In this paper we estimate the size of weekend effects for seven emotions and then explore their main determinants for the working population in the United States, using the Gallup/Healthways US Daily Poll 2008–2012. We first find that weekend effects exist for all emotions, and that these effects are not explained by sample selection bias. Full-time workers have larger weekend effects than do part-time workers. We then explore the sources of weekend effects and find that workplace trust and workplace social relations, combined with differences in social time spent with family and friends, together almost fully explain the weekend effects for happiness, laughter, enjoyment and sadness, for both full-time and part-time workers, with significant but smaller proportions explained for the remaining three emotions—worry, anger and stress. Finally, we show that workplace trust and social relations significantly improve emotions and life evaluations on both weekends and weekdays for all workers.
Citation: Helliwell JF, Wang S (2015) How Was the Weekend? How the Social Context Underlies Weekend Effects in Happiness and Other Emotions for US Workers. PLoS ONE 10(12): e0145123. https://doi.org/10.1371/journal.pone.0145123
Editor: Christopher M. Danforth, University of Vermont, UNITED STATES
Received: September 21, 2015; Accepted: November 29, 2015; Published: December 23, 2015
Copyright: © 2015 Helliwell, Wang. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited
Data Availability: Data have been deposited to Figshare: http://dx.doi.org/10.6084/m9.figshare.1618666.
Funding: The Canadian Institute for Advanced Research, through its Social Interactions, Identity and Well-being program, which JFH directs, supports the broad program of research that includes this paper. SW gratefully acknowledges financial support from the Korea Development Institute (KDI) School of Public Policy and Management. The funders have no role in study design, data collection, and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Emotions, both positive and negative, are key measures of subjective well-being , , . Recent empirical studies on emotions find evidence of variations through the week, often called a day-of-week effect. One study finds that those who are interviewed on Fridays report lower levels of mental stress than those interviewed in the middle of the week, using the British Household Panel Survey (BHPS) data . Some studies focus on the variations between weekends and weekdays, namely the weekend effect. For example, a recent study finds that people experience more positive emotions and fewer negative emotions during weekends and statutory holidays than on weekdays in the Gallup/Heathways US Daily Poll . A few other studies also find similar weekend effects , , , , .
Though the weekend effect seems to be well-proved, some recent studies pose challenges. Some researchers argue that the day-of-week or weekend effects appearing in some datasets might be due to possible selection bias in the choice of interview days , . In other words, the subjective well-being responses might be correlated with some observed or unobserved characteristics that affect individuals’ decisions to take interviews on specific days. The two studies are both based on the British Household Panel Survey data. One study draws the conclusion that the day-of-week pattern of job satisfaction is not substantially affected by the potential selection bias, although the day-of-week pattern for mental well-being does become less significant after adjusting for the likely bias . Another one emphasizes the potential selection bias originating from unobservables and suggests that the day-of-week patterns may vary across countries .
A few recent studies discover variations of weekend effects across demographic groups and social environments. One recent paper shows that weekend effects for emotions are smaller among older workers, particularly with respect to negative emotions . Another one finds that weekend effects vary with gender, marriage status, age, and working status . Some find that the variation of social time across days of week is an important determinant of weekend effects , . The quality of the social context at work has previously been shown to influence both life evaluations and emotions , .
Differing from previous studies, in this paper we focus more specifically on emotional weekend effects for the full-time and part-time working population using the Gallup/Heathways Daily Poll in the United States. In this way we are able to assess more directly, for a comparable set of respondents, the size of weekend effects for each of the seven emotions (happiness, enjoyment, laughter, worry, sadness, anger, and stress). We will also explore the varying explanatory power of the social context, both on and off the job, for each emotion, comparing full-time workers with part-time workers.
Seeing the possibilities of selection bias, we also deal with the selection issue in our data. Since we are studying weekend effects rather than day-of-week effects, we are mostly concerned about whether individuals’ answers depend on whether they were interviewed on weekends or weekdays, rather than whether they are interviewed on particular days of the week, as studied in , .
