Psychological Health Benefits of Companion Animals Following a Social Loss

Dawn C. Carr, PhD; Miles G. Taylor, PhD; Nancy R. Gee, PhD; Natalie Sachs-Ericsson, PhD


Gerontologist. 2020;60(3):428-438. 

In This Article


Data and Sample

Our data are drawn from the 2006–2014 waves of the Health and Retirement Study (HRS), a longitudinal panel study of adults over age 50 and their spouses in the United States that began in 1992 (see Juster & Suzman, 1995). We drew our sample based on completion of a 2012 experimental survey module focused on HAI that included questions about a range of issues related to CAs. In the HRS, experimental modules are completed by a random subsample of the larger HRS study. The vast majority of those with pets indicated having a cat or a dog. Given the small number of participants with other types of CAs (e.g., birds, fish, etc.), we excluded individuals with a pet other than dogs and cats from our analyses. In addition, given the research questions, we rely on data drawn from the psychosocial and lifestyle questionnaire, which is only given to a rotating half of the survey respondents every two waves (i.e., every 4 years). As a result, our data are structured as a pooled sample of individuals with baseline data determined based on the completion of the psychosocial questionnaire and Time 2 data collected 4 years later. For those who did not experience a major social loss, we rely on the most recent wave of data available. For those who experienced a loss, baseline data are based on the wave prior to the loss, and Time 2 is 4 years later.

Measures Defining Sample

CA ownership was determined from the HAI module using responses to questions regarding whether, and over what time period, individuals had a pet. Second, we identified a major social loss as anyone who (i) became divorced or (ii) became widowed. Our coding for social loss was limited by our sample size. It was not possible to examine different types of social losses separately, so our analysis combines these two types of major losses. Individuals who had a pet in the wave prior to reporting a loss were identified as a CA owner—with loss (loss/pet). Individuals who reported not having a pet for a minimum of 4 years prior to reporting a loss were identified as nonowners—with loss (loss/no-pet). We also included two "non-loss" control conditions. Specifically, these were individuals who were continuously married to the same individual over the course of the study period and who did not report any recent significant social loss (i.e., loss of a child, or parent). We divided non-loss participants by whether or not they had a CA over the study period (non-loss/pet, non-loss/no-pet). Loss of a child or a parent is also common in later life, but we do not examine these social losses because they could result in positive, neutral or negative effects on psychological health depending on the context of the loss. For example, individuals who provide direct care in their home to a parent or child are likely to have experienced greater social isolation and stress in association with the loss than individuals who experience the loss of a parent or child with whom they had very little interaction prior to the death. For these reasons, parental and child loss are not examined in this study. In addition, we excluded those in the "no-loss" groups who were exposed to a parental or child loss during the study period.

Our study uses two outcome measures—depressive symptoms and loneliness. Data about depressive symptoms are collected at every study wave among nonproxy respondents, and as noted earlier, loneliness is collected every other study wave. Based on the data structure and inclusion requirements, 437 individuals had complete data for depressive symptoms. A subset (75%) of this group also had complete data on loneliness (N = 332). Depressive symptoms were assessed 2 years before the social loss, and then 2 years after the social loss, resulting in a 4-year change score. A psychosocial questionnaire is completed on half the HRS sample each biennial wave, with the alternate half collected the following wave, and data have been collected every 4 years on a rotating basis since 2006. Due to the differences in timing of data collection for the psychosocial questionnaire, for loneliness, we included anyone in the sample with data on loneliness either one or two waves prior to the reported social loss, and again 4 years after the baseline loneliness assessment. Both outcome measures are a 4-year change score. In summary, the study includes four groups with sample sizes reported for depressive symptoms and loneliness, respectively: no-pet/no-loss (n = 130/102); loss/no-pet (n = 27/18); loss/pet (n = 43/30); and no-loss/pet (n = 237/182).

Outcome Measures

Depressive symptoms were assessed using a shortened version of the Center for Epidemiologic Studies Depression (CES-D) scale (Radloff, 1977). It comprised eight dichotomous measures based on whether respondents reported experiencing eight symptoms over the previous week: (i) felt depressed; (ii) everything was an effort; (iii) sleep was restless; (iv) was happy (reverse coded); (v) felt lonely; (vi) felt sad; (vii) could not get going; and (viii) enjoyed life (reverse coded). Our outcome measure for depressive symptoms is the 4-year change in number of depressive symptoms, controlling for baseline depressive symptoms. For participants who did not report a loss (i.e., no-pet/no-loss, pet/no-loss), we used the most recent data available.

For loneliness, we use a three-item composite measure located in the HRS psychosocial questionnaire which has been validated in the broader HRS (Hughes, Waite, Hawkley, & Cacioppo, 2004). Following Hughes and colleagues (2004), loneliness is calculated as the sum of three indicators, with response options of "1" (often), "2" (some of the time), and "3" (hardly ever or never). These indicators were how often do you feel (i) you lack companionship; (ii) left out; and (iii) isolated from others? Each item was reverse coded such that higher numbers indicate greater loneliness. Individuals who did not respond to all three questions were coded missing. Our outcome measure is a 4-year change in loneliness score, controlling for baseline (pre-loss) loneliness. For individuals without a loss (i.e., no-pet/no-loss, pet/no-loss), we used the most recent data available.

