Weather and Mobile Promotion
Abstract
Despite the ubiquitous impact of weather, little evidence exists, especially with field data, on
how consumers respond to weather and how this affects their responses to promotions. On the
basis of multiple field experiment datasets with over 10 million smartphone users, we
investigate how sunny and rainy weather affects consumers’ incremental purchase responses to
mobile promotions. Our field data can identify the causal impact because smartphone users
across a wide range of geographical locations and climates are randomly assigned to receive
mobile promotions with a prevention ad copy frame (treatment) and neutral ad copy frame
(control). The data can also identify the incremental purchase responses because there is a
baseline (holdout) group of smartphone users who are comparable to the treatment and control
groups but receive no promotion. Our data analyses account for customer self-selection bias in
multiple ways: (a) geographical locations in terms of different areas, (b) time variations in terms
of different days and hours of the day, and (c) individual usage behavior heterogeneity in terms
of activity bias in different weather condition. To further address endogeneity and self-selection
concerns, our identification strategies rely on the within-person changes in weather, via both
backward-looking historical weather and forward-looking forecasts, as well as deviations from
the normal expected weather of geographical regions. The main result is that consumers are
more responsive to SMS and APP promotions on sunshine days, but a prevention frame of ad
copy increases the effectiveness of promotions on rainy and inclement days. These results are
consistent with the affect-as-information theory (the sunny weather-induced positive mood
boosts consumer response likelihood and hazard) and mood congruency hypothesis (the rainy
weather-induced negative mood matches with the prevention frame of ad copy). In terms of the
odds ratio of purchase likelihood, sunny weather leads to about 1.21 times more purchases as a
result of the mobile promotions compared with cloudy weather, and rainy days leads to about
0.9 times less purchases than cloudy days. Hour-by-hour analyses with survival models suggest
the hazard rate of responses would be 73% faster in sunny weather but 59% slower in rainy
weather relative to cloudy sky. With over 150 million users of weather apps, there are a total of
2 billion checks of weather each day. For marketers, these staggering numbers may implicate
some new business opportunities with weather-targeted SMS and APP promotions for
smartphone users.
Keywords: Mobile, Targeting, Field Experiment, E-commerce, Advertising, Weather
1
Many companies notice the value of targeting weather (past, current, or future) in their ads
and promotions, as consumers buy products directly tied to weather (eMarketer 2014). A sunny
day is good for selling sunglasses and sport gears, while a rainy day necessitates buying an
umbrella. Over 200 major brands from automotive, customer package goods, insurance, travel,
and pharma industries have partnered with the Weather Company in the hopes of boosting sales
by leveraging weather information (Adweek 2015; Wall Street Journal 2014). Correlations
between weather and purchase are documented with big sales data from Walmart and P&G
matched with location data for 150 million mobile weather app users (BusinessWeek 2014).
Indeed, businesses are keen on how to make money off the weather, since “one third of the U.S.
economy is affected by forces of weather or climate” (Wharton Knowledge 2013). Despite such
managerial importance of weather targeting as depicted in popular trade press, little causal
evidence exists in the literature, especially with field data, on how consumers respond to
weather and how this affects their responses to promotions.
Therefore, this research investigates how sunny and rainy weather influences consumers’
incremental responses to promotions in the setting of mobile products. Consumers may
purchase more in sunny days, but they may not necessarily respond more favorably to
promotions or engender incremental purchases that would not have been realized otherwise. 1
We focus on the role of sunny and rainy weather conditions, after controlling for the effects of
temperature, humidity, visibility, air pressure, dew point, and wind. Literatures in economics
and psychology noted sunlight enables brain to produce more serotonin and leads people to
have positive mood. Thus, in sunny days people buy more stocks (Hirshleifer and Shumway
2003), apparels (Murray et al. 2010), and convertible cars (Busse et al. 2015). On the contrary,
in rainy weather people tend to have depressive symptoms, negative mood, and thus purchase
less (Hsiang, et al. 2013). Our setting here is unique because with mobile technologies
consumers are high on not only weather awareness (people check weather 3 times per day on
weather app), but also location awareness (e.g., via checking physical locations on maps on
smartphones). 2 In traditional retail settings, it is difficult to track location, weather, of
promotion responses simultaneously at individual consumer level. Nowadays, mobile
technologies allow for more precisely identifying individual people on where they are, what
weather they experience, and when they respond to promotions with the mobile setting.
Our investigation exploits a large-scale field data provided by a mobile telecom company
on over ten million smartphone users. The experiment design and unique features of this dataset
allow for identifying the causal incremental effects of weather on mobile promotion responses.
The company randomly sent two ad copies of mobile promotions to smartphone users. One ad
1 We acknowledge AE for this insight. While prior weather studies focus on purchase, we address effects on
promotion effectiveness, or responses to promotions.
2
Also, we acknowledge one anonymous reviewer that if weather is important and targetable even when the
advertised product on the mobile platform in our setting is not directly tied to weather like sunglasses, such
understanding can have non-trivial tangible consequences for marketers to plan promotional campaigns.
2
copy –‘treatment’—contains a prevention frame (“do not miss the opportunity to take
advantage of this deal”). The other ad copy—‘control’—contains a neutral frame with general
greetings (“dear respected customers). To assess incremental sales that would not have been
realized otherwise, our data also have a baseline—‘holdout’—group of smartphone users who
receive no promotion and are comparable to the treatment and control groups. Concurrently,
weather condition data are combined with the field mobile promotion responses. We have timestamped variations in weather conditions, customer locations, and responses to mobile
promotions. In other words, the matched field data capture not only consumer responses to two
different frames of mobile promotions but also weather-related attributes at the consumer
location corresponding to the time when promotions being sent. This allows us to directly test
the causal and incremental impact of sunny and rainy weather.
In an ideal experimental setting, people would be randomly assigned to rainy and sunny
weather conditions. However, because researchers cannot manipulate weather in the real world,
we rely on a quasi-experimental design that randomizes ad exposure (with treatment and control
promotions as described above). Since weather is naturally confounded with geographical
locations, time variations, and user behavior, our data analyses will isolate the effect of sunny
and rainy weather by controlling for customer self-selection bias in multiple ways: (a)
geographical locations in terms of different regions, states, and cities, (b) time variations in
terms of different days and hours of the day, and (c) individual usage behavior heterogeneity
(mobile usage) that may confound the results. To further address endogeneity and self-selection
concerns, our identification strategies rely on the within-person changes in weather (with both
forward-looking forecasts and backward-looking historical weather), as well as deviations from
the normal expected weather conditions.
The data analysis results indicate that consumers are significantly more likely to respond
to mobile promotions in sunny days compared with cloudy days, and significantly less so in
wet-weather days. In terms of economic significance with the odds ratio, sunny weather leads
to about 1.21 times more purchases as a result of the mobile promotions compared with cloudy
weather, and rainy days leads to about 0.9 times less purchases, holding other things constant.
Beyond the response rate (purchase or not), our field data can also test the response hazard, i.e.,
how quick or slow the response is conditional on making a purchase. Hour-by-hour analyses
with survival models suggest the hazard rate of purchase would be 73% faster in sunny weather
but 59% slower in rainy weather relative to a cloudy sky. In addition, the congruence between
ad copy and weather matters: ad copy with a prevention (versus neutral) frame boosts mobile
promotion effectiveness in rainy weather, but diminishes it in sunny days. Another key concern
is the generalizability in results of the mobile products. Thus, we obtained an additional field
experiment dataset from the telecom company with mobile ads promoting a different digital
product and without price discounts (thus only ads and no price promotions). The additional
data provide consistent and robust results for the impact of sunny and rainy weather.
3
Why would consumers respond to mobile promotions differently in sunny or rainy weather?
The psychological theory of “affect-as-information” provides a viable account. Specifically,
mood can be an information input, i.e., information about the advertising message and/or price
promotions, for product evaluation and decision-making (Schwarz and Clore 1983; Pham 2008;
Pham, Lee, and Stephen 2012). Compared to cloudy and rainy days, sunny weather with more
exposure to sunlight would enable brain to produce more serotonin, like in a “heliotherapy,”
and people may experience a better mood and have more favorable ad evaluation, thus more
and faster responses to mobile promotions. On the contrary, people in rainy days should
experience a worse mood state and more negative feelings due to precipitation and inclement
weather conditions, and less favorable evaluation of the advertising message and promotional
deals, thus fewer and slower response. Interestingly, mood congruency hypothesis of the theory
of affect-as-information (Mayer et al. 1992; Murray et al. 2010; Pham and Avnet 2009) also
holds that the rainy weather-induced negative mood matches with the loss-avoidance
prevention frame of ad copy, and thus a prevention frame of ad copy increases the effectiveness
of promotions on rainy and inclement days. The significant interaction between weather and ad
copy framing proffers some evidence for the psychological mood explanation for our results,
rather than a pure behavioral explanation, or activity bias of weather (e.g., users may be more
likely to purchase in good weather simply because they are more on the go and hence use their
phones more). This is because if the weather effect is purely driven by such behavioral
explanation, then consumers should not respond differently to prevention vis-à-vis neutral
frame ad copies under the same weather of the specific location.
Our research makes several contributions. This is the first study that examines the sales
effects of sunny and rainy weather on consumer responses to mobile promotions. Consumers
may purchase more in sunny days (weather not under the control of marketers), but do not
necessarily respond more favorably to promotions (marketing opportunity under the control of
marketers). In fact, critics may argue that promotions should be minimized in sunny days if
consumers are more likely to purchase anyway. Also, while promotions on sunny days drive
more purchases, they may actually be destructive if lowering incremental sales with the
promotions relative to without promotions. In this context, we find that compared with a
holdout group (that receives no ad or promotion), consumers respond more to mobile ads or
promotions on good sunshine days across two mobile products in two datasets, and such
responses indeed lead to more incremental sales that would not have been realized otherwise. 3
These findings of incremental purchases are non-trivial because we know relatively little about
how consumers respond to weather and because consumers nowadays are inundated with and
annoyed by irrelevant ads (Bart, Stephen, and Sarvary 2014). One key challenge is how to cut
through the ad clutter and gain customer attention. To attract consumer attention and responses,
weather-targeted ads and promotions may be effective. Such promotions are hyper-local and
3
We acknowledge SE and AE for this insight.
4
contextually relevant to the smartphone users, thus no longer interrupt the customer experience
but rather enhance it with incremental consumer purchase responses and thus firm sales revenue.
Related, we also advance the limited weather literature by investigating not only purchase
likelihood (Busse et al. 2015), but also the hazard, i.e., timing of responses to the promotions.
In this sense, our findings provide understanding of how soon weather may affect revenues
from mobile promotions. This is important because faster consumer responses to mobile
promotions mean accelerated cash in-flows for the firm with faster turnover ratios and lower
capital financing costs, thus more support for a steady stream of capital funding for timely
promotions and advertising programs.
Further, prior research involves products more directly tied to weather such as sunglasses,
while our work involves products relatively less directly tied to weather. Intuitively, many
companies understand the importance of weather environment when planning their ad
campaigns. It surprises no one by showing that firms can benefit from taking into account
weather for products (sunglasses or umbrellas) that are directly tied to sunny or rainy weather
(Busse et al. 2015; Murray et al. 2010; Pham and Avnet 2012). We show that weather conditions
are important even when the advertised product (mobile digital services) is less directly tied to
weather. This unintuitive finding can be accounted for by the psychological theory of affect-asinformation. That is, weather affects consumers’ mood states and thus their consumption of
digital services that are even less directly tied to the weather. Our results are important by
pointing out a critical path: move beyond the low-hanging fruits of promoting only weatherdependent products toward more advanced weather-sentiment analytics. That would allow for
more invigorating consumer targeting with weather on the burgeoning mobile platform.
Moreover, we extend the mobile targeting literature. Prior research provides insights into
targeting with consumer contexts in terms of time, location, and crowdedness in the
environment (Danaher et al. 2015; Kenny and Marshall 2000; Ghose, Goldfarb, and Han 2012;
Luo et al. 2013; Andrews et al. 2015). The weather has largely escaped research attention as a
contextual variable for mobile targeting. Ironically, weather and mobiles go hand in hand.
Checking weather is one of the most used native applications of smartphones. Consumers rely
on smartphones to check weather conditions before bedtime, as soon as waking up, before
leaving for work, and while deciding on evening plans. In fact, weather effects are often
transient, and it is difficult to dynamically track consumer response over time without mobile
technology. Smartphones are the most accessible and connected devices anywhere consumers
go. Weather is a much under-utilized yet highly pertinent environmental variable in the
marketing literature in general and mobile promotion literature in particular.
Our research also proffers several insights for managers. Marketers are faced with the
challenge of creating persuasive ads that generate sales. Old ‘spray and pray’ tactics are
increasingly ineffective (Bart, Stephen, and Sarvary 2014). Marketing is challenged to create
content to fit unique customer needs and environments. Our study informs managers that
weather may be a viable targeting instrument to increase sales impact of promotions. Everyone
5
discusses the weather. The weather is perhaps one of the most important determinants of
demand and supply for sectors from agriculture, building, energy, insurance, military, retailing,
travel, to tourism and beyond. Historically, marketing with weather has been limited to
seasonality applications, and managers can only avoid and cope with weather. However, with
modern mobile location technologies (Andrews et al. 2015), weather forecasts, and high
accessibility via weather apps, managers may plan and exploit weather for more effective
consumer targeting. Managers can now target customers by current locations and hyper-local
weather conditions to trigger the activation or suppression of ads and promotions. Familiar
brands such as Ace Hardware, Taco Bell, Delta airlines, Farmers insurance, Goodyear and
others are targeting local weather via mobile ad networks (MoPub, Jumptap, and TWC’s
smartphone app). During the 2014 Olympics, Burberry experimented a weather-based
campaign, which promoted products such as branded sunglasses and umbrellas according to
local London weather. Extending this, we illustrate how to boost the effectiveness of mobile
promotions by coordinating sunny and rainy weather with mobile services and products.
