specifies that suicides (S) for an age group (i) in a state (j) for a given year and quarter (t) is a function of antidepressant prescriptions (Aijt
), other determinants of suicides (Xjt
), quarter-year effects (γt
), state effects (μj
), state*time interactions (μ*τ) and an error term. The principal hypothesis tested is whether or not antidepressants are associated with suicides and if so, in which direction.
Given that suicides are counts, we use the Fixed Effects Poisson (FEP) estimator to estimate the models (Wooldridge, 2002
; Cameron & Trivedi, 1998). This estimator is a quasi-maximum likelihood estimator that includes parameters or “fixed effects” to account for unobserved heterogeneity across the units of observation (states, in our case). Estimates are consistent regardless of whether the counts actually have a Poisson distribution (Wooldridge 2002
). To permit overdispersion, a common feature of count data that is not accommodated by the Poisson MLE, standard errors are adjusted for heteroskedasticity of unknown form (Cameron & Trivedi 2005
). Each model includes the log of the relevant population as a right hand side variable to normalize for exposure. The coefficient on log population is constrained to equal one.
Data on completed suicides come from the Multiple Cause of Death public use files, provided by the Centers for Disease Control. From this data we calculate quarterly counts of suicides for 10–14 year olds and 15–19 year olds for each state in the years 1997–2003. We caution that the counts of suicides are probably underreported (Mohler and Earls 2001). However, so long as the underreporting is uncorrelated with our measures of antidepressant use, the underreporting will not bias the coefficients of interest. The only effect might be to raise the standard errors of the coefficients.
Data on antidepressants prescriptions come from IMS’s National Disease and Therapeutic Index (NDTI). The NDTI is a nationally representative sample of office based physicians in private practice drawn from a universe of all physicians in the United States. The sample is a randomly drawn, two-stage stratified cluster, where the stages are doctors and workdays. The sample of doctors is selected by primary specialty and the 9 census divisions. All primary specialties involved in direct patent care are included. Each physician reports information on all patients seen during two consecutive workdays in each calendar quarter. From the patient data, we create state-level quarterly counts of “drug appearances” (defined below) and prescriptions for antidepressants. We separate these into “new therapies” only and new and continued therapies combined (termed “all therapies”). New therapies mean that the drug therapy was started in the current visit. Continued therapies mean that the drugs were previously ordered and are continued. We examine new therapies separately from continued because research has shown that the risk of suicidal behaviors is the highest during the first few days of therapy (Jick et al., 2004
A drug appearance is a mention of a drug during a patient visit. In the NDTI data, drug appearances include prescriptions, samples, drugs sold or given to the patient from their stock, hospital orders, drug recommendations that were not accompanied by a prescription, and drugs that were not issued during the current visit (i.e. no prescription, no sample and no medication sold, but drugs were issued on a previous visit). Non-issues only appear for continued therapies. We exclude drug recommendations from the counts so that our count total represents patients who have obtained or can obtain the drug with a prescription. We also use a separate variable for each antidepressant that counts only the actual prescriptions.1
Actual prescriptions represent a range of 66–91 percent of the drug appearances, depending on the year and drug under consideration.
We generate drug appearance counts for new therapies and all therapies for patients aged 10–14 and 15–19. The counts are then divided by the number of total patient visits for the relevant age group in each state, year and quarter and multiplied by 1000. Therefore, our measures of antidepressant use are the number of drug appearances per 1000 patient visits (in the relevant age group) in a state for a specific year and quarter for 1997–2003. We also create similar rate for new therapies that take the form of prescriptions only. Means for these and all other variables are shown in .
Means and Standard Deviations
The NDTI identifies four groups of antidepressants: 1) Tricyclics and Tetracyclics (TCAs), 2) MAO Inhibitors (MAOIs), 3) SSRIs/SNRIs, 4) newer generation antidepressants (NGAs). Examples of TCAs include Elavil, Amitriptyline and Imipramine. Common SSRI/SNRIs include Zoloft, Lexapro, and Prozac. Wellbutrin is the most popular newer generation antidepressant. MAOIs are rarely prescribed to children. In the seven years of NDTI sample data, there is only one prescription of a MAOI for a child, therefore we exclude MAOIs as a separate category of antidepressants from this analysis. Annual trends in the three categories of antidepressants and suicides are show in and for ages 15–19 and and for ages 10–14.
