Physica A 493 (2018) 4753

Contents lists available at ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa

Does solar activity affect human happiness?

Ladislav Kristoufek
Institute of Economic Studies, Faculty of Social Sciences, Charles University, Opletalova 26, 110 00, Prague 1, Czech Republic

article info a b s t r a c t
Article history: We investigate the direct influence of solar activity (represented by sunspot numbers)
Received 14 September 2016 on human happiness (represented by the Twitter-based Happiness Index). We construct
Received in revised form 23 August 2017 four models controlling for various statistical and dynamic effects of the analyzed series.
Available online 4 November 2017
The final model gives promising results. First, there is a statistically significant negative
Keywords:
influence of solar activity on happiness which holds even after controlling for the other
Happiness factors. Second, the final model, which is still rather simple, explains around 75% of
Solar activity variance of the Happiness Index. Third, our control variables contribute significantly as
Sunspot numbers well: happiness is higher in no sunspots days, happiness is strongly persistent, there
Online data are strong intra-week cycles and happiness peaks during holidays. Our results strongly
Twitter contribute to the topical literature and they provide evidence of unique utility of the online
data.
2017 Elsevier B.V. All rights reserved.

1. Introduction

The influence of solar activity on human behavior and health has not only been a basis of various national sayings and
folklore, it has also been an active and controversial scientific topic for decades [14]. As reviewed by Lipa et al. [4], the solar
activity influences magnetosphere and ionosphere of the Earth both with short-lived disturbances such as solar flares and
with long-lived structural changes in the magnetic field. However, other effects of the solar activity on the Earth, its climate,
wildlife and human health and behavior have remained controversial until present days.
Mendoza & Diaz-Sandoval [5] study the relationship between solar activity and occurrence of myocardial infarctions
in Mexico City and find a weak positive relationship. In their following study, Mendoza & Diaz-Sandoval [6] enlarge the
dataset and they find that solar activity maxima are the most hazardous for myocardial infarctions deaths, mainly for age
groups above 65 years. Broadening the study outside of Mexico as well, Mendoza & de la Pea [7] focus on lower latitudes,
specifically Cuba and low latitudes of Mexico, and confirm the findings even for these conditions.
Cherry [8] argues and empirically shows (on the dataset from southeast Asia) that solar activity is correlated with
the Schumann resonances, deviations of which are detected by human brain and lead to health problems and possible
excessive death rates through the melatonin mechanism. Palmer et al. [9] review the topical literature focusing on the
field of heliobiology. They specifically discuss the Schumann resonances potential to explain the effect of solar activity on
human health and, in a light of the reviewed literature, find it a promising link. Babayev & Allahverdiyeva [10] study the
effect of geomagnetic variability on brain functioning using EEG and find that abrupt changes in geomagnetic variability
have strong negative effect mainly on emotional and vegetative parts of brain while personality characteristics are not
affected significantly. Some studies take the effect of solar activity even further and e.g. Mikulecky [11] finds that historical
revolutions culminated close to solar activity maxima while flourishing periods were near solar minima.
The recent studies thus suggest that increased solar activity has negative effect on humans, both from physical and
mental health perspective. However, most of these studies are fixed in time, either by being connected to a specific event

E-mail address: ladislav.kristoufek@fsv.cuni.cz.

https://doi.org/10.1016/j.physa.2017.10.031
0378-4371/ 2017 Elsevier B.V. All rights reserved.
48 L. Kristoufek / Physica A 493 (2018) 4753

Fig. 1. Evolution of daily sunspot numbers. Daily time series are available down to year 1818. The series follows a strongly cyclical path with a dominant
scale of around 11 years.

of an increased solar activity or by being laboratory/experimental. This is mainly due to a problematic availability of data,
specifically time series, of evolution of human health, be it physical or mental, that could filter out a possibility that the
reported results are due to some other factors which are characteristic for the given event.
Here we focus on the effect of solar activity on human happiness. As the human happiness is generally not easily
observable, we make use of the Happiness Index based on sentiment of Tweets (posts on Twitter) for a given day (more
details are given in the Data section). Utilizing online data, and specifically the Twitter activity, has already proven useful in
various disciplines [1220]. In our specific case, it allows us to quantify the relation between solar activity, represented by
sunspot numbers, and happiness. As both these series are available at daily frequency, it allows us to study the relationship
as a standard multivariate time series relationship. In the following section, we provide a detailed description of statistical
and dynamic properties of the analyzed series. The next section introduces a set of four models we use. And the last section
presents the results with some additional discussion.