Furthermore, marketers’ dream is to reach the customers in the right place, at the right
time, and with the right marketing message. Besides relevant to consumers with the hyper-local
weather, managers should also be creative with the right ad copy framing. Our results suggest
that compared to neutral frame ad copy, a prevention frame substantially assuages the negative
effect of rainy weather on responses to mobile promotions. Thus, managers may take actionable
steps in weather-sentiment analytics with appropriate ad copy, i.e., avoid prevention-framing
on sunny days (because of misalignment with good mood in sunshine days) and leverage it on
rainy days (because of alignment with negative mood in rainy days). This would help craft
effective mobile campaigns for “more bang of the buck.” In a nutshell, our targeting managerial
take-away is: in rainy days, managers should use the prevention frame ad copy—do not miss
the opportunity to take advantage of this special deal—for more mobile promotion purchases,
while in sunny days they should use the neutral frame ad copy—a general greeting of dear
respected customers—for more mobile promotion purchases. The weather undoubtedly affects
daily life, mood, and how consumers interact with the world. With over 150 million users of
weather apps, there are a total of 2 billion checks of weather each day (BusinessWeek 2014).
For marketers, this may implicate a whole new world of business opportunities with weathertargeted mobile ads and promotions.
Background and Hypotheses
Background on Mobile Promotion Effectiveness
Research has explored mobile promotion effectiveness via the role of contextual marketing
(Kenny and Marshall 2000; Luo et al. 2013). Contextual marketing through smartphone devices
is unique in its ability to provide targeted communication across virtually any geographic
location and time of day. Customers are more likely to respond to situation-congruent mobile
6
promotions when they are proximally located (Banerjee and Dholakia 2008). In a fast-food
study, customers were found to have a higher likelihood of coupon redemption when proximally
located to the restaurant (Spiekermann et al. 2011). Users have a higher likelihood of browsing
and interacting with stores that are more geographically-close to the customer (Ghose, Goldfarb,
and Han 2012; Molitor et al. 2013). Echoing this, Luo et al. (2014) found significant interaction
results between location and time in impacting movie ticket sales. Even the crowdedness of the
environment in which a promotion was received influences consumer involvement and
response to mobile ads (Andrews, Luo, Fang, and Ghose 2015). Danaher et al. (2015) uncovers
that time, location, product-category, and promotional specifications are critical to determining
mobile coupon success at shopping malls. Fong et al. (2015) finds that marketers may engage
in geo-conquesting mobile promotions to boost incremental sales to customers at a competitor’s
doorstep. The current study extends this stream of research by exploring the role of weather
conditions for consumer responses to mobile ads and promotions.
Background on Sunlight, Rain, and Mood
A substantial body of literature has examined the effects of weather. Over 60 years ago,
Steele (1951) illustrated the impact of weather on shopping volume for retail stores. Seasonal
effects of weather can impact new product introductions, media spending, shopping cycles, and
market competition (Busse et al. 2015; Simonsohn 2007). A central theme of prior weather
literature is that weather affects people’s mood and thus their shopping activities and other
outcomes. Persinger and Levesque (1983) found that weather accounted for 40% of the
variability in daily mood reports by subjects. With respect to sunny weather, Hirschliefer and
Shumway (2003) note that good mood resulting from sunny days positively influences stock
valuation. People may increase prosocial donation because sunny weather induces better mood.
Exposure to sunlight has a positive influence on mood like in a “heliotherapy” (Schwarz and
Clore 1983; Murray et al. 2010; Busse et al. 2015). Insufficient exposure to sunlight over long
periods of time sometimes leads to severe depression and even higher incidence of suicide.
On the contrary, rainy weather conditions are more likely to create negative feelings in
individuals (Klimstra et al. 2011). In poor weather days with rain, travel becomes burdensome
and dangerous, direct exposure to rain can leave individuals wet and uncomfortable in a lousy
mood state. Rainfall can negatively impact happiness and lead to tiredness and worse feelings
(Denissen, Butalid, Penke, and van Aken 2008). Indeed, in a comprehensive meta-analysis of
weather and climate, Hsiang et al. (2013) find that people in rainy weather experience more
depressive symptoms, less happiness, and more negative feelings such as anxiety, nervousness,
stress, and fear due to precipitation and the unpleasant, hazardous, wet, and inclement weather
conditions (Steele 1951; Hirshleifer and Shumway 2003; Conlin et al. 2007; Murray et al. 2010;
Zwebner, Lee and Goldenberg 2013). Also, in a series polls of weather app users on their
weather-sentiments, Rosman (2013) reports that people are unhappy and in bad mood states
around the time of rainy days and inclement weather such as hurricane. In contrast, people are
reported to experience better mood and are generally happier during pleasant sunny days. Thus,
7
extant literature ties not only sunlight to positive affect and feelings but also rainy weather to
negative affective feelings.
Hypothesis on Weather and Mobile Promotion Effectiveness
The theory of “affect-as-information” from psychology may account for the differences of
mobile promotion responses in sunny and rainy weather conditions. This theory holds that
mood can be an informational input, i.e., information about the advertising message as in Batra
and Stayman (1990) and/or price promotions of the products as in Hsu and Liu (1998) for
product evaluation and decision-making (Schwarz and Clore 1983). For example, “the
experience of positive feelings while thinking about a target object is generally interpreted to
mean that the target is desirable, attractive, valuable, etc., whereas the experience of negative
feelings is interpreted to mean that the target is undesirable, unattractive, not valuable, etc.”
(Pham 2008, p. 161). Indeed, prior research shows that affective feelings are informative for
consumer preferences and product promotion evaluations. Interestingly, one of the earliest
pieces of research exploring the use of HDIF (How-Do-I-Feel?) heuristic in decision-making
utilized sunny and rainy days as positive and negative mood manipulators, respectively
(Schwarz and Clore 1983). Since then, the wealth of knowledge examining the use of affectas-information in decision-making has grown considerably. In consumer psychology, numerous
studies have illustrated the use of mood as information for product evaluations (Stephen and
Pham 2008; Baggozi et al. 1999). People evaluate the same product more favorably when in a
good mood, and less favorably when in a bad mood (Hsu and Liu 1998 see Isen 2001 and
Gardner 1985 for reviews). Also, a state of high affect can lead to more favorable outcomes
(Kahn and Isen 1993) because consumers are more innovative and prone to variety seeking. In
financial economics, Hirschliefer and Shumway (2003) argue that as sunlight boosts people’s
moods, they tend to have a more favorable judgment of equity value and purchase more stocks
in sunny weather vis-à-vis cloudy or rainy and snowy weather. Consistent with this stream of
research, people in sunny (vs. cloudy) weather should experience a better mood and more
positive feelings that engender more favorable ad evaluation, thus leading to higher incremental
sales response to mobile promotions.
On the contrary, holding everything else constant, people in rainy weather experience more
depressive symptoms, less happiness, and more negative feelings such as anxiety, nervousness,
stress, and fear (Hsiang, et al. 2013; Gardner 1985; Klimstra et al. 2011). These negative
feelings engender less favorable evaluations of the same products and promotions, thus leading
to lower incremental sales response to mobile promotions in wet weather, compared to cloudy
skies or sunny weather.
Besides purchase likelihood, affect-as-information theory also predicts that people in
positive mood make faster choices due to a greater reliance on spontaneous heuristic decisionmaking, rather than slower deliberate processing (Pham et al 2001; Isen 2001; Schwarz and
Clore 1983). In other words, if sunny weather enhances mood, and positive mood leads to more
8
favorable promotion evaluations (Hirshleifer and Shumway 2003), then the theory of affect-asinformation would predict that consumers will not only have a higher likelihood but also faster
speed to respond mobile promotions in sunny weather. By the similar logic, if rainy weather
leads to negative feelings due to precipitation and the unpleasant, hazardous, we weather
conditions (Conlin et al. 2007; Murray et al. 2010), then the theory of affect-as-information
would predict that consumers will not only have a lower likelihood but also slower speed to
respond mobile promotions in rainy days. Overall, in line with the psychology literature,
because affect-as-information can be applied to ads as shown in the role of mood for advertising
effectiveness by Batra and Stayman (1990) and to price promotions as shown in the role of
mood in price promotions by Hsu and Liu (1998), we test the following hypothesis for both
mobile ads and promotions effectiveness:
H1: Ceteris paribus, sunny (rainy) weather will lead to higher and quicker (lower and
slower) incremental sales response to mobile ads and promotions, compared with
cloudy weather.
Further, the theory of affect-as-information indicates that feeling interpretations are subject
to a response-mapping dependency, or mood congruency contingency. The affect information
will be interpreted differently contingent upon matching or mis-matching, i.e. “query and
response-mapping dependency” (Pham 2008, p. 178). People in good moods find positive
information more salient and pay more attention to positive message, while in bad moods find
negative information more salient and pay more attention to negative message (Isen et al. 1978;
Mayer et al. 1992; Schwarz and Clore 1983). Indeed, Pham (2008) holds affective information
that matches (mismatches) with evaluative content produces more favorable (unfavorable)
framed cognitive appraisal. Also, Pham and Avnet (2009) find that mood heuristics interact
with message framing, such that the effects of positive mood on consumer preference is
attenuated by prevention frame (because positive mood mismatches with a negative lossavoidance prevention frame). In line with this strand of research, the mood congruence
contingency would suggest that rainy weather-induced negative mood matches the prevention
ad copy, while sunny weather-induced positive mood mismatches it. In other words, the
interaction between sunny weather and prevention (versus neutral) frame would be negative in
influencing consumer responses to mobile promotions, while the interaction between rainy
weather and prevention frame would be positive (i.e., negative multiplied by negative becomes
positive). So we test the following hypothesis:
H2: Ad copy with a prevention (vs. neutral) frame has a negative interaction with
sunny weather, but a positive interaction with rainy weather in influencing the
incremental sales response to mobile promotions.
9
Field Data, Identification, and Results
Field Data
The ideal test of weather effects is to conduct a randomized field experiment by
manipulating weather. Yet, this is impossible because weather conditions are not dictated by
researchers. No researcher can randomly assign one day as sunny and another day as rainy.
However, we can randomly assign mobile promotions and then match the data with weather
records, i.e., augmented field data. Thus, we conducted a quasi-field experiment to explore the
effect of weather on consumer responses to mobile promotions.
Our field data is provided by one large mobile telecommunications company. The mobile
campaigns promote discounted add-on services of video-streaming to smartphone users. The
basic mobile services are provided by the telecom company when users subscribe to a specific
cell phone plan such as call and SMS services. However, the add-on services must be purchased
additionally. This add-on service allows mobile users to watch videos on their smartphone
devices. Randomized promotions are sent to over 10 million (10,072,534) mobile users via
push-open SMS notification campaigns from September 2nd, 2013 through October 3rd, 2013.
There are two ad copies in the promotions, where one is with the prevention frame and the other
with neutral frame. 4 For the prevention frame ad copy (treatment), the SMS begins with “Do
not miss the opportunity to take advantage of this special deal!” This treatment ad copy
emphasizes on preventing the users from losing the deal (i.e., a negative tone with the wording
of not missing in ad copy priming) in order to promote and sell the mobile product. For the
neutral frame ad copy (control), the SMS begins with a general greeting of “Dear respected
customer.” This control ad copy emphasizes on a generic salutation (i.e., a neutral, non-negative
tone in ad copy priming) in order to promote and sell the mobile product. The rest of the SMS
promotions is the same across the treatment and control: “Subscribe to the mobile videostreaming service of [the Wireless Service Provider] for only ¥3 per month! Watch video
episodes of the most popular TV series on your mobile devices on-the-go! The regular price is
¥6. Purchase by replying “Yes” to this SMS within the next 48 hours!” All these SMS
campaigns are promoting the same mobile service, the discounted and regular prices are the
same, and all SMSs are sent to users at 9am everyday consistently. Interested users could
4 The mobile company was engaging in A/B tests to assess whether prevention frame leads to different results
from neutral frame in SMS promotions, so there is no promotion-framing due to managerial decisions. Also, a
promotion-framing might invite consumer reactance because that would directly cold sell the products to
consumers in their personal mobile devices. As manipulation checks, pilot tests and focus groups confirm that
users indeed interpret the wording of ‘do not miss the opportunity’ as loss-avoidance type negative message, i.e.,
preventing users from the negative outcomes of missing the deal, rather than gain-approach type positive message
of directly pushing users to grab the deal. Also, the neutral frame with the general greeting is was confirmed as
neutral rather than prevention frame in interpretations. Using these two ad copies, there are a total of 38 SMS
campaigns. Though not perfect balanced, 18 prevention vs 20 non-prevention SMS campaigns exist purely due to
the corporate agenda, rather than researcher intervention. Variations in message length could be a confound, given
the small screen sizes of mobile phones. However, the original SMSs in Chinese language were not significantly
different in length between the treatment and control conditions. Thus, promotion length confound is not a serious
threat. See Appendix B of the SMSs.
10
respond by purchasing the service with promoted price. 5 The purchase prices would be charged
immediately to the users’ wireless phone bill. All SMS campaigns are randomly sent. For the
randomization procedures, the SAS software’s random number generator and RANUNI
function are used to generate a random value from a uniform distribution. After that, the random
numbers are sorted in sequence to extract the sample to send to mobile users (Deng and Graz
2002). The targeted users did not previously subscribe to this video-streaming service, nor did
they receive a similar SMS promotion from the wireless service provider.
