Professional Documents
Culture Documents
The Effects of Creative Formats, Message Content and Targeting On Engagement
The Effects of Creative Formats, Message Content and Targeting On Engagement
R3
E. Optimal Reallocations
1
Web Appendix A -- Assessing Instrumental Variables
We conducted analyses to assess the strength of our instrumental variables. First, weakness would imply
that advertising is poorly correlated with these instruments. Thus we obtained the population R-squares
from the regression of log(1+ad impression) against them. The results are relatively high within the four
targets, ranging from 0.63-.73 (below Table Q1). Weakness would also imply that variations attributable
to the instruments would be small relative to variations in ad impressions; in this case, the effects of ads
() is expected to be highly correlated with the random error in the instrumental variable equation, and the
covariance of that error with the error in advertising impression (goodwill) equation (See Rossi and
Allenby, 2005). We calculated the correlation between and the ratio of this variance to covariance using
the posterior draws, and reported these in Table Q1. Note they are relatively small and several are not
significant, providing further evidence that our instruments can help identify ad effects.
Table Q1: IV Assessment: Correlation of draws with the ratio, goodwill error variance/ covariance of instrument
*Bold p-value less than 0.05. ** Population R-Squares: Retargeting (0.7281), Male (0.6961), Female (0.6313), Age (0.7067)
Reference:
Rossi, Peter, Greg Allenby, and Rob McCullogh (2005), Bayesian Statistics and Marketing.
Hoboken, N.J.: Wiley Series in Probability and Statistics
2
Web Appendix B -- Deriving an Approximate Optimal Importance Function
Recall that our goal is to approximate the most efficient proposal distribution for the Particle Filter, the
“optimal” importance function which has the form p( g t g t 1 , yt ) . The log of this distribution is
approximately:
B1) l ( g t ) ln p( yt g t ) p( g t g t 1 )
Given that our model is within the exponential family, l ( g t ) is twice differentiable; thus we take a second
order Taylor approximation of l ( g t ) at some point g :
1
l ( g t ) l ( g ) l ' ( g )( g t g ) ( g t g ) t l" ( g )( g t g )
B2) 2
1
l ' ( g )( g t g )
( g t g ) t l" ( g )( g t g )
2
B3) 1
C ( g t g m( g )) t 1 ( g )( g t g m( g )) , C a constant.
2
If we examine the last expression (B3), we can readily see it is a normal with a mean and variance; it thus
suggests a normal approximation for the optimal importance function:
B4) p( g t gt 1 , yt ) N m( g ) g, ( g )
~
B5)
p( gt gt 1, yt ) N gt* ,l" ( gt* )1
~
.
The mode is simulated using the Newton-Raphson algorithm.
References:
Doucet, Arnaud, Simon Godsill , Christophe Andrieu (2000), “On Sequential Monte Carlo
Sampling Methods for Bayesian Filtering,” Statistics and Computing, 10, 197-208.
Doucet, Arnaud, Nando de Freitas, and Neil Gordon, Editors (2001), Sequential Monte Carlo
Methods in Practice, New York: Springer.
3
Web Appendix C – Analysis of Data Spike
The spike potentially could overstate advertising effects, especially if ad response differs markedly during
the spike period. Thus, as a preliminary step we re-examined click-through rates at and outside the spike
in our sample. We found that click through rates at the spike was approximately 0.0550%, while average
rates after and before the spike were 0.0542% and 0.0547%, respectively; thus advertising effects might
not be substantively overstated due to the spike.
For a more formal analyses, however, we included seasonal dummies for the spike period and re-
estimated the proposed model. We note in Table R1 that the seasonal dummies are all insignificant, and
moreover our estimates remain largely unchanged, providing further evidence the spike in this sample is
unlikely to be a problem.
4
TABLE R1: ESTIMATES FROM PROPOSED MODEL: CONTROLLING FOR THE DATA SPIKE
5
Web Appendix D: Findings in Alternative GLMs and Linear Models
TABLE S1: LINEAR STATE SPACE MODEL NDLM – NO MEASUREMENT MODEL, NO INFLATION
6
TABLE S2: LOG- LINEAR STATE SPACE MODEL NDLM – NO MEASUREMENT
MODEL, NO INFLATION
7
TABLE S3: DYNAMIC POISSON – NO MEASUREMENT MODEL, NO ZERO INFLATION
8
TABLE S4: STATIC POISSON MODEL - ZERO INFLATED, NO MEASUREMENT MODEL
9
TABLE S5: DYNAMIC NEGATIVE BINOMIAL – NO MEASUREMENT MODEL, NO ZERO
INFLATION
10
Web Appendix E -- Optimal Reallocations
We optimally re-allocated impressions for one, two and four successive periods, given the information
available at t-1. Table T1 reports the % change in impressions in each case.
Table T1: ALLOCATIONS -- Actual Impressions (1.0e+8) and Model Based Allocation Changes (%)
Actual
Digital Format Exposure Impressions %(t-1|t)(1) %(t-1|t))(2) % (t-1|t)(4)
Targeted Segments
Message Content
Product 7.8150 -18.6164 -18.5501 -22.0608
Price 11.3920 12.7806 12.7266 15.1603
11
As a robustness check, we re-solved the one period (t-1|t)(1) problem over 1000 random draws from the
posterior, using a shorter period (T=80) given the computational complexity of solving P1. Table T2
reports the actual impressions and the mean (%) changes in impressions. Figure U1 shows the
distribution of the resulting % gains in total clicks.
Table T2: ALLOCATIONS: 1000 Random Posterior Draws: Model Re-Allocation Changes (%)
Actual T=80
Digital Format Exposure Impressions %(t-1|t)(1)
Targeted Segments
Message Content
Product 3.6553 -11.8092
Price 7.5417 5.7237
12
Figure U1: Distribution of the %changes in clicks
Reallocation: %Clicks - 1000 Draws - T=80; (t-1|t)(1)
140
120
100
80
60
40
20
0
-20 0 20 40 60 80 100 120 140 160 180
%Change Clicks
13