Impact of GKPro Sponsorship on OTB & GKPro Webtraffic

Side Projects Google Trends Public Data SEM Structural Equation Modeling Mediation Bootstrapping Plotting

An example of leveraging free data to benefit a company, with code.

Christopher Loan
2020-12-18

Background

With COVID-19 decreasing the amount of disc golf content in 2020, I found myself really watching a lot of “skins” matches hosted by GK Productions. Here’s a link to one of the last this season.

For those unfamiliar, these matches are more-or-less single-round match-play events. In this format, there is not an overall prize at the end; rather, players compete for cash prizes for taking the best (i.e., lowest) score on each hole. For example, if 3 of 4 players take par on hole 1, and only 1 player gets a birdie, that player gets $50. If 2 (or more) players tie the best score on a given hole, the cash “rolls over” to increase the value of the next hole. Typically, these are filmed around the time (and at the location of) a larger tournament or Disc Golf Pro Tour Event. They feature top pros showing their human side and are a blast to watch.

“Only The Best” Discs aka OTB Discs has sponsored these events all year, and they’ve offered free shipping on their products–on top of providing cash for the competitors–as a sponsorship promotion. I personally found myself fighting the urge to purchase a disc from them any time I watched a skins match (and often failed the urge). For that reason, I wanted to dive into the data and see what impact these had on OTB webtraffic. As a note, I would prefer to follow this up with sales data before being confident in any impact to their bottom line, but that’s not publicly available. We should note that someone may look at OTB’s website without purchasing anything.

show code
knitr::include_graphics(here::here("images", "discs.jpeg"))

Method

Data Sources:

GKPro’s YouTube:

Analytic Plan:

  1. Visually assess average web traffic for OTB on weeks with and without coverage (Figure 1 below).

  2. Plot web traffic over time for googling OTB’s website alongside people searching YouTube for GKPro (Figure 2 below).

  3. Plot the same graph as 2, but adding in trends for searching YouTube and Google for “Disc Golf” generally. This helps us contextualize these rises and falls in OTB webtraffic with seasonal trends. We may expect more googling of OTB during the summer when more people play regardless of the videos (Figure 3 below).

  4. Conduct a formal statistical analysis to see if the effect of a sponsored video significantly increased sales through GKPro videos, controlling for the direct effect of the sponsorship on OTB’s website and seasonal fluxuations in searches of YouTube and Google for “Disc Golf”. This goes one step farther than seeing if there is an association. It’s tough to explain without statistical terms, but this is called “mediation” and it checks if the variation in the link from \(Sponsorship --> GKPro_{webtraffic}\) is the same as the variation in \(GKPro_{webtraffic} --> OTB_{webtraffic}\). In total this provides a formal test of if the entire path exists, i.e.: \(Sponsorship --> GKPro_{webtraffic} --> OTB_{webtraffic}\). This value could be considered a minimum estimate of impact on webtraffic for sponsoring those videos.

To be clear, point #4 will be done controlling for the variation of webtraffic for “Disc Golf” generally in both YouTube and Google as a proxy for general disc golf popularity (although I know that’s an imperfect proxy).

Results & Discussion

Significant Mean Differences (Step 1)

show code

If you’re interested, data can be found on at this GitHub Repo

show code
dat <- 
  import("OTB_and_GKPro.csv") %>%
  clean_names() %>% 
  mutate(week = mdy(week),
         sponsored = 
           factor(sponsored_skins_video, labels = c('No','Yes')), 
         total_otb = 
           otb_discs_united_states + only_the_best_discs_united_states,
         Google_DG = disc_golf_united_states, 
         YouTube_DG = disc_golf_youtube_united_states)
dat_small <- 
  dat %>% 
  select(week, 
         sponsored, 
         total_otb, 
         gk_pro_united_states, 
         Google_DG, 
         YouTube_DG)

plot_dat <- dat_small %>% 
  pivot_longer(cols = total_otb:YouTube_DG, 
               names_to = "term", 
               values_to = "searches") %>% 
  
  mutate(`Search Term` = 
           factor(term, 
                  labels = 
                    c( "GKPro", 
                       "Google DG", 
                       "OTB", 
                       "YouTube DG")))