In order to identify the incremental sales effect of SMS promotions, 6 we randomly selected
20,000 mobile users who they are comparable to the treatment and control groups but didn’t
receive any promotion ads during our field experiment time window from the company. These
users are regarded as the baseline (holdout) group. Similar to mobile users received the SMS
promotions, users of the holdout group did not previously subscribe to this video-streaming
service, nor receive a similar SMS ad. Thus, the results can gauge incremental sales responses
due to mobile promotion, compared with the counterfactual of no mobile promotion.
To explore the weather effects, we collected the concurrent weather data at the user location
during the time the SMS ads were sent. We complement the mobile promotion data with a rich
database of weather variables across cities of mobile users at both daily level and hourly level.
We develop an algorithm based on machine learning techniques to automate the data scrapping
process to collect the weather data online from weather underground. 7 In this study, we focus
on the effect of sunny weather (all sunny and partial sunny), cloudy weather (cloudy or overcast
days without much sunlight or clear sky), and rainy weather (shower, rain and storm). Figure 1
presents the frequency of the individual-daily exposure of the distribution of weather conditions.
As shown in Figure 1, there are enough variations since sunny has over 3 million observations
and cloudy over 4.5 million observations, for instance. Because other weather conditions (snow
and fog) have very few cases, we mainly focus on sunny, cloudy, and rainy weather conditions.
[Figure 1]
Figure 2a and 2b show the initial model free evidence. Figure 2a presents the incremental
sales effect of mobile promotion used in our field experiment. The purchase rate of the holdout
group (without mobile promotions) is zero, which means mobile users will generally not
purchase this service without promotion no matter what is the weather condition in this context.
The zero purchase rate without promotion is reasonable here because the mobile company
provides more than 200 add-on mobile service packages. Without promotions, it is quite
difficult for users to notice the existence of a specific add-on service to make a purchase. In this
5
This field experiment data do not have a condition with only ads and without the promotional price discounts.
Thus, we conduct an additional field experiment on a mobile reading app platform with only ads (no discounts).
6
We acknowledge AE for this insight.
7 Per one reviewer, we have paid a handsome amount of cash and purchased the forecast weather data to test the
effect of unexpected change. The results are robust.
11
sense, the mobile promotion responses are incremental and would not capture otherwise. 8 This
also supports the notion that the results aren’t due simply to more mobile usage in sunny days.
Figure 2b provides the model-free evidence for the effect of weather on consumer responses
to mobile promotions. The response rate is the highest (1.05%) when the weather condition is
sunny, and the response rate of rainy weather is the lowest (0.64%). Thus, these findings suggest
initial evidence that sunny weather has a positive effect while rainy weather has a negative
effect on incremental responses to mobile promotions.
[Figure 2a, Figure 2b]
Identification Strategies
Because weather is naturally confounded with geographic locations, to conclude that
weather drives purchases, researchers need to rule out that weather is not simply correlated with
geographic locations and other unobserved drivers of demand. To enhance results validity and
rule out alternative explanations, we have multi-faceted identification strategies. First, we take
the user self-selection into account. Self-selection concerns arise due to the possibility that
people may self-select to the geographic location in which they live. Geographic locations with
certain latitude and longitude measures may be correlated with weather conditions and drive
the purchases. If so, then it is not weather but rather location, as a confound, that drives
consumer responses to mobile promotions. In other words, researchers might systematically
observe specific purchase types in specific locations due to geographic characteristics or the
characteristics of residents at these locations. And these locations may have systematically
different weather conditions. Specifically, researchers might observe non-purchases more in
rainy weather locations, but the correlation might be spurious. For instance, researchers may
observe rain in Seattle (not surprising where weather is often rainy) and few purchases. But this
might merely be caused by the fact that the type of people living in Seattle are less likely to buy
on their smartphones than the type of person living in, say, Los Angeles (where the weather is
usually sunny). To deal with this confound, our model controls for locations with fixed effects
of 31 providences covering 344 cities (see Appendix), which have different latitudes and
altitudes that may naturally correlate with weather. Second, we test the effects of sunny, rainy
or cloudy days, over and beyond the effects of a host of related variables. Such variables as
temperature, humidity, visibility, air pressure, dew point, and wind speed and directions are
controlled for in our model analyses. Third, although we test the psychological mood effect of
weather, we also need control for pure behavioral effect of weather (e.g., people more likely to
go about in sunny days and less so in rainy days). So we consider the differences between
weekdays and weekends. People behave differently on weekdays and weekends. On weekends
people may have more freedom to choose to go out if sunny and stay at inside if rainy outside.
8
Data in our second randomized field experiment suggest a non-zero holdout case. That is, without mobile
promotion, users will make some purchases. That helps generalize our findings, as discussed subsequently.
12
However, on weekdays people may have less such freedom if they need to go to work regardless
sunny or rainy weather. For this purpose, we only use the weekday data to more precisely
identify the effect of weather on mobile promotion responses. 9 Fourth, we consider the
confounding day effect. As our field experiment is implemented for one month, the
unobservable elements might change across the 31 days. Though we have considered the
difference between weekdays and weekends, we still need to take care of the day effect. For
instance, people may purchase more at the beginning (vs. end) of one month due to limited
budget. Hence, we control for the day fixed effect to ensure the validity of our model results.
Main Effect of Sunny and Rainy Weather Condition
Our econometric model estimates individual user’s likelihoodi to respond mobile
promotion. The latent likelihood is a logit function of weather conditions and covariates.
Following Guadagni and Little (1983) and Luo et al. (2013), our logit model has an i.i.d.
extreme value distribution of the error term:
Mobile Promotion Response Likelihoodiweather =
exp(𝑈𝑈𝑖𝑖 )
1+exp(𝑈𝑈𝑖𝑖 )
𝑈𝑈𝑖𝑖𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑒𝑒𝑒𝑒 = 𝛼𝛼0 + 𝛼𝛼1 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖 + 𝛼𝛼2 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 + 𝛽𝛽𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 + Δ𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 + 𝜇𝜇Ω𝐑𝐑(𝐢𝐢)
+ 𝜌𝜌𝐗𝐗 𝐢𝐢 + 𝜀𝜀𝑖𝑖
(1)
where i indicates the mobile user. The promotion response here equals to 1 if the user responds
to the SMS promotions by making a purchase, and 0 if otherwise. We use two dummies, Sunnyi
and Rainyi, for the three weather conditions of sunny, cloudy, and rainy to identify the weather
condition of the day a mobile user received the ad message, respectively. Sunnyi is equal to 1 if
the weather is sunny on the day that the mobile ads is exposed to individual i and 0 if otherwise.
Rainyi is equal to 1 if the weather is rainy on the day that the mobile ads is exposed to individual
i, and 0 if otherwise. Thus, when both Sunnyi and Rainyi are 0, the weather is cloudy, the
baseline weather. Framingi is equal to 1 if the SMS campaign includes prevention frame ad
copy and 0 if with the neutral frame ad copy. We also control for other weather-related variables,
i.e., highest temperature, temperature range 10, humility, dew point, air pressure, wind speed as
continuous variables and eight dummies for wind direction (all correlations among all variables
are not high, with largest correlation less than .15 in the data; Too windy with high wind power
can drive negative affect, Denissen, Butalid, Penke, and van Aken 2008). Ω𝑹𝑹(𝒊𝒊) is a vector of
30 geographical dummies (31 provinces) to control for location-based fixed effects. 11 In
9
Also, the effect of sunny weather and rainy weather may reasonably different between day and night. Accounting
for this, our subsequent hour-by-hour analyses with hazard models only use the hourly weather data of first 8 hours
(from 9 am to 5 pm) when most people are at work regardless of the weather conditions.
10 Temperature range equals to highest temperature minus lowest temperature in Celsius degree, indicating the
uncertainty of temperature in a day.
11
We also test and consistently support the weather effects by controlling for locational effects with 344 Cities (as
13
addition, we control for rural, a location-related dummy variable, which equals to 1 if the
mobile user live in rural areas and equals to 0 if the mobile user is from urban areas. This helps
control for different income differences between rural and non-rural areas. Finally, we control
for Xi, a vector of fixed effects for 31 days of the SMS promotion period.
Table 1 reports the summary statistics of the data. The mean value of consumer responses
shows the purchase rate during our observation window is about 0.78%. This purchase rate is
consistent with the rate between 0.6% and 2% for mobile coupons in Asia (eMarketer 2014).
[Table 1]
Table 2 presents the empirical estimates of weather effects. Column 1 is a simplified model
specification. The crucial coefficients of interest are Sunny and Rainy, which capture the impact
of weather on consumer response likelihood of mobile promotions. Column 2 includes dayfixed effects. Column 3 adds other weather variables including highest temperature,
temperature range, humility, dew point, air pressure, wind speed, wind direction. Column 4
enters both weather-related covariates and the day fixed effects, which is our full model.
Column 5 shows the model with robust standard error. As Table 2 shows, across all models,
sunny weather has a consistent and statistically significant positive effect, while rainy weather
has a consistent and statistically significant effect on the purchase of the service (all p < 0.05).
Thus, an increase in consumers’ likelihood to respond to mobile promotions in sunny weather
and a decrease in rainy weather, compared with the baseline cloudy weather as in H1. Besides
the effect of weather, the prevention frame of SMS leads to a higher purchase rate, which is
consistent with the regulatory relevance where prevention-message is more persuasive (Aaker
and Lee 2006) and the notion of negativity in motivation where negative priming is more
effective in changing and improving people’s actions (Finkelstein and Fishbach 2012).
Additionally, in terms of economic significance with the odds ratio, sunny weather leads to
about 1.21 (=e0.1880) times more purchases of mobile promotions compared with cloudy weather.
Rainy weather leads to about 0.9 (=e-0.1068) times less purchases compared with cloudy weather.
[Table 2]
Interaction Effect of Weather and Ad Copy Framing
This section explores the interaction effect of weather and ad copy framing. According to
the affect-as-information theory and query-mapping of mood (Pham 2008), we test whether
there are significant positive effect of 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 ∗ 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 , and negative interaction effect of
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖 ∗ 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 . The full interaction model is shown as below:
U iweather
= 𝛼𝛼0 + 𝜃𝜃1 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 ∗ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖 + 𝜃𝜃2 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 ∗ 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 + 𝛼𝛼1 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖 + 𝛼𝛼2 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖
+ 𝛾𝛾𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 + Δ𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 + 𝜇𝜇Ω𝐑𝐑(𝐢𝐢) + 𝜌𝜌𝐗𝐗 𝐢𝐢 + 𝜀𝜀𝑖𝑖
shown in Appendix) in addition to the providences.
14
(2)
Table 3 presents the estimated results. Column 1 reports the results of the interaction effect
between prevention framing and sunny weather effect, where the baseline is all other weather
conditions; Column 2 reports the results of the interaction effect between prevention framing
and rainy weather effect. Column 3 displays the two interaction items between the weather and
prevention framing together. These results show that when the weather is sunny, prevention
framing indeed leads to lowered likelihood to respond to mobile promotions (coef.=-0.2350,
p<0.05). However, when the weather is rainy, prevention framing leads to increased likelihood
to respond to mobile promotions by making a purchase (coef.=0.2239, p<0.05). These findings
are consistent with a plausible mood explanation of the weather effects. To the extent that sunny
weather induces good mood that is incongruent with negative ad messages and that rainy
weather induces bad mood that is congruent with negative ad messages (Pham 2008; Pham,
Lee, and Stephen 2012; Kahn and Isen 1993), it is sensible that ad copy with a preventionframing diminishes mobile promotion effectiveness in sunny weather but boosts it in rainy
weather as in H2. Thus, mobile ad framing can substantially regulate the weather effects
identified by this field dataset. 12 Also, the significant interaction between weather and ad copy
framing reveals the evidence, to some extent, for the psychological mood explanation for our
results, rather than a pure behavioral explanation or weather-based activity bias (e.g., users may
be more likely to purchase in good weather simply because they are more on the go and hence
use their phones more). This is because if the weather effect is purely driven by such behavioral
explanation, then there should be insignificant interactions between ad copies and sunny or
rainy weather conditions, i.e., consumers should not respond differently to prevention vis-à-vis
neutral frame ad copies under the same weather of the specific location.
[Table 3]
More Identification with Response Hazard
To further establish the effects of weather condition, we conduct hour-by-hour survival analyses
to test the effects of sunny and rainy weather on response hazard, or purchase speed. Mobile
users who were interested in the mobile video-streaming service should reply “Yes” within 48
hours due to the preset deal deadline in the mobile promotions. Thus, this timestamped field
data on the promotion purchases allows us to test how fast they respond to the mobile promotion.
As the weather condition may change each hour of a day, we collect weather data at the hourly
level and develop an hour-by-hour hazard model. Table 5 provides the purchase rate of the
hourly sunny, cloudy, and rainy weather. The promotion purchase rate is the highest (0.00022)
12
These results also contribute to literature on prevention-framing. Although prior findings illustrate the efficacy
of prevention-framing in goal attainment with lab studies (Aaker and Lee 2006), we add more evidence with a
large scale real-world field dataset. Our results not only corroborate lab studies and gauge its main effects on sales
for the firm, but also reveal that messages framed with an otherwise effective prevention framing actually weaken
mobile promotion impact on sunny days. These field evidences are critical for managers who may doubt the
generalizability of prevention-framing results from lab experiments to actual sales purchase field settings.
15
in hourly sunny weather and the lowest (0.00013) in hourly rainy weather, thus adding more
evidence for weather effects on consumer responses to mobile promotions at the hourly level.