Antidepressant Appearances, New Therapies, Ages 15–19
Antidepressant Appearances, All Therapies, Ages 15–19
Antidepressant Appearances, New Therapies, Ages 10–14
Antidepressant Appearances, All Therapies, Ages 10–14
The NDTI data has two limitations worth noting. First, the prescription data are from office based physicians so any prescriptions from in-patient facilities are not included. Second, we have no way of knowing whether or not a prescription was filled or if a prescription was written after a sample was given.
In generating the drug appearance rate, we do not distinguish among visits to different types of physicians. However, the doctors in the NDTI data can be identified by specialty and we use this information to create a variable representing the percent of antidepressant drug appearances that come from a psychiatrist. This variable is included in all models and will help to capture some of the characteristics of our sample of patient cases that may be correlated with suicides and with the drug appearance counts in the data, such as illness severity. For example, Burns, Costello, Angold, Tweed, Stangl, Farmer et al. (1995)
show that children with the most severe symptoms of mental illness are much more likely to visit a mental health specialist than any other source. These authors conclude, “For children with a diagnosis and/or impairment, the general health care system was rarely the sole source of mental health care.” p. 152. A study of adults found that individuals with more severe symptoms of major depressive disorder (MDD) were more likely to see psychiatrists than those with less severe symptoms (Kessler, Berglund, Demler, Jin, Koretz, Merikangas et al. 2003
). Further, adequacy of treatment for MDD, defined as either pharmacotherapy or psychotherapy visits above specific thresholds, was greater in the specialty mental health sector than other sectors.
We include in the models two variables representing the supply of psychiatric care in each state. The variables are 1) the number of psychiatrists per 100,000 population; 2) the number of psychiatric hospital beds per 100,000 population. These variables are important to include because the availability of psychiatric care can affect suicides by influencing the full price of receiving mental health treatment through the availability of services. The full price includes not only the monetary cost of care, but travel and time costs as well. In the case of health care, the latter are likely to be a large part of the full price. We caution that these variable could be endogenous if the observed level of psychiatric care reflects the overall mental health status of the state population. For example, holding all else constant, consumer demand theory predicts that there will be more psychiatrists demanded in states with higher rates of mental illness. However, we do not believe that the inclusion of these variables are problematic since models that omit these variables yield extremely similar results to those presented below.
A number of studies have shown that suicide rates are positively associated with higher rates of unemployment, lower income and lower education (Hamermesh and Soss 1974
, Huang 1996
; Marcotte 2003
; Ostamo, Lahelma & Lönnqvist 2001
), although these results are not upheld in some studies (Andres 2005
; see Rehkopf and Buka 2006
for a comprehensive review). Higher levels of female labor force participation have also been observed to be correlated with suicide rates (Huang 1996
; Burr, McCall & Powell-Griner 1997
). To represent these factors, we include in all models the state quarterly unemployment rate, quarterly real income per capita, the annual percentage of the population ages 25 years and over that has obtained a bachelor’s degree, and the female labor force participation rate. Unemployment and female labor force participation rates come from Department of Labor, Bureau of Labor Statistics. Per capita income comes from the Department of Commerce, Bureau of Economic Analysis, and educational attainment comes from the Department of Commerce, U.S. Census Bureau.
Higher suicide rates are often observed in rural areas (Middleton, Gunnell, Frankel, Whitley & Dorling 2003
; Singh and Siahpush 2002
; Wilkinson and Israel 1984
). A person’s religiosity is another factor that may predict suicidality (Huang 1996
; Greening and Stoppelbein 2002
). To represent these factors, we include the percentage of the state’s population living in rural areas and the annual percentage of each state’s population identifying with certain religions (Mormon, Southern Baptist, Protestant and Catholic). The rural population comes from the U.S. Census Bureau and the religion data come from Jones, Horsch, Houseal, Lynn, Marcum, Sanchagrin et al. (2002).
Substance use is also believed to be strong predictor of suicidal behaviors (Kelly, Lynch, Donovan & Clark 2002; Cutler, Glaeser, & Norberg. 2001
; and Borges, Walters & Kessler 2000
) In light of this relationship, state annual per capita beer consumption is included in all models following results from Markowitz, Chatterji & Kaestner (2004) showing a positive relationship between beer consumption and teenage suicides. Data on beer consumption come from the Brewers Almanac
and are expressed in gallons per capita.