You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/275835037

Effectiveness of Internet-Based Affect Induction Procedures: A Systematic


Review and Meta-Analysis

Article  in  Emotion · May 2015


DOI: 10.1037/emo0000035 · Source: PubMed

CITATIONS READS
35 1,433

3 authors:

Rebecca A Ferrer Emily Grenen


National Institutes of Health London School of Hygiene and Tropical Medicine
138 PUBLICATIONS   3,216 CITATIONS    7 PUBLICATIONS   87 CITATIONS   

SEE PROFILE SEE PROFILE

Jennifer M. Taber
Kent State University
73 PUBLICATIONS   936 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Rebecca A Ferrer on 04 November 2015.

The user has requested enhancement of the downloaded file.


Emotion
Effectiveness of Internet-Based Affect Induction
Procedures: A Systematic Review and Meta-Analysis
Rebecca A. Ferrer, Emily G. Grenen, and Jennifer M. Taber
Online First Publication, May 4, 2015. http://dx.doi.org/10.1037/emo0000035

CITATION
Ferrer, R. A., Grenen, E. G., & Taber, J. M. (2015, May 4). Effectiveness of Internet-Based
Affect Induction Procedures: A Systematic Review and Meta-Analysis. Emotion. Advance
online publication. http://dx.doi.org/10.1037/emo0000035
Emotion In the public domain
2015, Vol. 15, No. 2, 000 http://dx.doi.org/10.1037/emo0000035

Effectiveness of Internet-Based Affect Induction Procedures: A Systematic


Review and Meta-Analysis
Rebecca A. Ferrer, Emily G. Grenen, and Jennifer M. Taber
National Cancer Institute/National Institutes of Health

Procedures used to induce affect in a laboratory are effective and well-validated. Given recent method-
ological and technological advances in Internet research, it is important to determine whether affect can
be effectively induced using Internet methodology. We conducted a meta-analysis and systematic review
of prior research that has used Internet-based affect induction procedures, and examined potential
moderators of the effectiveness of affect induction procedures. Twenty-six studies were included in final
analyses, with 89 independent effect sizes. Affect induction procedures effectively induced general
positive affect, general negative affect, fear, disgust, anger, sadness, and guilt, but did not significantly
induce happiness. Contamination of other nontarget affect did not appear to be a major concern. Video
inductions resulted in greater effect sizes. Overall, results indicate that affect can be effectively induced
in Internet studies, suggesting an important venue for the acceleration of affective science.

Keywords: emotion induction, mood induction, Internet experiments, meta-analysis

Supplemental materials: http://dx.doi.org/10.1037/emo0000035.supp

Hundreds of studies have effectively induced affective or emo- online studies are paid small sums (e.g., $0.10 –$1) for answering
tional states among human participants in experimental psycho- a question online (for example, Amazon mTurk [https://www.mturk
logical studies (Lench et al., 2011; Westermann et al., 1996). .com/mturk/welcome], Survey Monkey Audience [https://www
These studies have resulted in important discoveries about the role .surveymonkey.com/mp/audience], and GfK Knowledge Panel),
of emotion for judgment and decision-making, sensation and per- have exponentially accelerated the number of experiments con-
ception, interpersonal attribution and interaction, and behavioral ducted online (Berinsky, Huber, & Lenz, 2012; Buhrmester,
action tendencies. Such studies have been conducted primarily in Kwang, & Gosling, 2011; Horton, Rand, & Zeckhauser, 2011;
controlled laboratory settings, and the effectiveness of different Paolacci & Chandler, 2014). In parallel with this acceleration,
affect inductions in the laboratory has been validated through software specific to conducting Internet research has been devel-
careful analysis (Lench et al., 2011; Westermann et al., 1996). oped by the research community (e.g., psiTurk 2.0) and industry
In recent years, methodological and technological advances in (for example, Qualtrics [http://www.qualtrics.com/], Survey Mon-
Internet research have shifted the focus from the laboratory as the key). Moreover, the growth of Internet research is reflected in
sole, or even most desirable, venue for experimental psychology conference workshops specifically aimed at identifying advances,
research. Indeed, experimental studies may be among the most challenges, and opportunities in experimental psychology con-
common type of psychological research that takes place online ducted via the Internet (e.g., Suri, Mason, & Goldstein, 2014).
(Skitka & Sargis, 2006). This shift has the potential to revolution- Given the advent of online research, it is important to understand
ize and positively disrupt the field of experimental psychology. whether emotion can be effectively induced using Internet meth-
The advent of readily available Internet panels, in which people odology.
who sign up to be in a pool of individuals selected to complete Internet-based experimental studies have several advantages.
Internet samples tend to be more representative of the general
public on characteristics such as age, income, and education than
student or convenience samples (Berinsky et al., 2012; Paolacci,
Chandler, & Ipeirotis, 2010; Ross et al., 2010), perhaps because
Rebecca A. Ferrer, Basic Biobehavioral and Psychological Sciences they can reach more geographically and otherwise diverse popu-
Branch, Behavioral Research Program, Division of Cancer Control and lations (Birnbaum, 2004; Reips, 2000). This is particularly true
Population Sciences, National Cancer Institute/National Institutes of when complex sampling strategies are used to yield nationally
Health; Emily G. Grenen and Jennifer M. Taber, Behavioral Research representative samples (e.g., providing computer and Internet ac-
Program, Division of Cancer Control and Population Sciences, National cess by companies such as GfK), but is even the case when the
Cancer Institute/National Institutes of Health.
sample is an Internet convenience sample or targeted population
Emily G. Grenen is now at ICF International.
Correspondence concerning this article should be addressed to Rebecca
(Birnbaum, 2004; Reips, 2000). Moreover, the lower costs and
A. Ferrer, Basic Biobehavioral and Psychological Sciences Branch, Be- greater efficiency of online experimentation allow more junior
havioral Research Program, Division of Cancer Control and Population researchers and faculty at smaller or less resourced universities to
Sciences, National Cancer Institute/National Institutes of Health, Rock- conduct research and contribute to psychological scholarship (Hor-
ville, MD 20850. E-mail: ferrerra@mail.nih.gov ton & Chilton, 2010; Kraut et al., 2004). In contrast to traditional

