Content uploaded by Junda Wang
Author content
All content in this area was uploaded by Junda Wang on Dec 16, 2021
Content may be subject to copyright.
Content uploaded by Junda Wang
Author content
All content in this area was uploaded by Junda Wang on Jul 04, 2021
Content may be subject to copyright.
How the COVID-19 Pandemic Has Changed Crowdfunding: Evidence
from GoFundMe
Junda Wanga,Xupin Zhangb,∗and Jiebo Luoc,∗
aDepartment of Computer Science, University of Rochester, Rochester, NY 14627.
bFaculty of Economics and Management, East China Normal University, Shanghai, 200241, China
cDepartment of Computer Science, University of Rochester, Rochester, NY 14627
ABSTRACT
While the long-term effects of the COVID-19 pandemic have yet to be determined, its immediate
impact on crowdfunding is nonetheless significant. This study adopts a computational approach to
better understanding this consequence. We aim to gain insight into whether and how the COVID-19
pandemic has changed crowdfunding. Using a unique dataset of all GoFundMe campaigns published
over the past two years, we explore the factors that have led to successfully funded crowdfunding
projects. In particular, we study a corpus of projects by analyzing cover images and other attributes
commonly found on crowdfunding sites. We first construct a classifier and a regression model to
assess the importance of features based on XGBoost. Next, we employ counterfactual analysis to
investigate the causality between features and the success of crowdfunding. Furthermore, sentiment
analysis and paired sample t-tests are performed to examine differences in crowdfunding campaigns
before and after the COVID-19 outbreak in March 2020. Findings suggest a significant racial disparity
in crowdfunding success. In addition, sad emotions expressed in a campaign’s description became
significant after the COVID-19 outbreak. This study enriches our understanding of the impact of the
COVID-19 pandemic on crowdfunding as well as the prevalence of discrimination in crowdfunding.
1. Introduction
The development of the Internet has introduced more
ways to raise money online in recent years. GoFundMe, an
American for-profit crowdfunding platform that encourages
people to create online crowdfunding projects for life events
such as illnesses and accidents, is a prime example. Despite
the increasing convenience of crowdfunding, campaigns’
success rates remain low. Yet little is known about the
“recipe” for a successful campaign; as such, uncovering the
factors contributing to successful crowdfunding constitutes
a key research aim (Kaartemo,2017).
Amid the COVID-19 pandemic, GoFundMe has be-
come a powerful online platform through which people
can raise or donate money. This digitally enabled process
has largely replaced offline crowdfunding. Researchers have
thus explored the factors influencing crowdfunding suc-
cess during the COVID-19 pandemic. For example, Ho,
Chiu, Mansumitrchai, Yuan, Zhao and Zou (2021) found
that campaign titles containing more characters could en-
hance crowdfunding success. Locations with more cases of
COVID-19 infections also tended to receive more donations
than places with lower infection rates. Rajwa, Hopen, Mu,
Paradysz, Wojnarowicz, Gross and Leapman (2020) con-
sidered the responsiveness of online crowdfunding during
the pandemic. However, neither study compared features
contributing to campaigns’ success before and after the
outbreak. The authors also did not investigate the potential
causality and associations among relevant features.
∗Corresponding author: Jiebo Luo, Tex.: +1(585)276-5784, Xupin
Zhang, Tex.: +1(585)967-0656
ORCID(s):
In this paper, we define a crowdfunding project as
successful when the amount raised is greater than the
target amount. We employ XGBoost to assess features’
importance. Moreover, target amounts are classified and
regressed. We use XGBoost to resolve regression or clas-
sification problems and provide a sequence of important
factors. Finally, we perform a counterfactual experiment to
analyze the influence of each factor and the impact of the
COVID-19 pandemic on these features. To the best of our
knowledge, this study is the first to investigate the impact of
the COVID-19 pandemic on crowdfunding campaigns.
2. Hypothesis
Life course theory explains individuals’ development
over time as a function of internal forces (agency) and
external influences (e.g., time and place) with an emphasis
on the social and historical trajectories that influence one’s
life course (Giele and Elder,1998). The life course refers
to social patterns in the timing, duration, spacing, and order
of events and roles. Individual time, also called ontogenetic
time, is based on a person’s chronological age; this concept
also assumes that periods of life (including childhood,
adolescence, adulthood, and old age) affect a person’s so-
cial positions, roles, and rights (Binstock, George, Cutler,
Hendricks and Schulz,2011). In contrast, generational time
draws attention to the experiences of groups or cohorts of
people based on age. For instance, many countries experi-
enced a “baby boom”— a faster-than-expected increase in
birth rates between 1946 and 1964 after World War II (Rice,
Lang, Henley and Melzer,2011). Whereas Baby Boomers’
consumer behavior has been researched extensively over
First Author et al.: Preprint submitted to Elsevier Page 1 of 9
Short Title of the Article
the last 70 years, scholars have paid less attention to other
cohorts.
