Beyond depression: Estimating 12-months prevalence of passive suicidal ideation in mid- and late-life in the Health and Retirement Study

Abstract

Objectives:

To provide valid estimates of the 12-month prevalence of passive suicidal ideation among older adults, without conditioning on depression status, using the Health and Retirement Study (HRS).

Methods:

Data come from the 2012 HRS (n= 17,434) and 2004/5 Baltimore Epidemiologic Catchment Area (ECA) Study (n=755). In the HRS, passive suicidal ideation (i.e., thought a lot about death – your own, someone else’s, or death in general) is only assessed on respondents who reported dysphoria/anhedonia; in the ECA, ideation is assessed on all respondents, regardless of depression. We compare two approaches to estimating the 12-month prevalence of passive suicidal ideation in the HRS without conditioning on depression symptoms: (i) a probit selection model within the HRS, and (ii) a prediction model developed using appended ECA data applied to the HRS.

Results:

Using observed data alone on those who screened positive for depression, 6% of older adults reported passive suicidal ideation in the past year. Depending on the approach used, between 5.4% and 9.2% of HRS respondents who screened negative for depression would have reported passive suicidal ideation had they been assessed. Correcting for this selection bias, between 10.9% and 13.4% of US adults aged >50 experienced passive suicidal ideation in 2012.

Conclusions:

Population surveillance of suicidal ideation among older adults is biased by survey approaches that only assess ideation in the context of depression.

Suicide is the 10^th leading cause of death in the US. Incidence of suicide is highest among older adults,¹ and rates have increased steadily over the past decade.² Suicidal behavior is often considered primarily in relation to mental disorders such as depression; however, several recent meta-analyses have concluded that depression only modestly predicts risk of completed suicide in the general population.^3,4 The links between depression and suicidal behavior may be even more nuanced among older adults. For example, late-life depression is expressed differently (sometimes called “depression without sadness”),⁵ which is not well captured by standard diagnostic criteria. Consequently, using depression status as a proxy for suicidal ideation, or only screening for suicidal behavior among older adults who endorse symptoms of depression, may underestimate suicide risk.

There are unresolved conceptual issues regarding appropriate assessment and clinical meaning of preoccupation with death among older adults. From a life course perspective, some degree of rumination about death may be normative in later life.⁶ However, a number of studies have shown that the distinctions between thinking about death, passive suicidal ideation and active suicidal ideation are fluid and contingent on the chronicity and frequency of these thoughts and emotions.^7–9 For example, both active and passive suicide ideation are similarly associated with history of suicide attempts.⁷ Frameworks such as the Interpersonal Theory of Suicide (ITS)^10,11 posit that completed suicide is not necessarily associated with active ideation, but rather is precipitated by factors such as sense of hopelessness, lack of fear of death, and acquired capability for self-harm. Taken together, existing research indicates that thinking about death may be an important factor in suicide risk.

Only a handful of community-based studies of have sufficiently large samples to examine the correlates and predictors of suicidal behavior among older adults. The overwhelming majority of studies of suicidal behavior in later life are based on clinical psychiatric samples,¹² which may not identify factors that contribute to changing population incidence of suicide over time. This lack of community-based studies may reflect, in part, the sensitive nature of asking about suicidal behavior.⁷ Surveillance surveys often intentionally exclude assessment of suicidal behavior (e.g., the Behavioral Risk Factor Surveillance System uses the eight-item Patient Health Questionnaire, only excluding the item on suicidal ideation).¹³ If suicidality is assessed, these items are often only asked of respondents thought to be at high-risk because of their answers to related items, almost always depressed mood.^14,15 Surveys that do assess suicidal ideation on all respondents (e.g., National Health and Nutrition Examination Survey (NHANES), National Survey on Drug Use and Health (NSDUH)) have limited numbers (often n<4000) of older adults. In sum, current population mental health surveillance efforts in the US have limited ability to estimate the prevalence and correlates of suicide risk among older adults.

Instruments such as the Composite International Diagnostic Interview (CIDI), represent the most well-validated assessments of major depression (MD) appropriate for general population surveys.^15,16 However, the CIDI MD module is limited as a means of assessing suicidality because instrument skip patterns dictate that the questions on suicidal ideation are only asked of respondents who screen into the module by reporting sadness or anhedonia.¹⁶ This means that respondents who do not endorse the screening symptoms are assigned a value of “missing” on all subsequent items, including ideation. The logic of this skip pattern is predicated on the notion that these screening symptoms are indeed highly sensitive measures of suicidal behavior; however, as discussed above, this assumption may not be tenable and may result in biased estimates of the prevalence of ideation, particularly for older adults.