[Table 4]
About the hazard of responses to mobile promotions, we first provide the model-free
evidence of Kaplan-Meier plot in Figure 3. The Kaplan-Meier plot suggests that the slope of
the line for the sunny weather is steeper than cloudy, and the slope of the line for the cloudy
weather is steeper than rainy. This suggests that in sunny (rainy) weather, people are not only
more (less) likely to respond but also take shorter (longer) time to respond mobile promotions.
Next, we build a hazard model to formally test the responses speed (fast or slow in purchasing).
[Figure 3]
The proportional hazard model is shown as below:
h(t) = h0 (𝑡𝑡)exp(𝑈𝑈𝑖𝑖𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑒𝑒𝑒𝑒 )
𝑈𝑈𝑖𝑖𝑖𝑖𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑒𝑒𝑒𝑒 = 𝛼𝛼0 + 𝛼𝛼1 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝑖𝑖𝑖𝑖 + 𝛼𝛼2 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝑖𝑖𝑖𝑖 + 𝛽𝛽𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 + Δ𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶
+ 𝜇𝜇Ω𝐑𝐑(𝐢𝐢) + 𝜌𝜌𝐗𝐗 𝐢𝐢 + 𝜀𝜀𝑖𝑖𝑖𝑖
(3)
where the dependent variable is h(t), which signifies the hazard rate for purchase response at
hour t. h(0) is the baseline hazard, indicating the hazard when all variables in the function of
U𝑖𝑖𝑊𝑊𝑊𝑊𝑊𝑊𝑊𝑊ℎ𝑒𝑒𝑒𝑒 are equal to zero, which represents the cloudy weather condition. HourlySunnyi is
equal to 1 if the weather is sunny at hour t that the individual i experienced and 0 if otherwise.
HourlyRainyi is equal to 1 if the weather is rainy at hour t that the individual i experienced and
0 if otherwise. The baseline of hazard model is hourly cloudy weather. Weather-related
covariates added to this model are also at hourly level. Other covariates, including framing,
location fixed effect, day fixed effect, are the same variables used in the previous logit model.
To further control for the behavioral effects of weather, we analyze data in the working time
period during the day of weekdays. That is, the impact of weather might differ by night versus
day, since there is no sunlight at night, and also people will be influenced less by rainy weather
when they are often indoor at night. Taking this into account, our hazard models only use the
hourly weather data of first 8 hours (from 9 am to 5 pm), which is the day time period and also
most people are at work during this time period from Monday through Friday regardless the
weather conditions.
Results of Cox proportional hazard models are presented in Table 5. As an hour-by-hour
analysis across individuals, the number of observations in hazard models (55,164,292) is
substantially larger than that in logit model (6,744,884). 13 Column 1 in Table 5 enters
Hourly_Goodweatheri, Hourly_Badweatheri and location fixed effect at the province-level.
Column 2 also enters the day fixed effects. Column 3 enters prevention framing and other
weather-related variables in the full model. The results in Table 5 indicate consistently and
statistically significant positive sunny day effects (coef.=0.5481, p<0.05) and negative rainy
13 In the hour-by-hour survival analyses with over 55 million rows in the hazard model, each run with HPC (high
performance computing) with parallel processing via university big machine costed around two weeks.
16
weather effects (coef.=-0.5247, p<0.05) on the hazard rate of responses to mobile promotions.
Specifically, according to the results in column 3, when the hourly weather is sunny, the hazard
rate of purchase is 73% higher, compared with the cloudy weather. The hazard ratio in rainy
weather is 59% lower, compared with cloudy weather. As such, as in H1, compared with cloudy
weather, sunny weather engenders not only higher but also quicker mobile promotion responses,
while rainy weather engenders both lower and slower responses in terms of hour-by-hour
hazard analyses.
[Table 5]
Further Identification with Unexpected Changes in Weather
One may surmise that it is not the weather per se, but rather the unexpected changes in
weather that may affect consumer mood and purchase. Indeed, Easton (2012) holds that
“immediately following a rain shower, when the sun bursts out and sparkles on puddles through
clean, fresh air, colors brighter and senses somehow keener, those moments are profoundly
exhilarating… Perhaps it is not the sunshine that matters so much as the pleasure we get when
our weather changes.” So we test whether the effects hold with unexpected changes in weather
conditions. We considered two types of unexpected changes: backward-looking (change from
daily weather condition of t-1 period to t period) and forward-looking (change from forecast
weather condition to real weather condition in t period).
We first consider the backward-looking unexpected change. More specifically, if the
weather yesterday is sunny, people may possibly suppose it will still be sunny today. However,
what if the weather condition changes today? We calculate the change of weather (ΔBackward
Change) with Weather Conditiont, minus Weather Conditiont-1, and its net value ranges from 2 to +2. Better_Backward equals 1 if △Backward Change is positive (change to better weather
than yesterday) and 0 if otherwise; Worse_Backward equals 1 if △Backward Change is
negative (change to the worse weather) and 0 if otherwise. The baseline of our model for this
section is No Change in weather.
The results in Table 6 confirm that when the weather at period t is better than period t-1,
the response likelihood of mobile promotion at period t is higher, as expected. Further, the
response likelihood at period t is indeed lower if the weather at period t is worse than period t1. Interestingly, consistent with the prospect theory, the effects of ‘worse’ weather are
significantly larger than those of ‘better’ weather in magnitude of impact. Additionally, column
3 in Table 6 shows that after controlling for the backward-looking unexpected changes in
weather, the effects of current sunny weather and rainy weather are still consistent and
significant (all p<0.05). Also, the incremental effect of ‘worse’ is actually significantly larger
than the effect of ‘rainy’, judged both by the effect size and the level of statistical significance.
This means that if yesterday it was rainy already, there is actually relatively little effect of rain
today. The key insight from this table is that the two things that matter most are ‘sunny’ and
‘not worse than yesterday’ concerning weather effects on mobile promotion effectiveness. Thus,
17
once again, these results based on surprises of weather reveal more empirical evidence for the
robust positive effects of sunny weather and negative effects of rainy weather on mobile
promotion responses.
[Table 6]
The forward-looking unexpected change is also of great importance. Though the forecast
is more and more accurate nowadays, the actual weather may still differ from the forecast. 14 If
people expect the weather of a specific day is sunny due to the forecast report, but the real
weather they experience at that day is actually rainy, they may have a worse mood, relative to
people who also experience the rainy weather and know the forecast is rainy accurately. Thus,
the forward-looking unexpected change from forecast to real weather may influence people’s
mood and their responses to mobile promotions. We calculate theΔForward Change with real
Weather Conditiont minus Weather Forecastt. Thus, Better_Forward equals 1 if △Forward
Change is positive (change to better weather than forecast). Worse_Forward equals 1 if △
Forward Change is negative (change to worse weather than forecast). The baseline is still No
Change Weather, meaning the forecast weather is accurate for t period. Table 7 presents the
results. The estimates affirm that if the real weather condition is worse than the forecast, people
are less likely to respond the mobile promotions. Again, results in Table 7 suggest that while
giving some weight to ‘rainy’ weather, the variable of ‘sunny’ is by far most important and
‘worse than predicted’ is also quite relevant in influencing responses to mobile promotions.
Importantly, column (3) still indicates robust positive sunny day effects (coef.=0.3325, p<0.05)
and negative rainy weather effects on responses to mobile promotions (coef.=-0.1008, p<0.05),
even after controlling for forward-looking unexpected changes.
[Table 7]
Even More Identification with Weather Deviation
Although generally sunny weather has positive effects and rainy weather has negative
effects, these weather conditions might have stronger or weaker effect in areas with different
normal weather. For example, the effect of rainy weather should be different in Seattle where
rainy is the normal weather, relative to Los Angeles where sunny is the normal weather. Across
Earth’s climatic Torrid, Temperate, and Frigid zones, it is well established that weather varies
systematically. 15 For example, locations in the subtropical zone have more rainfall, than those
in the temperate (tepid latitudes) zone. According to the climate conditions, China can be
divided into northern and southern region on the basis of the Qinling-Huaihe line. The northern
14
15
Per an anonymous reviewer, we have paid a substantial amount of fund to purchase the real forecast data.
In the Torrid or tropical Zone (Tropic of Cancer on the north and the Tropic of Capricorn on the south), the sun
seasonally passes directly overhead. The sun moves into the Southern Hemisphere right after the September
equinox, moves over the southern tropical regions, arrives the Tropic of Capricorn at the December solstice, and
returns northwards to the Equator. In the Temperate Zones (tepid latitudes), the sun never directly shines overhead,
so the climate is mildly warm or cold with four seasons: Spring, Summer, Autumn and Winter. In the Frigid Zones
(polar regions), the sun may shine at the midnight sun, and at the pole there are 6 months of daylight and 6 months
of night. These are the coldest parts of the earth covered with ice and snow.
18
region belongs to the temperate zone, while the southern region belongs to the subtropical zone.
Based on this classification, 31 provinces are divided into 15 in the south and 16 in the north.
Figure 4 depicts the frequency of sunny, cloudy and rainy weather across the regions. In the
south, the highest frequency is cloudy weather, followed by rainy and sunny weather. On the
contrary, in the north, sunny weather is the normal weather condition of highest frequency,
followed by cloudy and rainy weather. Thus, in the south region people might expect
precipitous wet weather and are accustomed to rainy days, so relatively rare events of sunny
days would have stronger positive effects on response to mobile ads, more so than in the north.
Table 8 column (1) and (2) present the estimates of weather effect in the south and north
separately. The coefficient of sunny weather of south (coef.=0.2904, p<0.01) in Column (1) is
larger than that of north (coef.=0.0406, p<0.05) in Column (2), which indicate the relatively
fewer events of sunny weather indeed have stronger effect on response to mobile ads in the
south where people commonly expect rainy weather. Also, the larger coefficient of rainy
weather in the north in Column (2) than the south in Column (1) also proves that the relatively
rare events of rainy weather have stronger negative effect in the north where people commonly
expect sunny weather.
[Figure 4, Table 8]
Additional Robustness Check
In this section we conduct various robustness checks. We first check different coding
schemes of the weather variables, with alternative measures of sunny and rainy weather and the
extent of sunshine in sunny days. We then test the interactions between sunny/rainy weather
and temperature.
Falsification Tests with Alternative Measures of Sunny and Rainy Weather
To affirm the positive sunny weather effects, the data should pass a falsification test. That
is, if we group cloudy and rainy weather as the baseline, the positive sunny weather effects
should still hold, if not stronger in effect size. This is because once combined together cloudy
and rainy weather still cannot induce better mood than sunny weather does. Similarly, if we
group cloudy and sunny weather as the baseline, the negative rainy weather effects should still
be statistically significant because combined together cloudy and sunny weather still cannot
induce worse mood than rainy weather. Table 9a reports the effects of sunny weather separately,
and Table 9b reports the effects of rainy weather separately. Columns 1-4 in Panel A replicate
model (1) with a minor change, in that sunny weather is the only weather dummy variable used
for estimating the model. The baseline of Columns 1-4 in Table 9a is all the other weather
conditions (cloudy and rainy weather), while the baseline of Columns 1-4 of Table 9b is all the
other conditions (sunny and cloudy) when testing the effects of rainy weather separately. These
estimates show robust positive sunny weather effects and negative rainy weather effects.
Interestingly, comparing the effect size, sunny weather effects are not only positive and
significant statistically, but also relatively stronger when the combined cloudy and rainy
weather is the baseline (coef.=0.2153, p<0.05, Table 9a) than when only cloudy is the baseline
19
(coef.=0.1880, p<0.05, Table 2). Similarly, rainy weather effects are not only negative and
significant statistically, but also relatively stronger when the combined cloudy and sunny
weather is the baseline (coef.=-0.1792, p<0.05, Table 9b) than when only cloudy is the baseline
(coef.=-0.1068, p<0.05, Table 2). Again, the results confirm that the sunny weather’s positive
effects and rainy weather’s negative effects are robust.
[Table 9a, Table 9b]
Robustness with All Sunny and Partially Sunny Conditions and Sunlight Hours
In addition, if sunny weather boosts consumer responses to mobile promotions because
sunlight induce better consumer mood, then people may have an even better mood when the
weather is sunny all day, relative to when it is only partially sunny. Thus, one should expect
the positive effects of sunny weather on mobile ads response to remain significant for both
sunny and partially sunny days, but the positive effects to be even stronger for all sunny days.
To test this, our data contains two types of sunny weather: “All Sunny” conditions marked by
consistent clear and sunny conditions throughout the entire day, and “Partially Sunny”
conditions marked by progressively sunny conditions following previously cloudy or inclement
conditions, with less sunlight than all sunny weather. Allsunnyi is equal to 1 if the weather is
sunny throughout the whole day and 0 if otherwise. Partialsunnyi is equal to 1 if the weather
has turned sunny from some other weather conditions and 0 if otherwise. The results are
displayed in Table 10. Column (1) presents the main effect with all sunny conditions and some
sunny conditions respectively. As expected, the coefficients of Allsunny (coef.=0.3525) and
Partialsunny (coef.=0.1246) are both significantly positive (p<0.05), which affirms the
robustness of the positive effect of sunny days on mobile promotion responses. The estimates
demonstrate that the positive effects of all sunny days are indeed stronger than partial sunny
days. In Column (2) of Table 10, we dropped all observations on partial sunny days, and
replicated the main effect model. The estimation result using the subsample without the all
sunny condition is shown in Column (3). Again, results demonstrate stronger positive effects
of all sunny weather comparing to partial sunny condition.
[Table 10]
To ensure the validity of our result, we also test the effect with sunlight hour of the day,
which is a continuous variable of sunshine, on sunny days. The results are presented in Table
11. The results suggest that longer sunlight hours indeed induce higher response likelihood to
mobile promotions, thus shoring up more evidence for the sunny weather effects.