Another principle of the life course theory posits that
individuals’ behavior can change due to geopolitical events
(e.g., war), geopolitical events (e.g., war), and economic
cycles (e.g., recessions) because people and families inter-
act within sociohistorical time. For instance, consumers’
attitudes are likely to be affected by economic up- and
down-swings (Katona,1974). We argue that the COVID-
19 pandemic has altered people’s decisions and behavior.
Clarifying the main factors that influence crowdfunding
campaigns, and the effects of the COVID-19 pandemic on
these factors, can aid scholars and policymakers in fighting
the pandemic.
In the current study, we analyze GoFundMe campaigns
to extract critical aspects contributing to crowdfunding
success. We collect data on 36,370 GoFundMe campaigns
and split the dataset into two parts: before the COVID-
19 outbreak (2019) and after the COVID-19 outbreak
(2020). We also consider whether relevant influencing
factors have changed against the backdrop of the pandemic.
Some scholars have contended that emotional elements
conveyed through text and facial expressions are likely to
attract donors (Rhue and Robert,2018). We hence extract
faces from pictures and judge the displayed emotion using
Baidu’s Application Programming Interface (API)1. In addi-
tion, we extract and infer individual campaign-level features
such as gender, race, age, beauty, target, location, followers,
shares, distinct donors, family status, facial attractiveness,
and crowdfunding duration. Text is a fundamental element
of information transfer; research suggests that textual fea-
tures, including descriptions, reviews, and emotion, heavily
mold crowdfunding’s success (Koch and Siering,2019).
Therefore, we incorporate text emotion into models through
a text scoring model that produces the scores as a feature.
In-depth analysis can then be performed based on the
aforementioned characteristics. If a campaign page visitor
sympathizes with certain project aspects, a longer visit
duration increases the probability of a personal donation and
hence successful crowdfunding (Koch and Siering,2019).
The aesthetic and technical scores of the cover image are
also thought to affect campaigns’ success (Zhang, Lyu and
Luo,2020). In light of the preceding discussion, we propose
three hypotheses:
•Hypothesis 1: The basic features of a crowdfund-
ing project and description significantly affect fund-
raising success.
•Hypothesis 2: Crowdfunding differs significantly be-
tween before and after the COVID-19 outbreak.
•Hypothesis 3: Social disparities in crowdfunding
success (in terms of race and compound factors)
reflect the impacts of the COVID-19 pandemic and
other social factors.
1https://cloud.baidu.com/doc/FACE/s/Uk37c1m9b
3. Related Work
Studies have begun to address the impact of the COVID-
19 pandemic on crowdfunding. Farhoud, Shah, Stenholm,
Kibler, Renko and Terjesen (2021) sought to understand
the effect of the COVID-19 pandemic on social enterprise
crowdfunding and outlined implications for crowdfunding
platforms. The authors discovered that social entrepreneurs’
crowdfunding success rates reflected the nature of the cam-
paigns and innovative ideas. Song, Cohen, Lui, Mmonu,
Brody, Patino, Liaw, Butler, Fergus, Mena et al. (2020)
compared patient and campaign characteristics between 250
users of complementary and alternative medicine (CAM)
and 250 non-CAM users. They observed that CAM users
were more likely to be women and to report more stage IV
cancer. In addition, Elmer, Ward-Kimola and Burton (2020)
noted that campaigns related to COVID-19 were likely to
raise more money and had more attractive descriptions than
other types of campaigns.
Most empirical studies on GoFundMe have involved
the medical setting. For example, Mattingly and colleagues
Mattingly, Li, Ng, Ton-Nu and Owens (2021) applied de-
scriptive statistics about campaign categories and features
to uncover potential associations among features. Notably,
they found that disclosing the virus source contributes to
higher donations Mattingly et al. (2021). Radu and Mc-
Manus Radu and McManus (2018) determined that victims
of intimate partner violence preferred to seek assistance
from informal social ties rather than official organizations.
Their findings conveyed the challenges of obtaining help
through traditional avenues.
Much of the literature on crowdfunding suffers from
several limitations. For example, most studies have referred
to visible website features when determining crowdfunding
success. More importantly, to our best knowledge, no study
has examined the impact of the COVID-19 pandemic on
crowdfunding using large-scale data.
4. Data and Features
4.1. Data sets
We focus on GoFundMe to analyze crucial factors con-
tributing to campaigns’ success. Specifically, we crawl the
36,370 crowdfunding campaigns on GoFundMe and divide
them into two parts: those collected before August 2019 and
after August 2020. This division was intended to indicate
whether the COVID-19 pandemic has affected people’s
attitudes towards crowdfunding along with whether and how
influential factors have changed. The dataset features are
summarized below and shown in Table 1.
