.94 (1.48, 2.53) 1.10 (0.79, 1.54) 1.05 (0.76, 1.44) Heart failure 1.21 (0.68, 2.17) 1.33 (0.71, 2.48) 2.05 (0.99, 4.23) 1.57 (0.66, 3.75) 1.60 (0.70, 3.65) Myocardial infarction 1.43 (0.89, 2.29) 1.45 (0.87, 2.41) 2.04 (1.13, 3.67) 1.22 (0.58, 2.53) 1.15 (0.55, 2.39) Mediators Proportion explained (%) Beta-blockers 25.4 Statins 18.0 Diuretics 8 Note: Heart failure: n = 93 (propensity-matched cohort); n = 309 (overall cohort). Myocardial infarction: n = 134 (propensity-matched cohort); n = 384 (overall cohort). Model 1a and Model 1b: Adjusted for follow-up time between Visit 1 and Visit 2. Model 2a: Adjusted for follow-up time between Visit 1 and Visit 2, sex, and baseline hemoglobin A1c. Model 2b: Adjusted for follow-up time between Visit 1 and Visit 2, age, sex, and baseline hemoglobin A1C. Model 3b: Adjusted for follow-up time between Visit 1 and Visit 2, age, sex, Hispanic/Latino background, cigarette smoking, alcohol use, physical activity level, healthy eating index (AHEI-2010), body mass index, family history of diabetes, hypertension, high cholesterol, metabolic syndrome, prediabetes, and baseline hemoglobin A1C. Interaction sex and CVD: P = 0.4273 (matched cohort). Abbreviations: AHEI 2010, Alternate Healthy Eating Index; CVD, cardiovascular disease.

Our results showed evidence of mediation of the effect of self-reported CVD through a pathway involving cardiovascular medication use (Table 3). The largest indirect effect was for beta-blockers (proportion explained = 25.4%), followed by statins (proporti