The National Hospital Discharge Survey (NHDS) consists of a continuous sample of hospital discharge records abstracted annually from medical records of patients treated at nonfederal short-stay hospitals in all fifty states and the District of Columbia.16
The NHDS datasets are available to the public through the Centers for Disease Control and Prevention website at http://www.cdc.gov/nchs/about/major/hdasd/nhds.htm
. Inclusion eligibility is restricted to hospitals having six or more beds and where the average length of stay for all patients is less than thirty days. The sample is updated periodically to reflect changes in eligibility. Medical diagnoses and surgical procedures contained in the NHDS are coded according to the International Classification of Disease, 9th
revision, Clinical modification (ICD -9-CM) codes. The NHDS adopts a complex, stratified, multistage probability design to ensure a representative national sampling. Variables in the NHDS dataset include data on age, gender, race, marital status, primary discharge diagnosis and six secondary discharge diagnoses, hospital bed size, hospital geographic location, hospital ownership, type of hospital admission, primary and secondary source of payment, discharge month and length of stay.17
For the purpose of this analysis, we merged the 2001, 2002 and 2003 NHDS datasets, and restricted our analysis to patients with a primary discharge diagnosis of CAD. The 2001 – 2003 NHDS datasets included 976,995 sampled hospital discharges. Patients younger than 65 years, those with pre-admission residence in NH, and those who died during their hospital stay were excluded ().
Primary Diagnosis of Coronary Artery Disease
Patients with a primary discharge diagnosis of CAD were identified by the ICD-9-CM codes 410, 411, 412, 413 and 414. Of the 976,995 patients in the 2001 -2003 NHDS datasets, 30,010 were aged ≥65 years and discharged with a primary discharge diagnosis of CAD (). Of these, 28,172 were discharged alive from the hospital. It is important to note that the NHDS is based on coded hospital discharge records and collected information does not allow identification of individual patients. There is a possibility that patients with multiple hospitalizations were captured more than once but these duplications are likely to be random. As such, we treated each discharge as representing a unique patient in our analyses.
Secondary Diagnosis of Depression
Out of the 28,172 individuals with CAD, a total of 686 patients were identified as having a secondary diagnosis of depression at time of discharge. These individuals were identified using ICD-9-CM codes 296 (affective psychoses, includes 296.0 – 296.9), 311 (depressive disorder, not elsewhere classified), and 300.4 (neurotic depression).
The primary outcome of interest was NH admission as ascertained at the time of hospital discharge, and identified from the “discharge status” variable in the datasets.
Other Secondary Diagnoses
The NHDS collected data on up to six secondary discharge diagnoses for each patient. Using this record, a list of co-morbid conditions based on ICD-9 codes was assembled. The list included heart failure (428), dysrhythmias (427), hypertension (401-405), diabetes mellitus (250), hypothyroidism (244), chronic obstructive pulmonary disease (491-492, 496), pneumonia (480-487), syncope (780.2), acute renal failure (584), iron deficiency anemia (280), urinary incontinence (788), urinary tract infection (599), and dementia (094, 290, 291, 292, 294 and 331). These other secondary diagnoses were chosen due to their known associations with either the predictor variable, a secondary diagnosis of depression, or the outcome variable, admission to a NH.
After descriptive analysis of the baseline characteristics for the pre-match cohort of patients with and without depression (, left hand panel), propensity scores were calculated to control for the imbalance in baseline covariates between patients. Matching by propensity score often balances all measured covariates and is superior to matching by individual covariates such as age, sex, race, etc. The propensity score is the conditional probability of receiving a particular exposure or treatment given a vector of covariates,18-20
and has been used in the literature to control for selection bias between two treatment groups.21-23
More recently, the technique has been used to control for the imbalance in baseline covariates between two groups of patients with and without a certain co-morbid condition.24
One of the key limitations of propensity score technique is, however, that unlike randomization, it cannot balance unmeasured covariates. However, as patients cannot be randomized to develop depression, that is less of a concern for the current analysis.
Baseline patients characteristics by the depression pre and post-matched with propensity scores
We calculated propensity scores for depression, using a non-parsimonious multivariable logistic regression model, with depression as the dependent variable, and all measured baseline characteristics as covariates. The covariates used in the model are presented in . The resulting propensity score for depression was used to match patients who had a secondary diagnosis of depression with up to three patients without the secondary diagnosis who had similar propensity scores. An SPSS macro was used to randomly match patients.25
Overall, 686 patients with depression were matched with 2058 patients without depression.
Baseline characteristics between the patients with and without depression in post-match cohort were compared, and absolute standardized differences on key covariates were estimated.26, 27
Bivariate and multivariable logistic regression analyses were conducted to assess NH admission for depressed patients compared with those not depressed. Covariates in the multivariable model were the same as those used in the model for propensity score. The effect of other covariates on NH admission was also examined using the same model, with age and length of stay as categorical variables. Odds ratios and their 95% confidence intervals were then converted into relative risks.28
The effects of depression on subgroups of patients based on age, sex, race, marital status, heart failure, diabetes, dementia, and hypothyroidism were examined. All tests were based on a 2-sided p value and p values of <0.05 were considered significant. All analyses were done using SPSS 13.2 for Windows.29