The mediating effect of engagement in physical activity over a 24-hour period on chronic disease and depression: Using compositional mediation model

Abstract

Background:

Popular mediation method only considers a single activity as a mediator instead of all 24-hour physical activity, such as a four-way decomposition method. We investigated the mediation of 24-hour movement continuum between chronic diseases (cardiovascular disease [CVD] and diabetes) and depression using a novel compositional mediation model.

Methods:

24-hour activity data measured by accelerometer were obtained from NHANES 2005–2006. Adjusted coefficient with 95% confidence interval (95% CI) for PHQ-9 total score and adjusted odds ratio (OR) with 95% CI for depression were computed from compositional mediation models.

Results:

In total, 2,375 participants aged ≥ 20 were included in our analysis. Both diabetes and CVD were associated with increased sedentary behavior (SB) and sleep and reduced moderate-to-vigorous physical activity (MVPA) and light-intensity physical activity (LPA), leading to an increased likelihood of depression. Although not all component indirect effects were associated with PHQ-9 total score and depression, the total indirect effect was significantly associated with both PHQ-9 total score (coefficient (95% CI) for diabetes: 0.162 (0.081, 0.261); coefficient (95% CI) for CVD: 1.139 (1.061, 1.240)) and depression (OR (95% CI) for diabetes: 0.235 (0.126, 0.362); OR (95% CI) for CVD: 1.200 (1.088, 1.346)) using the compositional mediation model.

Conclusion:

We developed a compositional mediation model for continuous and binary outcomes, which can handle entire compositional mediators as a unit. The mediation of 24-hour movement continuum mediated the association between diabetes, CVD, and depression. Our findings present potential interventions for reducing risk of depression among patients with CVD and diabetes.

1. Introduction

Depression is a common mood disorder that adversely affects daily life and contributes to poor physical health (Centers for Disease Control Prevention, 2010; Cho et al., 2019). The incidence of depression has increased markedly from 1990 to 2017 (Liu et al., 2020). The growing burden of chronic diseases has contributed to the rise in incidence of depression, and subsequent poorer quality of life (Cho et al., 2019; Hidaka, 2012; Simon, 2001). Therefore, it is of public health importance to identify depression risk factors and develop and effective low-cost preventative strategy.

Physical activity was suggested to influence chronic disease in relation to depression. Previous studies have reported that people with chronic diseases tend to engage in less physical activity, including moderate-to-vigorous physical activity (MVPA) and light-intensity physical activity (LPA), as well as more likely to engage in sedentary behavior (SB) (Barker et al., 2019; Sobngwi et al., 2002). Physical activity reduces the risk of depressive symptoms in all age groups, while SB increases the likelihood of depression (Dimeo et al., 2001; Doyne et al., 1983; Huang et al., 2021; McNeil et al., 1991). Several studies have reported that replacing SB with MVPA or LPA can significantly reduce depression (Kandola et al., 2020; Tully et al., 2020; Yasunaga et al., 2018a).

Epidemiological studies have observed that moderate recreational activity (MRA) mediates the effects of cardiovascular disease (CVD) and diabetes on depression using a four-way decomposition method (Huang et al., 2021; Xu et al., 2020). However, this approach considers a single activity as a mediator instead of measuring the full 24-hour physical activity. Therefore, by accounting for the full 24-hour activity as compositional data, changing any one activity will cause changes to other activities. For instance, within a fixed 24-hour period, increase in SB will lead to a compensatory decrease in the time of one or more other activities (MVPA, LPA or sleep). To better understand the mediating effects of physical activity and SB on the relationship between chronic disease and depression from a movement continuum perspective, it is necessary to consider the 24-hour activity as a unit in the mediation analysis.

In this study, we developed a compositional mediation model for both continuous and binary outcomes to deal with mediators that are compositional data, and explored the 24-hour movement continuum as a mediator for CVD and diabetes in association with depression. Findings will be helpful in proposing early interventions to reduce risk for depression among people with chronic diseases (CVD and diabetes).

2. Method

2.1. Data source and ethnic statement

The present study used data from the National Health and Nutrition Examination Survey (NHANES), a nationwide survey of adults and children in the U.S. NHANES is a cross-sectional study, where new samples are collected every two years. The study aims to monitor health and nutritional status, and represents the non-institutionalized population of the U.S. (National Center for Health Statistics, 2008). NHANES is a publicly available dataset approved by the National Center for Health Statistics institutional review board, where all included participants provided written informed consent. The present study used NHANES 2005–2006 data and included adults ≥ 20 years of age that wore an accelerometer. A total of 2375 participants (weighted=110,193,691) were included in our analysis.

