Journal of Happiness Studies (2021) 22:599–624

https://doi.org/10.1007/s10902-020-00242-8

RESEARCH PAPER

Does Volunteering Make Us Happier, or Are Happier

People More Likely to Volunteer? Addressing the Problem
of Reverse Causality When Estimating the Wellbeing Impacts
of Volunteering

Ricky N. Lawton1 · Iulian Gramatki1 · Will Watt2 · Daniel Fujiwara1

Published online: 17 March 2020

© Springer Nature B.V. 2020

Abstract
Evidence of the correlation between volunteering and wellbeing has been gradually accu-
mulating, but to date this research has had limited success in accounting for the factors that
are likely to drive self-selection into volunteering by ‘happier’ people. To better isolate the
impact that volunteering has on people’s wellbeing, we explore nationally representative
UK household datasets with an extensive longitudinal component, to run panel analysis
which controls for the previous higher or lower levels of SWB that volunteers report. Using
first-difference estimation within the British Household Panel Survey and Understanding
Society longitudinal panel datasets (10 waves spanning about 20 years), we are able to
control for higher prior levels of wellbeing of those who volunteer, and to produce the most
robust quasi-causal estimates to date by ensuring that volunteering is associated not just
with a higher wellbeing a priori, but with a positive change in wellbeing. Comparison of
equivalent wellbeing values from previous studies shows that our analysis is the most real-
istic and conservative estimate to date of the association between volunteering and subjec-
tive wellbeing, and its equivalent wellbeing value of £911 per volunteer per year on aver-
age to compensate for the wellbeing increase associated with volunteering. It is our hope
that these values can be incorporated into decision-making at the policy and practitioner
level, to ensure that the societal benefits provided by volunteering are better understood
and internalised into decisions.

Keywords Volunteering · Altruism · Subjective wellbeing · Wellbeing valuation ·

Compensating surplus · First difference

1 Introduction

Volunteering is an important element of civil society, defined broadly as “offering one’s

time for free… [including] organising or helping to run an event, campaigning, conser-
vation, raising money, providing transport or driving, taking part in a sponsored event,

* Ricky N. Lawton
ricky_lawton@hotmail.com
Extended author information available on the last page of the article

13
Vol.:(0123456789)
600 R. N. Lawton et al.

coaching, tuition or mentoring for no expense” (DCMS 2015, p. 62). Volunteering har-
nesses the time and labour of just under a quarter of all adults per year in the United King-
dom (UK), providing outcomes that would require the equivalent of around 1.3 million
full-time workers and delivering benefits of over £12.2bn to the UK economy annually.1
In addition to these economic impacts, volunteering has been shown to be associated with
benefits for the people who do it, in terms of their health and their wellbeing (Aknin et al.
2013; Konrath 2014). Self-worth, socialising and the feeling of doing something useful are
just a few of the factors which are related to this improvement in wellbeing (Deci and Ryan
2008; Ryan et al. 2008).
Volunteering may be organised formally through groups, clubs or organisations, or
be provided informally, in ways which are typically harder to evaluate. Previous research
shows that there is a ‘civic core’ of people who are more likely to volunteer (Mohan and
Bulloch 2012). The demographic composition of this core is based on where people live
(more people volunteer in rural areas than in urban), socioeconomic background (more
people from the well-educated middle class than poorer backgrounds), ethnic backgrounds
(more white people volunteer than people from other backgrounds) and age group (people
in mid-life are less likely to volunteer).
Previous research has shown that volunteering is associated with a number of positive
outcomes for volunteers, in terms of their health, sense of self-worth, and psychologi-
cal wellbeing (Binder and Freytag 2013; Meier and Stutzer 2008). In this study we focus
on subjective wellbeing (SWB) outcomes: people‘s subjective experiences of their own
wellbeing self-reported in national household surveys. There are a range of SWB ques-
tions–happiness, emotions, life satisfaction, worth/purpose in life, sadness, anxiety and
goal attainment—each tapping into different theoretical concepts of wellbeing (Kahneman
et al. 2003). By looking at SWB measures rather than standard economic preference meas-
ures we are looking at a more experiential level of wellbeing by assessing how outcomes
and activities actually impact on people’s wellbeing rather than how people predict they
will impact on their lives.
However, evidence on volunteering and wellbeing is mixed, with some studies finding
no relationship with wellbeing (Haller and Hadler 2006) and others finding a positive rela-
tionship (Greenfield and Marks 2004; Meier and Stutzer 2008). Bivariate analysis is only
able to state that, on average, volunteers have higher SWB than non-volunteers (Binder
and Freytag 2013; Dolan et al. 2008; Meier and Stutzer 2008). In other words, it may be
the case that those with pre-existing higher level of SWB have a higher propensity to vol-
unteer. This fails to account for the factors that are likely to drive self-selection into volun-
teering by happier or healthier people. Controlling for a range of other sociodemographic
factors that drive SWB helps mitigate this issue but cannot fully address the problem of
omitted variable bias.2

