1. Introduction
Sleep behaviors represent a fundamental yet underexplored determinant of workplace attitudes and organizational outcomes. While extensive research has examined traditional antecedents of job satisfaction—including demographic characteristics, job design, and organizational factors (
Kooij et al., 2010;
Aydin et al., 2012;
Judge et al., 2002), the role of strategic napping in enhancing employee wellbeing and work attitudes remains poorly understood. This gap is particularly significant given that job satisfaction, defined as an employee’s positive emotional state resulting from job appraisal, constitutes one of the most critical predictors of performance, turnover, and organizational effectiveness (
Weiss, 2002;
X. Liu et al., 2025).
The emerging literature on sleep and workplace outcomes reveals three critical limitations that this study addresses. First, existing research predominantly employs cross-sectional designs with limited samples, constraining causal inferences about the napping–satisfaction relationship (
De Bloom et al., 2017). Second, most studies focus on specific organizational contexts, particularly healthcare settings, limiting generalizability to broader work environments (
Deng et al., 2024;
Steven & Redfern, 2024). Third, and most importantly, previous research has largely treated napping as a binary phenomenon or assumed linear effects, overlooking the potentially crucial non-linear relationship between nap duration and workplace outcomes (
Leng & Yaffe, 2019).
This study adopts the work, non-work, and sleep framework (WNSF) as its theoretical foundation (
Crain et al., 2018). The WNSF conceptualizes sleep as a critical resource that influences work attitudes through energy restoration and cognitive enhancement mechanisms, while recognizing time as a finite resource requiring strategic allocation between work and recovery activities. However, the framework lacks a specification of the optimal sleep parameters and boundary conditions under which sleep behaviors maximize workplace benefits.
We theoretically propose and empirically test an inverted U-shaped relationship between nap duration and job satisfaction, grounded in three interconnected mechanisms. Physiologically, short naps (10–30 min) promote alertness and cognitive restoration without inducing sleep inertia, whereas longer naps trigger deeper sleep stages that result in temporary post-awakening impairment (
Brooks & Lack, 2006). Resource-theoretically, brief naps represent efficient recovery strategies that minimize time away from work activities, while excessive napping creates time conflicts and stress that counteract recovery benefits (
Crain et al., 2018). Psychologically, napping benefits follow diminishing returns patterns, with medium-length naps showing optimal effects on mood regulation and cognitive performance (
Yang et al., 2024).
This research makes four significant theoretical and practical contributions. Methodologically, we employ a quasi-natural experimental approach with longitudinal Chinese data and instrumental variable techniques to establish causal relationships between napping behaviors and job satisfaction, addressing critical limitations of previous correlational studies. Theoretically, we provide the first systematic empirical evidence for the curvilinear effect of nap duration on job satisfaction, identifying optimal nap zones that extend and refine the WNSF framework. Contextually, we demonstrate how sleep behaviors operate within cultural contexts where napping is institutionally normalized, offering insights for global organizational policy development. Practically, our rigorous analytical approach, incorporating sensitivity analyses and matching models, provides evidence-based foundations for workplace napping policies that can simultaneously enhance employee well-being and organizational outcomes.
2. Napping Behavior and Job Satisfaction
The work, non-work, and sleep framework (WNSF) provides the theoretical foundation for understanding how napping behaviors influence job satisfaction through systematic resource allocation and energy management processes (
Crain et al., 2018). The WNSF conceptualizes human energy as comprising physical energy (capacity to perform work) and energetic activation (vitality and enthusiasm), both of which are directly affected by sleep behaviors. Critically, the framework positions time as a finite resource requiring strategic allocation between work activities and recovery processes, establishing the theoretical basis for understanding why napping duration, rather than mere occurrence, determines workplace outcomes. This theoretical perspective suggests that the relationship between napping and job satisfaction operates through complex, interconnected mechanisms that collectively explain both the positive effects of strategic napping and the potential negative consequences of suboptimal nap durations.