2.2. Outcome measures

The Patient Health Questionnaire (PHQ-9) is a 9-item depression screening that asked about the frequency of depressive symptoms in the past two weeks to measure depression symptoms (trouble falling sleep, feeling down, overeating, feeling tired, little interest, feeling bad about yourself, trouble concentrating on things, moving or speaking so slowly that other people could have noticed, and thoughts of hurting yourself) (Kroenke et al., 2001). Each PHQ-9 question scored from 0 to 3, and the total PHQ-9 score ranged from 0 to 27. Depression was defined as PHQ-9 total score ≥5 recommend by Janssen et al. (2016) (Janssen et al., 2016). Although the National Quality Forum recommends 10 as the threshold, the threshold of 5 has the best sensitivity (92.3%) and acceptable specificity (70.4%), which is suitable for early intervening (Janssen et al., 2016; Kroenke et al., 2001). We used both PHQ-9 total score and depression as the outcomes in the mediation analysis.

2.3. Mediator – 24-hour physical behavior

Physical activity and sedentary behavior were assessed by using a uniaxial accelerometer (AM-7164, ActiGraph, LLC, Pensacola, Florida) for seven consecutive days. Each participant in the study was asked to wear an accelerometer in the morning, except for bathing or engaging in water activities (e.g., swimming). Participants that wore an accelerometer for a minimum of ten hours on any four days over a seven-day period were included in our study. Using data captured from the accelerometer and based on the definition of Theou and colleagues (2017), SB was defined as 0 to <100 counts/minute, LPA was defined as 100 to <2020 counts/minute, and MVPA was defined as ≥ 2020 counts/minute (Theou et al., 2017). Estimated mean time in minutes spent in each evaluation intensity per day was calculated by averaging all measures over the number of valid days. In order to avoid short periods having nothing to do with bedtime and long period that might be invalid data, we used the duration of the longest non-wearing time ranging from four to twelve hours in a 24-hour window as sleep activity (Urbanek et al., 2018). The time of each movement behavior in a 24-hour period was the ratio of each behavior (SB, LPA, MVPA, sleep) multiplied by 1440.

2.4. Assessment of diabetes

Diabetes was determined based on participant’s self-report and blood examination. Adhering to the American Diabetes Association, diabetes was defined if at least one of the following criteria were met: (1) fasting glucose concentration ≥ 126 mg/dL; (2) Glycosylated hemoglobin (HbA1c) ≥ 6.5%; (3) participant answered yes to the question “are you now taking diabetes medication to lower your blood sugar”; (4) participant answered yes to the question “have you ever been told by a doctor that you have diabetes”; and (5) participant answered yes to the question “are you taking insulin now.”

2.5. Assessment of cardiovascular disease

CVD was determined by self-reported questionnaire. Based on the American Heart Association, CVD was defined if at least one of following criteria were met: (1) participant answered yes to the question “have you ever been told you had congestive heart failure”; (2) participant answered yes to the question “have you ever been told you had coronary heart disease”; (3) participant answered yes to the question “have you ever been told you had angina/angina pectoris”; (4) participant answered yes to the question “have you ever been told you had a heart attack”; and (5) participant answered yes to the question “have you ever been told you had a stroke”.

2.6. Assessment of covariates

Included in our analysis as covariates were gender, age, race/ethnicity (Hispanic, White, Black, and other racial group), education level (less than 12 years, grade 12 or equivalent, and post high school), marital status (never married, married or living with partner, and others), general health condition (excellent, very good, good, fair, and poor), family poverty index ratio (a ratio of family income to poverty), smoking status (nonsmoker, former smoker, and current smoker), alcohol intake (no alcohol consumption, previous alcohol consumption, moderate alcohol consumption, and excessive alcohol consumption), body-mass index (BMI) was categorized based on the World Health Organization criteria(underweight if BMI<18.5, normal weight if BMI between 18.5 and 24.9, overweight if BMI between 25 and 29.9, and obese if BMI ≥ 30), and total energy used per day.

2.7. Statistical analysis

2.7.1. Demographic description

We applied weights to calculate the average total PHQ-9 score and the prevalence of depression corresponding to each influencing factor. We also compared the 24-hour activities (MVPA, LPA, SB, and sleep) of CVD and non-CVD, diabetic and non-diabetic, and depression and non-depression participants.

