Long-Run Trends in Antidepressant Use Among Youths After the FDA Black Box Warning

Abstract

Objective

In October 2004, the Food and Drug Administration (FDA) directed pharmaceutical companies to issue box warnings about the potential link between the use of antidepressants and suicidal ideation among children. The objective of this paper is to analyze the long-run trends in antidepressant use among children before and after the box warning for those with and without severe psychological impairment.

Methods

The analysis uses data from the Medical Expenditure Panel Survey (MEPS) for children aged 5–17, covering years 2000–2011. The paper compares the changes in the rate of any antidepressant use in the early post-warning years (2004–07) and late post-warning years (2008–2011) to the pre-warning years (2002–03) using multivariate probit models. Recycled predictions methods are used to estimate yearly predicted probabilities of use.

Results

There was a .5% statistically significant decline in the probability of using any antidepressants during the early post-warning years (2004–07) compared to pre-warning years adjusting for all covariates. In the long-run (2008–11), however, there was no statistically significant difference. Five years after the warning in 2009, the adjusted rates of use increased back up to their pre-warning levels (2.29% in 2003, 2.26% in 2009). The initial impact of the warning differed between the severe and non-severe populations, having a stronger and significant effect on the non-severe population.

Conclusions

The return to the pre-box warning rates raises concern that the impact of the warning may have dissipated over time. An important policy implication of these findings could be that more frequent updates of FDA risk warnings might be necessary.

Introduction

The Food and Drug Administration (FDA) is responsible for protecting public health interests by assuring the safety and security of medications.¹ To achieve that goal, the FDA undertakes comparative effectiveness research to determine risks and issue risk warnings when necessary. In October 2004, after conducting a series of meta-analyses and finding an increased risk of suicidality events among paroxetine-treated children, the FDA directed pharmaceutical companies to issue box warnings on antidepressants about the potential link between the use of antidepressants and suicidal ideation among children.^2–4 The goal of this paper is to analyze the long-run trends in antidepressant use among children before and after the box warning, and to see if an initial decline in use is followed by a return to rates of use before the warning.

The prescription of antidepressants for children was on the rise along with the diagnoses of depression until 2005, when both prescriptions and diagnoses rates dropped sharply due to the implementation of the box warning.^3,5 Initially, the box warning resulted in an overall decrease of antidepressant use for both adults and children.^6,7 The trends in antidepressant use in the years immediately following the warning have been studied extensively; however, most studies since then have focused on re-testing the effectiveness and risks of antidepressants for pediatric patients rather than analyzing the long-term effects of the box warning on antidepressant use.^8–12 One such study suggested that “meta-analyses fail to undermine evidence that antidepressants are associated with increased risk of suicidality in children”, and that the box warning on pediatric use of antidepressants are warranted.¹⁰ On the other hand, Lu et al. identified that while antidepressant use among youth declined, suicide attempts may have actually increased after the box warning,¹² suggesting that there may have been unintended negative consequences attributable to overall decline in treatment for depression,^3,5 and the decline in antidepressant prescription without the FDA-recommended monitoring and use of alternative treatments.¹³ Understanding national, long-run trends of antidepressant use acts as an indicator of the strength of the box warning in the midst of conflicting evidence, as well as how providers have reacted to the evolving uncertainty and complexity of youth depression treatment.

Other medication and product warnings have had varied impact on patterns of use.^14–17 For instance, FDA warnings regarding antipsychotic use for Parkinson’s disease and ADHD medication for pediatric patients did not result in change or decline of use¹⁸ possibly due to the lack of a media campaign.¹⁹ Other product warnings (e.g. artificial sweeteners and cigarettes), however, have resulted in sharp drops of product use shortly after a warning, only to be later followed by a renewed increase in product use.^20,21 Consumer studies show that warning compliance depends on a variety of consumer factors that are both design and non-design related,²² such as the fact that text labels are more difficult to remember,²³ and that for certain products it is more difficult to implement effective advisories.²⁴