[Table 11]
Robustness with Interaction between Weather and Temperature
Other variables such as temperature may also affect people’s behavior in sunny or rainy
weather. Sunshine is usually associated with nice weather, but when the temperature is too high,
sunlight may have a negative effect on mood (Cohen 2011; Zwebner, Lee, and Goldenberg
2014). Thus, as another falsification test, we check the robustness via the interaction between
20
weather conditions and temperature. 16 Results in column (1) of Table 12 show that the
interaction between sunny weather and highest temperature (highest degree of the day) is
significantly negative. This is reasonable because a sunny day if too hot would no longer be a
good weather that brings a good mood to lay consumers. Also, the results support the positive
interaction effect between rainy and highest temperature. This is also reasonable because the
warmer rainy day is a relative better weather condition and better mood when compared to
colder rainy day to lay consumers.
We then checked the interaction effects between weather and temperature range (highest
minus lowest degree of the day). This is also an important check because greater variations in
temperature ranges would instill more uncertainty about the weather in people. So it makes
sense to check whether the results still hold when accounting for temperature ranges. Column
(2) in Table 12 shows the insignificance of interaction effect between weather condition and
temperature range (p>0.1). After controlling for these interactions between weather and
temperature (high degree and high-minus-low range), results in columns (1) and (2) also
confirm the robust and consistent positive sunny weather effects and negative rainy weather
effects on consumer response to mobile promotions.
[Table 12]
Replication with A Different Mobile Product
To ensure the reliability and validity of the effect of weather on mobile promotion response,
we take replication and generalizability of results into consideration. 17 Our second experiment
uses a different type with digital products of book-reading digital service (similar to Kindle but
on the mobile platform). That is, this new experiment involves APP ads targeting digital book
reading services, while the first field experiment involves SMS promotions targeting video
streaming services. Also, the first experiment data do not have a condition with only ads, and
without the promotional price discounts. Thus, we conduct another field experiment to include
ads only (without price discounts) as a comparison benchmark to gauge incremental purchase
responses. A new feature in the second field experiment can help identify the psychological
mood effect of weather. That is, we use two different book recommendations in app ads (all
without price discounts) in our experimental manipulation: mood-related books vis-à-vis mood
neutral books. That is, among the APP campaigns, one is with mood-related romantic title
books (“Love after pain” or “Loneliness, a glorious failure”), while the other is with moodneutral biographical title books (“Team of Kun Chen” or “Story of old-time ladies”). As
manipulation checks, ratings from independent coders confirm that the titles with mood-related
16
The interaction effect between weather and wind power is not included in the robustness check as there is no
extreme value of wind power.
17
We acknowledge SE, AE, and one anonymous reviewer for this insight. While our results hold for two different
mobile products (video-streaming and book-reading), we caution broad generalization of results to other products.
See more discussion in the limitation section.
21
books (vs. the biographical books) are confirmed to trigger more affect feelings such as
romantic love or loneness as expected; see appendix B for the specific content of notification).
Hence, this field data allow us to test if the weather influences the mood of mobile users and
has an impact on their response to the mobile promotion, then the weather should have a
stronger impact on the mood-related books comparing with neutral-related books. Similar to
prior field data, we also have our data also have a baseline ‘holdout’ group of smartphone users
who are similar but receive no ads, so as to assess the incremental sales that would not have
been realized otherwise. Thus, the two field datasets differ from each other in terms of targeted
products (video-streaming vs. book-reading), marketing instruments (price discounts vs. pure
advertising), and delivery mobile technologies (SMS vs. APP). But both products are similar
(streaming and reading are both spare-time entertainment) and both somewhat weather-related
(consumers are more apt to stream or read in bad weather, which means that our results may be
more conservative). Still, the second study serves at least as a replication of the weather effects
and helps test the mood explanation of sunny and rainy weather effects on consumers’
incremental purchase responses to mobile promotions across two product categories.
The reading APP is one of the most popular mobile reading applications in China, providing
more than 400,000 mobile-books to over 300 million users. Randomized mobile promotions
were sent to 85,000 app users via app notifications nationwide on May 7th, May 19th and June
4th, 2015. All these APP campaigns promote in-app book purchase. The 85,000 randomly
selected mobile users will only receive the ad notification once during the experiment period.
After receiving the ad notification, interested users could respond to the ads by making an inapp purchase. The purchase prices would be paid immediately by charging to the users’ app
account. All campaigns are randomly sent and the randomization procedures are the same as
the first field experiment. We also randomly selected a holdout group of 15,000 mobile users
who didn’t receive the ad notification to identify incremental effects. Again, we collected the
concurrent daily-level weather data at the user location during the time the APP notification
promotions were sent. The effectiveness of mobile promotions in this new experiment is mobile
APP promotion response, which is measured as purchase amount. The multivariate linear model
is shown as below:
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 𝐴𝐴𝐴𝐴𝐴𝐴 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖
= 𝛼𝛼0 + 𝛼𝛼1 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖 + 𝛼𝛼2 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 + Δ𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 + 𝜇𝜇Ω𝐑𝐑(𝐢𝐢) + 𝜌𝜌𝐗𝐗 𝐢𝐢 + 𝜀𝜀𝑖𝑖
Ad Responsei, is a continuous variable, indicating the total amount of purchased books in 24
hours after the mobile users receiving the mobile promotion. Sunnyi and Rainyi are two
dummy variables indicating the weather condition of the day a mobile user received the ad
message respectively and the baseline is still cloudy weather. We also control for weather
related covariates. Furthermore, we control for the mobile usage behavior with Mobile App
Usage (how many times the mobile user opened the APP) and Page Views (how many pages
the mobile user clicked or read on the APP). Ω𝑹𝑹(𝒊𝒊) is a vector of 30 geographical dummies
22
to control for location-based fixed effects and Xi is a vector of day fixed effects for the 3
days. 𝜀𝜀𝑖𝑖 follows i.i.d. extreme value distribution.
To empirically identify the incremental impact of ads, we first test the impact of weather
on in-app book purchase without the mobile ads. Table 13 presents the effect of sunny and
rainy on the amount of books purchased by users in holdout group. The results indicate that
the weather condition has statistically insignificant impact on the in-app purchase without
mobile ads (all p>0.1).
[Table 13]
We then explore the effect of weather on consumer response conditional on receiving
mobile ads. Table 14 shows the weather effect on the amount of books purchased after
receiving mobile ads. Column (1) is the simple model with 2 dummy variables of weather,
location fixed effect, and book fixed effects. We entered other weather-related variables in
column (2) and day fixed effect in column (3). Model of column (4) includes mobile usage
covariates. Consistent with the first field experiment, results in Table 14 indicate that
compared with cloudy weather, the sunny weather (coef.=0.0254, p<0.05) boosts the amount
of books app users purchase after receiving the mobile ads, while rainy weather (coef.=0.0546, p<0.05) decrease it.
[Table 14]
Table 15 presents the results across the mood-related books and mood-neutral books.
Again, the results still consistently indicates positive sunny effect and negative rainy effect on
the mobile promotion response. In addition, results show relatively stronger weather effects
on the mood-related books, revealing some additional evidence (albeit weak evidence) that
the weather has a psychological mood impact on mobile promotion response. 18
[Table 15]
Conclusion
Firms are increasingly targeting weather conditions to increase the relevancy and
effectiveness of their promotions. This study documents the effect that the weather can have on
consumer response to mobile campaigns promoting digital services. Using large-scale field
experiments on consumer responses to two mobile digital services, we find that sunny and rainy
weather conditions have a first-order effect on consumer responses to mobile promotions. In
sunny weather, consumers are more likely to purchase the promotions, and at an accelerated
pace relative to cloudy days. In rainy weather, customers are less likely and more slowly to
respond the promotions, relative to cloudy days. These findings are robust across multiple
models and measures of weather.
Considering managerial implications, we note that the ad copy has an additional effect and
interacts with sunshine and rainfall weather variables. A prevention-framing hurts the initial
18 We also conduct a follow-up survey to explore the underlying mechanism. The survey evidence confirms that
weather influence people’s mood towards the mobile promotions and then influence their purchase intention (i.e.,
mood partially but significantly mediates the effects of weather conditions on mobile promotion responses).
23
promotion boost provided by sunny weather, but improves the initial promotion drop induced
by rainy weather, in line with the “affect-as-information” theory. Thus, managers may take
notice that ad copy creatives still matter for weather effects on mobile promotions.
There are a number of limitations of this study, which may serve as avenues for future
research. First, our study is limited to the setting of mobile digital services and may not be
generalizable to non-mobile settings such as physical products or even desktop platforms. The
additional mobile product (reading books on a mobile device) is still similar to the product in
the main study (video streaming service on a mobile device) in that it is also a product with
valuations that are likely to depend on weather (i.e., valuation for both products would be higher
if consumer believes that she will be more likely to spend time outside, where only mobile
devices can be used for watching videos or reading books). Consequently, while the additional
study helps to establish robustness of the findings, it would be important to support generalizing
results with other settings and other products that are weather-neutral in future research.
While we take measures to account for self selection bias, it is still possible to have more
marketing-responsive customers moving to better climates. Including fixed effects by region
helps mitigate this concern, but still that merely controls for the level of purchasing, not the
responsiveness to promotion. Thus, future research is needed to more rigorously control for self
selection bias.
In addition, our study design with field experiment datasets cannot directly gauge the
underlying psychological mechanisms. Thus, future research with lab studies may explore
whether and how in rainy days consumers may activate mood repair mindsets and engage in
more indulgent responses to promotions (Pham, Lee, and Stephen 2012; Isen 2001; Schwarz
and Clore 1983). As another example, it would be interesting for future research to test how
weather is related to consumer risk seeking or risk avoiding (Parker and Tavassoli 2000;
Reinholtz, Lee, and Pham 2014).
Furthermore, one distinct weather-related aspect of mobile promotions relates to weather
awareness, interpreted as ability to check weather multiple times using a mobile app. But people
are typically aware whether current weather is sunny, rainy or cloudy without checking with
their smartphones, especially during daytime (simply by looking out of the nearest window).
The unique difference of having easy access to mobile weather apps appears to be in enabling
consumers to having instant access to weather forecast – something that was not possible before.
Consequently, one suggestion for future research to strengthen the contribution on mobile
promotion effectiveness (rather than on weather-based geolocation promotions) would be to
emphasize how marketers may increase location-based promotional effectiveness by
combining information on both current and forecasted weather.
Moreover, consumers are generally aware of current weather (which is likely to be the case
regardless of prevalence of mobile devices), the proposed explanation based on “affect-asinformation” theory would also explain weather-based differences in promotional effectiveness
for other ad types, not necessarily delivered via mobile devices. For example, industry
24
practitioners long argued that advertisers need to integrate weather-related data in online display
programmatic environments. Another example would be billboard advertisements based on
real-time location-based data (including weather), such as the ones trialed by Google. As
weather may profoundly affect people’s daily life, advertising platforms such as Facebook and
Twitter (Wall Street Journal 2014), who earn the majority of their advertising revenue from
mobiles, are contemplating on how to effectively target smartphone users by their hyper-local
weather. If this is true, future research may explore mobile and other aspects to a broader
context of weather-based promotional effectiveness.
In conclusion, this research is one initial step to provide evidence for two key results: (1)
sunny days are better for advertising at least for the tested two product categories in mobile
platform, and (2) affects-as-information and mood congruency theory can be effectively used
in deciding marketing content on the sunny and rainy days. We hope this study can serve as a
springboard for more works to examine how consumers respond to weather for higher mobile
advertising and promotion effectiveness.
25
Figure 1 Distribution of Weather Conditions in Data
5000000
4573323
4500000
Cloudy
4000000
Sunny
3500000
3000000
2610439
Rainy
2500000
1779437
2000000
1500000
1000000
500000
547058
332798
0
26
138540 49507
28633
566
107
Figure 2a Response Rate of Promotion
0.012
Figure 2b Response Rate of Promotion in different weather condition
0.012
0.0095
0.01
0.01
0.008
0.008
0.006
0.006
0.004
0.004
0.002
0.002
0
0
Holdout without
Promotion
0.0105
0.0078
0.0064
0
Promotion
Sunny Weather with
Promotion
Cloudy Weather with
Promotion
27
Rainy Weather with
Promotion
Table 1: Summary statistics
Variable
Obs
Mobile Promotion Response
Sunny
Rainy
Framing
Wind Speed
Wind Direction
High Temperature
Temperature Range
Visibility
Dew Point
Humidity
Air Pressure
Rural
Province
Date
6,756,524
6,756,524
6,756,524
6,756,524
6,756,524
6,614,522
6,756,524
6,756,524
4,251,825
4,321,404
4,321,404
4,198,892
6,756,524
6,756,524
6,756,524
Mean
0.0078
0.2068
0.3184
0.5541
3.2465
3.0949
27.1813
9.7772
7.1485
60.9009
66.7555
29.9411
0.4873
18.2299
17.5310
Std. Dev.
0.0881
0.4050
0.4659
0.4971
0.5340
2.7234
3.3350
3.5308
3.6572
10.9377
15.0454
0.1076
0.4998
7.9071
10.5274
Min
Max
0
0
0
0
3
1
5
2
0
-14
7
29.08
0
1
1
1
1
1
1
7
10
37
23
19
78
98
30.37
1
31
31
Note: Mobile promotion response is a dummy variable, equal to 1 if the mobile user purchased the promoted mobile
service after receiving the SMS campaign, and 0 if otherwise. Sunny is equal to 1 if the weather is sunny on the
day that the mobile ads is exposed to individual i and 0 if otherwise. Rainy is equal to 1 if the weather is rainy
on the day that the mobile ads is exposed to individual i , and 0 if otherwise. Thus, when both Sunny and Rainy
are 0, the weather condition is the baseline cloudy weather. Framing is equal to 1 if the SMS campaign received
by the mobile user includes prevention framing (“Do not miss the opportunity to take advantage of this special
deal”), and 0 if otherwise. Wind speed, high temperature (℃), temperature range, visibility, dew point, humidity
and air pressure are continuous variables, rural or urban area is dummy variable, and nine dummies for wind
direction. There are 30 geographical dummies (31 provinces) to control for location-based fixed effects and 30
time dummies (31 days) to control for time-based fixed effects.
