Identifying Trajectories of Depressive Symptoms for Women Caring for Their Husbands with Dementia

Abstract

OBJECTIVES

To use an innovative statistical method, Latent Class Trajectory Analysis (LCTA), to identify and describe subgroups (called trajectories) of caregiver depressive symptoms in a national sample of wives providing informal care for their husbands with dementia.

DESIGN

Longitudinal.

SETTING

Community.

PARTICIPANTS

Respondents to the National Longitudinal Caregiver Survey were wife caregivers of veterans with dementia who were identified through Veterans Affairs hospitals nationally.

MEASUREMENTS

Mean number of depressive symptoms as measured using the Center for Epidemiologic Studies Depression scale (CES-D, 20-item scale).

RESULTS

Overall mean depressive symptoms of wife caregivers were 6.2 of 20, below the cutpoint (8 or 9/20) associated with clinical depression. Four distinct trajectories of caregiver depressive symptoms were identified. The trajectory with the highest number of symptoms (11.9 of 20), contained one-third of the sample. Another third had mean depressive symptoms virtually identical to the overall sample mean. The final third were divided between two trajectories, low depressive symptoms (mean CES-D, 3.0/ 20, 22% of sample) and very low (mean CES-D, 0.8/20, 14% of sample). Approximately two-thirds of the sample members were in a depressive symptom trajectory, with substantially higher or lower numbers of symptoms than the overall mean. Two subjective measures asked of wife caregivers (desire for more help, life satisfaction) were significantly associated with membership in the highest depressive symptom trajectory.

CONCLUSION

LCTA identified important depressive symptom subgroups of wife caregivers. A population-averaging method identified a mean effect that was similar to the effect in one-third of the cases but substantially different from that in two-thirds of the cases.

There is great interest in understanding how human processes (biological, social, behavioral) develop or change over time. Development of statistical procedures that take advantage of data sources with repeated measures has accelerated in recent years, allowing for empirical estimation of many such phenomenon.^1,2 This article applies such a statistical procedure, Latent Class Trajectory Analysis (LCTA), to the problem of identifying the effect of spousal caregiving on caregiver depressive symptoms. The effect of caregiving on the caregiver is important, because some 7 million persons in the United States provide informal care (unpaid help dealing with disability) to a family member aged 65 and older who is suffering from a long-term debilitating illness or disability, and virtually all of the more than 5 million Americans living with Alzheimer’s disease in the community receive such care.³ Caregiving has been found to be costly^4,5 and, on balance, harmful to caregivers,^6–8 and the prevalence of caregiving is anticipated to increase to 40 million persons caring for 28 million elderly disabled persons by 2050.⁹

Applying LCTA to analyze depressive symptoms of spousal caregivers is of interest for several reasons. First, although caregiving and its effect on caregiver depressive symptoms has been studied previously,^10,11 little is known about the nature of depressive symptom trajectories (empirically distinct patterns of change over time). LCTA is designed to identify the number and nature of trajectories that exist in a sample.¹² Second, most caregiver depression research has relied on statistical methods that identify variability around a common mean, presuming this measure of central tendency to be the key descriptor of the overall population, but unique distributions of trajectories across distinct subgroups could produce similar population averages.¹² LCTA is designed specifically to identify any such variability that is meaningful from a statistical standpoint, and there is reason to believe that groups of individuals respond differently to caregiving. Furthermore, the presence of distinct trajectories could have strong clinical and policy implications, making the use of this method potentially useful for identifying and responding to at-risk caregivers.

The purpose of this article is to use an innovative statistical method, LCTA, to determine the number and nature of distinct trajectories of caregiver depressive symptoms present in a national sample of wives caring for husbands with dementia, the most common reason for needing such care.

METHODS

Study Sample

The National Longitudinal Caregiver Survey (NLCS) is a national survey of informal caregivers of elderly veterans clinically diagnosed with Alzheimer’s disease or vascular dementia in the Veterans Affairs health system.⁴ The purpose of the NLCS survey was to document the amount and type of informal care provided to persons with dementia, as well as the consequences of providing such care. Respondents were surveyed annually (1999–2002), up to four times. Surveys of caregivers after care recipient death or institutionalization also were completed. The analyses were limited to wives caring for their husbands, because the sample of other caregivers (mostly adult children caring for their fathers) was too small for reliable estimates. The total number of caregiver–care recipient dyads used in the analyses was 1,580; 336 nonspousal caregivers and 15 caregivers without depressive symptoms data at any survey were excluded.