Individuals adhere to box warnings depending on the perceived risks and benefits of the product, which is true for both mental health patients²⁵ as well as providers.²⁶ These perceived risks are unrelated to the actual risk or even the proposed risk by a warning label.^21,27,28 The level of perceived risk is associated with factors such as the prescribers’ years of experience and gender.²⁶ Legal differences,²⁹ perceived liability,³⁰ and provider knowledge³¹ may also affect prescription behavior, regardless of the information provided by the box warning. Among consumers, a box warning for antidepressants does increase the perceived risk.³² On the other hand, consumers with prior continuous use might have a reduced perceived risk since continuous use of antidepressants reduces the risk of relapse and recurrence of depression.³³ In analyzing perceived risks over time, it is also important to consider the role of the pharmaceutical companies. In response to product sales decreases, pharmaceutical companies have been motivated to downplay warnings after their initial implementation.³⁴

Possibly due to changes in these perceived risks and behaviors, the 2004 box warning appears to have a differential effect on antidepressant use among children in the short-run and the long-run. Analyzing the trends of antidepressant use among children is the first step towards understanding the long-run response of consumers, providers, and manufacturers to risk warnings. In this study, we seek to examine the long-run effect of the box warning on antidepressant use and see whether the impact differs between the severe and non-severe populations. We divide the post-warning years into two time periods, early post (2004–07) and late post (2008–2011) to differentiate the short-run and long-run impact of the warning on rates of use. The cutoff year is chosen as 2007 due to FDA’s revision to the box warning that year, and to better accommodate the initial uncertainty that might have been associated with the 2004–07 time period.

Data

For our analysis, we use data for children aged 5–17 from the Medical Expenditure Panel Survey (MEPS), covering years 2000–2011. This age range corresponds to the population targeted by the FDA box warning. MEPS is a nationally representative survey of the non-institutionalized US population. Approximately 15,000 individuals are newly surveyed each year about their health care utilization, demographic characteristics, socio-economic status, and physical and mental health status. Our final sample includes 75,819 children aged 5–17.

Psychological impairment in our analysis is defined using the Columbia Impairment Scale (CIS) score, which is reported as part of the MEPS data. CIS is a 13-item global measure of impairment administered to parents by an interviewer. It addresses four areas of functioning: interpersonal relations, broad psychopathological domains, functioning in job or school, and use of leisure time, and has been used in prior trends studies among youth³⁵. Items are scored on a range of 0 to 4, with total scores ranging from 0 to 52. This is a parent-reported instrument that correlates well with the clinicians’ Children’s Global Assessment Scale score.³⁶ We use a cutoff score of CIS≥16 to differentiate children with severe psychological impairment from those without impairment.³⁷

The dependent variable of interest is any antidepressant use at a given year. Using the prescription names listed in the MEPS prescription files, we identified the children with any antidepressant use at a given year. The list of antidepressant prescription names included in our definition can be found in Appendix A1.^a,b

The additional covariates used in our models are age, race, gender, income level, insurance status, health status, and metropolitan statistical area (MSA). Age is calculated from date of birth and indicates age status as of December 31st of each year; it is categorized into three groups (5–9, 10–13, 14–17). Insurance status indicates the respondents’ insurance coverage by private insurance, Medicaid/SCHIP or Medicare, other public insurance, or uninsured. Income is denoted as a categorical variable: below poverty (<100% of the federal poverty level or FPL), near poverty (100–125% FPL), low income (125–200% FPL), middle income (200–400% FPL), and high income (>400% FPL). The race variable is coded using three Census-based categories: non-Hispanic White, Black, and Hispanic. Other race and ethnicity groups are excluded from the sample due to small sample size. Health status measures perceived health status, as the respondents were asked to rate their health according to the following categories: excellent, very good, good, fair, or poor.