4.2. Basic Features
The following campaign features can be directly ex-
tracted from the GoFundMe website: launch date, cover
image, description, category, current amount, target amount,
number of followers, number of shares, and number of
donors. We refer to these features as the basic features.
First Author et al.: Preprint submitted to Elsevier Page 2 of 9
Short Title of the Article
4.3. Inferred Features
•Quality Scores: We use the pre-trained model of neu-
ral image assessment (NIMA) to obtain the aesthetic
and technical scores of each cover image (Talebi and
Milanfar,2018).
•Text Features: First, we merge campaigns’ titles and
descriptions and use them as text data. Next, we
employ the Valence Aware Dictionary for Sentiment
Reasoning (VADER) (Hutto and Gilbert,2014) to
evaluate individual text data and obtain three pre-
dicted emotion scores: positive, negative, and neutral.
Finally, we examine the text using the Linguistic
Inquiry and Word Count (LIWC) program to obtain
more detailed text sentiment scores (i.e., sadness,
anger, and anxiety). Both campaign descriptions are
essential to successful crowdfunding. These potential
effects are difficult to measure directly; therefore, we
train an XLNet model to predict a campaign’s success
and discern its potential effects (Yang, Dai, Yang,
Carbonell, Salakhutdinov and Le,2019).
•Image Features: In terms of image features, we com-
pare the DeepFace API (Serengil and Ozpinar,2020)
with the Baidu API and consider research comparing
the Baidu API with competing APIs (Yang, Wang,
Sarsenbayeva, Tag, Dingler, Wadley and Goncalves,
2020). We eventually opt to use Baidu’s API, which
provides reliable face recognition services. This API
returns the facial attractiveness, age, race, emotion,
and gender of each face in a cover image (also
referred to as the profile image). For simplicity, we
calculate the mean attractiveness of faces when an
image contains more than one person. In terms of
age, we calculate the mean age as well as the num-
ber of children and number of older adults among
pictured people. People under age 15 are considered
children for this purpose, whereas those over age 60
are older adults. We regard the number of people of
different races as the characteristics variable. Baidu’s
API recognizes four races: Black, White, Asian, and
Other. We obtain the number of men and women to
classify gender. We also extract the emotion of each
face (happy, sad, grimace, neutral, or angry).
5. Methodology and Findings
5.1. Campaign Category
We identified 21 categories of campaigns and their
distributions before and after COVID-19, depicted in Fig-
ure 1. Some categories are related to campaigns’ success
rates, while others are not. Specifically, we analyze cate-
gories with the most significant effects on crowdfunding
success via t-tests. The impacts of certain categories on
crowdfunding success have changed over the past two
years. As it would be illogical to compare p-values directly,
we compare whether categories’ significance levels have
Table 1
Sources of (inferred) features.
Statistic Mean St. Dev. Source
basic features
Goal 176,738.900 10,635,100.000 Web crawler
Shares 844.223 2,439.918 Web crawler
Donors 235.444 1,190.212 Web crawler
Followers 234.578 1,153.083 Web crawler
Category Web crawler
days 322.139 288.000 Manually coded
text features
text_positive 0.175 0.071 Vader
text_neutral 0.779 0.076 Vader
text_negative 0.046 0.042 Vader
anx 0.178 0.353 LIWC2015
anger 0.227 0.509 LIWC2015
text_sad 0.487 0.817 LIWC2015
text_scores −0.501 0.468 XLnet
image features
age 27.235 10.952 Average of Baidu API
have_kid 0.534 1.159 Manually coded (by age feature)
old 0.028 0.179 Manually coded (by age feature)
facial attractiveness 35.963 12.408 the sum of Baidu API’s result
Black 0.382 1.346 the sum of Baidu API’s result
Asian 0.709 1.600 the sum of Baidu API’s result
White 2.062 2.725 the sum of Baidu API’s result
other 0.025 0.205 the sum of Baidu API’s result
happy 2.003 2.883 the sum of Baidu API’s result
sad 0.192 0.516 the sum of Baidu API’s result
grimace 0.012 0.113 the sum of Baidu API’s result
neutral 0.673 1.670 the sum of Baidu API’s result
angry 0.050 0.262 the sum of Baidu API’s result
male 1.819 2.703 the sum of Baidu API’s result
female 1.359 2.129 the sum of Baidu API’s result
Table 2
Paired sample statistics (N/A indicates not
available/applicable).