The goal of this study is to illustrate two analytic approaches to addressing the potential bias in population estimates of suicidal ideation among older adults in the context of non-ignorable missing data. This study has two objectives. First, to empirically identify factors that contribute to variation in the strength of the correlation between depressed mood/anhedonia and suicidal ideation in later life. Second, to estimate the 12-month prevalence of passive suicidal ideation among adults over the age of 50 in the US.

Methods

Data sources and sample descriptions.

Data for this project come from two studies that have identical measures of passive suicidal ideation: The Health and Retirement Study (HRS) and the Baltimore Epidemiologic Catchment Area Study (ECA). Both the HRS and ECA are longitudinal, population-based cohorts that have diagnostic assessments of MD (i.e., HRS: short-form CIDI; ECA: Diagnostic Interview Schedule (DIS)) which include identical measures of passive suicidal ideation. However, unlike the CIDI MD module, the DIS MD module asks all symptom items of all respondents, regardless of dysphoria/anhedonia. Additional details of the HRS and ECA cohorts are described elsewhere.^17,18

The HRS is an ongoing, nationally-representative, longitudinal survey of US adults aged >50 that began in 1992. Response rates for the ~20 000 respondents are between 85% and 90% at each biennial wave.¹⁷ For the present study, the HRS sample (RAND HRS data file, version P) consisted of 17 434 respondents who 1) completed the 2012 HRS core interview, 2) had non-zero survey weights, 3) completed the two depression screening items for the CIDI, and 4) had data on candidate predictors.

The ECA is a prospective cohort study drawn from households in East Baltimore and initially interviewed in 1981 (Wave 1; n=3481), then re-interviewed in 1982 (Wave 2, n=2768), 1993-1996 (Wave 3, n=1920) and 2004-2005 (Wave 4, n=1071). Response rates for each ECA wave are ~75%.¹⁸. For this analysis, the ECA was used as an external data source that was appended to the HRS, used to build a prediction model from a set of common variables, and then applied to the HRS to generate prevalence estimates. The ECA sample was limited to the 755 Wave 4 respondents aged ≥50 who had complete data on passive suicidal ideation and depression, to increase comparability with the HRS.

The HRS is approved by the Institutional Review Board (IRB) at the University of Michigan. The ECA is approved by the IRB at Johns Hopkins University. All respondents provided informed consent.

Measurements.

Outcome.

In both the HRS and ECA, passive suicidal ideation was assessed by asking respondents whether they “thought a lot about death, either your own, someone else’s, or death in general” (recorded as Yes/No) over a two-week period. This question was asked of all ECA respondents (i.e., regardless of other depressive symptoms) as part of the DIS MD module. In the HRS, this question was only asked to respondents who first screened into the CIDI MD module (Figure 1) by reporting feeling (a) either dysphoria (“felt sad, blue, or depressed”) or anhedonia (“lose interest in most things”) most of the day almost every day, and (b) that these symptoms lasted two weeks or more in the past 12 months; these respondents completed the full CIDI MD module, including the item about passive suicidal ideation. All other respondents were assigned a value of “missing” on this item.

Figure 1.

Skip pattern of the Composite International Diagnostic Interview (CIDI) Major Depression Module as implemented in the Health and Retirement Study

For the selection model, we generated a dummy variable (Yes/No) indicating whether a respondent screened into the CIDI based on the skip pattern illustrated by Figure 1. For the prediction model, we used parallel items in the DIS (i.e., symptoms of dysphoria and anhedonia in the past year) to create a corresponding dummy variable in the ECA to mimic the skip pattern of the CIDI. Additional details about the DIS and CIDI MD modules are provided in Supplemental Table S1.

Predictors.

Potential correlates of suicidal ideation based on existing studies and included sociodemographic characteristics, mental health status, functional status, history of medical conditions, and health behaviors.³ We generated a set of variables that were either operationalized identically or conceptually equivalent in both the HRS and ECA, detailed in Supplemental Table S2.

Sociodemographic characteristics.

These included age, sex, race/ethnicity, education and marital status.