2.7.2. Isometric log-ratio (ILR) and variation matrix

Let $M$ be a vector of m compositional mediators (24-hour activities in our study were mediator variables and treated as compositional data). Since $M$ is a $m$-simplex space, that is, $M=\left\{\left(M_1,\dots ,M_m\right):M_i>0,i=1,2,\dots ,m,\sum _{i=1}^m{M}_i=1\right\}$, modeling directly to $M$ may produce distorted results (Aitchison, 1982). To address this issue, isometric log-ratio (ilr) transformation technique which transform the compositional data from simplex $S^m$ to real Euclidean space $R^{m-1}$ was introduced by Egozcue et al. (2003).

Let denote ilr transformation coordinates by Z, that is, $Z=\left(Z_1,Z_2,\dots ,Z_{m-1}\right)$. For coordinate i,

$$Z_i=\sqrt{\frac{m-i}{m-i+1}}\ln \frac{M_i}{{\left({\prod }_{j=i+1}^m{M}_j\right)}^{\frac{1}{m-i}}}$$

The benefit of ilr is the ability to be expressed and interpreted as the average of pairwise log difference, that is,

$$\begin{gathered} Z_i=\sqrt{\frac{m-i}{m-i+1}}\text{ln}\frac{M_i}{{\left({\prod }_{j=i+1}^m{M}_j\right)}^{\frac{1}{m-i}}}=\sqrt{\frac{1}{(m-i+1)(m-i)}}\sum _{j=i+1}^m\text{ln}\left(\frac{M_i}{M_j}\right) \\ =\sqrt{\frac{1}{(m-i+1)(m-i)}}\sum _{j=i+1}^m\left(\text{ln}(M_i)-\text{ln}\left(M_j\right)\right) \end{gathered}$$

It is unable to construct directly for original compositions to measure variability of compositional data. Therefore, Aitchison introduced a variation matrix which is the variance of pairwise Log ratios (Aitchison, 1982). Let the compositional data $M=\left\{\left(M_1,\dots ,M_m\right):M_i>0,i=\mathrm{1,2},\dots ,m,\sum _{i=1}^m{M}_i=1\right\}$, the variation matrix is defined as

$$V=\left[\begin{array}{ccc}{v}_{11}& \cdots & v_{1m}\\ ⋮& \ddots & ⋮\\ v_{m1}& \cdots & v_{mm}\end{array}\right]$$

where $v_{jk}$ are sample variances of pairwise logratios between $M_j$ and $M_k$ for all $\mathrm{j},k=\mathrm{1,2},\dots ,\mathrm{m}$, and

$$v_{jk}=\mathit{Var}\left(\mathit{ln}\frac{M_j}{M_k}\right)$$

The closer $v_{jk}$ is to zero, the stronger the proportional relationship or co-dependence between $M_j$ and $M_k$

In this study, the time spent in SB, LPA, MVPA, and sleeping were converted to z, a three-dimensional coordinate using ILR, where

$$z=\left(\sqrt{\frac{3}{4}}\mathrm{l}\mathrm{n}\frac{MVPA}{{\left({\mathit{Sleep}}^{\mathrm{*}}S{B}^{\mathrm{*}}LPA\right)}^{\frac{1}{3}}},\sqrt{\frac{2}{3}}\mathrm{l}\mathrm{n}\frac{LPA}{{\left({\mathit{Sleep}}^{\mathrm{*}}SB\right)}^{\frac{1}{2}}},\sqrt{\frac{1}{2}}\mathrm{l}\mathrm{n}\frac{\mathit{Sleep}}{SB}\right)$$

Additionally, variation matrix reflecting co-dependence between SB, LPA, MVPA, and sleep in the general population, diabetes population, and CVD population were calculated. For 24-hour activities, two activities with strong co-dependence means that it may be easy to convert between each other.