t test
Year
2020 2019 Ratio
Travel & Adventure -0.082∗∗∗ -0.02∗-0.015
Environment 0.014 N/A N/A
Babies, Kids & Family 0.035∗∗∗ 0.044∗∗∗ 0.014
Sports, Teams & Clubs -0.040∗∗∗ -0.048∗∗∗ -0.055
Competitions & Pageants -0.092∗∗∗ -0.090∗∗∗ -0.136
Non-Profits & Charities 0.019∗∗ 0.007 0.003
Medical, Illness & Healing 0.046∗∗∗ 0.016 0.004
Volunteer & Service 0.010 0.011 -0.012
Business & Entrepreneurs -0.043∗∗∗ -0.080∗∗∗ 0.005
Weddings & Honeymoons -0.099∗∗∗ -0.051∗∗∗ 0.007
Funerals & Memorials 0.1254∗∗∗ 0.119∗∗∗ 0.011
Missions, Faith & Church -0.043∗∗∗ -0.041∗∗∗ -0.026
Education & Learning 0.001 0.011 -0.003
Animals & Pets -0.002 -0.007 0.008
Celebrations & Events -0.031∗∗ -0.031∗0.002
Creative Arts, Music & Film -0.035∗∗∗ -0.003 0.013
Accidents & Emergencies 0.052∗∗∗ 0.031∗∗∗ 0.016
Dreams, Hopes & Wishes -0.001 0.019∗0.016
Rent, Food & Monthly Bills 0.013 N/A N/A
Other 0.029∗0.014 0.190
Community & Neighbors 0.023 0.033∗∗ -0.003
ALL -0.003
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
changed; for example, those of the Travel & Adventure,
Non-Profits & Charities,Medical, Illness & Healing,Cel-
ebrations & Events,Creative Arts, Music & Film,Accidents
& Emergencies and Dreams, Hops & Wishes categories have
shifted. These categories’ success ratios have also increased,
except for Travel & Adventure category. People appear
First Author et al.: Preprint submitted to Elsevier Page 3 of 9
Short Title of the Article
Figure 1: Category distribution (best viewed by zoom-in on screen).
even more reluctant to donate to campaigns in the Travel
& Adventure category since the COVID-19 outbreak. In
particular, the Medical, Illness & Healing and Non-Profits &
Charities categories were not significant before the COVID-
19 outbreak but have become highly significant thereafter.
Regarding the Dreams, Hops & Wishes category, people
who had been more willing to donate are now reluctant.
For the other declining categories, among most of those
to which people were unwilling to donate before, people
have become even more hesitant to contribute(e.g, Sports,
Teams & Clubs;Missions, Faith & Church;Weddings &
Honeymoons;Competitions & Pageants). Of note, every
category’s success rate has declined since the COVID-19
outbreak.
5.2. Prediction
To analyze the contributions of influential factors to
campaigns’ success, we employ the XGBoost method and
divide features among basic features, text features, and
image features. We feed (1) basic features, (2) basic features
plus text features, and (3) all features into an XGBoost
model to obtain the accuracy, F1 score, precision, and recall,
respectively. We also train an XGBoost model to construct
a regression model with the ratio/percentage of success
as output. The inferred features significantly improve the
model’s performance (Table 3).
To extract statistically significant features, we construct
a logistic regression model for the classification problem;
results appear in Table 4and lend support to Hypothesis
1. To further analyze the impact of each feature on the
success of crowdfunding, we conduct several counterfactual
analysis experiments. Statistically significant features are
examined in a separate analysis due to our high number of
features.
6. Counterfactual analysis
We compare the effects of the abovementioned features
on crowdfunding success before and after the COVID-
19 outbreak and further analyze whether the pandemic
has altered people’s priorities. To better understand the
causality between these features and crowdfunding success,
Table 3
XGBoost performance metrics.
Classification Regression
Accuracy F1 Precision Recall R-square
Base 0.8239 0.7012 0.7563 0.6536 0.2859
Base+Text 0.8354 0.7244 0.7698 0.6839 0.4915
All 0.8523 0.7571 0.8009 0.7112 0.5445
we employ counterfactual analysis, a popular means of
comparative inquiry (Chang, Pierson, Koh, Gerardin, Red-
bird, Grusky and Leskovec,2021). We test our remaining
two hypotheses based on findings from the counterfactual
analysis experiments.
Removing the category factors to reduce the impact In
counterfactual analysis, we turn the factor indicator of each
category into 0 in each experiment as shown in Table 5. We
find that Creative Arts, Music & Film and Dreams Hopes
& Wishes had positive causal impacts on the success of
crowdfunding before the COVID-19 outbreak but a negative
causal effect after the outbreak. The positive effects of
Medical, Illness & Healing,Accidents & Emergencies and
Non-Profits & Charities on the success of crowdfunding
have increased dramatically since the COVID-19 outbreak.
The significance of these three categories in Table 5has
changed, suggesting the pandemic’s profound effects on
each. The COVID-19 pandemic has accordingly informed
individuals’ campaign preferences, with people focusing
more on medical or charity topics than dreams, arts, and
charity topics. These results validate Hypothesis 2. Notably,
the ratio in the table changes negligibly across multiple
categories; eliminating the impact of one category does not
significantly change the overall success rate, but this out-
come does not mean that these categories are not significant.