Outcome Measure

The outcome measure is the number of caregiver depressive symptoms, measured using the Center for Epidemiologic Studies Depression Scale (CES-D) short Boston form, consisting of 20 items scored in a yes-or-no format; higher scores indicate a higher number of depressive symptoms.¹³ Caregivers reported depressive symptoms at each of the four surveys.

Analytical Approach

Two broad types of statistical methods are used to analyze longitudinal data to identify important patterns of change. Hierarchical linear models (HLM)^14,15 and growth curve analyses (GCA) constitute one type.^16–18 Both provide an understanding of differences across population members in the development of a process (e.g., depressive symptoms) over time. Although HLM and GCA models differ in important ways, both share a fundamental assumption of a multivariate normal, continuous distribution of trajectories in the population.¹ They also assume variability around a grand population mean trajectory that is assumed to represent all trajectories in the sample or population.

The second type of longitudinal analysis is latent class trajectory analysis (LCTA).^1,12 LCTA has been used to study spousal depression after bereavement,¹⁹ childhood obesity,²⁰ and homelessness²¹ and in numerous areas of criminology such as hyperactivity and aggression²² and recidivism.²³ A key motivation behind the application of these methods is that population-averaging approaches may obscure substantive heterogeneity within a population, especially trajectories with qualitatively distinct developmental trends. LCTA does not assume an underlying normal distribution of population trajectories but instead assumes that there is a multinomial distribution of empirically distinct trajectories. (Hence the models are sometimes called “finite mixture models.”) Each estimated trajectory is viewed as being composed of a homogenous group of individuals sharing a developmentally distinct pattern of the outcome variable over time or age. Thus, the terms “trajectory” and “group” are sometimes used as synonyms in such models. Regardless of the language used, LCTA allows one to engage in the primary task of identifying the number and nature of discrete trajectories or groups in the data and determining which observed variables may be strong predictors of membership in a particular group.¹ This approach was chosen, because there is no a priori reason to expect that trajectories of depressive symptoms should be continuously and normally distributed about a common grand mean. The similarity and distinctiveness of trajectories (or lack thereof) is in itself an interesting empirical question.

A semiparametric Poisson²⁴ LCTA model was estimated using the software LatentGOLD (Version 4, Statistical Innovations, Inc., Belmont, MA). The Poisson variant of LCTA was chosen, because of the count nature of the dependent variable. The LCTA models used here can also be used for other types of dependent variables, including continuous, ordinal, and binomial.¹ Models identifying the number of trajectories of CES-D depressive symptoms were fit while including the following four predictors: age of caregiver at baseline; duration of caregiving at baseline (outlier values >95th percentile were recoded to the 95th percentile, which was 12 years); the care recipient’s score on the Behavior Rating Scale for Dementia²⁵ (a count of behaviors likely to be stressful for caregivers, measured as a time-varying covariate); and time, represented by the survey (waves 1–4). These measures were included, because they would be expected to influence the number of depressive symptoms present for caregivers, and their inclusion allowed for the estimation of the effect of these variables on the slope of CES-D depressive symptoms within each unique trajectory.

Models were first estimated using a different number of latent classes (first one, then two, and so on). Multiple iterations of each model were run using different starting values to reduce the chance of concluding that a model was wrongly identified as fitting the data the best. To do this, each iteration began with a different random starting estimate, and the consistency of final estimates was examined. Because the models with different numbers of latent classes are not nested, likelihood ratio statistics cannot be used to test for model improvement with additional classes; as such, the Bayesian Information Criterion was used to determine the number of trajectories present in the data.^1,12 After arriving at a final model solution, a final component of the analyses—assigning each individual to a group—was completed. Individuals were assigned to the trajectory group that most closely resembled their observed trajectory.¹

After trajectories were identified and individuals assigned to a group, the individuals who constituted each trajectory at each survey wave were compared using the following observed variables: caregiver demographic characteristics (education, race (African American, Asian, Latino, Native American, or White) and annual family income; care recipient characteristics (age and limitations in activities of daily living (range 0–7)); caregiving context, described according to the count of the tasks completed by caregivers for the care recipient, such as cooking, shopping, and dressing (range 0–22), and the total time spent (minutes per day) in direct provision of care; the number of hours per day that caregivers were responsible for their care recipients, including hands-on, direct care time; and two subjective caregiver measures (desire for more help in caregiving from family and friends and overall life satisfaction). Chi-square tests were used to compare categorical variables, and a Kruskal-Wallis test was used to compare continuous variables. These comparisons were designed to identify factors associated with belonging to a particular trajectory. A Bonferroni adjustment was used, because multiple tests were undertaken.