Figure 1 shows that OTB webtraffic is higher in weeks with a skins match (average webtraffic = 57.38), compared to weeks without a skins mathc (average webtraffic = 20.24). Because the standard error bars do not overlap, we can be confident that these are significantly different amounts of webtraffic

show code
plot_dat %>%
  filter(term == "total_otb") %>% 
  group_by(sponsored) %>% 
  summarize(`Average Searches` = mean(searches), 
            `Standard Error of Searches` = sd(searches)/sqrt(n())) %>% 
  ungroup() %>% 
  ggplot(aes(x = sponsored, y = `Average Searches`)) +
  geom_col(aes(fill = sponsored)) + 
  geom_errorbar(
    aes(ymin =`Average Searches`-1.96*`Standard Error of Searches`, 
        ymax =`Average Searches`+1.96*`Standard Error of Searches`), 
    width=.2,
    position = position_dodge(.9)) +
  labs(title = "Figure 1. Average Web Traffic on Weeks 
with and without GK Pro Skins Matches",
       caption = "Error Bars Show 95% Confidence Intervals",
       y = "OTB Web Traffic*",
       x = "Skins Match Same Week",
       fill = "Skins Match\nSame Week") 

Trend over Time (Steps 2 & 3)

Since we have this nice time-ordered data, I plotted these trends over time. Figure 2 shows Google Searches for OTB Discs alongside YouTube Searches for GK Pro. I highlighted the weeks with sponsored skins matches. The closely coupled peaks and valleys between these lines suggest people may be Googling OTB more often during weeks with a GKPro sponsored skins match; just what we expected!

Something important to note here is that OTB actually sponsors other videos, too, so we’d expect some noise here from other videos that are not skins matches (i.e., other GK Pro videos that aren’t skins matches, videos from their sponsored players, etc.). A better model would include those sources of variation, but I didn’t want to scour the internet for all their sponsored players’ videos.

We also see an early peak in the model that is before any sponsorsed skins match. A guess I have here is simply that this is the start of the professional disc golf season, and that watching professionial disc golf increases sales of discs. These early spikes that occur for both GK Pro and OTB are likely from the Las Vegas Challenge and the Memorial Championship, which both occured in those times.

show code
plot_dat %>% 
  filter(`Search Term` %in% c('OTB', 'GKPro')) %>% 
  mutate(sponsored_continuous = if_else(sponsored == "Yes", 100, 0)) %>% 
  ggplot() + 
  geom_line(
    aes(x = week, 
        y = searches, 
        color = `Search Term`)) +
  geom_col(
    aes(x = week, 
        y = sponsored_continuous), 
    position = "dodge", 
    fill = "hotpink", 
    alpha = 0.15)+
  labs(
  title = "Figure 2. OTB Discs Google Searches & 
GK Pro YouTube Searches Over Time",
caption = "Weeks of Sponsored Videos Highlighted with Vertical Bars",
y = "Web Traffic*", 
x = "Date") +
  scale_color_manual(values = 
                       c("cornflowerblue",
                         "coral3", 
                         "gray30", 
                         "tan"))+
  ylim(0,100)

Figure 3 shows the same webtraffic plot, but with the increases in Google and YouTube searches for “Disc Golf” to account for seasonal changes in trends. The early spikes which weren’t associated with sponsorship of a GK Pro match really seem to line up with seasonal trends. But later in the season, it looks like OTB webtraffic went way up compared to the small bump they got from a general increase from the season’s start.

show code
plot_dat %>%
  ggplot() +
  geom_line(aes(x = week, y = searches, color = `Search Term`)) +
  labs(title = "Figure 3. Webtraffic for OTB & GKPro, 
alongside general `Disc Golf` Searches",
       y = "Web Traffic*",
       x = "Date") +
  scale_color_manual(values = 
                       c("cornflowerblue", 
                         "gray30", 
                         "coral3", 
                         "tan")) +
  ylim(0,100)

I used a package called {lavaan} to do some inferential statistics to make sure these trends were significant. I was worried about having so few data points, so I actually did something called “bootstrapping” where I made a statistical model and then simulated the results 5000 times. The 5000 estimates are then used to give some confidence in your findings. This is a huge oversimplification, but I wanted to spare you the details.

Inferential Statistics Results (Step 4)

With 95% confidence, we can say that GK Pro videos increase webtraffic for OTB, even controlling for seasonal variation in disc golf popularity. Statistically, we can divide up the influence of these sponsored videos on OTB web traffic into:

You can think of the unique variation as “random noise” that we control for, and the “carried through” variability as the effect of these sponsored videos through GK Pro webtraffic to OTB’s webtraffic. There are explanations for the direct path that are not random noise, but if we’re interested in the mediated effect, this is noise to us. If the effect is in fact carried through GK Pro, we’d expect this to not be significant.