Improving the effect of features First, we divide the
remaining features into basic features, text features, and
image features after removing category features. Combined
with the counterfactual experiment, we analyze only sta-
tistically significant features. Table 6presents the results
of counterfactual analyses of such features using logistic
First Author et al.: Preprint submitted to Elsevier Page 4 of 9
Short Title of the Article
Table 4
Logistic regression model results for the relationship between campaigns’ features and success.
Dependent variable: success
models
basic model basic+text model text+image model aggregated model
Goal −0.00005∗∗∗ (0.00000) −0.00005∗∗∗ (0.00000) −0.00005∗∗∗ (0.00000)
Shares 0.00004∗∗∗ (0.00001) 0.00002∗(0.00001) 0.00002 (0.00001)
Donors 0.004∗∗∗ (0.0004) 0.003∗∗∗ (0.0004) 0.003∗∗∗ (0.0004)
Followers −0.0003 (0.0003) −0.0005 (0.0004) −0.001 (0.0004)
days −0.00002 (0.0001) −0.001∗∗∗ (0.0001) −0.001∗∗∗ (0.0001)
text_positive 44.048 (39.045) 39.543 (35.766) 46.768 (39.212)
text_neutral 43.616 (39.044) 39.151 (35.766) 46.565 (39.211)
text_negative 44.307 (39.045) 38.454 (35.766) 47.058 (39.212)
anx −0.105∗∗ (0.051) −0.060 (0.048) −0.097∗(0.052)
anger 0.027 (0.036) 0.050 (0.034) 0.031 (0.037)
text_sad −0.001 (0.023) 0.023 (0.021) −0.006∗(0.023)
text_scores 2.077∗∗∗ (0.049) 1.929∗∗∗ (0.045) 2.055∗∗∗ (0.049)
age −0.005∗∗∗ (0.002) −0.0001 (0.002)
have_kid −0.030 (0.020) 0.038∗(0.020)
old 0.040 (0.100) −0.049 (0.111)
facial attractiveness −0.002 (0.001) −0.00003 (0.002)
Black −0.052 (0.034) −0.078∗∗ (0.036)
Asian −0.064∗∗ (0.033) −0.065∗∗ (0.035)
White −0.043 (0.031) −0.018 (0.034)
other −0.257∗∗ (0.102) −0.152 (0.112)
happy 0.037 (0.031) 0.058∗(0.033)
sad 0.050 (0.048) 0.046 (0.052)
grimace 0.182 (0.151) 0.153 (0.164)
neutral 0.036 (0.035) 0.020 (0.037)
angry −0.056 (0.081) −0.095 (0.087)
female 0.005 (0.014) −0.021 (0.015)
male
aesthetic scores 0.003 (0.032) −0.027 (0.035)
technical scores −0.009 (0.032) 0.143∗∗∗ (0.036)
Constant −0.417∗∗∗ (0.028) −42.840 (39.043) −38.840 (35.766) −46.402 (39.211)
Observations 19,185 19,185 19,075 19,075
Log Likelihood −10,364.040 −9,166.808 −10,626.380 −9,080.760
Akaike Inf. Crit. 20,740.080 18,359.620 21,300.760 18,219.520
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
regression. Campaign success rates increase significantly
when increasing these features with their mean values.
•Goal Amount The counterfactual experiment re-
veals that a campaign’s goal amount has the most
significant causal relationship with success. The dis-
played funding goal signals the project’s complexity.
In general, the higher the crowdfunding goal, the
lower the project’s success rate Barbi and Bigelli
(2017). Our counterfactual experiment results are also
consistent with our previous findings.
•Duration The experimental results indicate that
the duration of crowdfunding campaigns positively
influences their success: the longer a project lasts, the
more likely people are to find and share it (𝑝 < 0.001).
Campaigns with a greater number of shares typically
receive more donations (𝑝 < 0.001).
•Donors The relationship between the number of
donors and success is significant (𝑝 < 0.001); that is,
the number of donors positively affects crowdfunding
success.
•Emotion of text Text scores obtained by XLNet
have had significant positive impacts on crowdfund-
ing success since the COVID-19 outbreak, implying
a statistically significant relationship between text
and success. Further analysis reveals a significant
relationship between emotion and campaign success
rate. For example, the anxiety (𝑝 < 0.1) and sadness
(𝑝 < 0.1) conveyed in text each negatively influence
success. This pattern corroborates our counterfactual
experiment.
•Technical Score Technical scores obtained via
the pretrained NIMA model have had a significant
positive impact on crowdfunding success since the
First Author et al.: Preprint submitted to Elsevier Page 5 of 9
Short Title of the Article
COVID-19 outbreak (𝑝 < 0.01). A statistically signif-
icant relationship thus seems to exist between image
quality and campaign success.