2.7.3. Compositional mediation model

Suppose we have a random sample size of $n$ from a population. Let $Y$ be an outcome, $M$ be a vector of m compositional mediators, $T$ be an exposure, and $X$ be a set of covariates that may affect the mediators,exposure, and outcome. Based on the ILR calculation we mentioned above, we proposed the compositional mediation model in two cases:

Case 1: If Y is continuous outcome, the steps of the compositional mediation model are as follow:

Step 1: Building linear regression Y on treatment T and ilr coordinates Z

$$Y={\beta }_0+\sum _{i=1}^{m-1}{\beta }_i{Z}_i+cT+\gamma X+\epsilon$$

where ${\beta }_0$ is the intercept, ${\beta }_i$ is the coefficient of $Z_i$, c is the coefficient of $\mathrm{T},\gamma$ is the vector of coefficient of X, and $\epsilon$ is an error term. By replacing $Z_i$, the equation will be simplified to

$$Y={\beta }_0+\sum _{i=1}^m{\beta }_i^{\prime }\text{ln}(M_i)+cT+\gamma X+\epsilon$$

where

$${\beta }_i^{\prime }=\left\{\begin{array}{c}{\beta }_i\sqrt{\frac{(m-i)}{(m-i+1)}}-min(i-\mathrm{1,1})\sum _{j=1}^{i-1}{\beta }_j\sqrt{\frac{1}{(m-j+1)(m-j)}},1\le i\le \mathrm{m}-1\\ -\sum _{j=1}^{m-1}{\beta }_j\sqrt{\frac{1}{(m-j+1)(m-j)}},i=\mathrm{m}\end{array}\right.$$

Step 2: Building linear regression $M_i$ on treatment T

$$\mathrm{l}\mathrm{n}\left(M_i\right)={\alpha }_{0i}+{\alpha }_{1i}T+{\gamma }_{i}X+{\epsilon }_i$$

where ${\alpha }_{0i}$ is the intercept, ${\alpha }_{1i}$ is the coefficient of T, ${\gamma }_i$ is the vector of coefficient of X, and ${\epsilon }_i$ is an error term.

Based on Step 1 and Step 2, we get the indirect effect of treatment T on Y through $M_i$ is ${\alpha }_{1i}{\beta }_i^{\prime }$. Further, the average indirect effect (AIE) and average total effect (ATE) are ${\sum }_{i=1}^m{\alpha }_{1i}{\beta }_i^{\prime }$ and ${\sum }_{i=1}^m{\alpha }_{1i}{\beta }_i^{\prime }+c$, respectively.

Case 2: If Y is binary outcome, the steps of the compositional mediation model are as follow:

Step 1: Building logistic regression Y on treatment T and ilr coordinates Z

$$(\text{Odd}(\text{Y}))={\beta }_0+\sum _{i=1}^{m-1}{\beta }_i{Z}_i+cT+\gamma X$$

where ${\beta }_0$ is the intercept, ${\beta }_i$ is the coefficient of $Z_i$, c is the coefficient of T, and $\gamma$ is the vector of coefficient of X. By replacing $Z_i$, the equation will be simplified to

$$(\text{Odd}(\text{Y}))={\beta }_0+\sum _{i=1}^m{\beta }_i^{\prime }\text{ln}(M_i)+cT+\gamma X$$

where

$${\beta }_i^{\prime }=\{\begin{array}{l}{\beta }_i\sqrt{\frac{(m-i)}{(m-i+1)}}-min(i-1,1)\sum _{j=1}^{i-1}{\beta }_j\sqrt{\frac{1}{(m-j+1)(m-j)}},1\le i\le \text{m}-1\\ -\sum _{j=1}^{m-1}{\beta }_j\sqrt{\frac{1}{(m-j+1)(m-j)}},i=\text{m}\end{array}$$

Step 2: Building linear regression $M_i$ on treatment T

$$\mathrm{l}\mathrm{n}\left(M_i\right)={\alpha }_{0i}+{\alpha }_{1i}T+{\gamma }_{i}X+{\epsilon }_i$$

where ${\alpha }_{0i}$ is the intercept, ${\alpha }_{1i}$ is the coefficient of T, ${\gamma }_i$ is the vector of coefficient of X, and ${\epsilon }_i$ is an error term.

From Step 1 and Step 2, we get the indirect effect odds ratio (OR) of treatment T on Y through $M_i$ is $e^{{\alpha }_{1i}{\beta }_i^{\prime }}$. Further, the average indirect effect OR (AIEOR) and average total effect OR (ATEOR) are $e^{{\sum }_{i=1}^m{\alpha }_{1i}{\beta }_i^{\prime }}$ and $e^{{\sum }_{i=1}^m{\alpha }_{1i}{\beta }_i^{\prime }+c}$, respectively.