1
2 FERRER, GRENEN, AND TABER

laboratory settings where the cost of experiment administration dations specific to epidemiology studies, such as the evaluation of
can be quite high, most online studies are automated to perform the the comparability of cases and controls).
activities that would normally require human labor: recruitment,
instruction delivery, experimental manipulation, and data collec-
tion. Moreover, the relative anonymity of the Internet can lead to Search Strategy
higher response rates when questions are perceived to be sensitive Relevant publications, in press manuscripts, dissertations, post-
(Birnbaum, 2004; Skitka & Sargis, 2006). Internet surveys are also ers presented at conferences, and raw data were identified through:
lower in measurement error and social desirability bias (Skitka & (a) searches of electronic databases (PsycINFO, PsycARTICLES);
Sargis, 2005, 2006), and online study participants have reported (b) email requests to professional organization listservs (Society
being intrinsically motivated to engage in studies (Ipeirotis, 2010; for Judgment and Decision Making, Society for Personality and
Paolacci et al., 2010; Ross et al., 2010). Social Psychology, Social Personality and Health Network, Soci-
Despite these well-demonstrated advantages, there have been ety for Medical Decision Making); (c) searching reference lists
concerns that experimental designs administered over the Internet from reviews or published studies; (d) emails to authors who have
lack control and precision (see Kraut et al., 2004; Skitka & Sargis, published in the field; and (e) a search of publicly available data on
2006). In a laboratory, researchers can control the research envi- the Time Sharing Experiments in the Social Sciences Website
ronment and paradigm administration. Thus, even experiments (TESS; http://www.tess.org). The keywords used for the initial
conducted on a laboratory computer may be fundamentally differ- database search were combinations of the terms [emotion; “posi-
ent than experiments conducted via the Internet, given that re- tive affect”; “negative affect”; affective; mood; stress; sadⴱ; hapⴱ;
searchers cannot control for elements in the participant’s physical fear; anxⴱ; angⴱ; gratⴱ; prideⴱ; disgustⴱ], [online; mturk; amazon;
environment (e.g., noise, lighting, or other people), and as such can “knowledge networks”; Web; Internet], and [manipⴱ; inductⴱ; elic-
have less confidence that participants are fully engaged in exper- itⴱ; influenⴱ].
imental paradigms. Furthermore, in the laboratory, an experi- Studies were included if they fit all of the following criteria: (a)
menter can monitor participants’ behavior to ensure that they are took place entirely online in a nonlaboratory setting; (b) included
taking the task seriously and understanding directions. However, experimental inductions of discrete emotions, mood, or stress; (c)
evidence suggests that Internet studies can replicate the results of contained manipulation checks that assessed participants’ self-
laboratory paradigms (Berinsky et al., 2012; Krantz & Dalal, reported affective state after the induction; and (d) included a
2000), although it is unknown whether the specific strategy of comparison condition (between-subjects design).1 All articles lo-
inducing emotion or manipulating participants’ emotional or af- cated in the search were published in English. Authors were
fective state can be effectively reproduced over the Internet. contacted directly when studies did not include sufficient statistical
Given the critical importance of carefully considering method- information to calculate effect size. Two studies were excluded
ological differences of Internet research compared with laboratory because the authors were unable to provide the necessary infor-
studies (Mathy, Kerr, & Haydin, 2003), one primary goal of this mation. Studies that fulfilled the above criteria as of May 1, 2014,
meta-analysis is to identify methods of induction that are particu- were included; the final sample included 26 studies, for which 89
larly promising for inducing affective states in Internet studies. independent effect sizes could be extracted. Figure 1 contains the
Thus, we defined and coded inductions based on categories that PRISMA diagram of the manuscript search, retrieval, and coding
have been examined in larger meta-analyses synthesizing primar- process (Moher, Liberati, Tetzlaff, & Altman, 2009).
ily laboratory studies (e.g., Lench et al., 2011; Westermann et al.,
1996).
We also sought to examine whether certain types of affect are Coding and Reliability
more difficult to induce over the Internet. Positive affect is typi-
Studies were coded for (a) publication status (i.e., published
cally more difficult to induce than negative affect (Augustine &
journal article vs. dissertation or unpublished data); (b) year of
Hemenover, 2009; Ito et al., 1998), and we expected this finding to
publication (or year study was conducted, if unpublished); (c)
replicate with Internet inductions. Finally, because of concerns
academic department of first author; (d) explicit description of the
over experimental control in Internet-based experimental manipu-
theoretical framework underlying study design; (e) the behavioral
lations (see Kraut et al., 2004; Skitka & Sargis, 2006), we exam-
or disciplinary domain and context of the study (e.g., financial,
ined whether affective states could be induced over the Internet
health, risk perception, or valuation); (f) reporting of use of a cover
without contamination of other affective states—that is, whether
story to blind participants to affective condition; and (g) reporting
an induction targeting one affective state induces only that affec-
of participant randomization to experimental condition. We also
tive state, but not others. If Internet-based inductions can have
coded characteristics of the participants in each sample, including:
sufficient experimental control and are effective, they should ex-
(a) mean age; (b) percent female; (c) percent White; (d) percent of
hibit significant overall effect sizes for target, but not nontarget,
sample currently living in the United States; (e) percent with some
affective states.

1
One between-subjects study with no neutral group and no clear com-
Method parisons was excluded (Egermann & McAdams, 2013); in this study, there
We followed consensus guidelines for conducting and reporting were 24 musical affect induction conditions, but no explicit hypotheses
about which conditions should be different from one another or whether
meta-analyses of observational studies in epidemiology (Stroup et each musical selection should induce positive or negative affect. As such,
al., 2000), as appropriate and adapted to the current meta-analysis we were unable to determine which conditions would be expected to differ
of experimental psychological research (e.g., omitting recommen- and how, and including it in analyses could introduce noise in effect sizes.
INTERNET-BASED AFFECT INDUCTION PROCEDURES 3

Potentially relevant study titles


identified by searches and
screened for relevance (3,741
studies) Studies excluded (2,841 studies)
Reasons: Search term “online” referred to brain
processes/ related behaviors
(neuroscience)
Studies involved clinical samples with
severe psychopathy
Abstracts read/ data examined for
more detailed evaluation (900
Studies excluded (488 studies)
studies)
Reasons: No affect induction
Study took place in laboratory
No comparison condition
Induction part of larger intervention

Studies obtained for more


detailed evaluation (412 studies)
Studies excluded (354 studies)
Reasons: No affect induction
Study took place in laboratory
No comparison condition

Potentially appropriate studies to


be included in meta-analysis (58
studies)
Studies excluded (32 studies)
Reasons: No manipulation check
Inadequate statistical information

Studies in final meta-analysis


(26 studies; k = 89)

Figure 1. PRISMA diagram of study selection.

college education or higher; and (f) population (e.g., Internet that made them feel particularly strongly on a specified affective
panel, student sample). state. Such inductions often involve instructing participants to list
The affective induction procedure in each study was coded for: 3–5 events that made them feel most emotional before choosing
(a) the affective state targeted in the experimental condition(s) as one to write about (e.g., Lerner & Keltner, 2001; Strack, Schwarz,
specified by the study authors; (b) the affective state targeted in the & Gschneidinger, 1985). Video elicitations involve presenting a
experimental condition(s) as specified based on coder judgment; film clip, usually excerpted from a feature length film and often
(c) the affective state targeted in the comparison condition (e.g., selected from a standardized set of clips (Gross & Levenson,
neutral affect) as specified by the study authors; (d) the affective 1995).
state targeted in the comparison condition as specified based on Picture inductions involve presenting previously validated (and
coder judgment; (e) type of affect induction (e.g., autobiographical often taken from standard databases) affectively laden images
recall); and (f) type of comparison induction. either supraliminally or subliminally (e.g., International Affective
Induction categories were selected so that they could be com- Pictures System [http://csea.phhp.ufl.edu/media.html]; Bradley &
pared with previous meta-analyses of affect induction and mood Lang, 1999). All studies identified here used supraliminal presen-
(e.g., Lench et al., 2011; Westermann et al., 1996). Categories tation. Real-life inductions involve using procedures that are os-
were not always mutually exclusive; for example, one study used tensibly outside study protocols to influence affect. In Internet
pictures and text inductions simultaneously (e.g., Boudewyns et research, real-life affect inductions often involve manipulating the
al., 2013). Web site interface in such a way that it is not meant to be
We identified a variety of techniques used to induce affect, and obviously part of the experiment. For example, such inductions
describe each in turn. A novel category we identified was supra- could involve purposely degrading the usability of the Web site
liminal word presentation, in which participants are exposed to a interface to induce negative mood (Hängii, 2004). Text inductions
series of positive or negative affectively laden words. Word pre- involve presenting paragraphs of text with affectively laden con-
sentation is supraliminal, which distinguishes it from subliminal tent, such as newspaper articles, jokes, or portions of textbooks or
priming in previous meta-analyses. In autobiographical recall, articles. Velten mood induction procedures involve instructing
participants are instructed to write in detail about an experience participants to put themselves in a target mood state, and to
4 FERRER, GRENEN, AND TABER