RESULTS

The mean age of wife caregivers at baseline was 69.9, and the mean age of the care-recipient husbands was 75.1. Approximately 82% of caregivers were white, 10% were African American, and the remaining caregivers were another race. The mean baseline income for couples was $23,369, and wives had been providing care an average of 4.5 years at that point. None of these variables were found to differ across the trajectories of depressive symptoms that are described below.

Four distinct trajectories of caregiver depressive symptoms were identified; fit statistics indicated that a 4-class model was most appropriate based on the failure to identify a unique solution for a 5-class model (different starting estimates led to different solutions) and the considerable improvement in the Bayesian (and Akaike) Information Criterion measure when moving from a 3- to a 4-class model (Table 1).¹ The four identified trajectories were characterized according to stability of observed and predicted levels of depressive symptoms within a given trajectory, meaning that the trajectories were approximately parallel to one another and did not intersect (Figure 1), although trajectories represented vastly different numbers of depressive symptoms. The overall CES-D sample mean is virtually identical in number of depressive symptoms and slope to the trajectory labeled as “moderate,” which contained approximately 34% (n =532, mean probability of assignment to the trajectory 0.75) of the sample and had mean CES-D at baseline of 6.2 (out of 20). This is lower than the number of depressive symptoms (8–9 out of 20)²⁶ found in past work to be indicative of clinical depression, although the other three trajectories differed greatly from the overall sample mean in terms of number of depressive symptoms.

Table 1.

Fit Statistics for Latent Class Trajectory Analysis Identifying Trajectories of Caregiver Depressive Symptoms

Class	Log Likelihood	Bayesian Information Criterion*	Akaike Information Criterion†	Parameters, n
1	−12,283.3	24,603.6	24,576.7	5
2	−9,815.9	19,712.9	19,653.8	11
3	−9,417.2	18,959.8	18,868.4	17
4	−9,360.0	18,889.7	18,766.0	23
5	MS	—	—	—
6	MS	—	—	—

^*A smaller value suggests a better model fit.

^†A smaller value with a unique solution suggests a better model fit.

MS = the presence of local maxima solutions suggesting that 5- or 6-class models are not appropriate. Local maxima occur when multiple solutions that are not unique can be found when testing for a given number of trajectories.

Figure 1.

Four trajectories of depressive symptoms according to the Center for Epidemiologic Studies Depression Scale (CES-D) compared with sample average.^*

^*The mean probability of class members belonging to that class was as follows: 0.87 for the high trajectory, 0.75 for moderate, 0.70 for low, and 0.80 for very low. Posterior probability of 0.70 or higher is indicative of a good model fit.^{22 †}Sample sizes were moderate (n= 532), high (n = 477), low (n = 348), and very low (n = 223).

The trajectory characterized by the largest mean number of depressive symptoms is labeled “high” (mean baseline CES-D, 11.9; mean probability of membership in trajectory, 0.87; n =477) and contained 30.2% of the sample. This level of depressive symptoms is well above the cutoff indicative of clinical depression.²⁶ The other two trajectories were named “low” (mean baseline CES-D, 3.0; mean probability of membership in trajectory, 0.70; n =348) and “very low” (mean baseline CES-D, 0.8; mean probability of membership in trajectory, 0.80; n =223).

At the end of the study, 715 (45.3%) of the 1,580 wife caregivers were still caring for their husbands. Approximately one-fifth of the caregiving dyads had been broken because of the death of the care recipient husband (n =364, 23.0%) or residence of the husband in a nursing home at Wave 4 (n =433, 27.4%). The remaining caregiving dyads (n =68, 4.3%) were lost to follow-up, or the caregiving wife died. A chi-square test marginally rejected the null hypothesis that the trajectories and the four follow-up states were independent (chi-square =16.9, degrees of freedom =9, P =.05, data not shown), although the differences in follow-up status across the trajectories at the end of the study were small.