2. Data

We work with two base time seriessunspot numbers and the Happiness Index. Sunspot numbers serve as a proxy
series for solar activity and it represents daily number of sunspots1 [21]. The series is obtained from the Solar Influence
Data Analysis Centre (SIDC) at http://www.sidc.be/sunspot-data and it is shown in Fig. 1. The daily dataset is available back
to 1818. It is evident that sunspot numbers follow a strong cyclical pattern which is well documented in the literature with
a dominant cycle of 11 years [22]. There are also days when no sunspot numbers were recorded, which might influence our
final regression and we thus need to take this aspect into consideration.
The Happiness Index series is obtained from http://hedonometer.org and it is available from 10.9.2008 to 27.5.2015,
which gives 2431 observations (the sunspot numbers series is constrained accordingly). The index is based on randomly
sampled Tweets (approximately 10% of all Tweets) for each given day. The happiness for a day is given by language processing
of the sampled Tweets with a use of the Amazons Mechanical Turk on the set of approximately 10,000 most frequently
used words based on collections of Google Books, New York Times articles, Music Lyrics, and Tweets. More details about
the selection procedure and lists of words can be found in Dodds et al. [2326]. For each of these words, a value between
1 (sad) and 9 (happy) has been attributed. Based on the Amazons Mechanical Turk, each of the 10,000 words gets an
average happiness value. Then, for each day using the Twitters Gardenhose feed, 10% of all English Tweets are recorded
and aggregated together. The Happiness Index is then a weighted average of all relevant (being one of the most frequently
used ones, i.e. with a score) words for a given day with a weight given by the happiness score of the given word.
The procedure has its limitationsthe index represents only the Twitter posts in English and there are rare specific cases,
which are questionable.2 Nonetheless, the index is very unique and it provides an irreplaceable source of information (at
least at this point3 ).

1 According to the Solar Influence Data Analysis Centre (SIDC) information, the daily total sunspot number is derived as N + 10 N where N is the
s g s
number of sunspots observed and Ng is the number of groups counted over the entire solar disk. More details and history of the measures can be found on
the provided webpage.
2 As a representative, we pick the day of Osama bin Ladens death (2 May 2011). This day is a very sad day based on the index because the Tweets
contain words dead, death, and killed which are considered to be negative. However, we can speculate that this was rather a happy day for the
English speaking countries which are represented by English Tweets. Nevertheless, this is one of only few exceptions when the index gives opposite than
expected happiness scores.
3 In an ideal situation, we would be able to obtain the list and content of the Tweets used for the index construction. The index could then be recreated
using a more rigorous natural language processing which would have better handled the outlying sentiment values as the ones mentioned in the previous
footnote. Unfortunately, these are not available.
 L. Kristoufek / Physica A 493 (2018) 4753 49

Fig. 2. Evolution of daily happiness index. Daily time series are available from 10.9.2008 to 27.5.2015 (2431 observations). The peaks of the series coincide
with holidays (mainly Christmas, New Years, and Easter) and other external events.

Fig. 3. Autocorrelation function of the Happiness Index. The autocorrelation shows a very slow decay with a power-law parameter estimated at 0.07, i.e. a
strongly persistent dynamics close to non-stationarity. Apart from the power-law correlation structure, we also observe a dominant time lag of 7 days,
i.e. one week. Happiness thus follows a strong weekly pattern.