Mental health status.

These included self-rated health, psychological distress, and memory. Psychological distress was indexed using the 8-item Center for Epidemiologic Studies – Depression (CESD) scale in the HRS and the 20-item General Health Questionnaire (GHQ) in the ECA. Standardized summative scores were generated for each instrument, with higher values indicating worse mental health. Memory was assessed using the immediate and delayed word recall tasks (HRS: 10-word recall, ECA: 20-word recall); these sum scores were also mean-standardized for comparability.

Functional status.

This was assessed by self-reported difficulties in activities of daily living (ADLs) (e.g., bathing, dressing and eating) and mobility (e.g., walking one block, walking several blocks, walking across a room, climbing one flight of stairs, and climbing several flights of stairs activities ).¹⁹

Medical conditions.

History of common medical conditions (e.g., diabetes, heart disease, cancer) were assessed by self-report.

Health behaviors.

Smoking status assessed by self-report.

Statistical analysis.

Selection bias occurs when there are unobserved heterogeneities affecting both selection (i.e., screening into the full CIDI MD module) and the outcome (i.e., passive suicidal ideation).²⁰ We employed two analytic approaches to address the selection bias introduced by the skip pattern of the CIDI in order to generate estimates of the population prevalence of passive suicidal ideation from the HRS. The first approach fit a sample selection model within the HRS;²¹ the second approach used data on passive suicidal ideation from the DIS to estimate a statistical prediction model²² in the ECA, which was then used to impute missing data values in the HRS.

Sample Selection Model in the HRS.

Figure 2 illustrates the four possible combinations of selection (depression screening) and outcome (passive suicidal ideation):

1. DS: screened in (+) and passive suicidal ideation positive (+);

2. $D\overline{S}$: screened in (+) and passive suicidal ideation negative (−);

3. $\overline{D}S$: screened out (−) and passive suicidal ideation positive (+); and

4. $\overline{D}\overline{S}$: screened out (−) and passive suicidal ideation negative (−).

We calculated the prevalence of DS and $D\overline{S}$ based on the observed data, and estimated the predicted probabilities of $\overline{D}S$ and $\overline{D}\overline{S}$ using a probit model with sample selection. The model allows a correlation between the unmeasured heterogeneities (error terms) that influence selection (screening into the full CIDI) and outcome (thinking a lot about death) (ρ ≠ 0), which indicates the existence of selection bias.^23,24 We fit the model using maximum-likelihood.

Figure 2. The joint relationship between depression and suicidal ideation using the Composite International Diagnostic Interview (CIDI) Major Depression Module: The 2012 Health and Retirement Study.

This figure illustrates how the skip pattern of the CIDI Major Depression Module impacts the estimation of the overall population prevalence of passive suicidal ideation in the HRS. The prevalence of passive suicidal ideation among those who screened into the full CIDI Module is known (DS) based on the observed data. The prevalence of passive suicidal ideation among those who did not screen into the full Module ($\overline{D}S$) is unobserved. We use two approaches (selection models within the HRS and prediction modeling using an external data source) to estimate $\overline{D}S$. We then combine these values of $\overline{D}S$ with the observed data to generate a revised estimate of the overall prevalence of passive suicidal ideation in the HRS that corrects for the selection bias introduced by the skip pattern. The inset heuristic Venn diagram illustrates the conceptual distinction between depression (D) and suicidal ideation (S) in the population overall.

Prediction Model in the ECA.

We used the ECA data to develop a logistic regression model of the probability of reporting passive suicidal ideation among those who were not depressed. We first imputed missing values of covariates in the ECA using multiple imputation by chained equations.²⁵ Next, we fit the prediction model using the same set variables as those from the selection equation of the sample selection model, and tested accuracy of the prediction model by conducting a 10-fold cross-validation.²² We then estimated the probability of passive suicidal ideation among those who did not screen into the CIDI in the HRS.

Using both approaches, we calculated the overall prevalence of passive suicidal ideation in the full HRS sample using the observed values for screened-in respondents and the estimated probability for screened-out respondents. We examined correlates of suicidal ideation in both the HRS and ECA using X² tests for categorical variables and Wald tests and Kruskal–Wallis tests for continuous variables.

Additional details of the methods are in the Appendix.