Data and Methodology
The data we use for this study come from the Gallup/Healthways US Daily Poll. From the beginning of 2008, Gallup has randomly interviewed about 1,000 American adults each day in the United States. By the end of 2012, the total number of respondents accumulated in the data is 1.77 million. The Daily Poll includes a set of questions on emotional well-being: for positive emotions we have happiness, enjoyment, and laughter, and for negative emotions we have worry, sadness, anger, and stress. Laughter is a yes or no response (yes = 1) to the question “Did you smile or laugh a lot yesterday?” Other six emotions are similar binary responses to the question “Did you experience the following feelings during a lot of the day yesterday?” Emotion questions were asked on every survey day, except for the stress question, which was not asked in 2011 and 2012.
The survey includes the labor force status for each respondent, so that we can derive a sample based on the working population. There are about 0.94 million working respondents, accounting for 58% of all respondents. Among the working respondents, nearly 80% are full-time paid workers.
The key explanatory variables are social hours and two measures of the quality of the workplace social context. The social hours variable is a response to the question “Approximately, how many hours did you spend, socially, with friends or family yesterday? Please include telephone or e-mail or other online communication.” This question was asked of all the respondents from 2008 to 2010, but in 2011 and 2012 only 10% and 5% of total respondents, randomly selected, were asked this question. Respondents report numbers between 0 and 24. Among the answers, there are about 5% of respondents reporting more than 16 social hours. To make this social hours variable more reliable, we replace any value greater than 16 by 16. There are two questions on workplace environment. One is “Does your supervisor always create an environment that is trusting and open, or not?” The answer to this question is binary, 1 for “yes” and 0 for “no”. 80% of respondents answer that the environment is trusting. Another question is “Does your supervisor at work treat you more like he or she is your boss or your partner?” The answer to this question is also binary, 1 for “partner” and 0 for “boss”. 62% report having a “partner-like” boss.
The survey also includes a number of socio-demographic variables, such as gender, age, marital status, level of education, number of children under 18, monthly household income, health insurance coverage and importance of religion. Monthly household income refers to before-tax income from all sources, including wages and salaries, remittances from family members living elsewhere, farming, and others. The response is categorical, in which zero to ten stands for no income, under $60, $60 to $499, $500 to $999, $1,000 to $1,999, $2,000 to $2,999, $3,000 to $3,999, $4,000 to $4,999, $5,000 to $7,499, $7,500 to $9,999, and $10,000 and over, respectively. We construct the numerical household income by replacing the categorical response by the mean of each non-top category, and $18,000 for the top income category. There are about 0.78 million working respondents reporting income. To reduce the impact of missing income on the number of observations, we assign a zero value to log income when income is missing, and a dummy variable which equals to 1 if the income is missing will be used together with log income in regressions. The summary statistics for all the variables are reported in Table 1.
To show the size of weekend effects, we estimate the following model for each emotion for full-time and part-time workers separately: (1) where i indexes individuals. The variable emotionit denotes one of the seven emotions, while weekendit is an indicator variable that equals to one if the emotions are for weekends or statutory holidays. The vector Xit denotes a set of individual- and household-level covariates, which include respondent’s gender, age, age squared divided by 100, marital status, education levels, household income, number of children, frequency of church attendance, an indicator variable that equals to one if having health insurance, and a dummy variable indicating the importance of religion in life. Xct is a vector for state-year fixed effects. εit is the error term.
To see how the weekend effect is varying with each respondent’s workplace environment, we estimate the following model for full-time and part-time workers respectively: (2) where the variable workit denotes the quality of the workplace environment. We have two measures for workplace quality: one is a dummy variable that equals to one if respondents report having a workplace is trustworthy and open, and the second is a dummy variable that equals to one if respondents report having a supervisor who is more like a partner than a boss. The two measures will first be included in the regression individually and then together. This is a typical difference-in-difference (DID) approach, where γ captures the difference in weekend effect by workplace environment.
Next we add the social time variables into Eq (2) to check how well weekend effects are further explained by social hours. The equation is as follows: (3) where Sit denotes a set of variables on social time, including log of social hours, a dummy variable for zero social hour, and a dummy for zero to one social hour.