subsequently read a series of self-referent statements (Velten, neutral) we divided the size of the comparison condition n in-
1968). cluded in calculations for each comparison by the number of
Each study was coded by the lead author (RF). A second comparisons in which the comparison condition’s data were used
coder (EG) independently coded 47 effect sizes (68.12% of k). (Borenstein et al., 2009). In subgroup analyses, we recalculated
Rater agreement was high (percent agreement ⬎93.62, ␬s ⬎ effect sizes to include the full n for comparison or experimental
.48). Discrepancies were resolved through discussion among groups, even if those n were divided in main analyses (e.g., when
the coders. The coding form can be obtained from the corre- an experiment had an anger, sadness, and neutral condition, the
sponding author. neutral condition n was halved in calculations for main analyses,
From these coding categories, we further coded: (a) whether the but was not divided in effect size calculations used in subgroup
affective state targeted was positively or negatively valenced; (b) analyses of studies with sadness inductions).
whether the affective state targeted was a discrete emotion or Previous meta-analyses examining the effectiveness of affect
mood state; (c) whether the author and coder ratings of affective inductions have handled the presence of nonneutral comparison
state targeted were discrepant; and (d) whether the author and conditions differently. Westerman and colleagues (1996) excluded
coder ratings of comparison affective state targeted were discrep- studies without a neutral comparison condition, given that changes
ant. in affective state postinduction cannot be isolated to an induction
of a specific affective state (i.e., changes could be either because
the experimental induction increased the target affective state, or
Study Outcomes and Effect Sizes
because the comparison condition reduced that affective state). As
Effect sizes were calculated based on the primary manipulation such, our primary analyses focus on studies that include a neutral
check item or scale (target affective state). Effect sizes were also comparison condition (Neutral Comparison Sample). However,
calculated for any secondary affective states assessed. A conser- consistent with the approach of Lench and colleagues (2011),
vative approach to quantifying contamination to other nontarget because the number of experiments inducing affective states via
affective states would take into account the nontarget affective the Internet was limited, we also report supplementary analyses
state most affected by the induction. Therefore, we calculated one that capitalize on the full sample of studies (Full Sample).
secondary outcome using the highest effect size for nontarget From the 26 studies included in the meta-analysis, 89 indepen-
affective state (highest nontarget affective state), as well as sec- dent comparisons (effect sizes, or k) of the main dependent vari-
ondary outcomes for each specific nontarget affective state. able (target affective state) were calculated. Of these, 59 effect
Effect sizes (Cohen’s d) were calculated as the mean differences sizes involved a neutral comparison condition (as opposed to
between the experimental and comparison condition divided by the another affective state comparison).
pooled SD. If means and SDs were not included in the publication,
other statistical information (e.g., t- or F values) was used to
Analysis Strategy
calculate d (Johnson & Eagly, 2000; Lipsey & Wilson, 2001). A
positive effect size indicated an increase in the target affect com- The Q statistic was calculated to examine whether effect sizes
pared with the comparison condition (in the one study that targeted were homogenous (i.e., whether all studies produced a statistically
a decrease in emotion, a positive effect size indicated a decrease of equitable effect size for the comparison of experimental condition
the target emotion). To correct for sample size bias, effect sizes to comparison condition on the target affective state manipulation
were weighted by the inverse of the variance of d (Hedges, 1981). check). Because heterogeneity in effect sizes was detected, and to
When a neutral condition was included in the design, effect sizes generalize our findings, random-effects procedures were used
were calculated by comparing each experimental condition to the throughout (Hedges & Vevea, 1998).
neutral condition. When more than one neutral condition was All analyses were performed using Comprehensive Meta-
included, the condition that most closely resembled the experimen- analysis (Biostat, 2014). First, we examined main effects of ex-
tal induction was used for comparison (e.g., a neutral video would perimental inductions across studies by examining the overall
be selected to compare with a video induction instead of a no effect size. Separate analyses were conducted on each outcome
treatment control when both were included). (target affective state, highest nontarget affective state, and each
When studies included multiple conditions and no neutral com- affective state for which a manipulation check was included).
parison condition, effect sizes were calculated based on the most These were conducted within the Neutral Comparison Sample and
appropriate comparison. For example, in a study with the follow- the Full Sample. Main effects analyses were also conducted within
ing four conditions: (a) positive affect induced with a Velten each subsample targeting a specific affective state (e.g., anger)
procedure; (b) positive affect induced with a picture induction; (c) when there were at least three studies targeting that affective state.
negative affect induced with a Velten procedure; and (d) negative Similarly, main effects analyses were conducted within each sub-
affect induced with a picture induction; Conditions 1 and 3 were sample targeting a specific affective state for highest nontarget
compared, as were Conditions 2 and 4. When a study included no affective state effect size when there were at least three studies
neutral comparison condition and no clearly delineated compari- targeting that affective state.
sons (as described), we calculated effect sizes for comparisons Because previous meta-analyses of emotion inductions reported
between each combination of experimental conditions. g (Lench et al., 2011) or r (Westermann et al., 1996), we also
To prevent interdependence of effect sizes included in the same calculated these statistics to facilitate direct comparisons between
analysis, when multiple comparisons in the study shared a com- our effects and those previously reported. Specifically, we calcu-
parison condition (e.g., an experiment had an anger, sadness, and lated both g and r for the overall effect size in both the Full Sample
neutral condition, and we compared both anger and sadness to and Neutral Comparison Sample. We also calculated g in main
INTERNET-BASED AFFECT INDUCTION PROCEDURES 5