The relationship between napping and job satisfaction operates primarily through three interconnected theoretical mechanisms that form an integrated framework for understanding workplace sleep behaviors. From a resource recovery perspective, napping represents a strategic intervention for replenishing depleted cognitive and emotional resources that accumulate during continuous work demands (
Dutheil et al., 2021). As employees experience resource depletion manifested as increased fatigue, decreased concentration, and reduced emotional regulation capacity—all directly affecting job satisfaction—napping activates parasympathetic nervous system responses, reducing stress hormone levels while promoting beneficial neurotransmitter activity (
Faraut et al., 2015;
Milner & Cote, 2009). Empirical evidence consistently demonstrates that brief naps of 10–20 min significantly reduce both subjective fatigue and objective stress indicators, enabling employees to evaluate their work environment more positively and restore cognitive resource reserves (
Brooks & Lack, 2006). This resource recovery process directly alleviates the negative effects of work stress, creating a pathway through which strategic napping enhances job satisfaction by enabling more positive workplace evaluations.
Simultaneously, napping significantly enhances multiple cognitive domains critical for job performance and satisfaction, including working memory capacity, information processing speed, decision-making quality, and creative thinking ability (
Lovato & Lack, 2010). These cognitive improvements enable employees to meet work challenges more effectively, reduce errors, and achieve higher performance standards, generating enhanced feelings of competence and achievement that directly translate into job satisfaction. According to cognitive appraisal theory, when employees perceive themselves as capable of meeting work demands through enhanced cognitive functioning, job satisfaction naturally increases. Neuroimaging studies reveal that strategic napping enhances prefrontal cortex activity and executive function without disrupting circadian rhythms, providing biological evidence for the cognitive enhancement pathway (
Wong et al., 2013). The third mechanism involves emotional regulation, where napping profoundly influences emotional well-being through neuroendocrine system modulation, particularly rebalancing limbic system and prefrontal cortex interactions (
Goldschmied et al., 2015). Research demonstrates substantial improvements in emotional states following napping, with negative affect reductions of 30–40% and positive affect improvements of 20–35% (
Taub & Berger, 1976;
Kaida et al., 2006). According to affective events theory, these enhanced emotional experiences accumulate and directly influence work attitudes, creating a direct pathway from emotional regulation to job satisfaction.
While napping provides substantial benefits through these mechanisms, the WNSF’s emphasis on time as a finite resource suggests that these benefits follow a curvilinear rather than linear pattern, supported by optimization theory and resource balance models. This theoretical prediction indicates that the relationship between nap duration and job satisfaction reflects a critical transition between increasing and diminishing returns phases. Within the optimal duration zone of 10–30 min, napping occurs primarily in light sleep stages, providing mental restoration with minimal sleep inertia and maximizing positive effects on work performance and satisfaction (
Brooks & Lack, 2006). This “efficient recovery interval” represents the optimal balance between resource restoration and time investment, where physiological and psychological benefits are proportional to the time allocated. However, beyond this optimal duration threshold, extended napping triggers multiple inhibitory factors that progressively diminish benefits. Sleep physiology research indicates that naps exceeding 30 min significantly increase deep sleep probability, resulting in post-awakening sleep inertia characterized by impaired cognitive and psychomotor function lasting 30–60 min (
Hilditch & McHill, 2019). Additionally, organizational time resource theory suggests that extended napping creates time conflicts and task accumulation pressures that offset recovery benefits, while social cognitive factors introduce perceived risks of negative peer and supervisor evaluations (
Mullins et al., 2014;
Takahashi et al., 2004).