•Race Among image features, race appears to have
differentially affected crowdfunding success since the
COVID-19 outbreak. Of all races, Black has the most
significant impact on success (𝑝 < 0.05). While Black
people are historically less likely to be funded than
people of other races, this effect was less significant
before COVID-19 (𝑝 < 0.1). In other words, Blacks
are even less likely to be funded during the pandemic.
Asian ethnicity also has a statistically significant im-
pact on the success of crowdfunding (𝑝 < 0.05) after
the COVID-19 outbreak in the full dataset; no statis-
tically significant impact applies before the outbreak.
These results support Hypothesis 3. We suspect that
the former finding might be related to the root cause
of the #BlackLivesMatter movement with the latter
being tied to the root cause of the #StopAsianHate
movement. These two directions would be interesting
to investigate in future work.
•Emotion of cover image We examine the rela-
tionship between the emotion of the cover image and
campaign success, identifying a significant difference
between them. The counterfactual experiments show
that sadness adversely affects success whereas happi-
ness has a positive impact (𝑝 < 0.1). People are more
willing to donate to campaigns whose cover images
carry positive valence than to those whose images
carry negative valence. This discovery is somewhat
surprising, as donors seem to prefer giving to individ-
uals who are optimistic about life.
•Children The logistic regression model indicates
that the impact of children on crowdfunding success
tends to be significantly positive if the campaign is
related to children. In addition, we confirm that the
effect is positive and consistent with the regression
model.
7. LDA Topic Generation and Analysis
We further analyze the relationships among crowdfund-
ing success and the topics in each category using latent
Dirichlet allocation (LDA) topic modeling (Blei, Ng and
Jordan,2003). We specifically employ LDA to divide each
category into five topics. We then build a regression model
to analyze which topics have the most positive impacts on
the success of crowdfunding and which have the most neg-
ative effects. We focus on categories featuring significant
changes before and after the COVID-19 outbreak: Accidents
& Emergencies,Travel & Adventure,Medical, Illness &
Healing,Celebrations & Events,Creative Arts, Music &
Film and Dreams, and Hopes & Wishes. The results are
shown in Figure 2, , where c denotes the correlation coeffi-
cient between success and topics. The first value represents
Table 5
Counterfactual analysis experiment for categories.
Year
2020 2019
Prediction Rate Prediction Rate
Travel & Adventure 1.014 .0141 1.030 .0296
Environment 1.000 N/A N/A
Babies Kids & Family 0.944 -.0060 0.990 -.0095
Sports Teams & Clubs 1.001 .0012 0.997 .0029
Competitions & Pageants 1.008 .0081 1.036 .0358
Non-Profits & Charities 0.997 -.0030 1.000 .0000
Medical Illness & Healing 0.951 -.0494 0.951 -.0491
Volunteer & Service 1.009 .0087 0.998 -.0019
Business & Entrepreneurs 1.000 .0030 1.005 .0033
Weddings & Honeymoons 1.004 .0045 1.011 .0110
Funerals & Memorials 0.972 -.0278 0.981 -.0186
Missions Faith & Church 0.999 -.0012 0.992 -.0081
Education & Learning 0.985 -.0150 0.981 -.0186
Other 0.994 -.0057 0.999 -.0010
Animals & Pets 1.018 .0183 1.012 .0119
Celebrations & Events 1.027 .0269 1.019 .0191
Creative Arts Music & Film 1.002 .00240 0.995 -.0048
Accidents & Emergencies 0.987 -.0129 0.988 -.0124
Rent Food & Monthly Bills 1.000 N/A N/A
Dreams Hopes & Wishes 1.000 .0003 1.000 -.0005
Community & Neighbors 1.000 -.0003 0.998 -.0019
Table 6
Counterfactual analysis experiments for important features.
Year
2020 2019
Prediction Rate Prediction Rate
Goal 0.002 -.9979 0.042 -.9577
Shares 0.984 -.0157 0.933 -.0665
Donors 1.420 .4204 1.677 .6766
days 1.007 .0071 1.014 .0144
have kid 1.007 .0071 1.037 .0368
Black 0.898 -.1021 0.902 -.0982
Asian 0.965 -.0350 0.977 -.0228
White 1.016 .0236 1.006 .0060
happy 1.023 .0235 1.023 .0233
anxiety 0.994 -.0059 0.995 -.0051
text sad 0.970 -.0297 0.995 -.0047
text scores 0.608 -.2853 0.775 -.2252
technical scores 1.025 .0249 1.013 .0130
the correlation coefficient before the COVID-19 outbreak,
while the second value represents the coefficient after the
outbreak. The frequencies of some words in the same cate-
gory are high, leading to substantial overlap between topics.
We therefore add the words whose frequencies are in the top
10 of all topics to the stop words list. We next apply an LDA
model again to generate topics and repeat this process until
no such words remain. In the end, the coherence of the topic
model is 0.34803. In addition, we observe some significant
changes in the Other category before and after the COVID-
19 outbreak. Because this category includes various topics,
we construct separate topic models for this category to
analyze the impact of each topic on crowdfunding success.