28
Table 2: Sunny (rainy) weather increases (decreases) consumer response to mobile promotion
Sunny
Rainy
(1)
(2)
(3)
(4)
(5)
0.3684***
0.1782***
0.4327***
0.1880***
0.1880***
(0.014)
(0.022)
(0.026)
(0.038)
(0.037)
-0.1243***
-0.0814
(0.017)
(0.031)
***
Framing
**
***
***
0.5238
0.7414
(0.017)
(0.070)
-0.4621
-0.1068
-0.1068*
(0.036)
(0.045)
(0.047)
0.8217
Wind Speed
High Temperature
Temperature Range
Visibility
Dew Point
Humidity
Air Pressure
Rural
***
*
0.7814
***
0.7814***
(0.034)
(0.084)
(0.087)
0.1542***
0.2812***
0.2812***
(0.034)
(0.040)
(0.039)
-0.0524***
-0.0191**
-0.0191**
(0.005)
(0.006)
(0.006)
0.0072
-0.0226***
-0.0226***
(0.006)
(0.007)
(0.007)
-0.0039
-0.0114**
-0.0114**
(0.004)
(0.004)
(0.004)
0.0238***
0.0006
0.0006
(0.002)
(0.002)
(0.002)
0.0005
0.0044***
0.0044***
(0.001)
(0.001)
(0.001)
1.1496***
-0.0214
-0.0214
(0.131)
(0.140)
(0.139)
0.2344***
0.0889**
0.0889**
(0.028)
(0.028)
(0.027)
-5.4551***
-5.0080***
-39.9486***
-4.9052
-4.9052
(0.062)
(0.178)
(3.962)
(4.263)
(4.216)
Wind Direction
No
No
Yes
Yes
Yes
Location Fixed Effect
Yes
Yes
Yes
Yes
Yes
Day Fixed Effect
No
Yes
No
Yes
Yes
N
6,744,884
3,432,022
4,058,481
3,321,575
3,321,575
Log Likelihood
-290404.56
-168922.52
-166772.56
-162413.8
-162413.8
AIC
580875.1
337923
333641.1
324939.6
324939.6
BIC
581328
338431.9
334275.5
325668.5
325668.5
LR chi2
37565.79
18012.36
20694.18
17339.23
17343.91
0.0000
0.0000
0.0000
0.0000
0.0000
Intercept
Prob > chi2
Note:
ap
< 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
29
Table 3 Mobile promotions with prevention-framing are less effective in sunny weather, but more
effective in rainy weather
Wind Direction
Location Fixed Effect
Day Fixed Effect
-5.5971***
(0.064)
No
Yes
No
(2)
0.8314***
(0.038)
-0.5841***
(0.059)
1.0151***
(0.044)
-0.7098***
(0.048)
0.2393***
(0.067)
0.2036***
(0.034)
-0.0397***
(0.006)
0.0023
(0.006)
-0.0091*
(0.004)
0.0199***
(0.002)
0.0028**
(0.001)
0.8092***
(0.133)
0.1798***
(0.028)
-30.3986***
(4.034)
Yes
Yes
No
(3)
0.3131***
(0.051)
-0.2829***
(0.072)
0.8113***
(0.086)
-0.2350**
(0.080)
0.2239*
(0.090)
0.2806***
(0.040)
-0.0180**
(0.006)
-0.0236***
(0.007)
-0.0144***
(0.004)
0.0004
(0.002)
0.0045***
(0.001)
-0.0222
(0.141)
0.0740**
(0.028)
-4.9523
(4.264)
Yes
Yes
Yes
(4)
0.3131***
(0.049)
-0.2829***
(0.074)
0.8113***
(0.090)
-0.2350**
(0.077)
0.2239*
(0.093)
0.2806***
(0.039)
-0.0180**
(0.006)
-0.0236***
(0.007)
-0.0144***
(0.004)
0.0004
(0.002)
0.0045***
(0.001)
-0.0222
(0.139)
0.0740**
(0.028)
-4.9523
(4.217)
Yes
Yes
Yes
N
Log Likelihood
AIC
BIC
LR chi2
Prob > chi2
6,744,884
-290193.29
580456.6
580936.9
37988.33
0.0000
4,058,481
-166627.46
333354.9
334015.7
20984.38
0.0000
3,321,575
-162404.18
324924.4
325679.3
17358.47
0.0000
3,321,575
-162404.18
324924.4
325679.3
17346.25
0.0000
Sunny
Rainy
Framing
Sunny*Framing
Rainy*Framing
(1)
0.6632***
(0.023)
-0.2452***
(0.034)
0.7001***
(0.023)
-0.5225***
(0.030)
0.1835***
(0.038)
Wind Power
High Temperature
Temperature Range
Visibility
Dew Point
Humidity
Air Pressure
Rural
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
30
Table 4 Response Rate of Hourly Weather Conditions
Purchase Rate
Number of Observation
Hourly Sunny
0.00022
2,643,566
Hourly Cloudy
0.00016
2,847,163
Hourly Rainy
0.00013
2,834,813
Total
0.00017
8,325,542
Figure 3 Response Hazard in Different Weather Conditions
0.99
1.00
Kaplan-Meier survival estimates
0
10
20
30
analysis time
Rainy
Sunny
31
40
Cloudy
50
Table 5 Hourly sunny (rainy) weather increases (decreases) response hazard (9am to 5pm)
Hourly Good Weather
Hourly Bad Weather
(1)
(2)
(3)
0.5485***
(0.018)
-0.7733***
(0.049)
0.6304***
(0.019)
-0.4465***
(0.051)
No
Yes
No
No
Yes
Yes
0.5481***
(0.022)
-0.5247***
(0.053)
-0.7244
(0.784)
0.0580***
(0.003)
-0.1518***
(0.005)
0.1264***
(0.004)
-0.0530***
(0.002)
-0.4277**
(0.155)
-0.0459***
(0.003)
-0.0584**
(0.019)
Yes
Yes
Yes
59,814,313
-409198.47
818460.9
818970
9962.50
0.0000
59,814,313
-398471.57
797025.1
797677.3
31416.30
0.0000
55,164,292
-367346.64
734817.3
735798.5
38775.47
0.0000
Framing
Wind Speed
Temperature
Dew Point
Humidity
Air Pressure
Visibility
Rural
Wind Direction
Location Fixed Effect
Day Fixed Effect
N
Log Likelihood
AIC
BIC
LR chi2
Prob > chi2
32
Table 6 Changes in Weather Matter (Backward Looking)
(1)
Better_Backward
0.1135
***
(3)
***
0.0869***
0.0960
(0.012)
Worse_Backward
(2)
(0.012)
-0.2086***
-0.2644
(0.012)
(0.014)
(0.012)
***
-0.1578***
(0.018)
0.2118***
Sunny
(0.015)
-0.0484*
Rainy
(0.023)
Weather Covariates
Yes
Yes
Yes
Location Fixed Effect
Yes
Yes
Yes
Day Fixed Effect
No
Yes
Yes
N
5,820,418
5,600,909
5,820,418
Log Likelihood
-360207.3
-343946.09
-360088.82
AIC
720482.6
687984.2
720251.6
BIC
720944.2
688607
720754
LR chi2
32236.01
31606.06
32472.95
0.0000
0.0000
0.0000
Prob > chi2
Note:
ap
< 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
Table 7 Changes in Weather Matter (Forward Looking)
Better_Forward
Worse_Forward
(1)
(2)
0.0767***
0.0679**
(3)
0.0332
(0.021)
(0.021)
(0.022)
-0.1459***
-0.1753***
-0.0960***
(0.022)
(0.022)
(0.024)
0.3325***
Sunny
(0.037)
-0.1008***
Rainy
(0.031)
Weather Covariates
Yes
Yes
Yes
Location Fixed Effect
Yes
Yes
Yes
Day Fixed Effect
No
Yes
Yes
N
2,088,283
1,414,148
1,414,148
Log Likelihood
-117074.77
-110874.42
-110830.69
AIC
234245.5
221850.8
221767.4
BIC
234848
222471.1
222412
LR chi2
12293.17
7226.75
7314.23
0.0000
0.0000
0.0000
Prob > chi2
Note:
ap
< 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
33
Figure 4 Distributions of Weather Conditions in North and South
Southern Region (Subtropical Zone)
Northern Region (Temperate Zone)
0.551
0.6
0.6
0.5
0.5
0.4
0.2
0.362
0.4
0.313
0.3
0.486
0.3
0.152
0.2
0.136
0.1
0.1
0
0
Sunny
Weather
Cloudy
Weather
Sunny
Weather
Rainy
Weather
Cloudy
Weather
Table 8 Weather Deviations Matter
South
North
0.2904***
0.0406**
(0.016)
(0.015)
-0.0665***
-0.0921***
(0.019)
(0.021)
1.7702***
1.7497***
(0.041)
(0.051)
-6.5763***
-5.9477***
(0.095)
(0.098)
Weather Covariates
Yes
Yes
Location Fixed Effect
Yes
Yes
Day Fixed Effect
Yes
Yes
N
4,788,301
5,009,603
Log Likelihood
-235383.55
-263560
AIC
470825.1
527182
BIC
471213.2
527598.2
LR chi2
22838.43
28421.56
Prob > chi2
0.0000
0.0000
Sunny
Rainy
Framing
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
34
Rainy
Weather
Table 9a Falsification Check of Sunny Weather vs. All Other Weather Conditions
Wind Direction
Location Fixed Effect
Day Fixed Effect
-5.4867***
(0.062)
No
Yes
No
(2)
0.4723***
(0.026)
0.7396***
(0.033)
0.1590***
(0.034)
-0.0345***
(0.005)
0.0146*
(0.006)
-0.0116***
(0.003)
0.0227***
(0.002)
0.0004
(0.001)
1.3790***
(0.130)
0.2145***
(0.028)
-47.4075***
(3.946)
Yes
Yes
No
(3)
0.2153***
(0.036)
0.7610***
(0.083)
0.2662***
(0.040)
-0.0151*
(0.006)
-0.0217**
(0.007)
-0.0129***
(0.004)
-0.0001
(0.002)
0.0044***
(0.001)
-0.0621
(0.139)
0.0835**
(0.028)
-3.7296
(4.236)
Yes
Yes
Yes
(4)
0.2153***
(0.035)
0.7610***
(0.086)
0.2662***
(0.039)
-0.0151*
(0.006)
-0.0217**
(0.007)
-0.0129***
(0.003)
-0.0001
(0.002)
0.0044***
(0.001)
-0.0621
(0.137)
0.0835**
(0.027)
-3.7296
(4.176)
Yes
Yes
Yes
N
Log Likelihood
AIC
BIC
LR chi2
Prob > chi2
6,744,890
-290431.88
580927.8
581366.9
37511.26
0.0000
4,058,482
-166858.05
333810.1
334431.3
20523.21
0.0000
3,321,575
-162416.67
324943.3
325659.2
17333.49
0.0000
3,321,575
-162416.67
324943.3
325659.2
17295.60
0.0000
Sunny
Framing
(1)
0.3890***
(0.013)
0.5202***
(0.017)
Wind Speed
High Temperature
Temperature Range
Visibility
Dew Point
Humidity
Air Pressure
Rural
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
35
Table 9b Falsification Check of Rainy Weather vs. All Other Weather Conditions
Wind Direction
Location Fixed Effect
Day Fixed Effect
-5.2735***
(0.062)
No
Yes
No
(2)
-0.5193***
(0.035)
0.6983***
(0.033)
0.1786***
(0.034)
-0.0525***
(0.005)
0.0493***
(0.005)
-0.0148***
(0.003)
0.0244***
(0.002)
0.0002
(0.001)
0.5837***
(0.126)
0.2149***
(0.028)
-23.0748***
(3.834)
Yes
Yes
No
(3)
-0.1729***
(0.042)
0.8411***
(0.083)
0.2962***
(0.040)
-0.0228***
(0.006)
-0.0155*
(0.007)
-0.0148***
(0.004)
0.0015
(0.002)
0.0040***
(0.001)
-0.0608
(0.140)
0.0827**
(0.028)
-3.6943
(4.246)
Yes
Yes
Yes
(4)
-0.1729***
(0.044)
0.8411***
(0.085)
0.2962***
(0.039)
-0.0228***
(0.006)
-0.0155*
(0.007)
-0.0148***
(0.003)
0.0015
(0.002)
0.0040***
(0.001)
-0.0608
(0.138)
0.0827**
(0.027)
-3.6943
(4.192)
Yes
Yes
Yes
N
Log Likelihood
AIC
BIC
LR chi2
Prob > chi2
6,744,890
-290759.21
581582.4
582021.6
36856.58
0.0000
4,058,482
-166917.78
333929.6
334550.7
20403.75
0.0000
3,321,575
-162426.24
324962.5
325678.3
17314.36
0.0000
3,321,575
-162426.24
324962.5
325678.3
17297.42
0.0000
Rainy
Framing
(1)
-0.2101***
(0.017)
0.3494***
(0.016)
Wind Speed
High Temperature
Temperature Range
Visibility
Dew Point
Humidity
Air Pressure
Rural
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
36
Table 10 Robustness Check of All vs. Partial Sunny Days