Converging evidence from multiple research domains supports this theoretical framework and demonstrates the complexity of napping effects on workplace outcomes. Studies consistently demonstrate that brief naps enhance alertness, mood, and cognitive performance, with optimal durations ranging from 10–30 min (
Brooks & Lack, 2006;
Petit et al., 2018). Workplace-specific research reveals that strategic napping serves as a vital recovery mechanism for replenishing psychological and physiological resources depleted by work demands (
Barnes, 2012). Longitudinal studies indicate that napping benefits follow diminishing returns patterns, with moderate-length naps showing optimal effects across multiple outcome domains (
J. Liu et al., 2019). However, research also reveals potential negative consequences of suboptimal napping behaviors, with studies indicating that naps exceeding 30 min are associated with increased sleep inertia, temporary cognitive impairment, and heightened cardiovascular risks (
L. Wang et al., 2022;
Tassi & Muzet, 2000). These findings suggest that the relationship between napping and workplace outcomes is more complex than previously assumed, requiring careful consideration of duration parameters rather than treating napping as a uniformly beneficial behavior.
The theoretical and empirical evidence reviewed above converges on several key conclusions that extend the WNSF framework and provide specific guidance for understanding napping–satisfaction relationships. First, napping influences job satisfaction through multiple interconnected mechanisms involving resource recovery, cognitive enhancement, and emotional regulation that operate simultaneously rather than independently. Second, these mechanisms operate within optimal boundaries determined by sleep physiology and organizational time constraints, creating threshold effects that determine whether napping enhances or impairs workplace outcomes. Third, the relationship between nap duration and job satisfaction follows an inverted U-shaped pattern, reflecting the transition from beneficial to potentially counterproductive effects as duration increases beyond optimal levels. This integrated framework extends the WNSF by specifying precise boundary conditions under which sleep behaviors optimize workplace outcomes, moving beyond general assertions about sleep’s importance to identify specific parameters for maximizing benefits. The framework also reconciles apparently contradictory findings in the literature by recognizing that napping effects depend critically on duration parameters rather than mere occurrence, providing a theoretical foundation for understanding when and why napping enhances job satisfaction.
Integrating these findings with the perspectives of the WNSF, we can infer a curvilinear relationship between napping habits and job satisfaction. Based on the aforementioned analysis, this study proposes the following hypotheses regarding the relationship between napping habits and job satisfaction:
Hypothesis 1a: Compared to the non-napping group, the napping group will exhibit higher job satisfaction.
Hypothesis 1b: The relationship between nap duration and job satisfaction is inverted U-shaped, implying the existence of an optimal point where nap duration maximizes its positive impact on job satisfaction.
3. Research Design and Model Selection
3.1. Model Specification
Under specific conditions, the β estimates in the fixed effects model may be highly dependent on the sample—that is, overly sensitive to random errors in the given data. Suppose there are few within-group observations, or the change in x relative to y is small. Then, the within-group effect estimate of x on y might exhibit substantial bias relative to the true effect due to random factors. A drawback of the fixed effects model is the need to estimate the coefficients for each group indicator variable. This significantly reduces the model’s efficiency and increases the standard errors of the coefficient estimates. The problem becomes more severe when the within-group sample size is small, as the group effects alone can explain a large portion of the variation in the dependent variable. The random effects model partially pools information across groups through the between-group component, rendering the β estimates less variable (
Gelman & Hill, 2007). Random effects estimates form a compromise between fixed effects models and mixed models, shrinking the α_j for deviating groups towards the average μ_α. This moves the β estimates away from the unstable fixed effects estimates and towards more stable (although possibly biased) estimates (
Clark & Linzer, 2015). Given that our sample data have relatively few within-group samples, a random effects model is more suitable.