However, one possibility I see is if YouTube recommended a GK Pro video, they won’t actually type it on YouTube, so the direct increase on OTB would look to occur without increased searching of “GK Pro”. I’ll ignore this unless this path is significant, but I don’t expect that

Quantifying the Increase: Is it significant?

Impact of number of GK Pro Searches on OTB Searches

The model shows that an increase in GK Pro viewership by 1% of their relative webtraffic corresponds to an increase of about 0.43% (95% confidence range = 0.14-0.70). This change is significant and is not small. If GK pro webtraffic went up by 50%, we would see about 21.5% increase on OTB webtraffic. I don’t know the scale of either of their webtraffic, but this seems like a substantial increase. Visually, this is \(GKPro_{webtraffic} --> OTB_{webtraffic}\).

Impact of OTB Sponsored Skins Match on GK Pro Searches

GK Pro is also (probably) feeling this benefit: if we compare the weeks with sponsored videos to weeks without, their webtraffic is up 28% (95% confidence range = 14.02-42.93) on weeks with sponsored videos compared to weeks without sponsored videos! This is visually \(Sponsorship --> GKPro_{webtraffic}\) . The researcher in me knows that this value is conflated with lots of spurrious associations we cannot assess, however, this estimate is controlling for the popularity of “Disc Golf” searches on both YouTube and Google during those weeks. With GK Pro being sponsored most weeks, though, it seems like their popularity may just rise and fall with release of a video (duh!), but that relies on someone sponsoring it (and it happens to be OTB).

The impact of Skins Match Sponsorship on OTB Searches through GK Pro Searches

This is what we are most interested in; it is the minimum impact that is actually felt by OTB through GK Pro. Only variation which is consistent through both variables is accounted for here. Visually, it looks like:

\[Sponsorship --> GKPro_{webtraffic} --> OTB_{webtraffic}\] This is the “mediated” effect we’re interested in. In essence, OTB is boosting their own webtraffic by 12.20% (95% confidence range = 2.64-26.6%) for every video they sponsor. They can use this to determine how “worth it” these videos are to them.

Total Impact of Sponsorship on OTB webtraffic (i.e., overall association)

It is worth noting that when we control for YouTube and Google searches for “Disc Golf” and the searches for GK Pro, we see no direct effect of sponsorship on OTB searches. This likely means that most people searching OTB are doing so because of GK Pro.

Appendix for Those Interested in Analytic Details:

I included minimal output from the regression below, and you can see it’s significant if the ci.lower and ci.upper have the same term (i.e., - or +), which represent a range of confidence in our findings at 95%. The defined parameter “mediated_effect” is the product of the paths a*b, which is the most common way to formally test mediation in structural equation modeling.

show code
mediation_mod1 <- '
total_otb ~ a*gk_pro_united_states + d*sponsored + Google_DG + YouTube_DG

gk_pro_united_states ~ b*sponsored+ Google_DG + YouTube_DG

mediated_effect := a*b
'

med_fit_boot<- sem(mediation_mod1, 
                   data = dat, 
                   estimator = "ML",
                   se = "bootstrap", 
                   bootstrap = 5000,
                   parallel = c("multicore"))

table <- parameterEstimates(med_fit_boot, 
                   level = .95, 
                   boot.ci.type = "perc", 
                   output = "pretty", 
                   stand = T) 

table %>% 
  select(-c(se:pvalue), 
         -c(std.lv:std.nox)
         ) %>% 
  filter(op!='~~')

Regressions:
                         Estimate ci.lower ci.upper
  total_otb ~                                      
    gk_pr_ntd_ (a)          0.431    0.144    0.697
    sponsored  (d)         14.476   -1.682   29.606
    Google_DG               0.163   -0.161    0.475
    YouTube_DG              0.292   -0.053    0.661
  gk_pro_united_states ~                           
    sponsored  (b)         28.326   13.681   42.869
    Google_DG               0.328   -0.067    0.705
    YouTube_DG             -0.358   -0.753    0.059

Defined Parameters:
                   Estimate ci.lower ci.upper
    mediated_effct   12.196    2.757   25.821