1
https://data.ncvo.org.uk/a/almanac17/economic-value-3/.
2
Recently, a number of studies have used more robust statistical strategies that provide a high degree of
confidence in the findings. For example, Stutzer and Frey (2004) analysed trends in volunteering and SWB
after controlling for other determinants of SWB and by assessing circumstances where volunteering status
essentially became random due to policy changes (natural experiments). Binder and Freytag (2013) ana-
lysed six waves of British Household Panel Survey (BHPS) by using statistical matching to estimate the
causal impact of volunteering on happiness. There is some evidence that in the context of charitable giving
the pathway of happier people giving more dominates the pathway of giving making people happier; how-
ever, the effect does go both ways (Boenigk and Mayr 2015).

13
Does Volunteering Make Us Happier, or Are Happier People More… 603

The main focus of the Binder study is on the impact of volunteering along the sub-
jective well-being distribution, applying quantile fixed effects regressions, to account for
time-invariant individual-specific heterogeneity by first estimating fixed effects and then
applying the Koenker–Basset quantile regression estimator on the time-varying error term
(after removing the influence of the time-invariant component of the error term). They find
that sustained volunteering (at least monthly) is significantly positively impacting on life
satisfaction over all deciles, but its impact is decreasing with life satisfaction (0.0969 at the
lowest and 0.0481 at the highest decile, with the median being close to the FE estimate).
The author concludes that the results show that even for happier individuals in the distribu-
tion, sustained volunteering still has a positive (if not smaller) effect.
Chen et al. (2014) use a univariate ordered probit model and simultaneous bivariate
ordered probit model to compare the potential endogeneity between volunteering and life
satisfaction on the Naiwan Survey of National Living Conditions with a sample of 4070.
An exogenous military service variable was included in the bivariate model to correct the
endogeneity of volunteering on life satisfaction. The results from the simultaneous bivari-
ate ordered probit model indicate volunteering has a significant positive effect on life sat-
isfaction (+ 0.812 in the ordered probit model for a response life satisfaction variable with
4 categories). While the authors conclude that failure to account for endogeneity appears
to underestimate the effect of volunteering on life satisfaction, it should be noted that this
effect size is unfeasibly large—given the value of the ordered probit cutoffs, it is enough
to move a respondent from ‘very dissatisfied’ to ‘satisfied’ in some situations. Further-
more, we can see that when running a more standard univariate ordered probit model, the
coe#cient of volunteering is insignificant. It is therefore likely that the main finding of
the authors is a result of overfitting an overly sensitive and complex econometric model
(maximum likelihood on two simultaneous equations). There is further evidence to support
this—namely that the coe#cients of the demographic control variables in the volunteering
equation and in the life satisfaction equation largely have opposite signs.
Two other papers focused on demographic subgroups of older or younger populations.
Hansen et al. (2018) explore the relationship between volunteering and well-being by using
OLS regressions on two waves of data of 18,559 individuals aged 50 and above from 12
European countries, analysing the life satisfaction impacts of change and stability in vol-
unteering status (+ 0.24 on a scale of 0–10, for aged 50 groups only) and in the intensity
(frequency) of volunteering, and exploring whether these impacts differ according to life
stage (age, employment status) and across countries. Findings show that net life satisfac-
tion is higher among longer-term, recent, and former volunteers than among stable (long-
term) non-volunteers. The authors find no significant life satisfaction differences between
the three groups with volunteer experience, while similar levels of life satisfaction are
observed among people who have increased and decreased their frequency of volunteer-
ing. The authors conclude that the positive effects of volunteering relate to the experience
rather than the dynamics of volunteering that is associated with well-being. It is also worth
noting that the authors defined volunteers as those respondents who volunteered at least
monthly, with the rest being defined as non-volunteers—this is a stricter definition than the
one used here and may partly be responsible for the higher coe#cient.
Truskauskaitė-Kunevi%ienė (2015) uses data from a two-wave Lithuanian longitudi-
nal youth community sample using structural equation modelling. The author finds that
increased life satisfaction does not lead to increase of volunteering frequency (at least for
girls), however, it leads to increased active participation (contribution) in one’s self’s, fam-
ily’s and community’s lives. The authors suggest that their findings are consistent with the
idea that both satisfaction and dissatisfaction with life can motivate volunteering.