Certainly, scholars have suggested utilizing the
Hausman (
1978) specification test to examine whether the assumption of the random effects model, which presupposes the orthogonality of explanatory variables and group effects, has been violated. When the test results are significant, it indicates a correlation between x and α_j, implying that the random effects model should be abandoned in favor of the fixed effects model. However, in most applications, the true correlation between covariates and group effects is not entirely zero. Therefore, if the Hausman test cannot reject the null hypothesis of orthogonality, it is likely not because the true correlation is zero. Conversely, the test probably lacks sufficient statistical power to reliably distinguish between small correlations and zero correlation; when using the random effects model, even if the Hausman test does not find significant results, there will still be bias in the β estimates; of course, in many cases, if the biased (random effects) estimator provides sufficient variance shrinkage compared to the unbiased (fixed effects) estimator, the biased estimator might be preferable (
Clark & Linzer, 2015). Thus, we compared the variance shrinkage of the fixed effects model and the random effects model and found that for most variables, the standard errors of the random effects model are smaller than those of the fixed effects model. This implies that the random effects model provides greater variance shrinkage for these estimates.
Therefore, to investigate whether napping and nap duration affect job satisfaction, the model designed in this study is as follows:
Wherein, represents the job satisfaction of individual at time . The core explanatory variable indicates whether individual takes a nap at time , and represents the nap duration of individual at time . is introduced to capture the curvilinear effect. is the control variable for individual at time . is the individual random effect, and is the unobservable residual term.
3.2. Sample Selection and Data Source
The independent variables, dependent variables, and data utilized in this study originate from the China Family Panel Studies (CFPS) database. CFPS, an open-access database, allows data acquisition through registration and application. Embarking in 2010 and conducting surveys biennially (with the exception of 2011), CFPS represents a large-scale, longitudinal study, comprising a nationally representative sample of Chinese adults and children, obtained through random sampling. Surveys from 2016 to 2020 encompass a set of measurements related to satisfaction, including inquiries about the work environment, income, safety, time, promotions, and overall job satisfaction, while comprehensive survey items regarding job satisfaction are absent in other survey years. Notably, questions about individual-level napping behaviors, specifically the occurrence and duration of naps, are present in all survey years from 2010 to 2020, excluding 2012. Furthermore, this research contemplates the issue of endogeneity within the model, selecting sunlight duration as an instrumental variable. The data for sunlight duration is retrieved from the Wind Database, which encompasses daily, weekly, monthly, quarterly, semi-annual, and annual data for over 3000 regions in China, including provinces, cities, and counties.
This study also engaged in meticulous data-cleaning processes to ensure the derivation of reliable results. Initially, we selected sample data that was consistently reported from 2016 to 2020, identified via the unique individual identifier “pid”. Furthermore, we refined the samples based on two variables from the questionnaire: “current employment status” (variable name in the data: employ) and “currently attending school” (variable name in the data: qc1), thereby isolating samples of individuals who were employed and not attending school during the period from 2016 to 2020. Upon examining the obtained samples, we found that independent variable 1 (to nap or not) did not exhibit missing values. For handling missing values, we adopted the following approach: if independent variable 2 (nap duration) was greater than 0, it indicated napping, ensuring a logical consistency between independent variables 1 and 2. The majority of the control variables did not exceed 1.5% of their own quantities in missing values. We did not initially impute missing values for control variables; first, we removed samples with missing dependent variables (while retaining samples with “pid” consistently reported over three years). The dependent variable, job satisfaction, had 568 samples with missing values (constituting 4.64% of the data), which we removed, further retaining samples consistently reported over three years. Consequently, the proportion of missing values in control variables also decreased. In this study, the primary control variables included residence (urban or rural), age, gender, marital status, type of occupation, and highest educational attainment. While we attempted to incorporate sleep duration on workdays and income into the control variables, their excessive missing values (47.86% for income and 23.07% for sleep duration on workdays) precluded their inclusion in the analysis. However, we will explore these in subsequent sensitivity analyses.
