Cardiovascular Disease Outcomes Associated with Three Major Inflammatory Dermatologic Diseases: A Propensity-Matched Case Control Study

All variables come from the MarketScan™ Commercial Claims and Encounters database for 2005–2007 [10]. No new studies with human or animal subjects were performed by any of the authors. The people included in the analysis were in the range of 30–64 years old and had at least 1 year of follow-up data and 6 months of baseline data. We chose the lower age limit because CVD is less common in individuals younger than this age. We chose the upper age limit to minimize loss to follow-up that could occur from 65 year olds switching from commercial insurance to Medicare.

The dermatologic diseases group was composed of three arms: rosacea (ICD-9 695.3), psoriasis (ICD-9 696.1), or AD (ICD-9 691.XX, except for 691.0X). None of the arms have people with any of the other two skin conditions. They were all defined with diagnosis codes, having at least 2 entries each. The dermatologic disease group was selected further using a washout period. We defined our washout period as at least a 180-day period after a person’s first diagnosis on record since the beginning of 2005 during which they received no treatment for their skin condition. Treatments identified during this washout period included biologics, systemic treatments (e.g., methotrexate and cyclosporine), vitamin D analogs, calcineurin inhibitors, topical corticosteroids, anthralin, coal tar, retinoids, and ammonium lactate for psoriasis; systemic treatments, antihistamines (oral and topical), calcineurin inhibitors, topical corticosteroids ± anti-itching or anti-fungal/bacterial agents, and anti-itch treatments (such as calamine lotion) for AD); and oral antibiotics, metronidazole, isotretinoin, retinoids, azelic acid, benzoyl peroxide, calcineurin inhibitors, sulfacetamide products, permethrin and oxymetazoline. This method allowed for creating a homogeneous set of recently diagnosed people with medication prescriptions for their illness.

The control group was comprised of people selected with a simple random sample from the database. None of the people in the control group had any of the three skin diseases. In order to match to the 1-year follow-up period for the dermatologic disease group, the control group was required to have complete data for the whole of 2006.

The cases of cardiovascular disease were defined as having at least 2 diagnosis codes (ICD 9-CM) or at least one diagnosis code plus a procedure code (CPT) for the disease.

We examined several covariates for explanation of variance in the outcome and for their intrinsic interest. Potential confounders included age and gender, number of inpatient days, number of days for which people had outpatient visits, and total health-care utilization costs. Since diabetes is a large risk factor for CVD, we included it as a predictor of CVD, defined by having at least 2 diagnosis codes between 2005 and 2007. Overall comorbid disease severity was examined with the Deyo–Charlson Comorbidity Index with the components related to the outcome removed. Those components were CVD, peripheral vascular disease (PVD), and old myocardial infarction (MI). We also examined the number of comorbid conditions, which could be as high as 14 (as opposed to the 17 conditions usually examined with Deyo–Charlson). In addition to the number of days spent in the hospital, we examined the total number of hospitalizations for the year but did not include this in the regression in order to reduce multi-collinearity. We also looked at the medication costs directly related to the skin conditions as defined in the washout period, for a total of 365 days post-treatment. In addition to summing the costs per person, we looked at the number of prescription drug refills related to these costs for that year. We further examined the total medical costs divided into yearly inpatient, outpatient, and medication costs.

The sample size was determined by the exclusion criteria and the eventual propensity score selection. The control group originally was comprised of 20,000 people and many were not included because they did not match well to the dermatologic disease group. Given the large sample size, we did not perform a power analysis prior to the study to estimate the effect size we could detect.

The study was designed to examine the independent risk of CVD among the dermatologic conditions compared with a general population of people in a commercial insurance database. Cardiovascular disease outcomes included ischemic heart disease (ICD-9: 410–414), transient cerebral ischemia (ICD-9: 435), heart failure (ICD-9: 428), occlusion and stenosis of pre-cerebral arteries (ICD-9: 433), and occlusion of cerebral arteries (ICD-9: 434). Cases were required to have two diagnosis codes or a diagnosis code plus a cardiovascular-related procedure code (Supplementary Table 1). We used a 365-day follow-up period for the exposure and control in order to standardize the results. Thus, variables such as cost and healthcare utilization will also be generalizable to other studies.