The Size of Weekend Effects
Table 2 gives a detailed summary of the three positive emotions—happiness, enjoyment, and laughter, and four negative emotions—worry, sadness, anger, and stress. In the table we report the number of observations, estimated means and standard errors for each emotion on weekends and weekdays, and weekend effects measured by the mean difference of each emotion between weekends and weekdays, and the percentage of change of emotion from weekdays to weekends. We can see generally that positive emotions are more prevalent than negative ones. Moreover, the prevalence of positive emotions is higher on weekends than weekdays, with the reverse applying to negative emotions. The weekend effect is statistically significant for each emotion. Moreover, the effect is sizable, with absolute values ranging from 0.014 to 0.150, and the percentage change ranging from 3.3% to 32.5%. The percentage improvements for negative emotions from weekdays to weekends are generally larger than for positive emotions, consistent with the findings in . For example, the reductions in stress, worry, and anger are -32.5%, -24.3%, -24.4% respectively, much larger than the improvement in enjoyment, 6.8%, the most improved among the positive emotions. These differences primarily reflect the fact that the average frequency is much less for negative than for positive emotions.
In this section we show the estimates of weekend effects using Eq (1). We run OLS regressions for each emotion, separately for full-time and part-time workers. Though all emotion variables are binary, we still run OLS for easier interpretation of the coefficients. However we have checked the results using logistic regressions and find similar results. For each group, we run two models, one is the model controlling only for state-year fixed effects, while the second includes state-year fixed effects and the full set of covariates described in Eq (1): gender, age, age squared divided by 100, marital status, education levels, household income, number of children, frequency of church attendance, an indicator variable that equals to one if having health insurance, and a dummy variable indicating the importance of religion in life. We report the coefficients of the weekend dummy in Table 3. We find that the weekend coefficients for each emotion are almost the same in the two models, which indicates that other covariates of weekend and weekday samples are well balanced, given the random sampling procedure. Moreover, the weekend effect for full-time workers is larger than for part-time workers. For happiness, enjoyment, anger and stress, the weekend effect for part-time workers is approximately half that for full-time workers. The relative size is two-thirds for laughter, three-quarters for worry, while almost equal for sadness.
Testing for selection effects
As shown in previous studies , , the size of day-of-the-week effects may be affected by the potential bias of respondents’ self-selection into specific days within a week, for surveys where only a single day is reported for each respondent, as is the case with our data. In this section we run a few tests to show that the impact of selection bias (being self-selected into weekends versus weekdays) in our data is fairly small, if not zero. We report the results only for paid workers, but the tests for full samples yield similar results.
In Table 4 we test the balance of social-demographic variables for respondents being surveyed on weekends and weekdays. Columns (1) and (2) report the mean of each variable for respondents being surveyed on weekends and weekdays respectively. Columns (3) and (4) show the mean difference between Weekends and Weekdays and the corresponding standard error. We will find small differences some variables, some of which are rendered statistically significant by the large sample size. In this case we may check the standardized differences of means, which is a common way to check sample balance in propensity score matching methods, following , . Column (5) reports the standardized differences of means calculated by the formula (4) where and are the sample means for weekends and weekdays, and Var(x1) and Var(x2) are the corresponding sample variances. We can see that the absolute value of each standardized difference is smaller than 2.2. This is much smaller than the often-used cutoff value 10. This confirms the similarity of the weekend and weekday samples.
We also illustrate in Fig 1 the propensity scores for being selected for weekend interviews. We use a probit model to estimate the propensity score. The variables used to estimate the propensity score of being selected into weekend include respondent’s gender, age, age squared divided by 100, marital status, education levels, household income, number of children, frequency of church attendance, an indicator variable that equals to one if having health insurance, a dummy variable indicating the importance of religion in life, and respondent’s full-time or part-time working status. The upper panel of Fig 1 shows the distribution of propensity scores for people who were actually interviewed on weekends and weekdays, and the lower panel shows the distribution of propensity scores for people who report emotions for weekends (who were interviewed on Sunday and Monday) and weekdays (who were interviewed on Tuesday to Saturday). In both cases, the distribution of propensity scores is very similar for the two groups, which suggests that each respondent has almost the same probability of being selected into weekends or weekdays.
Notes: The upper panel of Fig 1 shows the distribution of propensity scores for people who were actually interviewed on weekends and weekdays, and the lower panel shows the distribution of propensity scores for people who report emotions for weekends (who were interviewed on Sunday and Monday) and weekdays (who were interviewed on Tuesday to Saturday).