All analyses were conducted using Stata Version 14.2 using survey procedures and all prevalence estimates reflect the HRS sample weights (StataCorp, College Station, TX). All significance tests were evaluated at the level of P<0.05.

Results

Among 17,434 HRS respondents, 1,547 (8.9%) screened into the full CIDI MD module and therefore had data on passive suicidal ideation (Figure 2 and Table 1, left columns). Respondents who screened into the CIDI were younger, more likely to be female, less likely to be married, had higher psychological distress, more functional limitations, and were more likely to have history of smoking and most medical conditions. They did not differ in terms of race/ethnicity, educational attainment or memory. Among the 1,547 respondents who screened into the CIDI, 1,039 (67.2%) reported passive suicidal ideation (Table 1), which corresponds to an observed overall population prevalence of 6.0% (Figure 2). Among respondents who screened into the CIDI, those who reported ideation were older, more likely to be widowed, had higher psychological distress, more functional limitations and poorer health status than those who screened in but did not report ideation.

Table 1.

Sample characteristics of study participants from the 2012 Health and Retirement Study and the 2004-2005 Baltimore Epidemiologic Catchment Area follow-up study

	HRS (n=17,434)†						Baltimore ECA (n=755)‡
	CIDI major depression module screening status			Passive suicidal ideation among screened-in respondents			Passive suicidal ideation among all respondents
	Screened in (n=1,547)	Screened out (n=15,887)	Test statistics, df, P value	Yes (n=1,039)	No (n=508)	Test statistics, df, P value	Yes (n=56)	No (n=699)	Test statistics, df, P value
Age, mean (SE)	63.0 (0.4)	65.79 (0.3)	83.6, 1, <0.001	63.4 (0.4)	62.3 (0.5)	4.12, 1, 0.047	61.1 (1.5)	62.5 (0.4)	1.73, 1, 0.188
Female sex, %	63.8	53.5	38.01, 1, <0.001	65.9	59.9	2.67, 1, 0.108	75.0	63.1	3.19, 1, 0.074
Race, %			2.56, 1.79, 0.088			0.08, 1.87, 0.912			0.06, 2, 0.970
White	83.8	83.3		83.4	84.6		60.7	62.4
Black	8.6	10.4		8.8	8.3		35.7	34.2
Others	7.6	6.3		7.8	7.2		3.6	3.4
Education, %			1.50, 1.95, 0.227			2.22, 1.97, 0.115			1.30, 2, 0.522
Below high school	14.3	12.8		15.9	11.4		25.0	27.6
High school	33.3	32.1		34.2	31.6		28.6	33.6
Above high school	52.4	55.1		49.9	57.0		46.4	38.8
Marital status, %			39.57, 1.70, <0.001			12.94, 1.97, <0.001			0.53, 2, 0.767
Married/partnered	51.6	65.4		47.8	58.4		50.0	53.2
Separated/divorced/single	33.3	21.3		32.9	33.9		32.1	27.6
Widowed	15.1	13.3		19.2	7.7		17.9	19.2
Mental Health, mean (SE)
CESD	4.2 (0.1)	1.1 (0.0)	908.47, 1, <0.001	4.5 (0.1)	3.6 (0.2)	19.59, 1, <0.001	-	-	-
GHQ	-	-	-	-	-	-	20.2 (1.3)	15.4 (0.2)	19.94, 1, <0.001
Standardized CESD/GHQ	1.3 (0.1)	−0.2 (0.0)	908.47, 1, <0.001	1.4 (0.1)	1.0 (0.1)	19.59, 1, <0.001	0.7 (0.2)	−0.1 (0.0)	19.94, 1, <0.001
Memory, mean (SE)
Raw memory scores
Immediate recall	5.6 (0.1)	5.5 (0.0)	0.35, 1, 0.558	5.5 (0.1)	5.7 (0.1)	1.03, 1, 0.314	6.5 (0.4)	6.6 (0.1)	0.114, 1, 0.736
Delayed recall	4.6 (0.1)	4.6 (0.0)	0.01, 1, 0.913	4.5 (0.1)	4.6 (0.1)	0.64, 1, 0.426	5.1 (0.4)	5.0 (0.1)	0.00, 1, 0.964
Standardized memory scores
Immediate recall	0.2 (0.0)	0.1 (0.0)	0.35, 1, 0.558	0.1 (0.0)	0.2 (0.1)	1.03, 1, 0.314	−0.2 (0.1)	−0.1 (0.0)	0.114, 1, 0.736
Delayed recall	0.1 (0.0)	0.1 (0.0)	0.01, 1, 0.913	0.1 (0.0)	0.2 (0.1)	0.64, 1, 0.426	−0.1 (0.1)	−0.1 (0.0)	0.00, 1, 0.964
Functional status, %
Any ADL limitations	26.7	9.1	241.58, 1, <0.001	30.8	19.3	11.43, 1, 0.001	17.9	6.9	8.83, 1, 0.003
Any mobility limitations	68.2	41.4	199.67, 1, <0.001	71.3	62.4	10.36, 1, 0.002	51.8	35.6	5.81, 1, 0.016
Self-rated health status, %			199.45, 2.32, <0.001			6.48, 2.75, <0.001			9.36, 3, 0.025
Excellent/very good	18.6	47.2		16.4	22.5		8.0	14.9
Good	27.0	31.5		24.7	31.3		36.0	50.1
Fair	32.7	16.3		33.9	30.4		44.0	28.8
Poor	21.7	5.0		25.0	15.8		12.0	6.2
History of medical conditions, %
Psychiatric problems	58.2	15.1	1286.58, 1, <0.001	57.5	59.5	0.32, 1, 0.572	-	-	-
Heart disease	31.2	21.5	41.40, 1, <0.001	33.8	26.6	4.92, 1, 0.031	30.2	18.4	4.41, 1, 0.036
Diabetes	27.6	20.6	28.89, 1, <0.001	28.3	26.3	0.47, 1, 0.498	28.6	20.2	2.20, 1, 0.138
Hypertension	60.1	55.7	7.99, 1, 0.007	63.1	54.7	4.20, 1, 0.045	51.8	55.7	0.32, 1, 0.569
Stroke	10.9	6.6	24.77, 1, <0.001	10.8	10.9	0.00, 1, 0.972	21.8	5.2	23.66, 1, 0.10, 1, <0.001
Lung disease	22.0	8.6	148.16, 1, <0.001	23.8	18.8	3.97, 1, 0.051	-	-	-
Cancer	16.0	14.1	3.14, 1, 0.082	16.6	14.9	0.38, 1, 0.539	7.3	9.3	0.26, 1, 0.612
Arthritis	70.0	54.2	86.80, 1, <0.001	70.2	69.7	0.02, 1, 0.881	51.9	54.2	0.10, 1, 0.748
Smoking, %			40.55, 1.89, <0.001			2.49, 1.96, 0.088			0.70, 2, 0.704
Never smoked	35.3	44.8		34.3	36.9		57.1	60.9
Former smoker	40.0	41.8		39.0	41.9		14.3	15.5
Current smoker	24.7	13.4		26.7	21.1		28.6	23.6