In a compositional mediation model, we aimed to determine the effect of chronic disease (diabetes or CVD) on PHQ-9 total score or depression mediated by daily physical activity, after adjusting for gender, race/ethnicity, education, marital status, smoking status, alcohol consumption, family poverty index ratio, total energy used per day, health condition, and BMI. Adjusted regression coefficient with 95% confidence interval (95% CI) for PHQ-9 total score and adjusted odds ratio (OR) with 95% CI for depression were computed. All statistical tests were 2-tailed, and analyses were performed using R software version 4.0.3.

3. Results

Table 1 shows the average total PHQ-9 score and the prevalence of depression according to influencing factors. The overall weighted prevalence of depression was 16.98%. There were significant associations for average PHQ score and the prevalence of depression with gender, smoking status, educational attainment, health condition, BMI, diabetes, and CVD. Additionally, the prevalence of depression was significantly associated with racial/ethnic group and marital status.

Table 1.

Average total PHQ-9 score and prevalence of depression according to influencing factors.

	Overall Weighted N Use sample size instead of population size	Total PHQ-9 Score MEAN (SE)	p-value	Depression% (SE)	p-value
Gender
Male	55,114,947	2.00 (0.10)	<0.001	42.43% (2.82%)	0.005
Female	55,078,744	2.67 (0.12)		57.57% (2.82%)
Age (years)
20–44	50,065,438	2.37 (0.12)	0.066	49.50% (2.94%)	0.133
45–64	35,836,689	2.53 (0.17)		32.46% (2.82%)
65+	24,291,564	1.96 (0.12)		18.04% (1.97%)
Race
Hispanic	12,609,721	2.83 (0.24)	0.853	14.54% (1.77%)	0.037
Non-Hispanic White	80,832,170	2.21 (0.10)		67.57% (2.54%)
Non-Hispanic Black	11,176,106	2.54 (0.15)		11.10% (1.22%)
Others	5575,693	2.59 (0.35)		6.79% (1.57%)
Education
Less than year 12	16,201,143	2.82 (0.19)	0.003	20.97% (2.10%)	0.001
Grade 12 or equivalent	27,831,284	2.43 (0.18)		26.27% (2.53%)
Post high school	66,161,264	2.17 (0.10)		52.76% (2.89%)
Marital status
Never married	14,617,521	2.82 (0.22)	0.347	17.24% (2.11%)	0.003
Married or living with partner	76,389,002	2.07 (0.09)		61.32% (2.84%)
Widowed, divorced, or separated	19,187,168	2.98 (0.25)		21.44% (2.36%)
Smoking
Nonsmoker	58,279,619	2.16 (0.11)	0.001	45.62% (2.91%)	<0.001
Former smoker	29,329,699	2.15 (0.14)		24.31% (2.52%)
Current smoker	22,584,373	3.00 (0.21)		30.07% (2.62%)
Alcohol intake
No alcohol consumption	11,987,822	2.02 (0.22)	0.200	8.96% (1.60%)	0.125
Previous alcohol consumption	26,853,102	2.52 (0.17)		27.31% (2.50%)
Moderate alcohol consumption	54,120,951	2.16 (0.11)		45.14% (2.90%)
Excessive alcohol consumption	17,231,816	2.78 (0.23)		18.60% (2.28%)
Health condition
Excellent	12,279,658	1.02 (0.13)	<0.001	3.46% (1.07%)	<0.001
Very good	40,530,484	1.65 (0.10)		21.29% (2.48%)
Good	41,321,937	2.35 (0.12)		38.16% (2.84%)
Fair	14,414,952	4.53 (0.29)		30.70% (2.65%)
Poor	1646,660	9.06 (1.29)		6.39% (1.36%)
BMI
Underweight	1695,111	2.23 (0.43)	0.001	1.76% (0.70%)	<0.001
Normal	33,815,631	2.08 (0.14)		24.28% (2.55%)
Overweight	36,537,230	2.14 (0.13)		28.82% (2.54%)
Obesity	38,145,720	2.74 (0.16)		45.13% (2.91%)
Diabetes
Yes	10,965,701	3.05 (0.27)	0.006	86.61% (1.84%)	0.019
No	99,227,990	2.25 (0.08)		13.39% (1.84%)
CVD
Yes	101,996,722	3.10 (0.33)	0.014	88.87% (1.66%)	0.004
No	8196,969	2.27 (0.08)		11.13% (1.66%)

Note: p ≤ 0.05 was considered statistically significant.