We find significant changes in the Other category be-
fore and after the COVID-19 outbreak and analyze them
First Author et al.: Preprint submitted to Elsevier Page 6 of 9
Short Title of the Article
Figure 2: Topics under different categories. The first value represents the correlation coefficient before the COVID-19
outbreak, while the second value represents the coefficient after the outbreak. A green value indicates that the topic has
a positive impact on success, while a red value indicates that the topic has a negative impact on success. Note those topics
with changes in the polarity of the impact before and after the COVID-19 pandemic.
Figure 3: Topics under the Other category that had significant effect on success before the COVID-19 outbreak. We generate
four topics with the best coherence value: 0.35192.
individually. We start with a topic number with the best
coherence value. Topics in this category were relatively
scattered before the COVID-19 outbreak; we divide them
into four topics (coherence score: 0.35192; see Figure 3).
The topics are relatively concentrated after the COVID-19
outbreak and are thus only divided into two topics (coher-
ence score: 0.32625; see Figure 4). Before the COVID-
19 outbreak, the Other category included dreams, gifts,
children, travel, honeymoons & weddings, and other topics.
Combined with the above results, we find that most of these
topics negatively affect the success of crowdfunding. By
contrast, after the COVID-19 outbreak, most topics focus
on family, children, friends, and medical care; these topics
positively influence campaigns’ success. The significance
of the Other category appears to have changed given these
differences before and after the outbreak.
8. Conclusion and Discussion
Our study analyzes the changes in significant features
influencing crowdfunding success before and after the
COVID-19 outbreak and validates three hypotheses. The
results suggest a substantial difference in certain categories
before and after the outbreak. Although dreams, travel,
or other topics were less likely to be funded before the
COVID-19 outbreak, people have begun to donate to these
campaigns thereafter. People have also started to pay more
First Author et al.: Preprint submitted to Elsevier Page 7 of 9
Short Title of the Article
Figure 4: Topics under the Other category that had significant effect on success during the COVID-19 pandemic. We generate
two topics with the best coherence value: 0.32625.
attention to medical, accident, and charity projects. Some
categories have not changed before or after the outbreak. For
instance, campaigns including babies, family, funerals, and
memorials have always attracted donations relatively easily.
Conversely, campaigns involving sports, weddings, and
missions have received fewer donations. We also observe
significant differences in crowdfunding success by race.
The COVID-19 pandemic has made it more challenging for
Black and Asians to raise money because the pandemic has
exacerbated existing social disparities.
Consistent with research demonstrating that image at-
tributes influence the success of crowdfunding campaigns (Bretschnei-
der and Leimeister,2017;Hou, Zhang and Zhang,2020),
sadness, anxiety, and anger are found to have either negative
or no effects on crowdfunding in this study. Yet campaigns
with positive emotions conveyed through the cover image
are more likely to be funded than those showing negative
emotions. These findings suggest that fundraisers should
express optimistic attitudes towards life in addition to
describing their misfortune.
Our study also supports the prevalence of discrimination
in online marketplaces (Edelman, Luca and Svirsky,2017;
Farmaki and Kladou,2020). To the best of our knowledge,
this study is the first to reveal the presence of racial discrim-
ination on crowdfunding platforms.
9. Limitations and Future Research
Several limitations of this study merit attention. The first
limitation is that our findings may not generalize to other
countries because we only use U.S. data on GoFundMe.
However, the COVID-19 pandemic may have a considerable
impact on crowdfunding campaigns throughout the world.
Data collection is required in other countries to understand
the extent to which our findings are generalizable. Second,
we only consider stable features on the GoFundMe website.
We will account for other dynamic factors in the future,
such as the characteristics of each donation. Finally, we
do not differentiate between categories of crowdfunding
campaigns. Different categories may distinctly influence the
success of crowdfunding (Zhang et al.,2020), which is
worth investigating as well.
References
Barbi, M., Bigelli, M., 2017. Crowdfunding practices in and outside the us.
Research in International Business and Finance 42, 208–223.
Binstock, R.H., George, L.K., Cutler, S.J., Hendricks, J., Schulz, J.H.,
2011. Handbook of aging and the social sciences. Elsevier.
Blei, D.M., Ng, A.Y., Jordan, M.I., 2003. Latent dirichlet allocation. the
Journal of machine Learning research 3, 993–1022.
Bretschneider, U., Leimeister, J.M., 2017. Not just an ego-trip: Exploring
backers’ motivation for funding in incentive-based crowdfunding. The
Journal of Strategic Information Systems 26, 246–260.
Chang, S., Pierson, E., Koh, P.W., Gerardin, J., Redbird, B., Grusky, D.,
Leskovec, J., 2021. Mobility network models of covid-19 explain
inequities and inform reopening. Nature 589, 82–87.