Notes. ADL, activities of daily living; CESD, Center for Epidemiologic Studies Depression Scale; CIDI, the Composite International Diagnostic Interview; ECA, Epidemiologic Catchment Area; GHQ, general health questionnaire; HRS, Health and Retirement Study; SE, standard error. P values of continuous variables were from adjusted Wald tests for the HRS sample, and Kruskal–Wallis tests for the ECA sample. P values of categorical variables were from Pearson’s chi square tests for both samples. In the HRS sample, the degrees of freedom based on the fixed degrees of freedom rule for complex sample design was 56, which was calculated by subtracting the number of strata from the number of clusters.

^†Sample characteristics of the HRS sample were accounted for the survey design.

^‡The numbers of missing values in the ECA sample were 76 for self-rated health status, 75 for the GHQ, 75 for the immediate and delayed word recall task,, 17 for history of arthritis, 8 for history of hypertension, 6 for history of heart disease, 3 for mobility limitations, 3 for history of stroke, 3 for history of cancer, and 1 for history of diabetes.

Turning to the ECA, 9.4% (n=71) of respondents had depressed mood/anhedonia and 7.4% (n=56) reported passive suicidal ideation in the past year. Using the dummy variable of depressed mood/anhedonia to mimic the CIDI skip pattern, the conditional probability of passive suicidal ideation given depression (S|D) was 26.8%. The observed conditional probability of ideation among those who did not have depressed mood (S|$\overline{D}$ in Figure 2, which is unobserved in the HRS but known in the ECA) was 5.4%. ECA respondents who reported passive suicidal ideation had more psychological distress, more functional limitations and lower self-rated health than those who did not report ideation (Table 1, right columns), similar to the HRS.

