Psychological Targeting

What it is

That meta-analysis covers 41 studies across marketing, psychology and computer science, and splits the question into its two halves. On prediction, about five per cent of the variance in personality can be recovered from digital footprints once studies with data leakage are excluded. On targeting, only a third of eligible studies used a design capable of an unbiased estimate, and the difference is stark: three-quarters of the estimates from biased designs reached significance against under a fifth of those from the trait-by-treatment interaction design, whose pooled interval includes zero. The authors' reading is that previously reported targeting effects may have been capturing main effects of message design or of personality rather than any targeting benefit. Attenuating the targeting effect by realistic prediction accuracy gives an end-to-end effect close to zero, and their first practical recommendation is to prioritise advertisement quality over microtargeting.

In effect

The field's flagship result points the other way and is worth reading in full rather than in headline. Matz, Michal Kosinski, Gideon Nave and David Stillwell reached over 3.5 million individuals in three field experiments and reported up to 40 per cent more clicks and up to 50 per cent more purchases, which are the most favourable cells rather than the pooled estimates. Their own table shows the first study producing 390 purchases from over three million impressions and a flat null on clicks, the install effect in the second study driven only by the low-openness audience, and the third study comparing standard copy against one tailored message rather than matching anything. Two published critiques then attacked different things: Dean Eckles and colleagues showed the campaigns were not randomised, that platform optimisation produces imbalance in who sees what, and that biases of 25, 9 and 5 per cent would be enough to account for the three conversion results, while Byron Sharp and colleagues objected that two of the three studies compared targeted against deliberately mistargeted advertisements and that creative quality was never controlled. The authors replied to both in print.

What it does not say

It does not say psychological targeting does not work. The meta-analytic estimate under the identifying design cannot be distinguished from zero, which is a statement about what the evidence can resolve rather than a proof of absence, and the authors note that small positive effects could in principle accumulate at scale. It does not say Matz and colleagues were wrong: two sets of critics found two different problems, the authors answered both, and the meta-analysis excluded the largest study's arm on a stated statistical ground about converting rare-event odds rather than on any judgement about the work, which means the best-known study in the field sits outside the headline estimate. The honest sentence is that the flagship result is contested on design, on comparison and on analysis, and that nine years on, after the authors themselves invited replication in print, no direct replication could be located. The best positive result in reach, from Brahim Zarouali and colleagues, is political rather than commercial, measures persuasion on scales rather than money spent, and uses the congruent-against-incongruent comparison the critics object to.

---