In terms of the effects of the covariates included in the model, the Behavior Rating Scale for Dementia²⁵ had a statistically significant effect on the level of caregiver depressive symptoms within each trajectory. The effect of this measure was positive in all groups (data not shown), but the effect size was inversely related to the mean CES-D level of the latent trajectories; the “very low” group had the largest effect size (0.043), and the “high” group had the smallest effect size (0.012). A Wald test of the equality of the coefficients for this variable across the latent classes rejected the null hypothesis of no difference (P<.001), showing that the effect size varied across the groups. The effects of baseline age of the caregiver, duration of caregiving, and survey wave (1–4) on depressive symptoms within trajectories were small and not always statistically significant predictors of the slope of CES-D symptoms. Wald tests of the equality of the coefficients for these variables across the latent classes failed to reject the null hypothesis of no difference, showing that the variables did not have a significant effect on depressive symptoms across the trajectories.

Comparing Trajectories Using Observed Variables

Measures describing the care provided by wives and characteristics of wives and husbands were compared across trajectories (Table 2). Overall, there were few variables that differed across trajectories. For example, neither the number of tasks completed by caregivers for their husbands nor the minutes of care provided per day differed significantly according to trajectory. The amount of time each day that a wife was responsible for the care of her husband was significantly different at Wave 4, with those in the high trajectory reporting 15.1 hours per day, and those in the very low trajectory reporting 13.8 hours per day (P =.005).

Table 2.

Comparing Depressive Symptom Trajectories Using Characteristics of Wife Caregivers and Care Recipient Husbands (1999 and 2002)

Variable	High Trajectory		Moderate Trajectory		Low Trajectory		Very Low Trajectory		P-Value*
Caregiver wife
Care provided by wife, n (mean ± SD)
Caregiving tasks (0–22)†
Baseline	477	(15.5 ± 3.7)	532	(15.2 ± 3.7)	348	(15.0 ± 3.7)	223	(14.9 ± 3.6)	0.07
Wave 4	104	(15.2 ± 3.8)	136	(15.7 ± 3.6)	113	(15.5 ± 3.7)	60	(14.5 ± 3.6)	0.22
Time providing care, min/d
Baseline	453	(412 ± 252)	517	(392 ± 230)	335	(399 ± 225)	217	(373 ± 209)	0.34
Wave 4	74	(559 ± 323)	107	(509 ± 239)	87	(525 ± 264)	42	(523 ± 310)	0.89
Time responsible for husband, h/d
Baseline	474	(15.0 ± 4.4)	530	(14.6 ± 4.3)	347	(14.6 ± 4.5)	221	(14.7 ± 4.7)	0.28
Wave 4	101	(15.1 ± 4.6)	131	(15.2 ± 4.8)	111	(13.9 ± 4.3)	57	(13.8 ± 4.9)	.005
Subjective measures of wife
Very much wished she had more help caring for her husband, %‡
Baseline	475	37.3	529	24.0	346	19.9	222	12.6	<.001
Wave 4	103	38.5	134	24.3	113	15.9	60	9.4	<.001
Overall life satisfaction not very satisfying, %§
Baseline	472	39.6	531	14.3	346	2.6	222	2.3	<.001
Wave 4	104	44.2	136	6.6	112	0.0	60	0.0	<.001
Baseline demographics, n (mean ± SD)
Age of caregiver	477	(69.5 ± 7.8)	532	(69.9 ± 7.4)	348	(70.3 ± 6.8)	223	(70.1 ± 6.4)	0.71
Education of caregiver, years	477	(12.0 ± 2.5)	531	(11.9 ± 2.7)	348	(11.9 ± 2.7)	221	(12.6 ± 2.7)	0.03
Household income ($10,000)	457	(22.2 ± 11.1)	502	(23.8 ± 14.3)	329	(23.0 ± 12.9)	206	(25.4 ± 13.7)	0.07
Care recipient husband, n (mean ± SD)
Age at baseline	476	(74.7 ± 5.8)	529	(75.2 ± 5.9)	347	(75.4 ± 6.0)	223	(75.2 ± 6.3)	0.20
ADL limitations (0–7)||
Baseline	477	(3.0 ± 2.6)	532	(2.8 ± 2.5)	348	(2.8 ± 2.5)	223	(2.6 ± 2.6)	0.27
Wave 4	104	(3.2 ± 2.8)	135	(3.7 ± 2.6)	112	(3.4 ± 2.7)	60	(2.9 ± 2.8)	0.21

^*P-value is for chi-square tests for proportions and Kruskal-Wallis test for continuous variables.

^†Twenty-two tasks that caregivers could potentially do for their husbands such as dress them, bathe them, take them to medical appointments, and feed them.