Next we make a direct test for selection bias by exploiting the nice feature of our data that people report yesterday’s emotions. The emotions relating to Sunday, Monday and Tuesday which are reported on Monday, Tuesday, and Wednesday, respectively, are not affected by possible weekend selection bias. Column (1) of Table 5 shows the difference in emotions between Sunday and Monday and column (2) shows the difference in emotions between Tuesday and Monday. The values in column (2) are all equal to or very close to zero, suggest that emotions are stable across weekdays. The large differences between column (1) and column (2) suggest that real weekend effects exist. Thus the weekend effect we observe is not just driven by selection effects. We then compare the weekend effect measured by the difference between emotions on Sunday and Monday in column (1), with the regular weekend effect based on all data in column (3). We know that the former is not subject to selection bias, but the latter may be subject to certain level of selection bias. However, since the two results are so similar, the bias must be negligibly small. Note that our results suggest that there is no selection between weekends and weekdays in the Gallup/Heathways US Daily Poll, which may not mean the previous argument of respondents’ self-selection into specific days within a week in the British Household Panel Survey  is invalid, since the type of selection bias and the nature of data are different in the two cases.
Exploring the Determinants of Weekend Effects
In this section we explore the determinants of weekend effects. We first run OLS regressions following Eq (2) to examine how weekend effects vary with workplace social context, measured by reported workplace trust and type of supervisor (boss-like or partner-like). We further control social time variables following Eq (3). We report the estimated weekend effect for happiness in Tables 6 and 7 for full-time and part-time workers respectively. In each table, Models (1), (3) and (5) follow Eq (2) and Models (2), (4) and (6) follow Eq (3). We use the dummy for non-trusting workplace and its interaction with the weekend dummy in Models (1) and (2), the dummy for boss-like supervisor and its interaction with the weekend dummy in Models (3) and (4), and both workplace dummies and their interactions with the weekend dummy in Models (5) and (6).
From Model (1) in Table 6 we see that the weekend effect of happiness for full-time workers in high-trust workplaces is 0.025, while for those reporting non-trusting workplace environment it is 0.070, which equals to 0.025 plus the 0.045 coefficient on the interaction term. This implies that the weekend effect is almost three times larger for those working in a low-trust environment. Results in Model (3) using the alternative workplace environment variable confirm the finding: full-time workers reporting boss-like supervisor have weekend effects twice as do workers with partner-like supervisors. If we include both indicators of workplace social context, as shown in Model (5), the weekend effect of happiness for full-time workers reduces to 0.022, and that for unfavorable environment (both indicators equal to 1) rises to 0.073. Workers with a boss-like supervisor in a low trust workplace thus have weekend effects more than three times as large as for those having a partner-like supervisor in a higher trust work environment.
If we control also for the social time variable, the weekend effect for those full-time workers reporting favorable workplace environment is reduced to zero, as shown in Models (2) and (4), or even slightly negative (-0.003) as in Model (6). This suggests that weekend effects for respondents with favorable workplace environments are only due to the differing amounts of social time on weekends and weekdays. The weekend effects for those reporting unfavorable workplace environments are now 0.041, 0.039, and 0.034 in Models (2), (4) and (6) respectively, which are much smaller than the effects calculated without accounting for the difference in social time between weekends and weekdays.
We observe similar patterns for part-time workers for the weekend effect of happiness in Table 7. Without controlling for social time variable, the weekend effect for those reporting unfavorable workplace environments is always larger than for those reporting favorable environments (by about 0.013). Moreover, accounting for social time differences between weekends and weekdays reduces the weekend effect for those part-time workers reporting favorable workplace environment to a level not significantly different from zero.
The results for happiness in Tables 6 and 7 are summarized in Panel A of Table 8, in which we report weekend effects in different scenarios. To save space, we follow the format of Panel A to report the results for the remaining positive emotions in Panels B and C of Table 8, and for the four negative emotions in Panels A to D of Table 9.
The weekend effect for enjoyment is reported in Panel B of Table 8. Without controlling for the social time variable, the weekend effect for full-time workers is 0.042 for those reporting favorable workplace social contexts, compared to 0.140 for those in unfavorable workplace social contexts. Holding social time constant, the two corresponding values reduce to 0.011 and 0.103. Thus controlling for social time reduces the weekend effect by about three-quarters for those with good workplace environments, and by one-quarter for those with socially unfavorable workplaces. If we compare the weekend effect for those who report unfavorable workplace environments without accounting for social time, 0.140, with the effect, 0.011, for those who report favorable environments, after allowing for social time, the reduction is over 90%. The proportionate reduction is almost 90% (0.055 to 0.006) for part-time workers. In Panel C of Table 8 we summarize the weekend effects for laughter. We see that social time and the workplace social context together fully explain the weekend effect.