As the Happiness Index is our variable of interest, we check it more closely. Fig. 2 reveals that there are regular spikes
in the index, which coincide with Christmas holidays, New Years Eve and Easter. Other spikes can be treated as exogenous
as they coincide with movie blockbusters, and political and cultural events. The index is apparently strongly autocorrelated.
To check this in more detail, we examine its autocorrelation function in Fig. 3 and we find in fact the series is strongly
autocorrelated with a power-law decay with an exponent of 0.07, i.e. on the border of non-stationarity. In addition, we
observe that there is a strong cyclical pattern with a dominant lag of 7 days, i.e. one week. These properties need to be taken
into consideration when constructing a proper model.

3. Methods

We are interested in a possible effect solar activity could have on human happiness. For this purpose, we construct four
models, which, in ascending order, control for more statistical and dynamic properties of the analyzed series:

Happyt = 0 + 1 NoSpotst + 2 Spotst + t (1)

Happyt = 0 + 1 NoSpotst + 2 Spotst + 3 Happyt 1 + t (2)

6

Happyt = 0 + 1 NoSpotst + 2 Spotst + 3 Happyt 1 + 3+i Dayi,t + t (3)
i=1
6

Happyt = 0 + 1 NoSpotst + 2 Spotst + 3 Happyt 1 + 3+i Dayi,t + 10 Holidayt + t . (4)
i=1

Happy is the value of the Happiness Index, NoSpots is a dummy variable equal to 1 if there are no sunspots reported for the
given day, and 0 otherwise, Spots is the sunspots number for the given day, Dayi is a dummy variable for a given day of the
week (there are only six to avoid the dummy variable trap), and Holiday is a dummy variable equal to 1 if the given day is
50 L. Kristoufek / Physica A 493 (2018) 4753

Table 1
Comparison of models. Four models with an increasing number of explanatory variables are presented. Estimated effects are accompanied by t-ratios in
brackets. For the day of the week effect, the F -test is run to test a joint significance of specific days of the week (the presented number is thus not the
estimated effect but the F -statistic of the test). Significance is labeled by asterisks , , for significance at 10%, 5% and 1% level, respectively. Number of
observations N is the same for all models. For model quality comparison, we present the coefficient of determination R2 as well as the adjusted coefficient
of determination R2 , which controls for the number of explanatory variables in the model.
Variable Model #1 Model #2 Model #3 Model #4
0.101 0.033 0.122 0.102
Constant
[5.058] [2.232] [3.736] [3.654]
0.635 0.206 0.198 0.265
No sunspots dummy
[10.871] [4.720] [4.897] [7.637]
0.321 0.099 0.093 0.120
Sunspot numbers
[14.983] [6.098] [6.152] [9.225]
0.685 0.700 0.600
Lagged happiness
[46.301] [48.311] [46.657]
Day of the week effect 67.333 85.269
2.641
Holidays
[29.661]
N 2430 2430 2430 2430
R2 0.241 0.597 0.655 0.747
R2 0.240 0.596 0.653 0.746

either the Christmas Day or Christmas Eve or New Years Eve or New Years Day or Easter, and 0 otherwise. Both Happy and
Spots have been normalized (demeaned and standardized) for better reporting purposes.
Model #1 is a basic model which simply looks at the relationship between happiness and sunspot numbers while
controlling for days with zero reported sunpots. As the variables are normalized, 2 can be seen as a correlation after
controlling for the zero sunspot days. Model #2 adds the fact that the Happiness Index is strongly autocorrelated (Fig.
3). To control for any remaining autocorrelation and possible heteroskedasticity, we report the heteroskedasticity and
autocorrelation consistent (HAC) standard errors.4 Model #3 additionally controls for the 7-day cyclical component, also
referred to as the day-of-the-week effect. And Model #4 finally controls for the effect of holidays as these are usually
accompanied with positive spikes of happiness.
All explanatory variables on the righthand side (apart from the lagged Happiness Index) are exogenous, i.e. none of them
can be influenced by the explained variablehuman happiness. Specifically, the days of the week and holidays are given by
calendar and thus cannot be affected by human happiness and activity of the Sun is not affected by human feelings. Models
#1#4 can be thus consistently estimated using the ordinary least squares regression.5

