Thursday, April 6, 2017

The impact of "Did Not Play - Rest" on the NBA

TLDR: "... [G]enerally, NBA games in which a player does not play due to rest result in a 5% decrease in viewership."

"However, considering the size and confidence of the effect that resting players have on the ratings of NBA games, it may be a little early to introduce new rules to address for this potentially overstated issue."

The issue of National Basketball Association (NBA) teams choosing to have individual players sit out regular season games for "rest" is highly contested. I do not want to waste time reiterating much of the same arguments and opinions of others. Instead I want to offer some insight into the magnitude of the effect on demand when NBA players rest and suggest that the issue of resting may be overstated,

If having players rest were to have an effect on the demand for basketball, I would argue it would be most evident in the broadcast ratings. Recall that attendance is traditionally the number of tickets sold: if the announcement of players resting is made in short notice of a game, fans who have already purchased tickets cannot elicit their response though attendance (or lack thereof). Thus, in order to test the impacts that players resting has on demand, I start with building a data set on NBA national broadcasts from three sources:
  1. Television ratings from Show Buzz Daily;
  2. Games which listed a player officially marked in the boxscore as "Did Not Dress" or "Not With Team" from; and,
  3. FiveThirtyEight's CARM-ELO to give a relative value of skill of the teams in each match.
My final product becomes the ratings, number of viewers, and the teams' skill level of each nationally broadcast NBA game in the 2016-17 season (up to April 3rd). I identify 208 games with 68 occurrences of at least one player resting. I then use a model which I have described in a previous blog post for estimating NFL national broadcast viewership.

(1) viewershipi,j,t= f(DNP, Eloi + Eloj, Xt)

where Eloi + Eloj is the sum of the Elo scores of each team and Xt can be thought of as a set of game-specific variables including the day of week of the broadcast or what network the game aired on. Note that I have separately considered the viewership to be either the broadcast's ratings or the number of viewers (ratings is the percentage of all televisions - on or off - that are watching the broadcast, viewers in this case are the number of adults aged 18 to 49 watching the broadcast).

Finally, our variable of interest, DNP, is a dummy variable which takes a value of one if a player had sat out due to rest. I consider a player to be resting as follows:
  • Player must not be designated as inactive (i.e. not injured).
  • Player is either
    • listed as "Not With Team" and is a member of the away team.
    • listed as "Did Not Dress" (can be member of either team).
Through experimentation of model 1, I have found that there is some evidence that suggests the broadcast ratings do in fact suffer when players are rested. The robustness of such a model is certainly up for debate however I have tried several specifications and found generally, NBA games in which a player does not play due to rest result in a 5% decrease in viewership. Some of these results are significant up to a p-value cutoff of 0.1 to 0.5.

Below is a graphical depiction of one possible version of the model. Here I do not control for resting players and plot the predicted values against the actual. I have coloured the games with a resting player with orange (my attempt at little basketballs). Points that lie above the black 45-degree line suggest games in which we would have predicted more viewers than actually occurred (conversely, points below the black line had more viewers than expected).

Note the frequency of which these orange dots appear in relation to the black line. It appears that the majority reside above, but there are still many instances of observations on or below the 45-degree line which calls into question the causality of this correlation. This pattern exists when utilizing different variables for the regression, including team fixed effects, network fixed effects, day-of-week fixed effects, etc.

For a future study, a meaningful variation of this analysis would be to consider the effects of an injury holding a player out of the lineup in comparison to the team's decision to rest a player. Additionally, I have not considered the quality of the players who are sitting out. As Golden State Warriors player Kevin Durant suggested, "[fans] don't care if the 13th man on the bench rests ... it's only for like five players." (Although some of the players Durant went on to list do not often rest, if at all.) This analysis could use a measure of the resting player's productivity to test Durant's conjecture that the demand further declines when high-quality players rest.

I have identified some evidence that resting players may have a negative impact on the demand for basketball of a 5% decline from the predicted ratings for television broadcast of games where a team rests a player. These results have not yet been robustly tested and the estimated effect has quite a bit of variation, ranging from -10% to 0% impact on television ratings. Yet, the public reaction from the NBA suggests that they perceive any decline in demand to be detrimental to the game. The commissioner of the NBA, Adam Silver, has communicated with the 30 teams of the league that resting is "an extremely significant issue for our league." However, considering the size and confidence of that the effect of resting players has on the ratings of NBA games, it may be a little early to introduce new rules to address for this potentially overstated issue.

1 comment:

  1. this post is actually really good and uses logical statistics