The control variables we selected can have their missing values reasonably imputed from the values of the previous or following year. Naturally, for data across three periods within the control variables that are entirely absent (the missing values corresponding to “pid” are absent throughout the three years), we directly eliminate them. After operating in this manner, our independent variable 2 (nap duration) still had a 15.18% missing rate. We did not directly impute the missing values, but we eliminated all missing values of the core variables and ensured that the remaining data were reported for all three years (to ensure balanced panel data). We then examined the impact of the core independent variables on the dependent variable under the condition of having control variables. The results showed that the independent variables still have a significant impact on the dependent variable after such operations. To utilize the existing data for baseline regression analysis to the greatest extent, we performed imputation on the data without deleting the missing values of nap duration. In fact, whether to impute or not does not have a significant impact because we will match and introduce instrumental variables subsequently. After matching and introducing instrumental variables (since we have high requirements for the instrumental variables, any missing values in any year are unacceptable, i.e., our strict requirement is that the instrumental variable data for the individual must be complete for all three years), we found that the missing data for nap duration, imputed via model prediction method, were also basically deleted. Ultimately, in our baseline regression, we analyzed the data without matching and introducing instrumental variables (10,728 entries), while during the analysis of instrumental variable, sensitivity analysis, treatment effect model analysis, and PSM matching model, we used the data with matched and introduced instrumental variables for analysis (7401 entries). In this way, we can largely ensure the reliability and scientific nature of the model.
3.3. Variable Measurement
Independent variable 1: To nap or not. It is crucial to note that “noon break” typically refers to a period of rest or relaxation during midday, not necessarily involving sleep, whereas “nap” explicitly denotes a brief sleep during midday. These two may differ significantly in terms of time allocation and activity content, thereby potentially yielding different impacts and conclusions in research. The original questionnaire measured this by asking, “Do you currently have a habit of napping?” with 1 representing napping and 5 representing not napping. For ease of analysis, we subsequently use 0 to represent not napping. In the sample for our baseline regression, the proportion of samples not napping is 35.70% (N = 3830), while the percentage of those napping is 64.30% (N = 6898), totaling N = 10,728 samples.
Independent variable 2: Nap duration. The original questionnaire measured this with “How many minutes do you generally nap?” The unit of this variable is minutes, and for ease of interpretation, we will adjust its unit to hours in subsequent analyses. For samples without napping, we set the nap duration to 0. Nap duration ranges from 0 ≤ nap duration ≤ 240 min, that is, within a range of 0 to 4 h.
Dependent variable: Job satisfaction. The original questionnaire measured this with “Overall, how satisfied are you with this job?” (for studies using similar items, see
Demerouti et al., 2012;
Trougakos et al., 2008). Answer options range from 1 (very dissatisfied) to 5 (very satisfied). It is worth reiterating that the samples we retained are all in a working state (employed), and we did not include samples that are unemployed or job-waiting. Job satisfaction is the variable of our utmost concern, and we did not perform any imputation on it; we deleted any values with missing data, thus using the original data.
Control variables. In our research, there are a total of 6 available control variables, namely, residing in urban or rural areas (Replace with ‘Urban’ in the following), age, gender, marital status, occupation type, and highest educational level (Replace with ‘Education’ in the following). Past research indicates that age and educational level can impact job satisfaction (
Lee & Wilbur, 1985), and of course, gender, marital status, and occupation type (
Peng et al., 2022), as well as urban/rural and occupation type can also significantly impact job satisfaction (
Friend & Burns, 1977). Therefore, we controlled for these in our research.
Detailed information about the variables in the research can be viewed through the descriptive statistics in
Table 1.
3.4. Discussion on Endogeneity
This study primarily investigates the linear impact of napping (to nap or not) on job satisfaction and the curvilinear effect of nap duration on job satisfaction, recognizing that the aforementioned models may harbor endogeneity issues due to omitted variables. As previously mentioned, the substantial missing values for work income and weekday sleep duration prevent us from effectively incorporating them as control variables in the model for analysis. Both work income and weekday sleep duration can concurrently influence nap duration and job satisfaction. For instance, existing research indicates that higher performance rewards often correlate with elevated job satisfaction (
Heywood & Wei, 2006), and higher-income groups tend to exhibit higher sleep efficiency and longer sleep duration (
Sosso et al., 2021). The model may also omit unobservable variables (such as individual health status, lifestyle, and napping needs), which can simultaneously affect job satisfaction and the propensity to nap.
