Introduction

• After observing a surge in social media attention after the recent Paris attacks, 13th of November 2015, we wondered which factors determine web-based media coverage and indiviudal response of such incidents - and what determines the pervasiveness of interest from media and user alike

• We started out trying to explore the relationship between the media response of real life events and characteristics herof

• Specifally, our group set out to try and determine factors influencing the decline in media attention - we chose to use the term “decay rate”

Key findings

• Observations are very unevenly distributed
• The presented material in project text is not able to answer the posed question
• We found that we were not able to identify a significant causal relationship between categories/types of attack and decay rate
  • Supposedly due to the uneven distribution

• In utilizing Facebook data, we constructed a crawler of western media sites index; this is an inherently biased solution
• When identifying events in Facebook data, we encountered a trade-off between precision and sample size
• Technical complications in setting up 250 seperate string searches in a dataframe of 600.000 observations in Facebook data

Key findings

• Google Trends data show a high degree of similarity with data gathered using the Facebook API

Improvements

• We could include more years; 2014, 2013 etc.
• There are no apparent limitations to the Facebook API, in theory - though requesting takes time
• Google Trends does not allow for daily data more than -1 year; only weekly or monthly
  • This would reduce precision
• Increasing sample size would also require further web-searches for events
• Coarser search in Facebook data -> visit links and do text analysis on parsed html
  • Not all posts include links
  • Still custom search terms
  • Takes a lot of time
• Bootstrapping is not applicaple in this case

Improvements

• Machine Learning in small samples will have issues
  • Distribution of samples
  • Error rates & precision of predictions
• More extensive inspection of Goolge Trends v. Facebook data
  • Structured testing, correlation
  • More plots for visual inspection

Extended Analysis

• In attempting to answer the posed question we proceeded in applying additional statistical techniques:
  • Linear Modelling; backward selection with AIC (Akaike’s Information Criteria)
  • Machine Learning; decision trees
• We constructed geo-spatial plots in order to illustrate media attention in different parts of the world
• New variables were introduced to the dataframe
  • Mean of Google Trends index
  • Percentage of internet users in population for each country

Packages used

packagesused = c("MASS", "ggplot2", "plyr", "sandwich", "lmtest", "readr", "tidyr", 
                 "lubridate", "dplyr", "stringr", "readr", "rpart", "rpart.plot",
                 "class", "countrycode", "rvest", "maps", "countrycode")
lapply(packagesused, library, character.only =TRUE)

Linear Modelling

## Start:  AIC=812.01
## decay ~ Injuried + Category + Region + Dead
## 
##            Df Sum of Sq   RSS    AIC
## - Region    3   191.101 20052 807.61
## - Category  1     4.592 19865 810.05
## - Dead      1    86.342 19947 810.74
## - Injuried  1   134.117 19995 811.13
## <none>                  19861 812.01
## 
## Step:  AIC=807.61
## decay ~ Injuried + Category + Dead
## 
##            Df Sum of Sq   RSS    AIC
## - Dead      1    59.294 20111 806.10
## - Category  1    61.338 20113 806.12
## - Injuried  1   207.317 20259 807.33
## <none>                  20052 807.61
## 
## Step:  AIC=806.1
## decay ~ Injuried + Category
## 
##            Df Sum of Sq   RSS    AIC
## - Category  1     49.64 20161 804.51
## <none>                  20111 806.10
## - Injuried  1    532.73 20644 808.47
## 
## Step:  AIC=804.51
## decay ~ Injuried
## 
##            Df Sum of Sq   RSS    AIC
## <none>                  20161 804.51
## - Injuried  1    497.64 20658 806.59

• We see that the minimum information criteria is obtained by reducing the model to only including “Injuried”

Linear Modelling

## 
## Call:
## lm(formula = decay ~ Injuried, data = newdf)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -26.021  -7.921   1.916   5.579  39.556 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -6.64533    0.92048  -7.219 1.53e-11 ***
## Injuried    -0.02709    0.00949  -2.855  0.00483 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 11.02 on 175 degrees of freedom
##   (228 observations deleted due to missingness)
## Multiple R-squared:  0.04449,    Adjusted R-squared:  0.03903 
## F-statistic: 8.149 on 1 and 175 DF,  p-value: 0.00483

1. This states, that the only factor contained in our data to explain variation in decay rates is number of injured for a related event.No apparent signs of heteroskedasticity given BP-test.
2. This does have some intuitive elements - however the result is not interesting and we believe that there are interesting results to be found.

Linear Modelling

• Upon testing out different relationships, we found that the variable Increase has interesting and intuitive relationships with other variables contained in our data:
  • Increase = 1-day increase towards global maximum
• Conducting same Backward selection, the following variables minimize the AIC:
  • Region & Category

## 
## Call:
## lm(formula = increase ~ Region + Category, data = newdf)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -49.312 -15.816  -5.222  14.590  68.778 
## 
## Coefficients:
##                           Estimate Std. Error t value Pr(>|t|)  
## (Intercept)                 15.026     17.458   0.861   0.3904  
## RegionAsia                  -3.188      3.378  -0.944   0.3463  
## RegionEurope                12.902      6.239   2.068   0.0398 *
## RegionNorth America         14.590     16.280   0.896   0.3711  
## RegionSouth America          7.812     25.131   0.311   0.7562  
## CategoryNatural disasters   11.162     20.157   0.554   0.5803  
## CategoryTerrorism            6.385     17.267   0.370   0.7119  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 22.73 on 216 degrees of freedom
##   (182 observations deleted due to missingness)
## Multiple R-squared:  0.04477,    Adjusted R-squared:  0.01823 
## F-statistic: 1.687 on 6 and 216 DF,  p-value: 0.1253

Linear Modelling

• Given the descriptive statistics stated above, one might argue that the underlying relationship(s) are not properly explained by a linear model, given that the backward selection suggests two variables of interest
• Furthermore we see, that Regional factors may influence index increases

Machine Learning

• Due to the fact that we do not have the tools within the linear approach to specify a model
  • It is difficult to grasp underlying mechanisms
• It may prove useful to specify a model based directly on the data instead
• Desicion trees may help us understand the factors determining the increase variable

Machine Learning

• In the pruned decision tree, we observe that for all events in which number of deaths exceed 5,5:
  • European countries get significantly higher increase in searches
    • Except in the specific case where deaths are between 5,5-42 AND injuries between 14-20: here the increase is only slight larger in Europe
• We also see that for events with deaths below 5,5 all contributions to “increase” are minor in comparison
• Another peculiarity; for events where between 42-142 died the contribution is only a 6,1 increase
• A general characteristic for Africa, Asia and USA: in events with less than 42 deaths and 20 injuries, larger contributions to increase are observed than otherwise.

• Due to the fact that we observe very different patterns in the ML-approach compared to the LM, likely indicates the prescence of non-linearities in the acutal relationship
• We are not able to properly model the relationship due to a lack of theories within the field of economics to this particular approach
• The observed non-linearities might be improved upon by introduced otherwise omitted variables
• A variety of factors not included could possibly play a role in determining any increase in search-term frequency and media attention

Omitted factors

• We observed, in general, that a surge in searches after any event is less if the country in which the event happened were Asia or Africa
• One explanation could be e.g. differences in internet accessibility of citizens

EXTRA

EXTRA

EXTRA