effects analyses among studies targeting happiness, sadness, anger, .862, 95% (confidence interval) CI ⫽ .696, 1.027, p ⬍ .0001, g ⫽
and fear, consistent with Lench and colleagues (2011), and r for .822, 95% CI ⫽ .601, 1.04, r ⫽ .317; neutral comparison sample:
studies targeting general positive affect and general negative af- d ⫽ .828, 95% CI ⫽ .611, 1.044, p ⬍ .0001, g ⫽ .850, 95% CI ⫽
fect, consistent with Westermann and colleagues (1996). .687, 1.105, r ⫽ .355. These represent uniformly large effect sizes
Then, we examined whether coded factors moderated the mag- in accordance with standard interpretation of Cohen’s (1988) d.
nitude of the effect size, using metaregression techniques. Within Main effects analyses stratified by target affective state (see
each sample of studies, and for each outcome, each coded factor Table 1) indicated that, with the exception of happiness inductions,
was entered into a univariate metaregression (supplemental mate- inductions targeting all specific affective states were effective,
rials Appendix, Tables S2 and S3). Then (again within each producing significant effect sizes that ranged from medium (pos-
sample of studies and for each outcome), those factors that were itive affect, fear) to large (negative affect, disgust, anger, sadness,
significant predictors of effect size magnitude at the univariate or guilt). To facilitate comparisons to previous meta-analyses
level were simultaneously entered into comprehensive metaregres- (Lench et al., 2011; Westermann et al., 1996), Cohen’s g for
sions. Those coded factors that remained significant when entered relevant stratified analyses were: happiness g ⫽ .284, 95% CI ⫽
simultaneously were retained in the final models, while those that .099, .468; sadness g ⫽ .851, 95% CI ⫽ .728, .974; anger g ⫽
became nonsignificant were trimmed from the final models. .917, 95% CI ⫽ .484, 1.35; fear g ⫽ .643, 95% CI ⫽ .285, 1.00;
Metaregressions were undertaken within each subsample targeting r for relevant stratified analyses were: general positive affect r ⫽
a specific affective state when there were at least 10 studies that .207; general negative affect r ⫽ .366.
could contribute to a given analysis. Highest nontarget affective state. The overall effect size for
the highest nontarget affective state was nonsignificant for the full
Results sample: d ⫽ ⫺.016, 95% CI ⫽ ⫺.166, .135, p ⫽ .836, but
significant among the neutral comparison sample: d ⫽ .189, 95%
Overview of Included Studies CI ⫽ .074, .303, p ⫽ .001. These represent uniformly small effect
sizes (Cohen, 1988).
Included studies are denoted with an asterisk in the reference Main effects analyses stratified by target affective state (see
section, and specific details about each study are included in the Table 1) indicate a fairly low degree of contamination, such that
supplementary online materials (supplemental materials Appen- effect sizes for nontarget affective states range from nonsignificant
dix, Table S1). Studies were published or conducted from 2004 (positive affect, happiness, anger, sadness, fear, guilt, or disgust) to
through 2014. Of the 26 studies included in the meta-analysis, 15 negative and significant (negative affect).
(57.7%) were published; the remainder were dissertations or un-
published data. Of the 89 independent effect sizes, 61 (68.5%)
came from published studies. Samples were roughly equally pro- Moderator Analyses
portioned in gender (54.4% female across studies). Mean age Moderator analyses were conducted to examine the heterogene-
across studies was 31.12 (SD ⫽ 8.17). The majority of studies ity across studies. We first conducted univariate metaregressions to
were first-authored by researchers housed in psychology depart- test whether study and population characteristics moderated effect
ments (57.7%). Studies predominantly included samples residing sizes (results available in the online supplementary materials Ta-
in the United States at the time of participation (73.1%). Only 22% bles S2 and S3). Significant moderators identified in the univariate
of studies targeted a student sample; 8.7% included a nationally analyses were retained as predictors in multivariate metaregres-
representative sample. The proportion of independent effect sizes sions. Results of these metaregressions are reported here.
involving a given emotion target were as follows: general positive Target affective state. Full statistical analyses for the final
affect (24.7%); general negative affect (22.5%); anger (18.0%); metaregression models (that included only predictors that were
disgust (4.5%); sadness (9.0%); guilt (7.9%); shame (1.1%); fear significant in univariate metaregressions) are located in Table 2.
(9.0%); hope (1.1%); and happiness (4.5%). There were several notable findings regarding study design and
type of induction. First, across analyses, effect sizes for the target
Analyses of Homogeneity affective state were smaller when it was explicitly specified that
The overall effect sizes for the affective state induced in the participants were randomized to conditions. Second, also across
experimental conditions were heterogeneous for both the full sam- analyses, studies in which a video induction was used had signif-
ple: Q (88) ⫽ 1899.27, p ⬍ .0001; ␶2 ⫽ .55, I2 ⫽ 95.37, and the icantly higher effect sizes. Third, among the full sample, research
neutral comparison sample: Q (58) ⫽ 1738.23, p ⬍ .0001; ␶2 ⫽ in a health domain resulted in lower effect sizes for the target
.67, I2 ⫽ 96.66. The overall effect size for the nontarget affective affective state; however, this effect was not found in the neutral
state postinduction was also heterogeneous for both the full sample comparison sample.
of studies: Q (46) ⫽ 405.97, p ⬍ .0001; ␶2 ⫽ .20, I2 ⫽ 88.67; and There were also several notable findings regarding the type of
the neutral comparison sample: Q (25) ⫽ 78.105, p ⬍ .0001; ␶2 ⫽ affective state targeted. The pattern of analyses suggests that
.05, I2 ⫽ 67.99. These heterogeneous effect sizes support the inducing positive affective states produces lower effect sizes.
strategy of identifying moderators of the effect size magnitude. Specifically, among studies with a neutral comparison, targeting
any positively valenced emotion (including positive affect and
happiness) resulted in lower effect sizes; whereas among the full
Main Effects
sample of studies, targeting general positive affect resulted in
Target affective state. The overall effect size for the target lower effect sizes. Moreover, among the neutral comparison sam-
affective state was significant across analyses: full sample: d ⫽ ple, targeting guilt resulted in significantly higher effect sizes;
6 FERRER, GRENEN, AND TABER

Table 1
Main Effects Stratified by Target Affective State

Studies targeting general positive affect (PA) Studies targeting general negative affect (NA)
k d 95% CI p k d 95% CI p

Target affective state 21 .422 [.256, .589] ⬍.0001 20 .926 [.469, 1.384] ⬍.0001
Highest nontarget affective state 4 ⫺.017 [⫺.254, .220] .866 4 ⫺.411 [⫺.785, ⫺.036] .032
Nontarget NA 4 ⫺.017 [⫺.254, .220] .866 — — — —
Nontarget PA — — — — 4 ⫺.411 [⫺.785, ⫺.036] .032

Studies targeting happiness Studies targeting disgust


Target affective state 3 3.353 [⫺1.005, 7.710] .132 4 1.240 [.420, 2.061] .003
Highest nontarget affective state 3 ⫺2.388 [⫺5.719, .943] .160
Nontarget anger 3 ⫺2.455 [⫺5.621, .710] .128
Nontarget sadness 3 ⫺2.671 [⫺6.259, .918] .145

Studies targeting anger Studies targeting sadness


Target affective state 15 .924 [.487, 1.360] ⬍.0001 8 1.535 [.766, 2.304] ⬍.0001
Highest nontarget affective state 13 ⫺.109 [⫺.361, .143] .396 7 .407 [⫺.103, .918] .118
Nontarget fear 6 ⫺.235 [⫺.549, .079] .143
Nontarget anger — — — — 5 ⫺.146 [⫺.766, .473] .643
Nontarget sadness 7 ⫺.025 [⫺.407, .356] .897 — — — —
Nontarget guilt 4 ⫺.298 [⫺.584, ⫺.012] .041 — — — —
Nontarget happiness 5 ⫺.770 [⫺1.566, .026] .058 3 ⫺1.268 [⫺3.103, .568] .176
Nontarget hope 3 ⫺.075 [⫺.245, .095] .386 — — — —

Studies targeting fear Studies targeting guilt


Target affective state 8 .647 [.287, 1.007] ⬍.0001 8 1.017 [.070, 1.965] .035
Highest nontarget affective state 7 .063 [⫺.264, .390] .705 5 .162 [⫺.110, .433] .243
Nontarget anger 7 .039 [⫺.281, .359] .810 4 ⫺.181 [⫺.431, .070] .158

however, this effect was not uncovered in the analysis of the full Highest nontarget affective state. Full statistical analyses for
sample of studies (that includes nonneutral comparisons). the final metaregression models are located in Table 4. Across
The magnitude of effect sizes between the neutral comparison analyses, video inductions also increased the magnitude of the
sample and the full sample did not differ for anger inductions (B ⫽ effect size for nontarget affective states. However, beyond this
.063, p ⫽ .894), fear inductions (B ⫽ .389, p ⫽ .390), sadness consistency, findings about significant effect size moderators were
inductions (B ⫽ ⫺.606, p ⫽ .508), or negative affect inductions discrepant between analyses that included the full sample versus
(B ⫽ ⫺.084, p ⫽ .888). As such, metaregressions within studies those that included the neutral comparison sample. One potential
targeting these affective states were undertaken on all studies, reason for this discrepancy is the finding that having a neutral
regardless of the presence of a comparison condition (see Table 3). comparison condition significantly reduced the magnitude of the
Guilt inductions with neutral comparisons had significantly higher effect size, suggesting that there may be inherent differences
effect sizes than those using the full sample (B ⫽ 2.127, p ⫽ .001); between these samples. Among the neutral comparison sample, the
presence of a neutral comparison condition was the only signifi- only other significant moderator of effect size magnitude for
cant moderator in the final metaregression on guilt inductions. nontarget affective state in the final metaregression was targeting
Positive affect inductions with neutral comparisons had signifi- a discrete emotion, which increased the magnitude of effect sizes.
cantly lower effect sizes in univariate analyses (B ⫽ ⫺.660, p ⫽ Among the full sample of studies, discrete emotion induction was
.005); however, in the final regression model, the presence of a not a significant moderator; however, effect sizes in the full sample
neutral comparison was no longer a significant moderator and was were increased when the study was explicitly theory-based and
trimmed from analyses presented in Table 3. decreased for happiness inductions.
Results indicate that, among studies targeting positive affect,
inclusion of a cover story decreased the magnitude of the effect
Publication Bias
size, whereas using picture inductions increased effect size mag-
nitude. Among studies targeting negative affect, discrepant coder Publication bias analyses were undertaken among the full sam-
interpretation of target affective state significantly decreased the ple of studies, first by calculating fail-safe N (Rosenthal, 1979).
magnitude of the effect size. Among studies targeting anger, when The fail-safe N was 8,538, suggesting that even if a great number
randomization procedures were specified the effect size signifi- of additional relevant studies with null results were included, the
cantly decreased. Also among studies targeting anger, examining overall effect size would remain significant.
emotion in the context of valuation and reward was associated with However, because fail-safe N is biased toward overestimating
increased effect size. Finally, among studies targeting both anger the number of null studies necessary to render the overall effect
and sadness, video inductions significantly increased effect size. size nonsignificant (Orwin, 1983), we also generated a funnel plot
INTERNET-BASED AFFECT INDUCTION PROCEDURES 7