All Sunny
0.3802***
(0.027)
Wind Direction
Location Fixed Effect
Day Fixed Effect
Full
0.3525***
(0.025)
0.1246***
(0.027)
0.9063***
(0.061)
0.1559***
(0.026)
-0.0036
(0.004)
-0.0169***
(0.004)
-0.0129***
(0.002)
0.0003
(0.001)
0.0016*
(0.001)
0.0546
(0.096)
0.0541**
(0.017)
-6.3116*
(2.894)
Yes
Yes
Yes
0.9084***
(0.063)
0.1525***
(0.027)
0.0008
(0.005)
-0.0199***
(0.005)
-0.0162***
(0.002)
-0.0002
(0.001)
0.0018*
(0.001)
0.0170
(0.100)
0.0455*
(0.019)
-5.1314
(3.011)
Yes
Yes
Yes
0.1011**
(0.031)
0.9849***
(0.064)
0.0066
(0.029)
-0.0083
(0.006)
-0.0248***
(0.005)
-0.0094***
(0.002)
-0.0034*
(0.001)
0.0047***
(0.001)
-0.1869
(0.119)
0.0360
(0.019)
1.4245
(3.609)
Yes
Yes
Yes
N
Log Likelihood
AIC
BIC
LR chi2
Prob > chi2
5,691,938
-300310.17
600742.3
601569.2
28282.20
0.0000
5,472,803
-282469.82
565059.6
565870.6
25894.19
0.0000
3,805,112
-210947.68
422013.4
422789.3
21136.62
0.0000
All Sunny
Partial Sunny
Framing
Wind Speed
High Temperature
Temperature Range
Visibility
Dew Point
Humidity
Air Pressure
Rural
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
37
Partial Sunny
Table 11 Robustness Check of Sunlight Hours
Wind Direction
Location Fixed Effect
Day Fixed Effect
(1)
0.0965***
(0.019)
4.0140***
(0.775)
0.6159
(1.111)
0.0507
(0.060)
0.1151
(0.063)
0.0287
(0.025)
0.0260
(0.014)
0.0012
(0.011)
-2.0544**
(0.792)
-0.7590***
(0.190)
48.6490*
(24.136)
Yes
Yes
Yes
N
Log Likelihood
AIC
BIC
LR chi2
Prob > chi2
278,948
-21731.72
43533.44
43902.3
1367.76
0.0000
Sunlight Hours
Framing
Wind Speed
High Temperature
Temperature Range
Visibility
Dew Point
Humidity
Air Pressure
Rural
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
38
Table 12 Robustness Check of Interactions between Weather and Temperature
Sunny
Rainy
Sunny*High Temperature
Rainy*High Temperature
Sunny*Temperature Range
Rainy*Temperature Range
Framing
Intercept
Weather Covariates
Location Fixed Effect
Day Fixed Effect
N
Log Likelihood
AIC
BIC
LR chi2
Prob > chi2
(1)
1.3508***
(0.114)
-0.2827**
(0.092)
-0.0437***
(0.004)
0.0160***
(0.005)
(2)
0.2919***
(0.067)
-0.1117*
(0.051)
-0.0078
(0.005)
0.0107
(0.006)
1.1984***
1.1976***
(0.042)
(0.040)
***
32.7511
-6.0606***
(2.058)
(0.177)
Yes
Yes
Yes
Yes
Yes
Yes
6,428,845 6,642,306
-307275.45 -314137.84
614652.9
628369.7
615350.4
629014
28046.59
28253.25
0.0000
0.0000
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
39
Table 13 Weather has no impact on holdout group (mobile-book in second field experiment)
(1)
Sunny
0.0019
(0.024)
Rainy
-0.0005
(0.032)
Temperature
0.0047
(0.003)
Dew Point
-0.0074*
(0.003)
Humidity
0.0021
(0.002)
Visibility
-0.0015
(0.003)
Wind Speed
-0.0076*
(0.004)
Mobile App Usage
0.0032***
(0.000)
Page Views
0.0041***
(0.000)
Intercept
-0.0434
(0.139)
Location Fixed Effect
Yes
Day Fixed Effect
Yes
N
13231
R2
0.358
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
40
Table 14 9 Sunny (rainy) weather increases (decreases) consumer responses to mobile ads
(mobile-book in second field experiment)
Sunny
Rainy
(1)
(2)
(3)
(4)
0.0509***
0.0381**
0.0397**
0.0254*
(0.013)
(0.013)
(0.013)
(0.011)
***
-0.0469
-0.0850
-0.0691
-0.0546***
(0.014)
(0.015)
(0.015)
(0.013)
0.0027
0.0026
0.0008
(0.002)
(0.002)
Temperature
Dew Point
Humidity
Visibility
Wind Speed
***
***
(0.001)
-0.0055**
**
-0.0057
(0.002)
(0.002)
(0.002)
0.0027**
0.0030**
0.0019*
(0.001)
(0.001)
(0.001)
-0.0052***
-0.0036*
-0.0000
(0.001)
(0.001)
(0.001)
-0.0076***
-0.0106***
-0.0074***
(0.002)
(0.002)
(0.002)
-0.0031
0.0188***
Mobile App Usage
(0.000)
0.0021***
Page Views
(0.000)
0.4287***
0.5062***
0.4831***
0.3144***
(0.006)
(0.084)
(0.087)
(0.074)
Location Fixed Effect
Yes
Yes
Yes
Yes
Book Fixed Effect
Yes
Yes
Yes
Yes
Day Fixed Effect
No
No
Yes
Yes
N
84999
84156
84156
84156
R2
0.005
0.006
0.009
0.290
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
41
Table 15 Effect of weather on mood-related book VS neutral book
Mood
Neutral
0.0404*
0.0343a
(0.020)
(0.018)
-0.1107***
-0.0713**
(0.022)
(0.022)
0.0046
0.0032
(0.002)
(0.002)
-0.0071*
-0.0062*
(0.003)
(0.003)
0.0036*
0.0034*
(0.002)
(0.001)
-0.0038
-0.0032
(0.002)
(0.002)
-0.0065*
-0.0069**
(0.003)
(0.003)
0.4208**
0.4635***
(0.129)
(0.116)
Location Fixed Effect
Yes
Yes
Book Fixed Effect
Yes
Yes
Day Fixed Effect
Yes
Yes
N
38780
45376
R2
0.007
0.005
Sunny
Rainy
Temperature
Dew Point
Humidity
Visibility
Wind Speed
Intercept
Note: ap < 0.1, *p < 0.05, ** p < 0.01, *** p < 0.001
42
Appendix A: Provinces and Cities
Robustness Check of City-Level Locational Effects
Sunny Weather
Rainy Weather
(1)
(2)
(3)
(4)
0.2555***
0.1753***
0.2999***
0.2332***
(0.010)
(0.014)
(0.011)
(0.015)
-0.0534***
-0.1212***
-0.0507***
-0.1417***
(0.014)
(0.017)
(0.014)
(0.017)
0.1553***
0.4536***
(0.014)
(0.020)
Framing
Wind Power
High Temperature
Temperature Range
-0.2141***
-0.2845***
(0.023)
(0.024)
-0.0109***
-0.0591***
(0.003)
(0.004)
0.0315***
0.0661***
(0.003)
(0.004)
-4.9502***
-4.1559***
-4.9514***
-2.7294***
(0.092)
(0.142)
(0.092)
(0.158)
Wind Direction
Yes
Yes
Yes
Yes
City Fixed Effect
Yes
Yes
Yes
Yes
N
10058668
9807363
10058668
9807363
Log Likelihood
-516716.42
-499678.65
-516652.61
-499415.53
AIC
1034099
1000043
1033973
999519.1
BIC
1038802
1004879
1038691
1004369
50916.29
50523.91
51146.26
50682.42
0.0000
0.0000
0.0000
0.0000
Intercept
LR chi2
Prob > chi2
344 city names
Names of the Cities
Urban
Anqing, Anyang, Anshan, Bengbu, Baotou, Baoji, Baoding, Beihai, Beijing, Benxi, Binzhou, Cangzhou,
43
(167)
Changde, Changzhou, Chaozhou, Chenzhou, Chengdu, Chifeng, Chuzhou, Dalian, Daqing, Datong, Deyang,
Dezhou, Dongguan, Dongying, Erdos, Foshan, Fuzhou, Fushun, Ganzhou, Guangzhou, Guiyang, Guilin, Harbin,
Haikou, Handan, Hangzhou, Hefei, Hengyang, Hohhot, Hulunbuir, Huzhou, Huaian, Huainan, Huanggang,
Huangshi, Huizhou, Jixi, Jilin, Jinan, Jining, Jiamusi, Jiaxing, Jiangmen, Jiaozuo, Jinhua, Jingmen, Jingzhou,
Jiujiang, Kaifeng, Karamay, Kunming, Lhasa, Lanzhou, Langfang, Leshan, Lianyungang, Liangshan, Liaocheng,
Linfen, Linyi, Liuzhou, Longyan, Loudi, Luzhou, Luoyang, Luliang, Ma’anshan, Maoming, Mianyang,
Nanchang, Nanchong, Nanjing, Nanning, Nantong, Nanyang, Ningbo, Panjin, Pingdingshan, Putian, Qigihar,
Qinhuangdao, Qingdao, Qingyuan, Quanzhou, Rizhao, Xiamen, Shantou, Shanghai, Shangrao, Shaoxing,
Shenzhen, Shenyang, Shihezi, Shijiazhuang, Siping, Songyuan, Suzhou, Taizhou, Taiyuan, Taian, Taizhou,
Tangshan, Tianjin, Tongliao, Tongling, Weihai, Weifang, Weinan, Wenzhou, Wuhai, Urumqi, Wuxi, Wuhu,
Wuzhou, Wuhan, Xi’an, Xining, Xianyang, Xiangtan, Xiaogan, Xinxiang, Xinyu, Xingtai, Xuzhou, Xuchang,
Yantai, Yanan, Yancheng, Yangzhou, Yangquan, Yibin, Yichang, Yichun, Yinchuan, Yingkou, Yulin., Yulin,
Yueyang, Yuncheng, Zhanjiang, Zhangjiakou, Zhangzhou, Changchun, Changsha, Changzhi, Zhaoqing,
Zhenjiang, Zhengzhou, Zhongshan, Chongqing, Zhoushan, Zhuhai, Zhuzhou, Zibo, Zigong
Rural
Aba, Aksu, Alashan, Aletai, Ali, Atushi, Ankang, Anshun, Bayinnaoer, Bayinguoleng, Bazhong, Baicheng,
(177)
Baishan, Baiyin, Baise, Baoshan, Bijie, Bozhou, BoLe, Changdu, Changji, Chaohu, Chaoyang, Chengde,
Chizhou, Chuxiong, Dazhou, Dali, Greater Hinggan Mountains, Dandong, Dehong, Delingha, Dingxi, Douyun,
Ezhou, Enshi, Fangchenggang, Fuzhou, Fuxin, Fuyang, Gannan, Ganzi, Golmud, Gejiu, Gonghe, Guyuan,
Guangan, Guangyuan, Guichi, Guoluo, Hami, Haibeizhou, Haidong, Hainanzhou, Hanzhong, Haozhou, Hetian,
Hechi, Heyuan, Heze, Hebi, Hegang, Heihe, Hengshui, Huludao, Huaihua, Huaibei, Huaiyin, huangnan,
Huangshan, Hunchun, Gi’an, Jishou, Jieyang, Jinchang, Jinzhou, Jincheng, Jinzhong, Jingdezhen, Jinghong,
Jiuquan, Kashgar, Kaili, Kuytun, Laiwu, Lijiang, Lishui, Liaoyang, Liaoyuan, Linzhi, Lincang, Linxia, Lu‘an,
Liupanshui, Longnan, Luohe, Barkam, Maqin, Meishan, Meihekou, Meizhou, Mudanjiang, Naqu, Nanping,
Neijiang, Ningde, Panzhihua, Pingliang, Pingxiang, Puyang, Pu‘er, Qitaihe, Qinzhou, Qingyang, Quzhou,
Qujing, Shigatse, Sanmenxia, Sanming, Shannan, Shanwei, Shangluo, Shangqiu, Shaoguan, Shaoyang, Shiyan,
Shizuishang, Shuangyashan, Shuozhou, Simao, Suihua, Suizhou, Suining, Tacheng, Tianshui, Tieling, Tonghua,
Tongren, Tongchuan, Tongren, Turpan, Wenshan, Wulanchabu, Ulanhot, Wuzhong, Wuwei, Xilingol, Xilinhot,
Xiantao, Xianning, Shangri-la, Xiangfan, Xinzhou, Xinyang, Xingyi, Suqian, Suzhou, Xuancheng, Ya‘an,
Yangjiang, Yichun, Yili, Yiyang, Yingtan, Yongzhou, Yushu, Yuxi, Yunfu, Zaozhuang, Zhangjiajie, Zhangye,
Zhaotong, Zhongwei, Zhoukou, Zhumadian, Ziyang, Zunyi
44
Appendix B: Mobile Promotion Message
SMS in Field Experiment 1:
Message with Prevention Framing
Message without Prevention Framing
Original
Version
别错过占便宜的机会了！48 小时内回
复“是”购买视频礼包，每月仅需 3
元。当月购买后可在手机免费观看精
彩视频【公司名称】
尊敬的客户，您好！48 小时内回复
“是”购买视频礼包，每月仅需 3
元。当月购买后可在手机免费观看
精彩视频【公司名称】
English
Translation
Do not miss the opportunity to take
advantage of this special deal!