Control Measures

Our inclusion of control measures is informed by previous research identifying key factors associated with selection into CA ownership and nonownership (Carr, Taylor, et al., 2018). We use propensity score methods (described later) to account for selection factors and isolate the effect of CAs on psychological health at the time of a social loss. The methodology, inverse-probability weighted regression, incorporates two models: a propensity model, which estimates selection into CA ownership/nonownership, and a propensity adjusted regression model, which is used to predict the outcome measure. The measures used in both models are given in Table 1, with indicators to show inclusion in the propensity models and regression models.

With respect to physical health, we included three measures: functional limitations, self-rated health, and chronic pain. Functional limitations is a sum of six self-reported items, indicating the number of physical tasks an individual reports they were unable to perform: walking one block, climbing one flight of stairs, stooping or kneeling, lifting or carrying 10 pounds, picking a dime up off the ground, and pushing/pulling a large object (for further description, see Kail & Carr, 2017). Self-rated health is based on a five-point score ranging from "1" (poor health) to "5" (excellent health). Chronic pain is based on a positive response to the following question: "Are you often troubled with pain?"

We included five demographic control variables: gender, age, race, educational attainment, and suburban area. Women is a measure of gender, with women coded as "1." Age was measured as a continuous variable, ranging from 37 to 88. Due to sample size and the requirements of the propensity models, it was only possible to measure race as a dichotomous indicator of whether an individual was Non-Hispanic White (1) or a minority (0). Lives in a Suburban area is measured using the Beale Rural–Urban Continuum code. Individuals who lived in an area that included 250,000–1 million individuals were coded "1." Because our data were stacked based on timing of social loss, completion of the psychosocial questionnaire, and completion of the HAI module, there are three different baseline wave years in our study. We included a measure to account for whether the baseline year was 2010, which represents the majority of our sample.

We included two dichotomous measures related to active engagement. Works for pay is a measure indicating individuals engaged in paid work at baseline. Volunteers capture a positive response to the question: "Have you spent any time in the past 12 months doing volunteer work (coded "1") for religious, educational, health-related or other charitable organizations?" at baseline.

Based on our previous research findings (Carr, Taylor, et al., 2018), we included measures for personality from 26 items to calculate Big 5 traits (e.g., McCrae & Costa, 1997). Individuals are asked: "Please indicate how well each of the following describes you" with a list of factors such as reckless, moody, caring, nervous, broad-minded, etc. For each, individuals can respond: not at all, a little, some, or a lot. For more details, see Smith, Ryan, Fisher, Sonnega, and Weir (2017). We included three of the five domains: neuroticism, extroversion, and conscientiousness. These measures were assessed at either of the two waves prior to the potential social loss. For individuals without data in those two waves, we used personality measures drawn from earlier waves if available.

A series of additional measures were considered, but for the sake of parsimony, ultimately excluded in the final models because they were not significant and did not mediate or confound the relationship between pet–loss status and psychological health. These included health behaviors (e.g., exercise, alcohol, smoking), number of household members, a range of other health measures (e.g., number of chronic conditions, activities of daily living, and instrumental activities of daily living limitations), and grandparent status and time providing care.

Statistical Analysis

Although a randomized control study is the gold standard for identifying causal effects, it would be difficult to use this approach to study pets and social losses. Propensity methods seek to simulate the random assignment used in a randomized control study. Consistent with volunteering research, which faces similar selection problems (e.g., Carr, Kail, & Rowe, 2018), we use a type of inverse-probability weighted regression that allows us to compare multiple treatment groups at once (see Imbens & Rubin, 2010). Inverse-probability weighting calculates "potential outcomes" for an individual as though they were in a different pet–loss group (Abadie & Imbens, 2012a). Considered a doubly robust property of this approach, it calculates unbiased treatment effect estimates even when the propensity model or the outcome model (but not both) is miss-specified (Wooldridge, 2007). When there are more than two treatment groups, multivariate logits are used to generate the propensity calculations given a set of selection variables (Cattaneo, 2010). Specifically, we tested four pet-loss groups as noted earlier: pet/no-loss, pet/loss, no-pet/no-loss, and no-pet/loss.

To carry out our analyses, we first adjusted for the propensity of being in each of these treatment groups. As mentioned, the propensity model is the first part of this approach, followed by a regression model of the outcome that is adjusted for the propensity estimates. This involves the calculation of a potential outcome mean (POM), which is the predicted outcome given the propensity to select into each treatment group relative to all other groups (Abadie & Imbens, 2012b). Based on the POMs, an average treatment effect (ATE) was calculated, determined by the difference in the outcome between each treatment group and each assigned control group (e.g., pet/no-loss, pet/loss, no-pet/no-loss, and no-pet/loss). This allowed us to assess statistical differences between each group in relation to all alternative pet/loss groups. In sum, the adjusted regression predicts the outcome (i.e., change in depressive symptoms/loneliness), controlling for observable factors that influence the outcome measure while adjusting for selection into the pet/loss groups.

Change in depressive symptoms/loneliness is calculated by determining the difference between baseline (wave t) and 4 years after baseline (wave t + 2), whereby a larger number indicates more depressive symptoms/higher loneliness. After calculating the average change in psychological health for each treatment and the associated error (based on adjustment from the propensity model), we report the ATEs for each group relative to each alternative "control" group, reporting the average change in depressive symptoms and loneliness and statistical differences between each group. The ATEs are interpreted as the difference between the expected effect for being in a different pet–loss group relative to the average effect associated with being in the assigned group.