4. Results and discussion

We estimate four models to quantify the relationship between solar activity (proxied by sunspot numbers) and human
happiness (measured by the Tweets-based Happiness Index). More specifically, we are interested in the direct influence of
solar activity on happiness. Such strong claim can be done as the models meet the exogeneity assumption, which is justified
in the previous section. The four models control for specific statistical and dynamic properties of the analyzed time series.
We work with nested models so that a model is an extension of the previous one. The estimated models are summarized in
Table 1.
Model #1 can be considered as a basis model where we study the direct relationship between the two main variables
without controlling for other factors apart from the days with zero number of sunspots. We always report the estimated
effect, t-ratios and possible statistical significance. For the basis model, we observe a significant negative relationship
between the solar activity and happiness, i.e. the increased solar activity has a negative influence on human happiness at
practically any reasonable level of significance (with t = 14.983, p 0.01). The effect itself is estimated at 0.321 and
as the variables of interest are normalized, an increase of solar activity by a single standard deviation leads to a decrease of
happiness by 0.321 of its standard deviation. The basis model has R2 = 0.24, which can be considered low but it also implies
that around a quarter of the happiness variability can be explained by the solar activity, which is not negligible.
Model #2 additionally controls for persistence of the Happiness Index by including the lagged happiness measure. This
majorly improves quality of the model up to R2 = 0.60. The persistence level is high with the lagged coefficient estimated at
0.685 with t = 46.301 (p 0.01), i.e. again significant at any reasonable significance level. The effect of sunspot numbers
is weakened down to 0.099. Nevertheless, the effect is still significant at any standardly used significance level with
t = 6.098 (p 0.01). Note that the weakening of the effect after including the lagged happiness is not surprising as

4 Models with more lagged values of the Happiness Index have been estimated as well (and also for the model specifications #3 and #4) but are not
reported here as they do not provide additional informative value and the results remain qualitatively the same. We thus prefer parsimony over overfitting.
5 Exogeneity is a crucial assumption for time series ordinary least squares regression. Weak memory and heteroskedasticity is controlled for by HAC
standard errors. Potential strong memory of residuals is rejected in all four models using Augmented DickeyFuller test [27]. All necessary assumptions are
thus met and the least squares estimates are thus reliable.
 L. Kristoufek / Physica A 493 (2018) 4753 51

Fig. 4. In-sample predicted values of the happiness index. The predicted values of Model #4 are shown here and we find a high correlation of r = 0.864
between the happiness index and its predicted values implied by Model #4.

evidently happiness and solar activity are negatively correlated and happiness is persistent (i.e. positively autocorrelated)
which leads to a negative bias of Model #1. Not controlling for the persistence would thus lead to an exaggeration of the
influence of solar activity on human happiness.
Models #3 and #4 add calendar variables days of the week and holidays, respectively. Both effects are statistically
significant, again at any reasonable level, which has two straightforward interpretations. First, human happiness varies
strongly during the week and specific days are on average yielding more happiness. A more detailed inspection of the data
shows a not so surprising pattern that the most popular days are Fridays, Saturdays and Sundays. Second, people are happier
during holidays, which is again not surprising. The final model (Model #4) has a high coefficient of determination R2 = 0.75
with the adjusted R2 at practically the same level. We have thus arrived at, still rather simple, model of human happiness
which is able of explain around 75% of its variation. In the final model, the effect of solar activity remains strongly significant
with t = 9.225 (p 0.01) and the estimated effect of 0.12, i.e. an increase of solar activity by one standard deviation
leads to a decrease in human happiness by 0.12 of its standard deviation. Quality of the final model is further represented
by Fig. 4 which shows the relationship between the Happiness Index (measured in standard deviation units) and its values
implied (in-sample predicted) by our final model. The correlation of r = 0.864 only translates from the high R2 = 0.75. We
observe that the model fits the actual data very well and there are no strong patterns of deviation which could be seen for
the restricted models (not reported here6 ).
The presented results are not without some challenges and issues. First, even though the Twitter data provide an
unprecedented source of information, they are limited by the site existence. It is thus impossible to go deeper into past
without using another platform. However, there is no such platform with comparable volume of activity. And even if there
were, it would not provide much longer series anyway. Second, there are variables which are relevant for human happiness
but are not in the analyzed dataset. These are mainly economic, social, sociological and cultural factors which are either
hard to measure and/or unavailable at such a high frequency as one day. As long as these factors are not strongly correlated
with solar activity, their omission should not cause any troubles. However, we can only speculate whether this is the case
or not. Third, the Happiness Index takes only English language Tweets into consideration. It is rather just a matter of time
before an index for other languages is introduced as well and our analysis and findings can be validated. Fourth, the detected
relationship might be spurious due to a strong long-term component of the sunspot numbers series. As reported already
by Pittock [22], the dominant scale of the sunspot numbers is around 11 years. Length of the analyzed time series is under
7 years so that it is shorter than the sunspot numbers dominant scale. It is possible that the decreasing part of the sunspot
numbers cycle coincides with the increasing part of the Happiness Index thence forming the spurious relationship. There
are several arguments against this possibility. Referring back to Fig. 2 and to the comparison of the Happiness Index and
the sunspots number in Fig. 5, we observe that the index is not in a monotonous trend. There is no statistical evidence of
persistence of the models residuals, existence of which would suggest spuriously found relationship. In addition, our final
model (Model #4) explicitly controls for persistence of happiness itself. These suggest that the found relationship between
solar activity and human happiness is genuine. And fifth, the concerns can be raised against the Happiness Index itself as
discussed in the Data section. However, the index is quite unique in its availability and direct applicability. Our study can be
thus seen as a first step in the direction of finding relationship between directly measured happiness and solar activity, and
it can be further expanded on.

