Facebook offers brand-lift tests to help you determine the impact of your ads. These studies compare groups of people who've seen your ads with groups of people who haven't to determine if there's a statistically significant difference in brand resonance between the groups. A statistically significant difference means that the difference in results is large enough that it is likely to be due to the true effect of ad exposure and unlikely to be caused by random chance. If there is a statistically significant difference, then we can use that information to determine the incremental lift in brand awareness that occurred as a result of your ads.

The level or degree of statistical significance is important because it can give you a corresponding confidence level that can then be used to help inform your marketing decisions based on the results of your lift study. The higher the confidence level, the more likely your results are due to the effect of your ads and not random chance. For example, if you specified a 90% confidence level and ran the same lift study ten times, a statistically significant study may show the same results nine out of ten times. (Note: Brand-lift studies have a fixed confidence level.)

For brand-lift studies, we use industry-standard practices to measure statistical significance. We poll the audience that saw your ads in the test group, along with the audience that didn't see your ads in the control group, and then use statistical modelling (using a logistic regression model) to analyse the number of responses and difference between the two groups.

Brand lift with a single test group

Brand-lift studies start by using a randomisation framework to divide your ad's target audience into a test group and control group. The test group contains people who will possibly be exposed to your ads, and the control group contains people who won't be exposed to ads. If these individuals were not randomly placed into the control group, they would have been eligible to see the ads. The people within the test group that actually see your ads are referred to as the "exposed group".

When there is one test group, we poll the exposed group and control group, and the difference in responses between the two groups is analysed. Facebook uses the following logic to determine whether there was a statistically significant difference between the two groups:

1. Facebook uses a logistic regression model, a type of analysis for studies with binary results, to analyse all the data. The covariates – variables that could be used to predict the outcome of the test – are age, gender and whether or not they were exposed to the ad (the treatment condition). The treatment condition is set to:
   • One for users in the exposed group
   • Zero for users in the control group
2. Facebook will estimate coefficients for each covariate. The variable of whether or not an individual was exposed to your ad is assigned a coefficient, which is used to determine the effect of the ad itself. There is a probability value associated with the differences observed between exposed and unexposed users, which is used to assess whether the result is statistically significant.

Brand lift with multiple test groups

When a brand-lift study has more than one test group, the logic above is not suitable as it only determines statistical significance between one test group and one control group.

When an advertiser wants to run a brand-lift study with more than one test group, Facebook uses the following logic to help determine the winning strategy:

1. Facebook orders the test groups by point estimates of their cost per brand lift and picks the statistically significant group with the lowest positive amount as the candidate winner.
2. We find the next-best performing group other than the candidate winner using the brand-lift point estimates. In most cases, this will be the group with the next-lowest positive brand lift. Note: In rare cases, it will be the group with the least negative brand lift, and in this scenario, we will not show results for the second-best performing group.
3. Facebook runs a simulation to compare the cost per brand lift for these two groups.
4. If the confidence of the candidate winner is above 60%, Facebook deems it is the winning cell. If the confidence level is not above 60%, we can't confidently declare a winner. (We use 60% because it is a trade-off between the number of studies where we can declare a winner and how precise we want to be in our winner declaration.)

For more information on brand lift, please contact your Facebook account representative.

* Nguồn: Facebook