Significant difference between groups determined by X² test (depression) or ANOVA-test (PHQ-9 score).

Abbreviations: CVD, cardiovascular disease; BMI, Body-mass index; SE, standard error; PHQ, patient health questionnaire.

Table 2 presents physical activity by diabetes status, CVD, and depression. We observed that time spent engaging in SB, LPA, MVPA, and sleep were significantly different between diabetes and non-diabetes groups. Participants with diabetes had less MVPA and LPA, and more SB and sleep in a 24-hour period compared with non-diabetic participants. SB, LPA, MVPA, and sleep were similarly associated with CVD and depression.

Table 2.

A: Engagement in physical activity by diabetes status.

	Diabetes	Non-diabetes	p-value
Sedentary behavior	39.33% (0.61%)	35.90% (0.21%)	<0.001
Light-intensity physical activity	22.94% (0.50%)	26.26% (0.18%)	<0.001
Moderate-to-vigorous physical activity	1.25% (0.10%)	2.34% (0.05%)	<0.001
Sleeping	36.47% (0.38%)	35.45% (0.14%)	0.012
Table 2B Engagement in physical activity by cardiovascular disease status
	CVD	Non-CVD	p-value
Sedentary behavior	40.73% (0.68%)	35.88% (0.21%)	<0.001
Light-intensity physical activity	20.92% (0.54%)	26.33% (0.18%)	<0.001
Moderate-to-vigorous physical activity	1.23% (0.16%)	2.37% (0.05%)	<0.001
Sleeping	37.12% (0.38%)	35.42% (0.14%)	<0.001
Table 2C. Engagement in physical activity by depression status
	Depression	Non-depression	p-value
Sedentary behavior	37.32% (0.50%)	36.02% (0.21%)	0.017
Light-intensity physical activity	24.52% (0.41%)	26.22% (0.19%)	<0.001
Moderate-to-vigorous physical activity	1.80% (0.10%)	2.38% (0.05%)	<0.001
Sleeping	36.36% (0.33%)	35.38% (0.14%)	0.007

Fig. 1 shows the variation matrix between SB, LPA, MVPA, and sleep in the general population, diabetes population, and CVD population. We observed that the co-dependence between SB and LPA or the co-dependence between sleep and LPA were stronger than co-dependence between SB and MVPA or the co-dependence between sleep and MVPA in any population. In addition, the co-dependence between LPA and MVPA was weak.

Fig. 1.

A: The variation matrix between SB, LPA, MVPA, and sleep – All Population. B: The variation matrix between SB, LPA, MVPA, and sleep – Diabetes Population. C: The variation matrix between SB, LPA, MVPA, and sleep – CVD Population.

Fig. 2 presents the effects of diabetes on the total PHQ-9 score and the prevalence of depression mediated by daily physical activity. We observed that diabetes increased the time spent engaging in SB and sleep, and reduce time spent on MVPA and LPA, thereby increasing depression and PHQ-9 total score. Although the indirect effect through sleep was not significant on PHQ-9 total score, and the indirect effects through SB, LPA, and sleep were not significant on depression, the total indirect effect was significantly associated with both PHQ-9 total score (coefficient = 0.162, 95% CI 0.082–0.261) and depression (adjusted OR = 1.139, 95%CI 1.061–1.240) (Table 3).

Fig. 2.

A: presents the effect of diabetes to PHQ-total score mediated by daily engagement in physical activity. B: presents the effect of diabetes to depression mediated by daily engagement in physical activity.

Table 3.

The effects of diabetes or CVD on PHQ-total score and depression mediated by daily engagement in physical activity.

	Diabetes on PHQ score Coefficient (95% CI)	Diabetes on Depression OR (95% CI)	CVD on PHQ score Coefficient (95% CI)	CVD on Depression OR (95% CI)
Total Indirect Effect	0.162 (0.081,0.261)	1.139 (1.061,1.240)	0.235 (0.126,0.362)	1.200 (1.088,1.346)
Direct Effect	−0.040 (−0.563,0.499)	0.890 (0.556,1.395)	−0.139 (−0.767,0.520)	1.015 (0.612,1.649)
Total Effect	0.121 (−0.395,0.655)	1.014 (0.639,1.569)	0.096 (−0.531,0.754)	1.217 (0.736,1.971)

Fig. 3 shows the effects of CVD on total PHQ-9 score and depression mediated by daily physical activity. We observed that CVD significantly increased the time spent engaging in SB and sleep, and reduced time spent for MVPA and LPA, thereby increasing depression and PHQ-9 total score. Similarly, the indirect effect through sleep was not associated with total PHQ-9 score, the indirect effect through SB, LPA, and sleep were not associated with depression, but the total indirect effect was associated with both total PHQ-9 score (coefficient = 0.235, 95% CI 0.126–0.362) and depression (adjusted OR = 1.200, 95% CI 1.088–1.346) (Table 3).