Table 2
Target Affective State Effect Size—Final Meta-Regression Models

Studies with neutral comparison All studies


Moderator B 95% CI p B 95% CI p

Participant randomization specified 59 ⫺1.244 [⫺1.643, ⫺.846] ⬍.0001 89 ⫺1.244 [⫺1.615, ⫺.873] ⬍.0001
Positively valenced affective state targeted 59 ⫺.543 [⫺.905, ⫺.182] .003
General positive affect targeted 59 89 ⫺.442 [⫺.767, ⫺.117] .008
Guilt targeted 59 1.001 [.256, 1.746] .008
Video induction 59 1.307 [.703, 1.911] ⬍.0001 89 1.513 [1.046, 1.980] ⬍.0001
Health domain 59 89 ⫺.560 [⫺.985, ⫺.136] .010

of the standard error by the standard mean difference (see Figure whether there are systematic reasons why some studies or induc-
2). The distribution is asymmetrical, suggesting the potential of tions may be able to influence happiness via the Internet, whereas
publication bias. However, trim-and-fill calculations (Duval & others are not. Future research is necessary to test whether happi-
Tweedie, 2000) indicate that removing the 24 effect sizes contrib- ness can be effectively induced via the Internet.
uting to the asymmetry does not render the effect size nonsignif- Similarly, it has been hypothesized that positive affect may be
icant (d ⫽ 0.348, 95% CI ⫽ .152, 545). particularly difficult in Internet inductions, as individuals may be
in better moods at home when completing a study on their own
Discussion timeline, than when completing a study in a laboratory (Göritz et
al., 2007). Consistent with this hypothesis, positively valenced
The present meta-analysis suggests that emotion can be effec-
affective state inductions had lower effect sizes than negatively
tively experimentally induced via the Internet. Overall, manipula-
valenced affective state inductions. Indeed, the effect sizes for
tions effectively induced the target affective state, and this finding
Internet inductions of positive affect in the present analyses are
was consistent across all main effects analyses. The effect sizes for
lower than those reported in a previous meta-analysis of mood
Internet inductions seem comparable with those calculated in
inductions largely administered in the laboratory (Westermann et
previous meta-analyses that examined inductions in primarily lab-
al., 1996; r ⫽ .41 compared with our r ⫽ .207). The lower effect
oratory settings (Lench et al., 2011 g ⫽ 0.83 compared to our
overall g ⫽ .822 and .850; Westermann et al., 1996 r ⫽ .48 size for positive affect inductions is also consistent with the
compared with our r ⫽ .317 and .355). assertion that positive affective states are less differentiated than
The inductions were effective overall for most affective states. negative affective states (Shiota et al., 2014), and as such, positive
The one exception was the body of studies that attempted to induce affect may be more difficult to precisely induce. Targeting discrete
happiness; among these studies, the overall effect size was positive positive emotions (e.g., pride, gratitude) may produce higher effect
but not significant. It may be that it is more difficult to induce sizes. However, the previously described findings concerning the
happiness via the Internet than other affective states. This stands in difficulty of effectively inducing one specific positive emotion,
contrast with previous meta-analysis of laboratory inductions, happiness, do not support this hypothesis. It may be that inducing
where happiness inductions had effect sizes comparable with those any positive affective state, general or discrete, is more difficult
targeting discrete negative emotions (Lench et al., 2011). How- than inducing negative affect because of relatively low levels of
ever, lower effect sizes for happiness are consistent with hypoth- differentiation among positive affect.
eses regarding negativity bias, which posit that negative affect is The same may also be true for negative affect induction: It may
easier to induce than positive affect (Augustine & Hemenover, be more difficult to induce negative affect via the Internet than in
2009; Ito et al., 1998). It is also possible that the small number of the laboratory (Westermann et al., 1996; r ⫽ .53 compared with
studies targeting happiness (k ⫽ 3) did not yield adequate power to our r ⫽ .366). However, discrete negative emotions induced via
detect an effect size. The overall effect size was influenced by one Internet manipulations yielded similar effect sizes to those induced
study with a negative effect size. Without additional power, it is in a laboratory (Anger: g ⫽ .80 vs. .917; Sadness: g ⫽ .73 vs. .851;
impossible to determine whether this study is an exception, or for Lench et al., 2011 compared with our analyses, respectively).

Table 3
Highest Nontarget Affective State Effect Size (Final Models)

Studies with neutral comparison All studies


Moderator k B 95% CI p k B 95% CI p

Theory-based 47 .323 [.047, .599] .023


Discrete emotion targeted 26 .390 [.154, .626] .001
Happiness targeted 47 ⫺1.074 [⫺1.619, ⫺.529] .001
Video induction 26 .571 [.261, .881] .001 47 .450 [.102, .798] .011
Neutral comparison condition — — — — 47 ⫺.479 [.228, .730] .001
8 FERRER, GRENEN, AND TABER

Table 4
Final Meta-Regressions Stratified by Target Emotion (DV ⫽ Target Emotion)

Studies targeting general positive Studies targeting general negative


affect (PA) affect (NA)
Moderator k B 95% CI p k B 95% CI p

Participant randomization specified — — — 20 ⫺2.871 [⫺4.029, ⫺1.712] ⬍.0001


Discrepant coder interpretation of target affective state — — — ⫺.880 [⫺1.749, ⫺.011] .047
Cover story 21 ⫺.367 [⫺.679, ⫺.054] .021 — — —
Picture induction 21 .455 [.086, .824] .016

Studies targeting anger Studies targeting sadness


Participant randomization specified 15 ⫺.108 [⫺1.600, ⫺.416] .0009 8 — — —
Video induction 1.640 [1.010, 2.269] ⬍.0001 1.614 [.216, 3.013] .024
Valuation domain .842 [.253, 1.431] .005 — — —
Note. No significant moderators of fear effect size magnitude uncovered in analyses.