Table 2 presents the regression results of the selection model (left columns) and the prediction model (right columns) at identifying correlates of passive suicidal ideation in the HRS and ECA, respectively. In both cases the overall model fit was significant relative to the null. In the selection model, the correlation of the error terms was 0.67 (95% confidence interval [CI], 0.13-0.90), consistent with our expectations that there were unobserved heterogeneities influencing both depression and passive suicidal ideation and that selection bias was present. The factor significantly associated with suicidal ideation in the selection model was depressive symptoms and marital status. In the prediction model, the mean cross-validated C-statistic was 0.70 from the training sets and 0.43 from the validation sets, indicating that the predictive model had fair to good discrimination at identifying passive ideation in the ECA. The factor most strongly correlated with ideation in the prediction model was history of stroke (β=1.21, 95% CI: 0.04, 2.38).

Table 2.

Results from the selection and prediction models for estimating passive suicidal ideation

	HRS (n=17,434) Selection model (bivariate probit)		ECA (n=683) Prediction model (logit)
	Screened into the CIDI Coef. (95% CI)	Passive suicidal ideation Coef. (95% CI)	Passive suicidal ideation Coef. (95% CI)
Age	−0.02 (−0.02, −0.01)	(exclusion restriction)	−0.01 (−0.05, 0.03)
Female sex	0.11 (0.03, 0.20)	0.19 (−0.00, 0.37)	0.20 (−0.58, 0.98)
Race
White	Reference	Reference	Reference
Black	−0.32 (−0.46, −0.18)	−0.24 (−0.49, 0.00)	0.32 (−0.48, 1.12)
Other	−0.20 (−0.34, −0.06)	−0.12 (−0.46, 0.22)	−0.13 (−2.25, 1.99)
Education
Below high school	Reference	Reference	Reference
High school	0.18 (0.06, 0.30)	0.08 (−0.22, 0.38)	−0.08 (−1.10, 0.95)
Above high school	0.27 (0.15, 0.40)	0.09 (−0.21, 0.38)	0.57 (−0.39, 1.53)
Marital status
Married/partnered	−0.09 (−0.18, −0.01)	−0.18 (−0.33, −0.02)	Reference
Other status	Reference	Reference	0.00 (−0.75, 0.75)
Mental health (CESD/GHQ)	0.23 (0.21, 0.25)	0.17 (0.10, 0.23)	0.38 (−0.04, 0.79)
Memory
Immediate word recall	0.04 (0.00, 0.08)	0.03 (−0.02, 0.08)	−0.03 (−0.66, 0.61)
Delayed word recall	0.01 (−0.02, 0.04)	0.01 (−0.04, 0.07)	0.19 (−0.45, 0.84)
Self-rated health status
Excellent	Reference	Reference	Reference
Good	0.15 (0.03, 0.27)	0.03 (−0.23, 0.29)	0.06 (−1.17, 1.30)
Fair	0.26 (0.12, 0.39)	0.19 (−0.09, 0.47)	0.34 (−0.95, 1.62)
Poor	0.35 (0.23, 0.47)	0.28 (−0.03, 0.60)	0.48 (−1.39, 2.36)
ADL limitations
None	Reference	Reference	Reference
Any	0.06 (−0.06, 0.18)	0.20 (−0.02, 0.42)	0.13 (−1.14, 1.40)
Mobility limitations
None	Reference	Reference	Reference
Any	0.08 (−0.02, 0.19)	0.06 (−0.11, 0.23)	0.15 (−0.72, 1.01)
History of medical conditions
Psychiatric problems	0.60 (0.51, 0.68)	0.14 (−0.15, 0.42)	N/A
Heart disease	0.07 (−0.02, 0.16)	0.11 (−0.06, 0.28)	0.32 (−0.58, 1.22)
Hypertension	−0.09 (−0.18, −0.01)	0.07 (−0.12, 0.26)	−0.32 (−1.06, 0.42)
Lung disease	0.16 (0.04, 0.29)	0.12 (−0.07, 0.30)	N/A
Cancer	−0.00 (−0.10, 0.10)	0.00 (−0.21, 0.22)	−0.05 (−1.32, 1.22)
Diabetes	0.05 (−0.04, 0.14)	(exclusion restriction)	−0.16 (−1.07, 0.74)
Stroke	0.11 (−0.02, 0.23)	(exclusion restriction)	1.21 (0.04, 2.38)
Arthritis	0.06 (−0.03, 0.14)	(exclusion restriction)	−0.14 (−0.93, 0.65)
Smoking
Never	Reference	Reference	Reference
Ever	0.01 (−0.09, 0.12)	−0.06 (−0.22, 0.09)	0.18 (−0.80, 1.16)
Current	0.04 (−0.10, 0.18)	0.08 (−0.12, 0.27)	0.09 (−0.76, 0.94)