On the other hand, the model may also be subject to self-selection issues. Generally, nap duration is often related to organizational policies, working hours, and personal habits. Regardless of whether the organization has stringent regulations or individuals have good napping habits, the choice to nap and nap duration often vary. For example, in a relatively liberal organization, especially those with flexible working arrangements, whether employees choose to nap and the duration of naps are often at their discretion. Moreover, individuals who nap regularly may be more inclined to select work environments that permit napping, thereby attaining higher job satisfaction (
Baxter & Kroll-Smith, 2005). This implies that, under certain circumstances, employees may choose whether to nap or determine the duration of naps based on organizational requirements and characteristics, or their own needs.
To tackle the endogeneity concerns stemming from reverse causation and omitted variables in our model, we utilize the duration of daylight in an individual’s residing city as an instrumental variable. It is well-documented that daylight exposure can extend continuous sleep duration. This is because reduced exposure to daylight and extended periods of darkness correlate with a prolonged biological night, attributed to the extended secretion of melatonin, leading to longer sleep spans (
Stothard et al., 2017;
Wehr, 1991;
Wehr et al., 1993). Furthermore, the quality of sleep is intrinsically tied to daylight exposure. Pertinent studies have shown that exposure to white light, rich in short wavelengths during the day, correlates with enhanced evening fatigue and improved sleep quality (
Viola et al., 2008), diminished sleep onset latency, and an increased accrual of slow-wave sleep (
Wams et al., 2017). Notably, the timing of such light exposure appears to have a bearing on sleep, with findings suggesting that individuals exposed to >10 lx of light have more frequent nocturnal awakenings and reduced slow-wave sleep (
Wams et al., 2017). The WNSF also postulates that macro-level variables, such as light exposure, have a profound impact on sleep quality (
Crain et al., 2018). Digging deeper, evidence suggests that sleep deprivation augments negative emotions in the workplace the subsequent day (
Zohar et al., 2005), and the choice to take a nap during the day, as well as the nap duration, is significantly influenced by the sleep quality and duration of the preceding night (
C. Wang et al., 2019). Thus, it is reasonable to infer that the daylight duration of the previous day might influence an individual’s decision to nap and the duration of the nap on the following day. Yet, the duration of daylight remains orthogonal to the unobserved variables that influence job satisfaction, validating the scientific rationale behind using city daylight duration as an instrumental variable for napping. To further mitigate the concerns of omitted variables in our model, we have also incorporated a sensitivity analysis.
To alleviate the estimation bias induced by self-selection issues, this study employs both treatment effect models and propensity score matching (PSM) for estimation. The estimation principle of the two-step method in treatment effect models is similar to that of the two-stage least squares (2SLS), but it requires the endogenous variable to be a dummy variable. The first step of the two-step estimation involves utilizing the probit model to estimate the probability of an individual entering the napping group. The second step employs ordinary least squares (OLS) regression to obtain the coefficient estimation value for job satisfaction. PSM, proposed by
Rosenbaum and Rubin (
1983) to address the measurement issue of distance between individuals during matching, calculates the average treatment effect (ATE) through the following steps: ① Incorporate variables affecting both “to nap or not” and job satisfaction into the matching variables. ② Utilize the logit model to estimate the conditional probability of an individual entering the napping group, i.e., the propensity score, and test whether the matching results pass the balance test and whether the fit between the treatment group and the control group is optimized after matching. ③ Employ the propensity score values for kernel matching, calculating the average treatment effect on the treated (ATT) for “to nap or not”.
3.5. Validation of Control Variables Based on the LASSO Model
In pursuit of a more robust and reliable model, we scrutinized the efficacy of our chosen control variables through the LASSO model. Upon executing the LASSO regression and tuning the parameter lambda, we determined the variables that necessitated control. As can be discerned from
Table 2, under the stipulation of lambda = 0.01, the control variable ‘age’ is excluded. Even when lambda is adjusted to 0.001, ‘age’ remains unincorporated. Consequently, in subsequent analyses, we did not include ‘age’ among the control variables.