The exception was fear, which may be more difficult to induce via emotion among some participants, contributing to a lower effect
the Internet (Lench et al., 2011 g ⫽ .1.05 compared with our .643). size. However, because some studies may have used randomiza-
We also examined whether other study, design, and sample tion but failed to specify in the article, it is difficult to interpret this
characteristics were associated with significantly increased (or association. Unlike in previous meta-analyses, the proportion of
decreased) effect sizes. Video inductions were uniformly associ- women (Lench et al., 2011) and presence of cover story techniques
ated with higher effect sizes for the target affective state, particu- (Lench et al., 2011; Westermann et al., 1996) did not influence the
larly for anger and sadness, consistent with Westermann et al. magnitude of the overall effect size. However, the presence of a
(1996). Videos are considered to be a robust means of eliciting cover story did produce reductions in the overall magnitude of
affective states in laboratory research (Kring & Gordon, 1998; effect size for studies targeting positive affect.
Rottenberg, Ray, & Gross, 2007; see Lench et al., 2011). However, More important, among negative affect inductions, studies
contamination was more likely when using video inductions, con- where the coders rated the experimental condition as discrepant
sistent with previous work on emotion elicitation using films, from the authors’ report (e.g., rated a negative affect induction as
which found that films elicit a range of unintended emotions relevant to anger) were associated with lower effect size magni-
(Gross & Levenson, 1995). Thus, even though videos were most tudes. This result makes sense intuitively; when coders rated an
effective at inducing the target affective state, the goals of the induction as less precise than authors classified it to be, it follows
study should be carefully considered when selecting video induc- that it might be less effective in inducing the target affective state.
tions for Internet research on affect, particularly when hypotheses These results underscore the importance of carefully considering
rely on isolating the influence of one particular affective state. the nature and target of the induction when writing and interpreting
Studies had lower effect sizes when the authors specified that an article using an Internet negative affect induction. However,
the study involved randomization. It is possible that more highly this imprecision seemed to be a problem only for negative affect
controlled studies are more adequately balanced in terms of par- inductions in the present meta-analysis; although this variable was
ticipants’ baseline affect, and as such, are less able to induce associated with smaller effect sizes at the univariate level among

Funnel Plot of Standard Error by Std diff in means


0.0

0.2

0.4
Standard Error

0.6

0.8

1.0

-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6

Std diff in means

Figure 2. Funnel plot of full sample.


INTERNET-BASED AFFECT INDUCTION PROCEDURES 9

studies with a neutral comparison condition (see online supple- cate publication bias (Lau, Ioannidis, Terrin, Schmid, & Olkin,
mentary materials), the moderator coefficient was no longer sig- 2006). Finally, in the context of interpreting publication bias, it is
nificant in the final meta-regression models. For negative affect important to note that we are not making the case that all studies
inductions in particular, imprecision may be a problem because that try to induce mood via the Internet are effective. Rather, this
there are more distinctions in negative than positive affective states meta-analysis suggests that it is possible to successfully induce
(Shiota et al., 2014), potentially leading to greater likelihood that mood online.
different negative affect inductions will unintentionally target very This synthesis highlights several gaps in knowledge that should
different discrete emotions. be addressed in future research. No studies reviewed used sublim-
There was some degree of contamination of inductions. That is, inal priming or imagination inductions, only one used a music
the effect sizes for nontarget affective states were not always null induction (in combination with other components), and only one
or negative, which would be expected if an induction influenced study attempted to induce stress. Future research should examine
the target affective state but not other nontarget affective states. the feasibility of a wider variety of affect inductions, consistent
Specifically, the overall effect size for nontarget affective states with the extant laboratory-based literature. Moreover, future re-
among studies with a neutral comparison condition was signifi- search is necessary to identify potent means of inducing affective
cant, although small. Although contamination is a concern, re- states that do not involve contamination; although video inductions
search suggests that contamination may also be an issue in labo- uniformly increased the magnitude of effect sizes for target affec-
ratory studies (Gross & Levenson, 1995). Moreover, the small tive states, they also increased effect sizes for nontarget affective
magnitude of such effect sizes for highest nontarget affective state states. This may be a particular problem when a study hypothesis
suggests this is not a critical concern for future Internet research involves very specific predictions about different affective states.
involving affective inductions. Future research should also address the shortcomings of Internet
However, note that null effect sizes for nontarget affective states inductions for general positive affect, general negative affect, and
should be interpreted with caution, as they depend on the measures sadness; although effect sizes were lower for Internet studies than
authors included in their study. Only 52% of the comparisons for the extant laboratory literature, it is possible that different types
included a manipulation check for a nontarget emotion. Thus, of Internet inductions would be more effective in successfully
these analyses can be a guide, but should be interpreted with inducing these states.
caution. For example, analyses indicate that positive affect induc- This study has important implications for future affective sci-
tions had nonsignificant effect sizes for other nontarget affective ence research. Because of recent methodological and technological
states. However, most positive affect inductions included only advances, Internet research has revolutionized the process of con-
measures of positive and negative affect. As such, we cannot ducting psychological experiments (Berinsky, Huber, & Lenz,
determine whether positive affect inductions would involve con- 2012; Buhrmester, Kwang, & Gosling, 2011; Horton, Rand, &
tamination of discrete emotion states such as pride or hope. Zeckhauser, 2011; Paolacci & Chandler, 2014; Skitka & Sargis,
The present meta-analysis has a number of limitations in addi- 2006). However, skepticism remains concerning researchers’ abil-
tion to the caution over interpreting effect sizes for nontarget ity to achieve control and precision in Internet experiments (see
affective states described above. A chief limitation concerns the Kraut et al., 2004; Skitka & Sargis, 2006). The present meta-
self-reported nature of manipulation checks. Although self-report analysis suggests that affective scientists can confidently acceler-
remains the predominant method for establishing the effectiveness ate their programs of research by leveraging the Internet to take
of an affect induction across the extant literature (Lench et al., advantage of important benefits such as increased sample repre-
2011), research suggests that individuals may have some difficulty sentativeness (Berinsky et al., 2012; Paolacci et al., 2010; Ross et
identifying and labeling affective states (e.g., Buck, 1990; Buck, al., 2010), cost and resource conservation (Kraut et al., 2004),
Miller, & Caul, 1974), and often self-reported affective states do reduced measurement error and social desirability bias (Skitka &
not correspond with other indicators of affect, such as physiolog- Sargis, 2005, 2006), and increased speed in data collection
ical arousal or facial expressions (Barrett, 2006; e.g., Cacioppo et (Buhrmester, Kwang, & Gosling, 2011).
al., 2000; Lindquist et al., 2012). This limitation is applicable to
the present meta-analysis, but also more broadly to the studies of References
Internet-delivered emotion inductions that contribute to the anal-

ysis (as well as studies involving emotion inductions in laborato- References marked with an asterisk indicate studies included in the
ries). meta-analysis.
Another limitation concerns publication bias; although trim- Augustine, A. A., & Hemenover, S. H. (2009). On the relative effective-
and-fill analyses indicate that removing the 24 effect sizes con- ness of affect regulation strategies: A meta-analysis. Cognition and
tributing to the asymmetrical funnel plot does not render the Emotion, 23, 1181–1220. http://dx.doi.org/10.1080/0269993080
overall effect size nonsignificant, such trimming does reduce the 2396556

magnitude of the effect size. Thus, conclusions about the magni- Baron, J. (April 3, 2014). Provided via personal email communication.
tude of effect size for affect inductions administered via the Unpublished data.
Barrett, L. F. (2006). Are emotions natural kinds? Perspectives on Psy-
Internet should be interpreted with some caution. However, this
chological Science, 1, 28 –58. http://dx.doi.org/10.1111/j.1745-6916
limitation is offset by the inclusion of a high proportion (40%) of .2006.00003.x
unpublished studies (dissertations and unpublished data), as well Berinsky, A. J., Huber, G. A., & Lenz, G. S. (2012). Evaluating online
as the finding that publication status did not moderate effect size. labor markets for experimental research: Amazon.com’s Mechanical
Moreover, asymmetry in effect sizes by sample size can result Turk. Political Analysis, 20, 351–368. http://dx.doi.org/10.1093/pan/
from inherent differences in the types of studies rather than indi- mpr057
10 FERRER, GRENEN, AND TABER