Edelman, B., Luca, M., Svirsky, D., 2017. Racial discrimination in
the sharing economy: Evidence from a field experiment. American
economic journal: applied economics 9, 1–22.
Elmer, G., Ward-Kimola, S., Burton, A.G., 2020. Crowdfunding during
covid-19: An international comparison of online fundraising. First
Monday .
Farhoud, M., Shah, S., Stenholm, P., Kibler, E., Renko, M., Terjesen, S.,
2021. Social enterprise crowdfunding in an acute crisis. Journal of
Business Venturing Insights 15, e00211.
Farmaki, A., Kladou, S., 2020. Why do airbnb hosts discriminate?
examining the sources and manifestations of discrimination in host
practice. Journal of Hospitality and Tourism Management 42, 181–189.
Giele, J.Z., Elder, G.H., 1998. Life course research: Development of a field.
Methods of life course research: Qualitative and quantitative approaches
, 5–27.
Ho, H.C., Chiu, C.L., Mansumitrchai, S., Yuan, Z., Zhao, N., Zou, J., 2021.
The influence of signals on donation crowdfunding campaign success
during covid-19 crisis. International Journal of Environmental Research
and Public Health 18, 7715.
Hou, J.R., Zhang, J.J., Zhang, K., 2020. Pictures that are worth a
thousand donations: How emotions in project images drive the success
of crowdfunding campaigns? an image design perspective. An Image
Design Perspective (October 20, 2020) .
Hutto, C., Gilbert, E., 2014. Vader: A parsimonious rule-based model
for sentiment analysis of social media text, in: Proceedings of the
International AAAI Conference on Web and Social Media.
Kaartemo, V., 2017. The elements of a successful crowdfunding campaign:
A systematic literature review of crowdfunding performance. Interna-
tional Review of Entrepreneurship 15, 291–318.
Katona, G., 1974. Psychology and consumer economics. Journal of
Consumer Research , 1–8.
Koch, J.A., Siering, M., 2019. The recipe of successful crowdfunding
campaigns. Electronic Markets 29, 661–679.
Mattingly, T.J., Li, K., Ng, A., Ton-Nu, T.L., Owens, J., 2021. Exploring
patient-reported costs related to hepatitis c on the medical crowdfunding
page gofundme®. PharmacoEconomics-Open 5, 245–250.
Radu, M.B., McManus, L., 2018. A qualitative analysis of requests for
financial help via gofundme by victims of intimate partner violence.
Sociological Spectrum 38, 312–325.
Rajwa, P., Hopen, P., Mu, L., Paradysz, A., Wojnarowicz, J., Gross, C.P.,
Leapman, M.S., 2020. Online crowdfunding response to coronavirus
disease 2019. Journal of general internal medicine 35, 2482–2484.
Rhue, L., Robert, L.P., 2018. Emotional delivery in pro-social crowdfund-
ing success, in: Extended Abstracts of the 2018 CHI Conference on
Human Factors in Computing Systems, pp. 1–6.
Rice, N.E., Lang, I.A., Henley, W., Melzer, D., 2011. Common health
predictors of early retirement: findings from the english longitudinal
study of ageing. Age and ageing 40, 54–61.
First Author et al.: Preprint submitted to Elsevier Page 8 of 9
Short Title of the Article
Serengil, S.I., Ozpinar, A., 2020. Lightface: A hybrid deep face recog-
nition framework, in: 2020 Innovations in Intelligent Systems and
Applications Conference (ASYU), IEEE. pp. 23–27. doi:10.1109/
ASYU50717.2020.9259802.
Song, S., Cohen, A.J., Lui, H., Mmonu, N.A., Brody, H., Patino, G., Liaw,
A., Butler, C., Fergus, K.B., Mena, J., et al., 2020. Use of gofundme®
to crowdfund complementary and alternative medicine treatments for
cancer. Journal of cancer research and clinical oncology 146, 1857–
1865.
Talebi, H., Milanfar, P., 2018. Nima: Neural image assessment. IEEE
Transactions on Image Processing 27, 3998–4011.
Yang, K., Wang, C., Sarsenbayeva, Z., Tag, B., Dingler, T., Wadley, G.,
Goncalves, J., 2020. Benchmarking commercial emotion detection
systems using realistic distortions of facial image datasets. The Visual
Computer , 1–20.
Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R., Le, Q.V.,
2019. Xlnet: Generalized autoregressive pretraining for language
understanding. arXiv preprint arXiv:1906.08237 .
Zhang, X., Lyu, H., Luo, J., 2020. What contributes to a crowdfunding
campaign’s success? evidence and analyses from gofundme data. arXiv
preprint:2001.05446 .
First Author et al.: Preprint submitted to Elsevier Page 9 of 9