Methods

To estimate the how the impact of warning on antidepressant use among kids changes over time, we first divide the entire sample period (2000–2011) into four time periods: early pre-warning (2000–2001), pre-warning (2002–03), early post (2004–07) and late post-warning (2008–11). Then, we estimate the short-run and long-run impact of the warning on antidepressant use using a multivariate probit model:

$$\text{Prob }(y=1|X,\mathit{\text{Time}})=G({\beta }_1\mathit{\text{Time}}+{\beta }_{2}X)$$ (1)

where y is any antidepressant use at a given year; X is vector of explanatory variables including gender, age, race, income, health status, insurance status, and MSA; Time is the categorical variable that denotes the time periods of early pre, pre, early post, and late post-warning years, and G(․) is the CDF of standard normal distribution. The main coefficient of interest is β₁, which denotes the short-run (early post) and long-run (late-post) impact of the warning compared to the pre-warning years.

Sample years before the warning are divided into two categories early-pre (2000–01) and pre (2002–03), and the pre-warning period (2002–03) is used as the reference category. This is mainly due to the noticeable difference in antidepressant use rates between these two periods, with years 2002–03 right before the warning being more representative of the antidepressant use trends right before the warning.

We used the coefficient estimates from the model in Equation 1 to generate recycled predictions, also called the predictive margins method, to estimate the regression adjusted yearly predicted probability of use.³⁹ The standard errors of the predicted probabilities were estimated using bootstrap methods (N=500).⁴⁰ In all of our analysis, we adjusted for the sampling design used in the MEPS, including the estimation weight, sampling strata, and primary sampling unit.⁴¹ We estimated the probit model in Equation 1 for two subsamples (children with and without psychological impairment) to explore whether the long run effects were different between the severe and non-severe populations. Goodness of fit in all models were verified using the Hosmer-Lemeshow goodness-of-fit test (Appendix A2).³⁸

Results

Table 1 presents the descriptive characteristics of the final sample that includes 75,819 children aged 5–17. The children in the sample are balanced in terms of gender (48.9% female), predominantly white (42.0%), and are mostly insured (90.4%). Only 27.7% of the children live in households below the federal poverty level, and 1.5% reported the use of any antidepressants. Among our final sample, 19% of the children (n=14,523) are considered to have severe psychological impairment (CIS≥16), and the remaining 81% are considered non-severe (CIS<16). The prevalence of antidepressant use differs noticeably between the severe and non-severe populations. The prevalence of antidepressant use among the severe and non-severe children are 5.0% and .7%, respectively.

Table 1.

Descriptive Statistics for Children Aged 5–17 (2000–2011)

	Severe (CIS*>=16)		Non-severe (CIS<16)		p- value	Total
	N=14,523		N=61,296			N=75,819
	n	%	n	%		n	%
Female	6,653	45.8%	30,402	49.6%	<0.001	37,055	48.9%
Race
white	6,529	45.0%	25,349	41.4%	<0.001	31,878	42.0%
black	3,527	24.3%	12,689	20.7%		16,216	21.4%
hispanic	4,467	30.8%	23,258	37.9%		27,725	36.6%
Age
5–9	4,762	32.8%	23,311	38.0%	<0.001	28,073	37.0%
10–13	4,545	31.3%	19,749	32.2%		24,294	32.0%
14–17	5,216	35.9%	18,236	29.8%		23,452	30.9%
Insurance
private	5,228	36.0%	28,369	46.3%	<0.001	33,597	44.3%
medicare/medicaid	7,990	55.0%	26,649	43.5%		34,639	45.7%
other	75	0.5%	233	0.4%		308	0.4%
uninsured	1,230	8.5%	6,045	9.9%		7,275	9.6%
MSA	11,850	81.6%	51,189	83.5%	<0.001	63,039	83.1%
Income
below poverty (<100% FPL**)	4,889	33.7%	16,121	26.3%	<0.001	21,010	27.7%
near poverty (100–125% FPL)	1,182	8.1%	4,951	8.1%		6,133	8.1%
low income (125–200% FPL)	2,769	19.1%	12,005	19.6%		14,774	19.5%
middle income (200–400% FPL)	3,651	25.1%	17,303	28.2%		20,954	27.6%
high income (>400% FPL)	2,032	14.0%	10,916	17.8%		12,948	17.1%
Perceived health status
excellent	5,612	38.6%	30,744	50.2%	<0.001	36,356	48.0%
very good	4,354	30.0%	17,404	28.4%		21,758	28.7%
good	3,371	23.2%	10,891	17.8%		14,262	18.8%
fair	777	5.4%	1,501	2.4%		2,278	3.0%
poor	163	1.1%	159	0.3%		322	0.4%
Any antidepressant use	733	5.0%	421	0.7%	<0.001	1,154	1.5%

^*CIS stands for Columbia Impairment Scale.