3.6. Analytical Strategy and Methodological Approach
To ensure scientific rigor and causal validity in our inferences about the relationship between napping and job satisfaction, this study constructs a multi-layered, progressive analytical framework. This systematic strategy not only addresses various potential biases but also establishes a robust chain of evidence through cross-validation between methods.
First, baseline random effects regression models serve as the foundation of our analysis, establishing preliminary correlations between napping and job satisfaction by incorporating key control variables (including residential area, gender, marital status, education level, and occupation type). The choice of random effects models is based on sample characteristics and Hausman test results, effectively balancing the trade-off between estimation efficiency and bias. However, despite including multiple control variables, baseline models still face endogeneity challenges, particularly unobserved individual characteristics (such as health status and sleep preferences) and reverse causality that may lead to estimation bias. Therefore, relying solely on control variable methods cannot establish reliable causal relationships, prompting us to employ more advanced econometric methods.
Second, the application of instrumental variable (IV) methods is crucial for addressing endogeneity issues. The selection of sunlight duration as an instrumental variable is based on sufficient theoretical foundations and empirical testing, satisfying both relevance and exclusion conditions. This method isolates variations in napping induced by exogenous factors, thereby resolving reverse causality and omitted variable bias. We ensure the validity of instrumental variables through rigorous econometric tests (including F-statistics, Cragg–Donald Wald tests, and Kleibergen–Paap rk Wald tests). This approach is decisive in establishing the causal relationship between napping behavior and job satisfaction, particularly in verifying the existence of an inverted U-shaped relationship.
Third, sensitivity analysis is an indispensable component of our analytical framework, systematically assessing the sensitivity of results to unobserved confounding factors. By constructing a comparative benchmark (using occupation type as a reference), we can quantify how strong potential omitted variables would need to be to alter our research conclusions. This transparent methodological assessment is critical for verifying the robustness of causal inferences, especially considering that we cannot include certain potentially relevant variables (such as work income and workday sleep duration) in the basic model due to their high missing rates. Without this analysis, we would be unable to effectively evaluate the extent to which results are influenced by omitted variables.
Fourth, treatment effect models and propensity score matching (PSM) methods offer complementary methodological approaches to addressing self-selection issues. Treatment effect models explicitly model the napping decision process through two-stage estimation, while PSM creates matched samples based on observable characteristics to minimize selection bias. The combined application of these two methods is particularly important because they are based on different statistical principles and can cross-validate the consistency of results. The core advantage of PSM is that it does not rely on linear assumptions but directly compares individuals with similar characteristics, thus providing an important supplement to parametric models. If both methods show that napping has a significant impact on job satisfaction, the credibility of our inferences is greatly enhanced.
Finally, we employ multiple robustness testing strategies to systematically assess the stability of results. Quantile regression analysis allows us to test whether the effect of napping is consistent at different points in the job satisfaction distribution, revealing potential effect heterogeneity. Gender-based group analysis examines the applicability of findings across male and female populations, providing key information for understanding the universality of napping effects. Additionally, by expanding the set of control variables (including work income, nighttime sleep duration on workdays, and life stress), we can comprehensively assess the sensitivity of results to model specifications. In particular, the inclusion of these additional control variables compensates for factors omitted from the baseline model due to data limitations, further enhancing the reliability of causal inferences.
This hierarchical, multi-faceted analytical strategy is necessary because no single method can adequately address the complex challenges of causal inference. Various methods are based on different identifying assumptions and address different types of bias: IV methods handle endogeneity, PSM and treatment effect models resolve self-selection issues, sensitivity analysis evaluates the impact of unobserved heterogeneity, and multiple robustness tests verify that results do not depend on specific model specifications.