Subscribe to the mobile videostreaming service or only ¥3 per
month! Watch video episodes of the
most popular TV series on your mobile
devices on-the-go! The regular price is
¥6. Purchase by replying “Yes” to this
SMS within the next 48 hours
[Company Name]
Dear respected customer,
Subscribe to the mobile videostreaming service of for only ¥3 per
month! Watch video episodes of the
most popular TV series on your
mobile devices on-the-go! The
regular price is ¥6. Purchase by
replying “Yes” to this SMS within
the next 48 hours [Company Name]
SMS in Field Experiment 2:
Message with mood-related books
Message with neutral books
Original
Version
一叶便知秋，一书看世界！
【公司名
称】百万电子书等您来挑选。百精
彩新书上架，不要错过买新书的好
机会！
一叶便知秋，一书看世界！
【公司名
称】百万电子书等您来挑选。百精
彩新书上架，不要错过买新书的好
机会！
English
Translation
Knowing the season from the change
of plant, learning the world from
reading books. [Company Name]
provides more than one million mobile
e-books. Hundreds of new e-books
are available now. Don’t miss the
good opportunity to purchase new
books!
Knowing the season from the change of
plant, learning the world from reading
books. [Company Name] provides
more than one million mobile e-books.
Hundreds of new e-books are
available now. Don’t miss the good
opportunity to purchase new books!
45
References
Aaker, J. L., & Lee, A. Y. (2006). Understanding Regulatory Fit. Journal of Marketing Research, 43(1), 15–19.
Adweek (2015), With Turn-by-Turn Directions, Google's Waze App Wants to Win Mobile Advertising Dunkin' Donuts,
Panera Bread target drivers, March 24, by Lauren Johnson.
Allen, M. A., & Fischer, G. J. (1978). Ambient Temperature Effects on Paired Associate Learning. Ergonomics, 21(2),
95–101.
Alter, A. L., & Kwan, V. S. (2009). Cultural sharing in a global village: Evidence for extracultural cognition in European
Americans. Journal of Personality and Social Psychology, 96(4), 742.
Andrews, M., Luo, X., Fang, Z., & Ghose, A. (2015). Mobile Ad Effectiveness: Hyper-Contextual Targeting with
Crowdedness. Marketing Science, forthcoming.
Bagozzi, R. P., Gopinath, M., & Nyer, P. U. (1999). The Role of Emotions in Marketing. Journal of the Academy of
Marketing Science, 27(2), 184–206.
Bahng, Y., & Kincade, D. H. (2012). The relationship between temperature and sales. International Journal of Retail &
Distribution Management, 40(6), 410–426.
Bakhshi, S., Kanuparthy, P., & Gilbert, E. (2014). Demographics, Weather and Online Reviews: A Study of Restaurant
Recommendations. In Proceedings of the 23rd International Conference on World Wide Web (pp. 443–454). New York,
NY,
Banerjee, S. S., & Dholakia, R. R. (2008). Mobile advertising: does location based advertising work? International
Journal of Mobile Marketing.
Bart Y, Stephen A, Sarvary M (2014). Which products are best suited to mobile advertising? A field study of mobile
display advertising effects on consumer attitudes and intentions. J. Marketing Res. 51(3):270-285.
Bassi, A., Colacito, R., & Fulghieri, P. (2013). ’O Sole Mio: An Experimental Analysis of Weather and Risk Attitudes in
Financial Decisions. Review of Financial Studies, 26(7), 1824–1852.
Busse, M. R., Pope, D. G., Pope, J. C., & Silva-Risso, J. (2015). The Psychological Effect of Weather on Car Purchases.
The Quarterly Journal of Economics, 130 (1): 371-414.
BusinessWeek (2014), The Weather Channel’s Secret: Less Weather, More Clickbait. BusinessWeek: Technology, by
Suddath, C. (2014, October 9).
Cawthorn, C. (1998). Weather as a strategic element in demand chain planning. Journal of Business Forecasting Methods
& Systems, 17(3), 18.
Chiang, K.-P. and Dholakia, R.R. (2003) “Factors driving consumer intention to shop online: an empirical investigation”,
Journal of Consumer Psychology, 13 (1/2), 177-183.
Cohen, A. (2011). The photosynthetic President: Converting sunshine into popularity. The Social Science J, 48(2), 295–
304.
Danaher, P. J., Smith, M. S., Ranasinghe, K., & Danaher, T. S. (2015). Where, When and How Long: Factors that
Influence the Redemption of Mobile Phone Coupons. Journal of Marketing Research, forthcoming.
Daurer, S., Molitor, D., Spann, M., & Manchanda, P. (2013). The Impact of Smartphones, Barcode Scanning, and
Location-based Services on Consumers’ Search Behavior.
Deng, C., & Graz, J. (2002). Generating randomization schedules using SAS programming (pp. 267–270). Presented at
the Proceedings of the 27th Annual SAS Users Group International Conference.
Easton, M. (2012, July 25). Does sunshine make us happier? from http://www.bbc.com/news/uk-18986041
eMarketer. (2014, March 19). Driven by Facebook and Google, Mobile Ad Market Soars 105% in 2013 - eMarketer.
Ghose, A., Goldfarb, A., & Han, S. P. (2012). How Is the Mobile Internet Different? Search Costs and Local Activities.
Information Systems Research, 24(3), 613–631.
Google patents weather-based ads. (2012). Retrieved from
46
http://www.telegraph.co.uk/technology/google/9161262/Google-patents-weather-based-ads.html.
Guadagni, P. M., & Little, J. D. C. (1983). A Logit Model of Brand Choice Calibrated on Scanner Data. Marketing
Science, 2(3), 203–238.
Higgins, T (1987), “Self-Discrepancy: A Theory Relating Self and Affect,” Psychological Review, 94 (3), 319–40.
Hirshleifer, D., & Shumway, T. (2003). Good Day Sunshine: Stock Returns and the Weather. The Journal of Finance,
58(3), 1009–1032.
Hsiang, S. M., Burke, M., & Miguel, E. (2013). Quantifying the influence of climate on human
conflict. Science, 341(6151), 1235367.Johnson, E. J., & Tversky, A. (1983). Affect, generalization, and the perception
of risk. Journal of Personality and Social Psychology, 45(1), 20–31. doi:10.1037/0022-3514.45.1.20
Isen, Alice (2001), “An Influence of Positive Affect on Decision Making in Complex Situation, Theoretical issues with
Practical Implications, Journal of Consumer Psychology, 11(2), 75-85.
Barbara E. Kahn and Alice M. Isen (1993), The Influence of Positive Affect on Variety Seeking Among Safe, Enjoyable
Products, Journal of Consumer Research, 20 (2), 257-270.
Keller, M. C., Fredrickson, B. L., Ybarra, O., Côté, S., Johnson, K., Mikels, J., … Wager, T. (2005). A Warm Heart and a
Clear Head: The Contingent Effects of Weather on Mood and Cognition. Psychological Science, 16(9), 724–731.
Kenny, D., & Marshall, J. F. (2000). Contextual marketing: the real business of the Internet. Harvard Busi Rev, 78(6),
119–125.
Klimstra, Theo A., Frijns, Tom, Keijsers, Loes, Denissen, Jaap J. A., Raaijmakers, Quinten A. W., van Aken, Marcel A. G., Koot,
Hans M., van Lier, Pol A. C., Meeus, Wim H. J. (2011), Come rain or come shine: Individual differences in how weather affects
mood. Emotion, 11(6), 1495-1499.Lambert, G., Reid, C., Kaye, D., Jennings, G., & Esler, M. (2002). Effect of sunlight
and season on serotonin turnover in the brain. The Lancet, 360(9348), 1840–1842.
Lee, Angela Y. and Jennifer L. Aaker (2004), “Bringing the Frame into Focus: The Influence of Regulatory Fit on
Processing Fluency and Persuasion,” Journal of Personality and Social Psychology, 86(2), 205-218
Luo, X., Andrews, M., Fang, Z., & Phang, C. W. (2013). Mobile Targeting. Management Science, 60(7), 1738–1756.
Marshall, J. (2014, June 11). Weather-Informed Ads are Coming to Twitter. Retrieved from
http://blogs.wsj.com/cmo/2014/06/11/weather-informed-ads-are-coming-to-twitter/
Mayer, J. D., Gaschke, Y. N., Braverman, D. L., & Evans, T. W. (1992). Mood-congruent judgment is a general effect.
Journal of Personality and Social Psychology, 63(1), 119–132.
McShane, B. B., Bradlow, E. T., & Berger, J. (2012). Visual influence and social groups. J Marketing Res, 49(6), 854–
871.
Molitor D, Reichhart P, and Spann M (2013). Location-based advertising: measuring the impact of context-specific
factors on consumers’ choice behavior. Working paper, Munich School of Management.
Murray, K. B., Di Muro, F., Finn, A., & Popkowski Leszczyc, P. (2010). The effect of weather on consumer spending.
Journal of Retailing and Consumer Services, 17(6), 512–520.
Fong, N. M., Fang, Z., & Luo, X. (2015). Geo-conquesting: Competitive locational targeting of mobile
promotions. Journal of Marketing Research, forthcoming.
Parker, P and N. Tavassoli (2000), Homeostasis and consumer behavior across cultures, International Journal of
Research in Marketing 17 (1), 33-53
Persinger, M. A., & Levesque, B. F. (1983). Geophysical variables and behavior: xii. the weather matrix accommodates
large portions of variance of measured daily mood. Perceptual and Motor Skills, 57(3), 868–870.
Pham, M. T., Cohen, J. B., Pracejus, J. W., & Hughes, G. D. (2001). Affect monitoring and the primacy of feelings in
judgment. Journal of consumer research, 28(2), 167-188.
Pham, M. T. (2008). The Lexicon and Grammar of Affect as Information in Consumer Decision Making, The GAIM. In
Social Psychology of Consumer Behavior. Psychology Press.
47
Pham, Michel Tuan, and Tamar Avnet (2009). Rethinking Regulatory Engagement Theory. Journal of Consumer
Psychology, 19(2), 115-123.
Pham, Michel Tuan, Leonard Lee, and Andrew T. Stephen (2012), “Feeling the Future: The Emotional Oracle Effect,”
Journal of Consumer Research, 39 (3), 461-477.
Radas, S., & Shugan, S. M. (1998). Seasonal Marketing and Timing New Product Introductions. Journal of Marketing
Research, 35(3), 296–315.
Reinholtz, Nicholas, Leonard Lee, and Michel T. Pham (2014), “Sunny Days, Risky Ways: Exposure to Sunlight
Increases Risk Taking,” Working paper at Columbia Business School.
Roll, R. (1992). Industrial structure and the comparative behavior of international stock market indices. The Journal of
Finance, 47(1), 3-41.
Rollier, A. (1927). Heliotherapy: Its Therapeutic, prophylactic and social value. The American Journal of Nursing, 815823.Roslow, S., Li, T., & Nicholls, J. A. F. (2000). Impact of situational variables and demographic attributes in two
seasons on purchase behaviour. European Journal of Marketing, 34(9/10), 1167–1180.
Rosman Katherine (2013), Weather Channel Now Also Forecasts What You’ll Buy, The Wall Street Journal, Aug 14,
2013. http://www.wsj.com/articles/SB10001424127887323639704579012674092402660.
Saunders, E. M., Jr. (1993). Stock Prices and Wall Street Weather. The American Economic Review, 83(5), 1337–1345.
Simonsohn, U. (2007). Clouds make nerds look good: field evidence of the impact of incidental factors on decision
making. Journal of Behavioral Decision Making, 20(2), 143–152.
Simonsohn, U. (2009). Direct risk aversion evidence from risky prospects valued below their worst outcome.
Psychological Science, 20(6), 686–692.
Sinclair, R. C., Mark, M. M., & Clore, G. L. (1994). Mood-related persuasion depends on (mis) attributions. Social
Cognition, 12(4), 309–326.
Smith, A. (2013). Smartphone Ownership 2013. http://www.pewinternet.org/2013/06/05/smartphone-ownership-2013/
Spiekermann, S., Rothensee, M., & Klafft, M. (2011). Street marketing: how proximity and context drive coupon
redemption. The Journal of Consumer Marketing, 28(4), 280–289. doi:http://dx.doi.org/10.1108/07363761111143178
Steele, A. T. (1951). Weather’s Effect on the Sales of a Department Store. Journal of Marketing, 15(4), 436–443.
Stephen, Andrew T. and Michel Tuan Pham (2008), “On Feelings as a Heuristic for Making Offers in Ultimatum
Negotiations,” Psychological Science, 19 (10), 1051-1058.
Schwarz, Norbert, and Gerald L. Clore (1983). Mood, misattribution, and judgments of well-being: informative and
directive functions of affective states. Journal of Personality and Social Psychology, 45(3), 513.
Wall Street Journal (2014), Weather-Informed Ads are Coming to Twitter,
http://blogs.wsj.com/cmo/2014/06/11/weather-informed-ads-are-coming-to-twitter/
Wharton Knowledge. (2013). Today’s Forecast for the Weather Business: Increased Revenues and a Focus on
Innovation. Retrieved October 17, 2014, from http://knowledge.wharton.upenn.edu/article/todays-forecast-for-theweather-business.
Zhao Guangzhi and Cornelia Pechmann (2007), “The Impact of Regulatory Focus on Adolescents' Response to
Antismoking Advertising Campaigns,” Journal of Marketing Research, 44 (November), 671-687.
Zwebner, Y., Lee, L., & Goldenberg, J. (2014) “The Temperature Premium: Warmer Temperatures Increase Object
Valuation.” Journal of Consumer Psychology, 24(2), 251-259.
48