In the four panels of Table 9 we report the weekend effects for worry, sadness, anger, and stress respectively. Compared to our previous results for positive emotions, the explanatory power of workplace environment and social time is lower for all negative emotions except sadness. Specifically, in the case of full-time workers, social time and workplace environment together explain about half of the weekend effects for worry (-0.110 to -0.056), 80% for anger (-0.088 to -0.018) and 40% for stress (-0.202 to -0.123). For part-time workers the part of the weekend effect explained by the social context variables is one-third for worry (-0.079 to -0.051), 50% for anger (-0.026 to -0.013), and 20% for stress (-0.095 to -0.078).
If weekends are better than weekdays for those with poor social contexts in their workplaces, could bad jobs provide a way of getting better weekends? The answer is “no”. Taking “happiness” as an example, we can see this point clearly from Table 6, which shows that the coefficients of “dummy for non-trusting workplace” in all columns are negative and statistically significant. We highlight in Figs 2 and 3 our evidence showing clearly that full-time workers in jobs with good social contexts have more positive emotions, fewer negative emotions, and higher life evaluations on both weekends and weekdays. Fig 2 shows the results for three positive emotions and life evaluations. The workplace impacts on happiness on weekends and weekdays indicated by “trust” can be calculated from the coefficient of “dummy for non-trusting workplace” and “dummy for non-trusting workplace*weekend” in column (2) of Table 6. Similarly, the impacts indicated by “supervisor” can be calculated by “dummy for boss-like supervisor” and “dummy for boss-like supervisor*weekend” in column (4), and the impacts indicated by “both” are from the all these four variables in column (6). The results for other emotions and life evaluations are taken from parallel regressions reported in the online appendix. For all three emotions, the gains from a good workplace social context are significantly greater on weekdays than on weekends, but weekend emotions are nonetheless heavily dependent on the memories and anticipations of what has and will happen during the working week. Some of these effects may be due to the fact that many respondents work on weekends, but our sample of employed respondents is not asked about their days of work, beyond whether they work part-time or full-time.
Notes: The vertical bars indicated by “trust” show the difference in positive emotions or life evaluations between those reporting “trusting workplace” and those reporting “non-trusting workplace”. The bars indicated by “partner” show the difference between those reporting “partner-like supervisor” and those reporting “boss-like supervisor”. The bars indicated by “both” show the difference between those reporting both “trusting workplace” and “partner-like supervisor” and those reporting “non-trusting workplace” and “boss-like supervisor”. 95% confidence interval is also shown for each bar.
Notes: The vertical bars indicated by “trust” show the difference in negative emotions between those reporting “trusting workplace” and those reporting “non-trusting workplace”. The bars indicated by “partner” show the difference between those reporting “partner-like supervisor” and those reporting “boss-like supervisor”. The bars indicated by “both” show the difference between those reporting both “trusting workplace” and “partner-like supervisor” and those reporting “non-trusting workplace” and “boss-like supervisor”. 95% confidence interval is also shown for each bar.
Fig 3 shows the results for four negative emotions. Consistent with the results for positive emotions, the gains from a good workplace social context are significantly greater on weekdays than on weekends. The corresponding figures for part-time workers, reported in an online appendix show results consistently smaller, as we have already seen in Tables 6–9, than those for full-time workers.
Life evaluations are heavily influenced by the social context at work, but to an amount that is the same on weekdays and weekends. When people are asked about their lives as a whole, what goes on at work is clearly a very important determinant of their answers. Since they are being asked about their lives as a whole, and not about the previous day, their answers do not depend, and nor should they, on the day of the week when the question was asked. The results at the bottom of Table 2 show nonetheless for the employed population as a whole that overall life evaluations are slightly but significantly higher on weekends than on weekdays, by an amount equal to about 0.2%, a tiny fraction of the difference shown for any of the emotions, which is consistent with . This finding helps to validate both emotional reports and life evaluations, since the former are intended to reveal day-to-day changes, while the latter are intended to look beyond the day-to-day variations in experiences to provide a broader measure of subjective well-being.