The propensity scores were generated for the three dermatologic disease arms versus the control group, and certain people were selected from the randomly created control group using Mahalanobis matching [11]. All variables, except for the exposure variable, were included in the propensity score calculation. In order to look for a balanced propensity score set, we compared the covariates individually with effect sizes and with a global test designed by Hansen and Bowers [11, 12]. Continuous data were reported as mean (standard deviation) and categorical data were reported as n (%). Analyses were performed in SAS 9.4 (SAS Institute, Cary, NC, USA) and R [12].

This article does not contain any new studies with human or animal subjects performed by any of the authors.

The sample sizes for the disease and control groups were arrived at separately. For the dermatologic disease group, we began with people having either outpatient or inpatient visits with ages inclusive between 30 and 64 years (n = 21,801,147). After excluding patients that lacked diagnosis criteria, sufficient wash-out period, at least 1 year of follow-up, and at least one outpatient record, there were 4263 remaining patients (Fig. 1). For the control group, we began with a simple random sample of 20,000 people from the 2006 database. After excluding those of age less than 30 years or greater than 64 years, with less than 12 months of continuous enrollment, or with a diagnosis of the dermatologic diseases investigated, there were 14,162 patients. Following the propensity score matching with the combined dermatologic disease group, the control sample size was 4263 patients. Our sample consisted of n = 8526 people, matched 1:1 between the dermatologic diseases and control groups.

Our sample was comprised of people between the ages of 30–64 years old with an average age of 49.1 years [standard deviation (SD) 9.0]. Women comprised 67% of our total sample, and 13% of the population had either type I or type II diabetes. The sample had an average of 15.7 days of outpatient visits (SD 15.1), 0.3 days of inpatient visits (SD 1.5), and 0.08 hospitalizations (SD 0.34) per year. The average number of comorbidities was 0.28 (SD 0.57) and the average score of the modified Charlson Comorbidity index was 0.34 (SD 0.8). For health-related expenses, there were an average US$8508 of total costs (SD 16,069), $1049 of inpatient costs (SD 6756), $3925 of outpatient costs (SD 8735), and $1987 of medication costs (SD 3927). For patients diagnosed with AD, psoriasis, or rosacea, an average of 3 dermatologic medications (SD 3) were filled, resulting in an average annual cost of $840 (SD 3261). All covariates, except the number of inpatient and number of outpatient days, had statistically significant differences between the groups (Table 1).

The propensity score matching reduced the Hansen and Bowers global test for balance between treatment and control populations from a Chi-square value of 1155 (p < 0.0001) to 12.4 (p = 0.0528).

We considered first an effects modification model, and found that the overall test for the variable was insignificant (F = 3.5118, p = 0.3912), and we reported only the main effects model. The model was highly predictive of the outcome with an AUC value of 0.7935. None of the three dermatologic disease groups had statistically significant increased odds of experiencing a CVD event at one year compared to the control group, after controlling for all other covariates (Table 2). In univariate analysis, the unadjusted odds of cardiovascular disease was higher in patients with psoriasis.

The large sample size gave our study power to find even small clinical effect size. There were several prominent potential confounders for which the database does not collect data, including obesity and smoking and drinking history. Consequently, we expect a degree of residual confounding that may bias our estimate upward from the null. The washout period can potentially select a population with lower severity because it looks for new treatment patients, but this may be balanced by the fact that they are at least being treated by prescription medications. The study had a short follow-up period of 1 year, but there was a cross-section of many patient ages. With younger patients, 1 year may be insufficent to detect an increased risk of CVD. We used a large commercial claims database that is applicable to a large proportion of the population. Because this claims database did not include Medicaid patients, the magnitude of effect may not be generalizable for indigent populations. Insurance claims data do not provide information on systemic inflammatory burden beyond the diagnosis. Consequently, we have no way of measuring the systemic inflammation burden of skin disease.