^‡Caregivers were asked, “Do you wish that you had more help caring for your loved one from your family or friends?” Response ranged from 1 (very much) to 3 (not at all). Smaller number is associated with more burden for the caregiver.

^§Caregivers were asked, “On the whole, how satisfied are you with your life currently?” Responses ranged from 1 (not satisfying) to 3 (very satisfying). Smaller number is associated with more burden for the caregiver.

^||Count of the number of activity of daily living (ADL) items for which the husband needed help completing or that he could not do at all.

SD =standard deviation.

Two self-reported subjective measures (caregiving wives wishing that they had more help caring for their husbands and a measure of overall life satisfaction of the wife) were the only variables that differed statistically across trajectories at all four study waves. (Waves 1 and 4 only are shown in Table 2.) Caregivers in the high depressive symptoms trajectory were much more likely to report that their life was not very satisfying. (One hundred eighty-seven in the high depressive symptoms trajectory answered this way (39.6%), vs 76 for the moderate depressive symptoms trajectory, 9 for the low, and 5 for the very low, P<.001.) Similarly, persons in the high depressive symptoms trajectory were much more likely to report that they very much wished that they had more help in caring for their husbands (n=177, 37.3%) than those in the very low trajectory (n=28, 12.6%, P<.001, Table 2).

DISCUSSION

Longitudinal methods such as the one used to identify distinct trajectories can help improve understanding of the developmental trend of an outcome across time, as well as help identify covariates associated with those trends. Four distinct trajectories of caregiver depressive symptoms in wives caring for their husbands with dementia were identified. Although the overall average CES-D score (6.2/20) was below the cutpoint associated with clinical depression (8 or 9/20),²⁶ approximately one in three wife caregivers had a number of depressive symptoms throughout the study period consistent with clinical depression. The number of depressive symptoms varied greatly across the four trajectories, but the developmental pattern was stable (flat) within and across the groups over the 3-year study period. Rather than assuming a common trajectory across time for a sample, the statistical method employed in this study allows the data to “speak for itself” on that question; this is a primary strength of this method. This method is not necessarily advocated as a replacement for typical methods of analyzing longitudinal data but rather as an additional statistical tool that can be used to examine and uncover key trends present in longitudinal data.

There are several study limitations. First, a clinical measure of depression was not available, although the measure of depressive symptoms used has been found to be linked to greater risk of clinically significant depression.²⁶ Similarly, a measure of history of depression or depressive symptoms from earlier in the life course, which would be expected to be a strong predictor of depressive symptoms in later life, was not available. Other potential explanations of why wives exposed to similar circumstances had different trajectories of depressive symptoms could include (unmeasured) explanatory style,²⁷ social support, or respite care. All of these potential explanations are worth exploring. Second, the sample consisted only of caregiving wives, so it was not possible to investigate any sex differences that may exist. Third, the sample included dementia caregiving only. Although dementia is the most common reason for caregiving, the effect of caregiving for other reasons could differ in terms of effect on caregiver depressive symptoms.⁶ Fourth, only limited measures of caregiver stress and burden were available with which to provide insight into the types of caregiver wives who were members of the identified trajectories. Fifth, only four waves or observations of data were available; more-repeated measures always improve estimates. Finally, the study population consisted of elderly veterans only. Such a population is certainly distinct in many ways, but veterans have substantial racial and socioeconomic heterogeneity, so it is likely that the sample is more broadly representative of the characteristics of wives in the United States than are some samples previously used to study caregiving.

Depressive symptoms are only a public health problem to the extent that an increase in symptoms is associated with a greater risk of clinical depression. This has been found to be the case.²⁶ As the U.S. population ages and there are fewer younger persons for each elderly person, spousal caregiving is likely to become even more prevalent, perhaps increasing the risk of depression among caregivers. More work is needed to determine the need for and feasibility of screening for depressive symptoms in caregivers. Screening for caregiving and depressive symptoms has been undertaken in a rural primary care setting,²⁸ and the American Medical Association has developed a caregiver screening tool.²⁹ It was found that two simple questions were consistently correlated with membership in the highest depressive symptoms trajectory (lower life satisfaction and desire for more help from family and friends). Simple measures such as these when coupled with a straightforward question about caregiving and screening for depressive symptoms may help identify caregivers at risk of depression. This needs to be tested in divergent settings. In conclusion, analytical methods such as LCTA are promising tools for analyses of mood and other quantitative phenomena that change over time and may be fruitfully used as supplements to traditional longitudinal analysis methods.