In this paper we estimate the size and sources of weekend effects for seven emotions for the working population in the United States, using the Gallup/Healthways US Daily Poll 2008–2012. We first find that weekend effects, measured as the difference in fractions of the population reporting each emotion between weekends and weekdays, are statistically and economically significant for all seven emotions. Moreover, full-time workers have larger weekend effects than do part-time workers for all emotions except for sadness, where the effects are similar for full-time and part-time workers. We also show that the weekend effects we find are not driven by the bias of self-selection into weekends versus weekdays.
We then explore the sources of weekend effects and find that the social quality of the workplace is a key correlate of emotions and life evaluations for US employed workers on all days of the week, positive for positive emotions and life evaluations and negative for negative emotions. Moreover, we find that the effects of workplace quality are larger on weekdays than on weekends. This also means that the weekend effects are much smaller for workers with good workplace social contexts, as indicated by high workplace trust and a partner-like boss. Moreover, social time can largely or even entirely explain the remaining weekend effects for positive emotions and sadness for workers with favorable workplace social contexts. The workplace social environment and social time together almost completely account for the weekend effects for happiness, laughter, enjoyment and sadness, for both full-time and part-time workers. The explanatory power is lower for the remaining negative emotions. Taken together, the quality of the social contexts on and off the job are the primary forces behind weekend effects in the subjective well-being of the working population of the United States.
The Canadian Institute for Advanced Research, through its Social Interactions, Identity and Well-being program, which JFH directs, supports the broad program of research that includes this paper. SW gratefully acknowledges financial support from the Korea Development Institute (KDI) School of Public Policy and Management. The funders have no role in study design, data collection, and analysis, decision to publish, or preparation of the manuscript. We thank the Gallup Organization for access to data from the Gallup/Healthways US Daily Poll.
Conceived and designed the experiments: JFH SW. Analyzed the data: JFH SW. Wrote the paper: JFH SW.
- 1. Durand M, Smith C. The OECD approach to measuring subjective well-being. In: Helliwell JF, Layard R, Sachs J, editors. World happiness report 2013. New York: UN Sustainable Development Solutions Network; 2013. p. 112–137.
- 2. Helliwell JF, Wang S. Weekends and subjective well-being. Social Indicators Research. 2014; 116(2): 389–407.
- 3. Kahneman D, Krueger A. Developments in the measurement of subjective well-being. Journal of Economic Perspectives. 2006; 20(1): 3–24.
- 4. Taylor MP. Tell me why I don’t like Mondays: investigating day of the week effects on job satisfaction and psychological well-being. Journal of the Royal Statistical Society: Series A (Statistics in Society). 2006; 169(1): 127–142.
- 5. Kennedy-Moore E, Greenberg MA, Newman MG, Stone AA. The relationship between daily events and mood: the mood measure may matter. Motivation and Emotion. 1992; 162(2): 143–155.
- 6. Ryan RM, Bernstein JH, Brown KW. Weekends, work, and well-being: psychological need satisfactions and day of the week effects on mood, vitality, and physical symptoms. Journal of Social and Clinical Psychology. 2010; 29(1): 95–122.
- 7. Stone AA, Hedges SM, Neale JM, Satin S. Prospective and cross-sectional mood reports offer no evidence of a “Blue Monday” phenomenon. Journal of Personality and Social Psychology. 1985; 49(1): 129–134.
- 8. Stone AA, Schneider S, Harter JK. Day-of-week mood patterns in the United States: on the existence of ‘Blue Monday’, ‘Thank God It’s Friday’ and weekend effects. Journal of Positive Psychology. 2012; 7(4), 306–314.
- 9. Young C, Lim C. Time as a network good: evidence from unemployment and the standard workweek. Sociological Science. 2014; 1: 10–27.
- 10. Tumen S, Zeydanli T. Day-of-the-week effects in subjective well-being: does selectivity matter? Social Indicators Research. 2014; 119(1): 139–162.
- 11. Helliwell JF, Huang H. How’s the job? Well-being and social capital in the workplace. Industrial and Labor Relations Review. 2010; 63(2): 205–227.
- 12. Lee W. Propensity score matching and variations on the balancing test. Empirical Economics. 2013; 44(1): 47–80.
- 13. Rosenbaum P, Rubin D. Constructing a control group using multivariate matched sampling methods that incorporate the propensity score. American Statistician. 1985; 39(1): 33–38.