Fig. 3.

A: presents the effect of CVD to PHQ-total score mediated by daily engagement in physical activity. B: presents the effect of CVD to depression mediated by daily engagement in physical activity.

4. Discussion

In the present study we introduced a compositional mediation model for continuous and binary outcomes, which is the extension of mediation model for dealing with a situation when the mediators are compositional data. Although previous research tried to use the four-way decomposition method to analyze each compositional data feature as a mediator (Discacciati et al., 2019; VanderWeele, 2014; Xu et al., 2020), to the best of our knowledge, this is the first study to include all compositional data features as mediators in a single mediation analysis. Compared with the four-way decomposition method, our method includes the entire compositional mediators as a unit, that is, it jointly estimates the effect of exposure on whole compositional mediators, instead of each mediator separately, thereby making the indirect and direct effects more accurate.

After adjusting for potential confounders, we observed that diabetes and CVD increased the time spent engaging in SB and sleep, and compensatorily decreased the time spent in LPA and MVPA. Our findings are consistent with result from previous studies that individuals with a chronic condition, such as diabetes or CVD, are less likely to engage in physically activity (Durstine et al., 2013; Sobngwi et al., 2002). For instance, previous studies reported that older adults with diabetes engaged in less physical activity and more SB than those without diabetes (Loprinzi, 2014), and people with CVD preferred to increase SB and reduce any type of physical activity (Xu et al., 2020). There are two possible reasons for these associations: 1) people with chronic diseases are at greater risk for mental disorders, but mental disorders can affect the homeostatic system involved in the stress response, resulting in reduced physical activity (Campbell and Turner, 2018; Correll et al., 2017; Verhaak et al., 2005); and 2) some chronic diseases are accompanied by physical disabilities, resulting in greater engagement in SB and lower physical activity. However, previous studies have examined separately the effects of chronic diseases on physical activity and SB (Huang et al., 2021; Xu et al., 2020), ignoring the complex nature of the interrelationship between chronic disease and both physical activity and SB. Our findings from 24-hour physical activity behavior as a unit revealed that having a chronic disease reduced physical activity and increased SB from the perspective of time usage throughout the day. Behavior changes are interchangeable, in addition to the increase in SB directly caused by chronic diseases, the decrease in physical activity caused by chronic diseases also compensatory increases SB, and vice versa. Our results suggest that compared with non-diabetics, the average daily SB time of patients with diabetes increased from 35.90% to 39.33% (an increase of 3.43%, approximately 50 min), while the average LPA and MVPA time decreased from 28.60 a day to 24.19 (reduced by 4.41%, approximately 64 min). CVD behavior time changes was more severe, SB time on average increased from 35.88% a day to 40.73% (an increase of 4.85%, approximately 70 min), while LPA and MVPA reduced from 28.70% to 22.15% (decreased by 6.55%, approximately 94 min). As the severity of the disease increased, the level of physical activity among patients with chronic disease will decrease further, and SB will increase further. The accumulation of levels for dual risk factors will lead to further deterioration of the physical and psychological conditions of patients with chronic diseases.