Note: ADL, activities of daily living; CESD, Center for Epidemiologic Studies Depression Scale; CIDI, the Composite International Diagnostic Interview; ECA, Epidemiologic Catchment Area; GHQ, general health questionnaire; HRS, Health and Retirement Study. The prediction model was fit among respondents who were not depressed in the ECA sample.

We then used these two approaches to estimate the prevalence of passive suicidal ideation in the HRS. Using the selection model, we estimate that 9.2% (Standard Error [SE]= 0.07) of respondents who screened out of the CIDI would have reported passive suicidal ideation had they been asked (S|$\overline{D}$ in Figure 2). Overall, the selection model estimated that 13.4% (SE=0.22) of HRS respondents experienced passive ideation in the past year (DS+$\overline{D}S$ in Figure 2). Applying the prediction model of suicidal ideation developed in the ECA to the HRS, the estimated prevalence of passive suicidal ideation among those who did not screen into the CIDI ($\overline{D}S$ in Figure 2) was 5.4% (95% CI: 0.053, 0.055). Overall, the prediction model estimated that 10.9% (95% CI: 0.104, 0.114) of HRS respondents experienced passive ideation in the past year (DS+$\overline{D}S$).

Because the outcome thinking about death could be indicative of recent bereavement, we conducted a post-hoc analysis stratifying the model by marital status. As shown by Table 3, passive ideation was higher among those widowed and separated/divorced/never married relative to currently married. This suggests that while widowhood may contribute to the likelihood of reporting preoccupation with death, this does not entirely explain our findings.

Table 3.

Probability of past-year passive suicidal ideation in the HRS, stratified by marital status, from the selection and prediction models

	N	Selection model			Prediction model
		$\overline{D}S$	Total	$\overline{D}S$	Total
Main analysis	17,434	0.092	0.134	0.054	0.109
By marital status
Widowed	3,030	0.088	0.150	0.054	0.134
Married/partnered	10,807	0.058	0.092	0.051	0.092
Separated/divorced/never married	3,597	0.074	0.142	0.063	0.143

Note: $\overline{D}S$, screened out and passive suicidal ideation positive; HRS, Health and Retirement Study.

Conclusions

This study illustrates a novel application of two distinct analytic approaches to reduce bias in the surveillance of suicidal ideation. To our knowledge these findings represent the first attempt of quantifying the population prevalence of passive suicidal ideation in the HRS, the largest surveillance survey of health and aging in the US. We estimate that between 10.9% and 13.3% of US adults aged >50 experienced passive ideation within the past year. The variability of these estimates reflects differences in model assumptions, but both approaches indicate that relying on the observed data alone underestimates the prevalence of passive ideation among older adults. Given that late-life suicide attempts are characterized by a high degree of lethality,²⁶ accurate estimates of suicide risk factors are needed to inform prevention efforts targeting older adults.

To put these estimates in context, the 2012 NSDUH reported that 2.4% of US adults aged ≥50 experienced active ideation (i.e., “seriously think about trying to kill yourself?”) in the past year.²⁷ In the 2011/12 NHANES, 3.8% of adults aged 40-59 and 2.7% aged ≥60 reported currently experiencing (i.e., in the past two weeks) ideation (i.e., “thoughts that you would be better off dead or of hurting yourself in some way?”).²⁸ Neither assessment was conditional on depression status. The indicator of passive suicidal ideation used in the current study (i.e., thinking a lot about death) is a far less conservative measure of suicidality than either of these metrics, and therefore we expect our prevalence estimates to be higher.