13
Does Volunteering Make Us Happier, or Are Happier People More… 605

Table 2 Panel structure of Number of waves Total In regression

USoc + BHPS volunteering data
Number of Percent Number of Percent
observationsa observations

1 26,148 10.66 14,863 7.34

2 31,264 12.75 23,392 11.56
3 79,929 32.6 68,281 33.73
4 8360 3.41 6798 3.36
5 10,590 4.32 8905 4.40
6 11,532 4.70 9744 4.81
7 20,321 8.29 17,785 8.79
8 19,640 8.01 17,752 8.77
9 8019 3.27 7317 3.61
10 29,400 11.99 27,599 13.63
Total 245,203 100 202,436 100

The number of individuals will be equal to the number of observations

divided by the number of waves (but may be slightly larger in the ‘in
regression’ subsection because some observations may be dropped due
to missing data)

allow for much better causal attribution in the results.4 USoc preserves most of the meth-
odology and variable encoding used in the BHPS, with n = 11,781 BHPS respondents car-
rying over into USoc (out of the total of 83,741 unique individuals in the combined data-
set), providing a longer time series in combination and thus increasing the precision of
our panel data estimation techniques, and providing a larger sample than the BHPS-only
sample previously used by Binder (2015).
Note that the merged panel dataset, although not fully balanced, has a solid panel struc-
ture. The number of individuals that appear in only one wave is 26,148, which is only
10.66% of the combined sample size of 245,203, or 31.22% of the number of unique indi-
viduals in the data. Furthermore, if we restrict ourselves only to those observations that
have data for all the necessary variables to be included in the main wellbeing regression
(using life satisfaction as the outcome), the number of irrelevant observations for panel
data estimation (appearing in only one wave) further drops to 14,863 out of 202,436 obser-
vations (7.34%, or 21.47% of the 69,216 unique individuals). Table 2 provides more details
on the panel structure by tabulating an auxiliary variable listing how many different waves
the individual is present.

3.2 Data Analysis

A key issue that we face in this analysis is that any observed relationship between SWB
and volunteering may be due to a host of factors aside from volunteering. For example,
healthier or richer people may be more likely to volunteer and they will also have higher

4
Panel data estimation techniques (FE and FD) control for time-invariant factors, including long-term
health and psychological factors, as well as prior trends of pre-existing SWB.

13
Does Volunteering Make Us Happier, or Are Happier People More… 607

measure of health status in the general population (Chandola and Jenkinson 2000; Pinquart
2001; Wu et al. 2013). Self-rated health may also be better at predicting more serious,
chronic illnesses better than less serious illnesses (Doiron et al. 2015), or have a stronger
association with physical health than functional health (Pinquart 2001). However, in the
absence of an objective assessment of a person’s physical health, the general self-rated
health question provides a useful indicator of health impacts within household panel sur-
veys, which has been used in numerous studies previously (Borgonovi 2008; Brooks 2006,
2007).
For the volunteering variable Vit , our main focus is on one of the simplest possible
measures—a binary indicator (dummy variable) of whether the respondent volunteered in
the past 12 months. Information on volunteering also enables us to analyse different fre-
quencies of volunteering in the Understanding Society + BHPS data, similarly to Binder’s
(2015) approach.
In the vector Xit we control for as many of the main determinants of health and sub-
jective wellbeing as possible, as set out in Fujiwara and Campbell (2011). These include
household income, marital status, number of children, employment status, self-rated health,
age, gender, broad ethnicity, geographic region, and education.6
OLS regressions are performed on Eq. (1) as is. In the fixed effects specification, all var-
iables in Eq. (1) are demeaned7 to cancel out individual-specific factors (the time-invariant
components of each variable in the model).8
Attrition bias, where respondents drop out of the panel, is a prominent obstacle to robust
fixed effects and first differences estimation. Although some authors (e.g. Schmidt and
Woll 2017) attempt to combat attrition bias by designing and applying different version of
longitudinal sampling weights, the effectiveness of this approach has been debated. Cum-
ming and Goldstein (2016) argue that it “is statistically ine#cient, because it drops incom-
plete data records, is inflexible, and in practice gives rise to undue complexity involving
a proliferation of weighting systems for different analyses.” Earlier, Vandecasteele and
Debels (2007) found that longitudinal weights do manage to reduce attrition bias to some
extent, but there are also situations where they increase the size of the bias. These mixed
conclusions in the literature, together with rather solid panel structure of our dataset (where
only 7.43% of the observations are rendered irrelevant in a panel context due to attrition)
led us to the decision not to use longitudinal weights in an attempt to correct for attrition
bias.

3.2.2 First Differences Estimation

First differences estimation is another approach to cancelling out unobserved but time-
invariant individual-specific factors. Instead of using the individual’s current level of well-
being and volunteering status as in pooled OLS, we use the change between the current

6
We also control for time-specific effects and seasonality by including dummy variables for the wave of
the survey and the month of the interview and for local area characteristics via region dummies [unfortu-
nately more detailed geographic information is not available in the BHPS data].
7
The means over time of the respective variables for that particular individual are subtracted from each
observation.
8
Use of OLS and fixed effects regressions assumes that the SWB reporting scale (1–7) is cardinal.
Research shows that the cardinal models (OLS regressions) and ordinal models (ordered latent response
models, such as ordered logit/probit) give remarkably similar results, and hence for ease of interpretation
we assume cardinality, as is standard in much of the literature (Ferrer-i-Carbonell and Frijters 2004).

13