^**FPL stands for Federal Poverty Level.

Note: p-values are computed using the Pearson's chi-squared statistic.

The probit regression results are presented in Table 2, which includes estimates for the total sample in addition to the subsamples of severe and non-severe children. Among all children, there is a significant decline of .5% in the probability of using any antidepressants during the early post-warning years (2004–07) compared to pre-warning years after adjusting for all covariates (coef=−.13, p=.01). In the long-run, however, there was no statistically significant difference. In late post-warning years (2008–01), the probability of use was not significantly different than the pre-warning years at any conventional level (p=.33).

Table 2.

Impact of post-warning years on antidepressant use among children (aged 5–17)

	(1) All		(2) Severe (CIS>=16)		(3) Non-severe (CIS<16)
	Coef	SE	Coef	SE	Coef	SE
Time (pre: 2002–03)
early pre: 2000–01	−0.16**	0.06	−0.24**	0.08	−0.09	0.08
early post: 2004–07	−0.13**	0.05	−0.10	0.07	−0.16**	0.06
late post: 2008–11	−0.05	0.05	−0.08	0.08	−0.03	0.07
Female	−0.08*	0.04	−0.06	0.05	−0.08	0.05
Race (white)
black	−0.58***	0.06	−0.54***	0.08	−0.64***	0.07
hispanic	−0.44***	0.06	−0.44***	0.07	−0.44***	0.08
Age (5–9)
10–13	0.49***	0.05	0.51***	0.08	0.47***	0.08
14–17	0.76***	0.05	0.73***	0.08	0.79***	0.08
Insurance (private)
medicare/medicaid	0.18***	0.06	0.19*	0.08	0.18*	0.08
other	0.18	0.21	0.08	0.35	0.35	0.30
uninsured	−0.44***	0.09	−0.62***	0.14	−0.29**	0.11
Income (poverty)
near poverty	−0.01	0.07	0.07	0.10	−0.10	0.11
low income	−0.02	0.06	0.02	0.08	−0.07	0.07
middle income	0.10	0.06	0.24**	0.09	−0.04	0.08
high income	0.12	0.07	0.23*	0.10	−0.01	0.09
Health status (excellent)
very good	0.18***	0.04	0.23***	0.07	0.15**	0.06
good	0.43***	0.05	0.52***	0.07	0.33***	0.06
fair	0.65***	0.08	0.68***	0.10	0.67***	0.11
poor	0.75***	0.16	0.87***	0.20	0.53*	0.22
MSA	0.00	0.05	−0.02	0.07	0.03	0.06
Constant	−2.91***	0.10	−2.25***	0.14	−2.81***	0.12
Severe (CIS>=16)	0.76***	0.04	-	-	-	-
Observations	75,819		14,523		61,296

^***p<0.001,

^**p<0.01,

^*p<0.05

Note: Marginal effects for time variable in Model 1 are as follows:

2. early pre: 2000–01 −0.006

3. early post: 2004–07 −0.005

4. late post: 2008–11 −0.002

The results also show that short-term impact of the warning differs between the severe and non-severe populations. Among children with severe psychological impairment, there was no significant decline in the probability of using antidepressants during the early post-warning years (2004–2007) compared to the pre-warning years (2002–2003) (p=.17). However, non-severe children were .3% less likely to use antidepressants in the early post-warning years and the effect was significant (coef=−.16, p=.01). Neither the severe nor the non-severe populations showed significant decline in antidepressant use in the late post-warning years (2008–2011) compared to the pre-warning years (2002–2003). In other words, antidepressant use in the late post-warning years did not significantly differ from the pre-warning years in either population.