Biostat. (2014). Comprehensive meta-analysis software. Retrieved from Göritz, A. S., & Moser, K. (2006). Web-based mood induction. Cognition
http://www.meta-analysis.com/index.php and Emotion, 20, 887– 896. http://dx.doi.org/10.1080/026999305
Birnbaum, M. H. (2004). Human research and data collection via the 00405386

internet. Annual Review of Psychology, 55, 803– 832. http://dx.doi.org/ Göritz, A. S. (2007). The induction of mood via the WWW. Motivation
10.1146/annurev.psych.55.090902.141601 and Emotion, 31, 35– 47. http://dx.doi.org/10.1007/s11031-006-9047-4
Borenstein, M., Hedges, L. V., Higgins, J. P. T., & Rothstein, H. R. (2009). Gross, J. J., & Levenson, R. W. (1995). Emotion elicitation using films.
Introduction to meta-analysis. West Sussex, United Kingdom: Wiley. Cognition and Emotion, 9, 87–108. http://dx.doi.org/10.1080/
http://dx.doi.org/10.1002/9780470743386 02699939508408966
ⴱ ⴱ
Boudewyns, V., Turner, M. M., & Paquin, R. S. (2013). Shame-free guilt Hängii, Y. (2004). Stress and emotion recognition: An internet experiment
appeals: Testing the emotional and cognitive effects of shame and guilt using stress induction. Swiss Journal of Psychology, 63, 113–125. http://
appeals. Psychology & Marketing, 30, 811– 825. http://dx.doi.org/ dx.doi.org/10.1024/1421-0185.63.2.113
10.1002/mar.20647 Hedges, L. V. (1981). Distribution theory for Glass’s estimator of effect
Bradley, M. M., & Lang, P. J. (1999). International affective digitized size and related estimators. Journal of Educational Statistics, 6, 107–
sounds (IADS): Stimuli, instruction manual and affective ratings (Tech. 128. http://dx.doi.org/10.2307/1164588
Rep. No. B-2). Gainesville, FL: The Center for Research in Psychophys- Hedges, L. V., & Vevea, J. L. (1998). Fixed- and random-effects models in
iology, University of Florida. meta-analysis. Psychological Methods, 3, 486 –504. http://dx.doi.org/
Buck, R. (1990). Rapport, emotional education, and emotional compe- 10.1037/1082-989X.3.4.486

tence. Psychological Inquiry, 1, 301–302. http://dx.doi.org/10.1207/ Helweg-Larsen, M., & Vernon, D. (June 4, 2014). Provided via personal
s15327965pli0104_4 email communication. Unpublished data.
Buck, R., Miller, R. E., & Caul, W. F. (1974). Sex, personality, and Horton, J. J., & Chilton, L. (2010). The labor economics of paid crowd-
physiological variables in the communication of affect via facial expres- sourcing. Proceedings of the 11th ACM conference on electronic com-
sion. Journal of Personality and Social Psychology, 30, 587–596. http:// merce. http://dx.doi.org/10.1145/1807342.1807376
dx.doi.org/10.1037/h0037041 Horton, J. J., Rand, D. G., & Zeckhauser, R. J. (2011). The online
Buhrmester, M., Kwang, T., & Gosling, S. D. (2011). Amazon’s Mechan- laboratory: Conducting experiments in a real labor market. Experimental
ical Turk: A new source of inexpensive, yet high-quality, data? Per- Economics, 14, 399 – 425. http://dx.doi.org/10.1007/s10683-011-9273-9
spectives on Psychological Science, 6, 3–5. http://dx.doi.org/10.1177/ Ipeirotis, P. G. (2010) Demographics of Mechanical Turk (Tech. Rep. No.
1745691610393980 CeDER-10 – 01). New York, NY: New York University. Retrieved from
Cacioppo, J. T., Berntson, G. G., Larsen, J. T., Poehlmann, K. M., & Ito, http://hdl.handle.net/2451/29585
T. A. (2000). The psychophysiology of emotion. In R. Lewis & J. M. Ito, T. A., Larsen, J. T., Smith, N. K., & Cacioppo, J. T. (1998). Negative
Haviland-Jones (Eds.), The handbook of emotion (2nd ed., pp. 173–191). information weighs more heavily on the brain: The negativity bias in
New York, NY: Guilford Press. evaluative categorizations. Journal of Personality and Social Psychol-
Carter, L. E., McNeil, D. W., Vowles, K. E., Sorrell, J. T., Turk, C. L., ogy, 75, 887–900. http://dx.doi.org/10.1037/0022-3514.75.4.887
Ries, B. J., & Hopko, D. R. (2002). Effects of emotion on pain reports, Johnson, B. T., & Eagly, A. H. (2000). Quantitative synthesis of social
tolerance and physiology. Pain Research & Management, 7, 21–30. psychological research. In H. T. Reis & C. M. Judd (Eds.), Handbook of

Chen, R., & Sakamoto, Y. (2013). Perspective matters: Sharing of crisis research methods in social and personality psychology (pp. 496 –528).
information in social media. Paper presented at the Hawaii International London: Cambridge University Press.

Conference on System Sciences. Koblitz, A. K., Ferrer, R. A., & Klein, W. M. P. (2012). Self-affirmation

Choi, Y. (2013). Jurors’ subjective certainty and standards of proof: The and affect: Effects on implementation intentions. Unpublished data.
role of emotion and severity of carge in subjective probability judgment. Retrieved from: http://www.tessexperiments.org/data/koblitz279.html,
Theses, Dissertations, and Student Research: Department of Psychology, April 9 2015.
Paper 61. http://digitalcommon.unl.edu/psychdiss/61. Krantz, J. H., & Dalal, R. (2000). Validity of Web-based psychological
Cohen, J. (1988). Statistical power analysis for the behavioral sciences. research. In M. H. Birnbaum (Ed.), Psychological experiments on the
Hillsdale, NJ: Erlbaum. Internet (pp. 35– 60). San Diego, CA: Academic Press. http://dx.doi.org/

Cohen, P. M., & Knobe, J. (April 6, 2014). Provided via personal email 10.1016/B978-012099980-4/50003-4
communication. Unpublished data. Kraut, R., Olson, J., Banaji, M., Bruckman, A., Cohen, J., & Couper, M.

Coleman, M. D. (2010). Sunk cost, emotion, and commitment to educa- (2004). Psychological research online: Report of Board of Scientific
tion. Current Psychology, 29, 346 –356. http://dx.doi.org/10.1007/ Affiars’ Advisory Group on the conduct of research on the internet.
s12144-010-9094-6 American Psychologist, 59, 105–117.

Coleman, M. D. (2011). Emotion and the self-serving bias. Current Kring, A. M., & Gordon, A. H. (1998). Sex differences in emotion:
Psychology, 30, 345–354. http://dx.doi.org/10.1007/s12144-011-9121-2 Expression, experience, and physiology. Journal of Personality and

Coleman, M. D. (2013). Emotion and the ultimate attribution error. Social Psychology, 74, 686 –703. http://dx.doi.org/10.1037/0022-3514
Current Psychology, 32, 71– 81. http://dx.doi.org/10.1007/s12144-013- .74.3.686
9164-7 Lau, J., Ioannidis, J. P., Terrin, N., Schmid, C. H., & Olkin I. (2006). The

Denson, T. F. (April 9, 2014). Provided via personal email communica- case of the misleading funnel plot. British Medical Journal, 333, 597–
tion. Unpublished data. 600. http://dx.doi.org/10.1136/bmj.333.7568.597

Duval, S., & Tweedie R.. (2000). Trim and fill: A simple funnel-plot-based Lee, G., Barrowclough, C., & Lobban, F. (2011). The influence of positive
method of testing and adjusting for publication bias in meta-analysis. affect on jumping to conclusions in delusional thinking. Personality and
Biometrics, 56, 455– 463. Individual Differences, 50, 717–722. http://dx.doi.org/10.1016/j.paid
Egermann, H., & McAdams, S. (2013). Empathy and emotional contagion .2010.12.024

as a link between recognized and felt emotions in music listening. Music Lee, S. Y. (2010). Framing moral responsibility: The influence of moral
Perception, 31, 139 –156. http://dx.doi.org/10.1525/mp.2013.31.2.139 emotions on persuasive health messages. Unpublished doctoral disser-

