What Makes an Image Worth a Thousand Words? 
A Content Analysis of #guncontrol-related Image Characteristics That 
Predict Sharing Behavior 
• Dr. Mike Egnoto, Visiting Assistant Professor, Media Arts, Sciences and Studies, Ithaca College 
• Weiai (Wayne) Xu, PhD Candidate, Department of Communication, SUNY-Buffalo 
• Dr. Gregory D. Saxton, Associate Professor, Department of Communication, SUNY-Buffalo 
• Dr. Michael A. Stefanone, Associate Professor, Department of Communication, SUNY-Buffalo
Not all images are created equal.
Why Study Images and Virality? 
• Network 
• Content 
• Source 
Textual characteristics 
Visual characteristics 
What Image Characteristics Predict Sharing Behavior ?
A Typology of Image Characteristics 
Appeal 
fear 
One/two 
sided 
sex 
metaphor 
threat 
emotional 
rational 
ethos 
humor 
Frame 
Valence 
Attribute 
Goal 
other 
Intensity 
No 
Low 
Medium 
High 
Intended 
valence 
No 
Negative 
Positive 
Human 
presence 
No 
Yes
Examples of Appeals 
Fear One/two sided Emotion (other than fear)
Examples of Appeals 
Metaphor Sex Rational
Examples of Appeals 
Ethos 
Humor 
Threat
Examples of Frames 
Goal frame 
Attribute frame 
Risky choice frame
Research questions 
RQ1: What proportion of image-based appeals are emotional, rational, or mixed? 
RQ2: Which image-based appeals are most effective in terms of message propagation? 
RQ3: What is the proportion of risk, attribute, and goal framing in these images? 
RQ4: Which frames are most effective in terms of message propagation? 
RQ5: What is the proportion of positive, neutral, and negative emotional valence in these 
images? 
RQ6: Which emotional valences are most effective in terms of message propagation? 
RQ7: What is the proportion of low, medium, and high emotional intensity images? 
RQ8: Is there an optimum level of emotional intensity regarding the propagation of these 
images?
Data Description 
• Timeframe: October 1st through 15th of 2013 
• Twitter hashtag: #guncontrol 
• 8,306 of which were original tweets 
• 486 tweets contain image 
• 138 images were selected, which yielded 101 usable images for coding
Results: frequency count 
Appeals Frequency Combined total 
Fear 9 
Emotional 12 
Ethos 2 23 
Threat 0 
Rational 28 
Metaphor 6 
1 / 2 sided argument 4 38 
Humor 23 
Sex 2 25 
Other / no appeal 15 
All frequencies n =101.
Results: frequency count 
Frame Frequency Combined total 
Risk frame 2 
Attribute frame 17 
Goal frame 24 
Other/no frame 58 
Valence Frequency Combined total 
Negative 42 
Positive 23 
Neutral 36 
Intensity Frequency Combined total 
Low 56 
Medium 9 
High 0 
No valence 36 
All frequencies n =101.