Figures 1–3 present the regression-adjusted yearly trends in antidepressant use in 2000 to 2011 among the three samples of interest: all children aged 5–17, severe and non-severe children, respectively. Five years after the warning in 2009, the adjusted rates of use increase back up to the pre warning level in 2003 (2.29% in 2003, 2.26% in 2009). The initial decrease in antidepressant use is not prominent among the severe children, as represented by the early-post area in Figure 2. However, we observe an almost perfect U-shaped trend among non-severe children in Figure 3 where antidepressant use rate as of 2008 (1.09%) is almost identical to the pre-warning usage rate of 1.10% in 2003.

Figure 1.

Regression Adjusted Trends in Any Antidepressant Use - All

Figure 3.

Regression Adjusted Trends in Any Antidepressant Use - Non-severe (CIS<16)

Figure 2.

Regression Adjusted Trends in Any Antidepressant Use - Severe (CIS>=16)

As a sensitivity check, we also examined the same trends among the adult population and found no spillover effects of the warning on the adult population. There was no decline in antidepressant rates in the early-post or late-post periods in the adult sample. On the contrary, adults’ rates of any antidepressant use were significantly higher (not lower) than the pre-period in both early-post (2004–07) and late post (2008–11) periods at p<0.001.

Discussion and Conclusions

We found a .5% statistically significant decline in antidepressant use in the years immediately following the box warning, suggesting that the diffusion of risk information was effective at reducing rates of use in the short run. However, the initial impact of the warning seemed to fade away in the long-run, as the prevalence of antidepressant use returned to the pre-warning year levels by 2009. We also found that the initial impact of the warning differed between the severe and non-severe populations. The warning appeared to have a stronger effect on children without severe psychological impairment. As shown by the regression-adjusted annual trends, the U-shaped trend of immediate decrease in use followed by a return to the pre-warning levels is most prominent among children without severe psychological impairment. Among the psychologically impaired population, there was a slight decline in antidepressant use until 2008, but the change in prevalence did not appear as dramatic and was not significant.

These findings suggest that providers and families of youth may have reacted to the box warning in an appropriate manner, weighing the warning in with the risks and benefits of the treatment, and deciding not to use the medication when psychological impairment was low, but choosing to use the medication when impairment was high and the benefits outweighed the risks. A return to the pre-box warning rates raises concern that this thoughtful accounting of the risks and benefits may have dissipated over time. The period of increased caution might have been followed by a return to previous levels after analyzing the newly available information.

An important policy implication of our findings could be that more frequent updates of FDA risk warnings might be necessary to prevent the patients and the providers from “forgetting” the potential risks outlined in the original warning. Methods have been suggested to increase the long-run effectiveness of warnings for both medical and nonmedical products. For example, a study of the warning labels for cigarettes found that designs should be rotated regularly and that they should be simple text accompanied with images.⁴³ Complex advisories are likely to be ignored by consumers and effective advisories should be tailored to specific consumers.⁴⁴ Label design has a significant impact on warning label effectiveness.⁴⁵ A careful review of the successes and failures of different consumer advisories can show how to best increase the long-run effectiveness of pediatric antidepressant risk warnings.

One limitation of this study is the methodology used to determine the severity of depression. Even though the Columbia Impairment Scale is commonly used in clinical settings, it is not specific to the assessment of depression severity. The severe psychological impairment indicator we used that was based on the CIS score only serves as a proxy to the respondents’ actual level of severity and need for antidepressants.

An interesting future research question that is not within the scope of this paper is whether the short-run and long-run trends observed in antidepressant use was driven by provider or patient behavior. Some research has suggested that both patient and provider behaviors have changed due to the warning,^46–48 but studies have focused mostly on the decreased treatment of depression,⁵ or the lapses in heeding the box warning in ambulatory services.¹¹ Another related research area is the appropriate clinical reaction to this specific FDA box warning. Rather than simply writing fewer prescriptions, some researchers have advised that the box-warning should concurrently result in increased monitoring of children on antidepressants or an increase in alternatives such as cognitive behavioral therapy, interpersonal counseling,⁴⁹ and psychosocial treatment.⁵⁰