Gadarian, S. K. (2008). Elicit cues and emotion. Unpublished data. tation, University of Wisconsin-Madison.
Retrieved from: http://www.tessexperiments.org/data/gadarian568.html, Lench, H. C., Flores, S. A., & Bench, S. W. (2011). Discrete emotions
April 9 2015. predict changes in cognition, judgment, experience, behavior, and phys-
INTERNET-BASED AFFECT INDUCTION PROCEDURES 11

iology: A meta-analysis of experimental emotion elicitations. Psycho- Ryan, T. J. (2012). What makes us click? Demonstrating incentives for
logical Bulletin, 137, 834 – 855. http://dx.doi.org/10.1037/a0024244 angry discourse with digital-age field experiments. The Journal of Pol-

Lerner, J. S., Gonzalez, R. M., Small, D. A., & Fischhoff, B. (2003). itics, 74, 1138 –1152. http://dx.doi.org/10.1017/S0022381612000540

Effects of fear and anger on perceived risks of terrorism: A national field Scheibehenne, B., & von Helversen, B. (2015). Selecting decision strat-
experiment. Psychological Science, 14, 144 –150. http://dx.doi.org/ egies: The differential role of affect. Cognition & Emotion, 29, 158 –167.
10.1111/1467-9280.01433 Shiota, M. N., Neufeld, S. L., Danvers, A. F., Osborne, E. A., Sng, O., &
Lerner, J. S., & Keltner, D. (2001). Fear, anger, and risk. Journal of Yee, C. I. (2014). Positive emotion differentiation: A functional ap-
Personality and Social Psychology, 81, 146 –159. http://dx.doi.org/ proach. Social and Personality Psychology Compass, 8, 104 –117.

10.1037/0022-3514.81.1.146 Sinclair, M., Ashkanasy, N. M., & Chattopadhyay, P. (2010). Affective

Lerner, J. S., Li, Y., & Weber, E. U. (2013). The financial costs of antecedents of intuitive decision making. Journal of Management &
sadness. Psychological Science, 24, 72–79. Organization, 16, 382–398. http://dx.doi.org/10.5172/jmo.16.3.382
Lindquist, K. A., Wager, T. D., Kober, H., Bliss-Moreau, E., & Barrett, Skitka, L. J., & Sargis, E. G. (2005). Social psychological research and the
L. F. (2012). The brain basis of emotion: A meta-analytic review. internet: The promise and peril of a new methodological frontier. In Y.
Behavioral and Brain Sciences, 35, 121–143. http://dx.doi.org/10.1017/ Amichai-Hamburger (Ed.), The social net: The social psychology of the
S0140525X11000446 Internet (pp. 1–26). Oxford, United Kingdom: Oxford University Press.
Lipsey, M. W., & Wilson, D. B. (2001). Practical meta-analysis. London, Skitka, L. J., & Sargis, E. G. (2006). The internet as psychological
United Kingdom: Sage laboratory. Annual Review of Psychology, 57, 529 –555. http://dx.doi
Mathy, R. M., Kerr, D. L., & Haydin, B. M. (2003). Methodological rigor .org/10.1146/annurev.psych.57.102904.190048
and ethical considerations in internet-mediated research. Psychotherapy: Strack, F., Schwarz, N., & Gschneidinger, E. (1985). Happiness and
Theory, Research, Practice, Training, 40, 77– 85. http://dx.doi.org/ reminiscing: The role of time perspective, affect, and mode of thinking.
10.1037/0033-3204.40.1-2.77 Journal of Personality and Social Psychology, 49, 1460 –1469. http://
Moher, D., Liberati, A., Tetzlaff, J., & Altman, D. G., & the PRISMA dx.doi.org/10.1037/0022-3514.49.6.1460
Group. (2009). Preferred reporting items for systematic reviews and Stroup, D. F., Berlin, J. A., Morton, S. C., Olkin, I., Williamson, G. D.,
meta-analyses: The PRISMA statement. PLoS Medicine, 6, e1000097. Rennie, D., . . . Thacker, S. B. (2000). Meta-analysis of observational
http://dx.doi.org/10.1371/journal.pmed.1000097 studies in epidemiology: A proposal for reporting. Meta-analysis Of

Morgan, M. (2013). Playing against the odds: The role of mood, experi- Observational Studies in Epidemiology (MOOSE) group. Journal of the
ence and risk in decision-making. (Doctoral dissertation). Retrieved American Medical Association, 283, 2008 –2012. http://dx.doi.org/
from http://hdl.handle.net/10292/23806 10.1001/jama.283.15.2008
Orwin, R. G. (1983). A fail-safe N for effect size in meta-analysis. Journal Suri, S., Mason, W. A., & Goldstein, D. G. (2014). The 2nd workshop on
of Educational Statistics, 8, 157–159. http://dx.doi.org/10.2307/1164923 crowdsourcing and online behavioral experiments. Workshop con-
Paolacci, G., & Chandler, J. (2014). Inside the Turk: Understanding Me- ducted at the American Conference on Economics and Computation,
chanical Turk as a participant tool. Current Directions in Psychological Palo Alto. Survey Monkey. Retrieved from http://www.surveymonkey
Science, 23, 184 –188. http://dx.doi.org/10.1177/0963721414531598 .com

Paolacci, G., Chandler, J., & Ipeirotis, P. G. (2010). Running experiments Tsai, M.-H., & Young, M. J. (2010). Anger, fear, and escalation of
on Amazon Mechanical Turk. Judgment and Decision Making, 5, 411– commitment. Cognition & Emotion, 24, 962–973. http://dx.doi.org/
419. psiTurk. Retrieved from http://gureckislab.org/blog/?p⫽4318 10.1080/02699930903050631
Reips, U.-D. (2000). The Web experiment method: Advantages, disadvan- Velten, E., Jr. (1968). A laboratory task for induction of mood states.
tages, and solutions. In M. H. Birnbaum (Ed.), Psychological experi- Behaviour Research and Therapy, 6, 473– 482. http://dx.doi.org/
ments on the Internet (pp. 89 –114). San Diego, CA: Academic Press. 10.1016/0005-7967(68)90028-4

http://dx.doi.org/10.1016/B978-012099980-4/50005-8 Verheyen, C., & Goritz, A. S. (2009). Plain texts as an online mood-
Rosenthal, R. (1979). The “File Drawer Problem” and tolerance for null induction procedure. Social Psychology, 40, 6 –15. http://dx.doi.org/
results. Psychological Bulletin, 86, 638 – 641. http://dx.doi.org/10.1037/ 10.1027/1864-9335.40.1.6
0033-2909.86.3.638 Westermann, R., Spies, K., Stahl, G., & Hesse, F. W. (1996). Relative
Ross, J., Irani, L., Silberman, M. S., Zaldivar, A., & Tomlinson, B. (2010). effectiveness and validity of modd induction procedures: A meta-
Who are the crowdworkers? Shifting demographics in mechanical turk. analysis. European Journal of Social Psychology, 26, 557–580. http://
In CHI ’10 Extended Abstracts on Human Factors in Computing Sys- dx.doi.org/10.1002/(SICI)1099-0992(199607)26:4⬍557::AID-
tems. (pp. 2863–2872). Paper Presented at ACM Special Interest Group EJSP769⬎3.0.CO;2-4
on Computer-Human Interaction, Atlanta, GA, 10 –15 April New York,
NY: Association for Computing Machinery.
Rottenber, J., Ray, R. D., & Gross, J. J. (2007). Emotion elicitation using Received July 9, 2014
films. In J. A. Coan & J. J. B. Allen (Eds.), The handbook of emotion Revision received October 7, 2014
elicitation and assessment (pp. 9 –28). London: Oxford University Press. Accepted October 8, 2014 䡲

View publication stats

You might also like