Our findings suggest that the effects of diabetes and CVD on both PHQ-9 score and depression were mediated by 24-hour daily physical activities. Several studies reported that increasing any form of physical activity including LPA and MVPA, and reducing sedentary time reduced the probability of depression (Cahuas et al., 2020; Dinas et al., 2011; Vallance et al., 2011). LPA and MVPA not only significantly reduced depression in the general population (Schuch et al., 2018) but also increased both, while reducing SB and sleep, can lower the risk of depression among individuals with either CVD or diabetes (Mekary et al., 2013; Rethorst et al., 2017; Yasunaga et al., 2018a). Additionally, previous studies have reported that physical activity interventions had a beneficial effect on reducing depression symptoms (Conn, 2010; Hu et al., 2020). A meta-analysis of ninety-two studies reported that physical activity reduced depression by a medium effect (Rebar et al., 2015). A potential explanation is that physical activity increases endorphins and other natural brain chemicals that can enhance well-being, while buffering negative thought cycles that leads to depression (Centers for Disease Control Prevention, 2010). Also, after physical activity, the lower layer of granular cells in the dentate gyrus of the hippocampus will produce a large number of new neurons, because the number of new neurons in a certain range is beneficial to the brain and reduces certain depressive symptoms (Kodali et al., 2016; Tunc-Ozcan et al., 2019). Another meta-analysis observed that SB was associated with increased risk of depression (pooled relative risk (PR) = 1.31; 95% CI: 1.16 to 1.48 in cross-sectional studies, and pooled RR = 1.14; 95% CI: 1.06 to 1.21 in longitudinal studies) (Zhai et al., 2015). This is potentially due to SB contributes to lose vitality, and increase blood pressure and cerebrovascular resistance, which are known to negatively impact brain health long-term (Maasakkers et al., 2020).

In this study, we used 24-hour physical activity as a unit, that is, a change in one indicator causes an inverse change in other indicators. At the same time, we observed that the magnitude of the mediating effects from large to small were MVPA, LPA, SB, and sleep for the impact of CVD and diabetes on depression. The World Health Organization recommend substituting SB with physical activity, especially MVPA for optimal health (Bull et al., 2020). Additionally, Tully et al. (2020) reported greater benefits when replacing SB with MVPA than SB with LPA among the older adult population using isotemporal substitution model (Tully et al., 2020). Therefore, substituting SB engagement with MVPA is a theoretically optimal intervention model. However, based on our variation matrix in different subpopulations (general, diabetes, non-diabetes, CVD, and non-CVD), there were greater difficulty for replacing SB with MVPA than SB with LPA (co-dependence between SB and LPA: 0.297, and co-dependence between SB and MVPA: 1.714). Changing inherent lifestyles and behaviors is a complex, continuous, and gradual process. More specifically, it may be more difficult for people who do not have an established exercise habit to initiate and maintain than to increase some of their daily activities in some inadvertent ways. Yasunaga et al. (2018) reported that substituting thirty minutes of SB with LPA can reduce risk of depression (Yasunaga et al., 2018b). Another systematic review documented that after adjusting for MVPA, LPA was associated with improved health outcomes (Amagasa et al., 2018). Therefore, a more practical and feasible intervention plan should be hierarchical and step-by-step, that is, for those who engage in SB for a long period, it is necessary to design a plan that encourages replacing some SB with LPA first, and for those that engage in LPA for long period are encouraged to replace some LPA with MVPA.

5. Strength and limitation

To the best of our knowledge, this is the first study to include all compositional data features as mediators in a single mediation analysis. However, several limitations must be noted. First, there is potential for selection bias as people with depression or certain chronic diseases might be less likely to participate in the survey. Second, depression was defined based on the PHQ-9 total score with no clinical diagnosis, potentially overestimating or underestimating depression. Finally, this is a cross-sectional study, therefore the causal relationships between chronic disease, physical activity, and depression cannot be determined. Prospective cohort studies are needed to confirm our findings.

Our study limitations are offset by notable strengths. First, the survey questionnaire was carefully designed to capture important sociodemographic and behavioral characteristics, such as family poverty index ratio, smoking, and alcohol consumption. Second, the analysis was based on a large national representative sample. Third, the 24-hour physical activity data were measured by an accelerometer reducing risk of potential measurement bias. Finally, using the compositional data method to determine the coefficients of variation matrix can more accurately estimate the co-dependence of the four behaviors (LPA, MVPA, SB, and sleep) compared with traditional methods.

6. Conclusion

The present study presents a compositional mediation model for continuous and binary outcomes, which is the extension of compositional mediation model for dealing with entire compositional mediators as a unit. We observed that diabetes and CVD were associated with increased SB and sleep, and reduced MVPA and LPA, which could lead to an increased likelihood of depression from the perspective of all-day physical activity. The strong co-dependence of SB and LPA in various populations suggests the need to develop a stratified and gradual preventative measure to reduce depression in patients with CVD and diabetes, such as increasing engagement in LPA and reduce SB, while maintaining current engagement in MVPA. Future preventative measures to reduce depression in patients with CVD or diabetes should consider the compositional feature of 24-hour physical activity.