# Retweets 
Results: retweet count 
count mean sd min max 
All messages 101 1 2.149 0 18 
Valence Categories 
Negative Valence 42 .929 1.257 0 5 
Neutral Valence 36 .722 1.466 0 7 
Positive Valence 23 1.565 3.764 0 18 
Frame Categories 
No Frame 58 .862 1.1615 0 4 
Valence Frame 2 1 1.414 0 2 
Attribute Frame 17 2.059 4.507 0 18 
Goal Frame 24 .583 1.213 0 5 
Intensity Categories 
none 36 .722 1.466 0 7 
low 56 1.179 2.57 0 18 
medium 9 1 1.5 0 4
Results: Retweet Count
Results: Retweet Count
Results: Retweet Count
Results: Retweet Count
Results: Correlation Matrix 
1 2 3 4 5 6 7 8 9 10 
1. Retweet count 1 
2. Risky Choice Frame 0 1 
3. Attribute Frame 0.22* -0.06 1 
4. Goal Frame -0.11 -0.08 -0.25* 1 
5. Negative Valence -0.03 -0.12 -0.004 0.33*** 1 
6. Positive Valence 0.14 0.26** 0.01 -0.30** -0.46*** 1 
7. Follower count 0.15 -0.07 0.07 -0.08 -0.06 0.08 1 
8. Human Presence 0 -0.14 0.13 -0.05 0.27** -0.12 0.30** 1 
9. Hashtag Count 0.01 0.02 0.09 0.07 0.15 -0.07 -0.03 0.06 1 
10. Mentions Count 0.00 
5 
-0.11 -0.02 -0.10 -0.0004 -0.01 0.27** 0.18 -0.14 1 
Zero-Order Correlation Matrix 
t statistics in parentheses 
* p < 0.05, ** p < 0.01, *** p < 0.001
Results: Negative Binomial Regression 
# Retweets for different types of Frames #Retweets for different types of Valence 
Valence Frame 0.16 
(1.23) 
Attribute Frame 0.77+ 
(0.46) 
Goal Frame -0.48 
(0.47) 
Negative Valence 0.39 
(0.44) 
Positive Valence 0.91+ 
(0.49) 
# followers 0.00 
(0.00) 
0.00+ 
(0.00) 
Human Presence in image -0.18 
(0.39) 
-0.20 
(0.42) 
# hashtags 0.05 
(0.07) 
0.06 
(0.07) 
# user mentions 0.04 
(0.19) 
0.04 
(0.20) 
_cons -0.43 
(0.39) 
-0.83+ 
(0.49) 
N 101 101 
Pseudo R2 0.084 0.067 
Model Significance (2) 8.83 7.00 
Log likelihood -131.69 -132.61
The big picture 
• A theory-guided coding framework for images 
• An exploratory predictive model for image diffusion based on image 
characteristics 
Supported by the grant from Air Force Office of Scientific Research (AFOSR) 
Title: Socio-Cultural Media Sharing as Conversations: Sensing and Modeling 
Behavior in Response to Environmental Changes
THANK YOU! 
Dr. Mike Egnoto, megnoto@ithaca.edu 
Weiai (Wayne) Xu, weiaixu@buffalo.edu 
Supported by the grant from Air Force Office of Scientific Research (AFOSR) 
Title: Socio-Cultural Media Sharing as Conversations: Sensing and Modeling 
Behavior in Response to Environmental Changes

What makes an image worth a thousand words NCA2014

  • 1.
    What Makes anImage Worth a Thousand Words? A Content Analysis of #guncontrol-related Image Characteristics That Predict Sharing Behavior • Dr. Mike Egnoto, Visiting Assistant Professor, Media Arts, Sciences and Studies, Ithaca College • Weiai (Wayne) Xu, PhD Candidate, Department of Communication, SUNY-Buffalo • Dr. Gregory D. Saxton, Associate Professor, Department of Communication, SUNY-Buffalo • Dr. Michael A. Stefanone, Associate Professor, Department of Communication, SUNY-Buffalo
  • 2.
    Not all imagesare created equal.
  • 3.
    Why Study Imagesand Virality? • Network • Content • Source Textual characteristics Visual characteristics What Image Characteristics Predict Sharing Behavior ?
  • 4.
    A Typology ofImage Characteristics Appeal fear One/two sided sex metaphor threat emotional rational ethos humor Frame Valence Attribute Goal other Intensity No Low Medium High Intended valence No Negative Positive Human presence No Yes
  • 5.
    Examples of Appeals Fear One/two sided Emotion (other than fear)
  • 6.
    Examples of Appeals Metaphor Sex Rational
  • 7.
    Examples of Appeals Ethos Humor Threat
  • 8.
    Examples of Frames Goal frame Attribute frame Risky choice frame
  • 9.
    Research questions RQ1:What proportion of image-based appeals are emotional, rational, or mixed? RQ2: Which image-based appeals are most effective in terms of message propagation? RQ3: What is the proportion of risk, attribute, and goal framing in these images? RQ4: Which frames are most effective in terms of message propagation? RQ5: What is the proportion of positive, neutral, and negative emotional valence in these images? RQ6: Which emotional valences are most effective in terms of message propagation? RQ7: What is the proportion of low, medium, and high emotional intensity images? RQ8: Is there an optimum level of emotional intensity regarding the propagation of these images?