Our findings are consistent with a 2018 CDC report that more than half of suicide decedents did not have evidence of psychopathology at the time of their death.²⁹ This is particularly relevant for older adults;²⁶ older adults who die by suicide are more likely to have medical problems or chronic pain, but are less likely to have psychiatric symptoms or a clinical diagnosis, relative to younger suicide decedents.³⁰ Our findings demonstrate that surveillance surveys which rely on depressive symptoms as a screening indicator for suicidality systematically miss cases of ideation among older adults.

Findings should be interpreted considering study strengths and limitations. The HRS is a well-characterized nationally representative cohort of older adults. By employing multiple analytic approaches, each with distinct strengths, limitations, and assumptions, we aimed to triangulate on a range of plausible estimates for passive ideation. Within a triangulation framework, we both gain insight into the reliability of our results and generate insights regarding the direction and magnitude of unmeasured biases.³¹ Limitations include the identification and robustness of the selection model;^32,33 while we empirically identified variables for exclusion restrictions, they may not be completely exogenous. Selection models assume that the estimates generated using observed data from the screened-in respondents are applicable to those unobserved, and this method is highly sensitive to model specification.³² Despite these limitations, selection models have been recently used in epidemiology to address missing data in a variety of contexts (e.g., estimate population HIV seroprevalence from partial data).²⁴

As a counterbalance to the weaknesses of the selection model, we employed a predictive modeling approach using an external dataset. The HRS and ECA are generally comparable in terms of sample characteristics and measures. However, they differ in four potentially important ways. First, the timing of data collection (HRS in 2012; ECA in 2004/5). Second, the ECA is a single-site study whereas the HRS is nationally-representative. Third, the samples differ in terms of racial/ethnic and socioeconomic composition, with the ECA having both a larger proportion of non-white respondents and a larger proportion of respondents with lower education, although these characteristics were not significantly associated with passive suicidal ideation in this analysis. Finally, the skip pattern we mimicked in the ECA was not identical to the HRS, because the latter additionally required that the depressive symptoms were of sufficient intensity, which was not assessed in the DIS. These factors may have contributed to the difference in prevalence estimates of ideation generated by the prediction modeling approach.

These findings have reverberating implications for research, clinical care, and policy. For research, misclassification is a double-edged sword that both reduces the statistical efficiency of studies and has the potential to produce spurious associations when the misclassification is correlated with case status (e.g., depression), muddying the scientific literature with false positive findings. For clinical care, these findings inform the debate regarding the utility of screening for suicidal behavior, separate from depression, in primary care. The US Preventative Task Force states that evidence is “insufficient to assess the balance of benefits and harms of screening for suicide risk” and it specifically cites a need for research on the epidemiology suicide risk.³⁴ Our findings emphasize the importance of functional limitations as a correlate of suicidal ideation and thus could inform targeting of such screening efforts. These findings support health policy efforts to embed and expand the availability of comprehensive mental health care (e.g., collaborative care models, interprofessional healthcare teams) for older adults.³⁵ A combination of approaches (e.g., physician training,³⁶ integrated team-based care practices,³⁷ machine learning algorithms³⁸) are likely needed to improve current suicide prevention efforts.

Finally, it is important to contextualize these findings within the broader discussion how best to measure suicidality in the general population. We conducted a post-hoc analysis assessing the correlation of thinking about death with another DIS item, “felt so bad that yon wanted to die,” that was not included in the CIDI. These two items were strongly correlated (19.3% of those who endorsed “thinking a lot about death” also endorsed “wanting to die” vs. only 0.9% who did not (OR: 27.8; 95% CI: 9.8, 78.6)), showing empirically, at least in the ECA, respondents who endorse this item are not solely thinking about deaths of loved ones (e.g., recent widowhood)). These findings illustrate the importance of including comprehensive measures of suicidality in large population health surveillance and informs, but by no means resolves, the ongoing debate on how to measure the heterogeneous experiences like suicidal behavior.³⁹

HIGHLIGHTS.

1. What is the primary question addressed by this study? How common is, and what are the predictors of, passive suicidal ideation among older adults in the US?

2. What is the main finding of this study? After correcting for selection bias, between 10.9% and 13.4% of older US adults experienced passive ideation in the past year, as compared to 6% based on the observed data alone. History of stroke and current functional limitations predict passive suicidal ideation, above and beyond depressed mood.

3. What is the meaning of this finding? Passive suicidal ideation is not limited to older adults reporting depressed mood, and is associated with health history and functional status.