facebook marketing science professional

experiments

Randomized control trials

A/B test

test vs control group

no control group

cannot monitor lift

test different versions of ads

Lift Tests

conversion lift

split testing

brand lift

uses a poll system

measure incremental conversions

which ad has lowest cost per result

multi-cell lift test

limitations

  1. not enough statistical power

Reasons

limited frequency/duration

other overlapping campaigns running

too small control group vs test

how to resolve

keep testing

use results

collect more data

  1. hard to isolate variables

person can be in both test & control group

ex. people located in same household

you can't track all data/unknowns

ex. cash transactions

ex. device switching

test/control groups may have outliers

windsorization - upper/lower bound setting for outliers

Observational method

Definition

done when intervention is not possible

does not interfere with data

limitations

isolating the variable

types

exact match

propensity score

regression adjustment

natural experiments

data availability

click to edit

how to resolve

ave. the diff. observation model results

conduct RTA's regularly and use DDA

collect many data points

how well ad resonated with people

3 ways to design lift test

nest test

single-cell test

multi-cell test

ex. fb ads vs no fb ads

ex. # of frequency exposure

ex. creatives

measurement tools

A/B test

brand lift

conversion lift

marketing mix model

facebook attribution

performing an analysis

Type of attribution model

single-touch attribution model

multi-touch attributionmodel

first click of visit

100% of credit of conversion
gives to the first click

last touch (including last click)

100% credit to the last click,
visit, impression or view

even credit

evenly split across touchpoints

positional

higher credit to first and last click

ex. positional 30. 30% first and last,
the rest is evenly split

time-decay

gives more credit closer to
conversion event

data-driven attribution model