6 The most visible deviation of the restricted models is the inability to sufficiently fit the peaks of happiness during holidays, which is controlled for in
Model #4.
52 L. Kristoufek / Physica A 493 (2018) 4753

Fig. 5. Comparison of daily happiness index and sunspot numbers. Daily time series are available from 10.9.2008 to 27.5.2015 (2431 observations).

In summary, we have investigated the direct influence of solar activity on human happiness. We have built four models
controlling for various statistical and dynamic effects of the analyzed series. Our final model gives promising results. First,
there is a statistically significant negative influence of solar activity on happiness which holds even after controlling for the
other factors. Second, the final model, which is still rather simple, explains around 75% of variance of the Happiness Index.
Third, our control variables contribute significantly as wellhappiness is higher in no sunspots days, happiness is strongly
persistent, there are strong intra-week cycles and happiness peaks during holidays.

Acknowledgments

The research leading to these results was supported by the People Programme (Marie Curie Actions) of the European
Unions Seventh Framework Programme FP7/20072013/ under REA grant agreement number 609642. The author further
acknowledges financial support from the Czech Science Foundation (grants number 1600027S).

References

[1] J.M. Wilcox, The interplanetary magnetic field. Solar origin and terrestrial effects, Space Sci. Rev. 8 (1968) 258328.
[2] K.F. Novikova, The effect of solar activity on the development of myocardial infarction and mortality resulting therefrom, Kardiologia 4 (1968) 109112.
[3] M. Feinleib, E. Rogot, P.A. Sturrock, Solar Activity and Mortality in the United States, Int. J. Epidemiol. 4 (3) (1975) 227229.
[4] B.J. Lipa, P.A. Sturrock, E. Rogot, Search for correlations between geomagnetic disturbances and mortality, Nature 259 (1976) 302304.
[5] B. Mendoza, R. Diaz-Sandoval, Relationship between solar activity and myocardial infarctions in Mexico City, Geofisica Int. 39 (1) (2000) 5356.
[6] B. Mendoza, R. Diaz-Sandoval, Effects of Solar Activity on Myocardial Infarction Deaths in Low Geomagnetic Latitude Regions, Nat. Hazards 32 (2004)
2536.
[7] B. Mendoza, S.S. de la Pena, Solar activity and human health at middle and low geomagnetic latitudes in Central America, Adv. Space Res. 46 (2010)
449459.
[8] N. Cherry, Schumann resonances, a plausible biophysical mechanism for the human health effects of solar geomagnetic activity, Nat. Hazards 26 (2002)
279331.
[9] S.J. Palmer, M.J. Rycroft, M. Cermack, Solar and geomagnetic activity, extremely low frequency magnetic and electric fields and human health at the
Earths surface, Surv. Geophys. 27 (2006) 557595.
[10] E.S. Babayev, A.A. Allahverdiyeva, Effects of geomagnetic activity variations on the physiological and psychological state of functionally healthy