  • 10.
    Data Description •Timeframe: October 1st through 15th of 2013 • Twitter hashtag: #guncontrol • 8,306 of which were original tweets • 486 tweets contain image • 138 images were selected, which yielded 101 usable images for coding
  • 11.
    Results: frequency count Appeals Frequency Combined total Fear 9 Emotional 12 Ethos 2 23 Threat 0 Rational 28 Metaphor 6 1 / 2 sided argument 4 38 Humor 23 Sex 2 25 Other / no appeal 15 All frequencies n =101.
  • 12.
    Results: frequency count Frame Frequency Combined total Risk frame 2 Attribute frame 17 Goal frame 24 Other/no frame 58 Valence Frequency Combined total Negative 42 Positive 23 Neutral 36 Intensity Frequency Combined total Low 56 Medium 9 High 0 No valence 36 All frequencies n =101.
  • 13.
    # Retweets Results:retweet count count mean sd min max All messages 101 1 2.149 0 18 Valence Categories Negative Valence 42 .929 1.257 0 5 Neutral Valence 36 .722 1.466 0 7 Positive Valence 23 1.565 3.764 0 18 Frame Categories No Frame 58 .862 1.1615 0 4 Valence Frame 2 1 1.414 0 2 Attribute Frame 17 2.059 4.507 0 18 Goal Frame 24 .583 1.213 0 5 Intensity Categories none 36 .722 1.466 0 7 low 56 1.179 2.57 0 18 medium 9 1 1.5 0 4
  • 14.
  • 15.
  • 16.
  • 17.
  • 18.
    Results: Correlation Matrix 1 2 3 4 5 6 7 8 9 10 1. Retweet count 1 2. Risky Choice Frame 0 1 3. Attribute Frame 0.22* -0.06 1 4. Goal Frame -0.11 -0.08 -0.25* 1 5. Negative Valence -0.03 -0.12 -0.004 0.33*** 1 6. Positive Valence 0.14 0.26** 0.01 -0.30** -0.46*** 1 7. Follower count 0.15 -0.07 0.07 -0.08 -0.06 0.08 1 8. Human Presence 0 -0.14 0.13 -0.05 0.27** -0.12 0.30** 1 9. Hashtag Count 0.01 0.02 0.09 0.07 0.15 -0.07 -0.03 0.06 1 10. Mentions Count 0.00 5 -0.11 -0.02 -0.10 -0.0004 -0.01 0.27** 0.18 -0.14 1 Zero-Order Correlation Matrix t statistics in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001
  • 19.
    Results: Negative BinomialRegression # Retweets for different types of Frames #Retweets for different types of Valence Valence Frame 0.16 (1.23) Attribute Frame 0.77+ (0.46) Goal Frame -0.48 (0.47) Negative Valence 0.39 (0.44) Positive Valence 0.91+ (0.49) # followers 0.00 (0.00) 0.00+ (0.00) Human Presence in image -0.18 (0.39) -0.20 (0.42) # hashtags 0.05 (0.07) 0.06 (0.07) # user mentions 0.04 (0.19) 0.04 (0.20) _cons -0.43 (0.39) -0.83+ (0.49) N 101 101 Pseudo R2 0.084 0.067 Model Significance (2) 8.83 7.00 Log likelihood -131.69 -132.61
  • 20.
    The big picture • A theory-guided coding framework for images • An exploratory predictive model for image diffusion based on image characteristics Supported by the grant from Air Force Office of Scientific Research (AFOSR) Title: Socio-Cultural Media Sharing as Conversations: Sensing and Modeling Behavior in Response to Environmental Changes
  • 21.
    THANK YOU! Dr.Mike Egnoto, megnoto@ithaca.edu Weiai (Wayne) Xu, weiaixu@buffalo.edu Supported by the grant from Air Force Office of Scientific Research (AFOSR) Title: Socio-Cultural Media Sharing as Conversations: Sensing and Modeling Behavior in Response to Environmental Changes