humans: Some results of Azerbaijani studies, Adv. Space Res. 40 (2007) 19411951.
[11] M. Mikulecky, Solar activity, revolutions and cultural prime in the history of mankind, Neuro Endocrinol. Lett. 28 (6) (2007) 749756.
[12] J. Bollen, H. Mao, X. Zeng, Twitter mood predicts the stock market, J. Comput. Sci. 2 (2011) 18.
[13] P.T. Metaxas, E. Mustafaraj, Social media and the elections, Science 338 (2012) 472473.
[14] D. Mocanu, A. Baronchelli, N. Perra, B. Goncalves, Q. Zhang, A. Vespignani, The twitter of babel: Mapping world languages through microblogging
platforms, PLoS One 8 (4) (2013) e61981.
[15] D. Garcia, C.J. Tessone, P. Mavrodiev, N. Perony, The digital traces of bubbles: feedback cycles between socio-economic signals in the Bitcoin economy,
J. R. Soc. Interface 11 (2014) 20140623.
[16] P.A. Grabowicz, J.J. Ramasco, B. Goncalves, V.M. Eguiluz, Entangling mobility and interactions in social media, PLoS One 9 (3) (2014) e92196.
[17] C.M. Alis, M.T. Lim, H.S. Moat, D. Barchiesi, T. Preis, S.R. Bishop, Quantifying regional differences in the length of twitter messages, PLoS One 10 (4)
(2015) e0122278.
[18] H.S. Moat, C.Y. Olivola, N. Chater, T. Preis, Searching choices: Quantifying decision-making processes using search engine data, Topics in Cognitive
Science 8 (3) (2016) 685696.
[19] W. Zhang, X. Li, D. Shen, A. Teglio, Daily happiness and stock returns: Some international evidence, Physica A 460 (2016) 201209.
[20] X. Li, D. Shen, M. Xue, W. Zhang, Daily happiness and stock returns: The case of Chinese company listed in the United States, Econ. Modell. 64 (2017)
496501.
[21] SILSO World Data Center, The International Sunspot Number, Int. Sunspot Number Monthly Bull. Online Catal. (1818-2015). http://www.sidc.be/silso/.
[22] B. Pittock, A critical look at long-term sunweather relationships, Rev. Geophys. Space Phys. 16 (3) (1978) 400420.
[23] P.S. Dodds, C.M. Danforth, Measuring the happiness of large-scale written expression: Songs, blogs, and presidents, J. Happiness Stud. 11 (4) (2010)
441456.
 L. Kristoufek / Physica A 493 (2018) 4753 53

[24] P.S. Dodds, K.D. Harris, I.M. Kloumann, C.A. Bliss, C.M. Danforth, Temporal patterns of happiness and information in a global social network:
Hedonometrics and twitter, PLoS One 6 (12) (2011) e26752.
[25] M.R. Frank, L. Mitchell, P.S. Doods, C.M. Danforth, Happiness and the patterns of life: a study of geolocated tweets, Sci. Rep. 3 (2013) 2625.
[26] P.S. Dodds, E.M. Clark, S. Desu, M.R. Frank, A.J. Reagan, J.R. Williams, L. Mitchell, K.D. Harris, I.M. Kloumann, J.P. Bagrow, K. Megerdoomian, M.T.
McMahon, B.F. Tivnan, C.M. Danforth, Human language reveals a universal positivity bias, Proc. Natl. Acad. Sci. USA 112 (8) (2015) 23892394.
[27] Davdi A. Dickey, Wayne A. Fuller, Distribution of the estimators for autoregressive time series with a unit root, J. Amer. Statist. Assoc. 74 (1979) 427431
Taylor & Francis.