Multi-channel campaigns with attribution models
Models of attribution and causal inference [1/5]
These blogs will be built by collaborators of Grupodot, to give continuity to the series of articles.
This article is the first of a serie built by collaborators of Grupodot, presenting the basic concepts, such as methodologies and tools available for the implementation of attribution models and causal inference, aimed at understanding and promoting specific conversions (actions) of your customers, especially in the advertising interaction.
From marketing, attribution is the practice of tracking and assigning values to the online marketing contact points, (or in various channels) that lead to a defined conversion.
Rather than dealing with a casual issue, having efficient models of both, attribution and causality, allows us to know the causes that trigger the customer’s actions, being an imperative for any company that is committed to a digital transformation process or wants to win a competitive advantage, relying on interactions with customers in physical and digital channels.
We consider that employing these two complementary concepts, analyzes and constructions can be carried out around the client’s behavior, obtaining an alternative strategy to answer common questions and assess the moments of interaction that can contribute the most to the objectives and intentions of a company in front of your target audiences.
The basic concepts will be presented from the social psychology, helping to have a different perspective in front of the understanding of cause / effect models, its application to marketing and analysis of user behavior; In the same way, the concepts of “hard” sciences related to models of causal inference will be addressed.
The first problem: Putting correlation analysis and prejudice in place
Perhaps it is due to our linear formation that we tend to give a high value to the results based on averages and correlations, for this, we attribute the totality of the causes of the events to the correlations that have qualitative and quantitative values between them.
We can see dramatic examples such as those presented byspurious-correlations
After seeing these graphs, surely you can think that it isn’t far from situations in which the bias prevails over objectivity, either by pressure, need to “win” a customer or a place within a company. Perhaps, it’s because of the subjective natural tendency that we often give wrong use to correlation tools.
Therefore, an essential condition for the successful inclusion of the concepts of attribution models to the practices of the teams of: Marketing, Data Engineer, among others, is to give rise to correlations and prejudice, focusing on the work that it requires causal analysis and additional data processing, since the correlation between two variables does not imply that one causes the behavior of the other.
Two complementary work routes are presented: theory of attribution and causal inference, the first one comes from social psychology and the second from basic sciences (statistics, mathematics and related).
An anticipated conclusion and first prerequisite suggested in this type of modeling recipes, is that before entering into the preparation of a causal data model, it’s essential to evaluate if the correct questions are being asked about the user segments, since the traditional segmentation, based on the value-spending-monetary capacity of the client, tends to hide in the eyes of analysts and marketing teams the forms of vital grouping, such as those that focus on the consistency of user behavior for decisions making.
Attribution Theory by Fritz Heider(1958)
The Psychology of Interpersonal Relations
by Fritz Heider
Link: http://a.co/d/6MNKhee
Heider’s attribution theory tries to analyze the people’s behavior and the events of life. In social psychology it’s called the attributive process.
Although in his book and later work, Heider didn’t fully develop the methodology but laid the foundations for it, now he focuses on two possible causes that can account for the behavior of others:
Internal attribution: The process of assigning cause of behavior based on the person’s internal traits such as personality, knowledge, skills, desires and intentions, rather than external agents.
External attribution: Also called situational attribution, refers to the interpretation of person’s behavior immersed in a context such as climate and influence of others.
H. Kelley’s covariation model (1967-1973)
One of the main problems that these definitions and attribution theory have, is that for many it is common sense, and that’s why they assume biased conditions in what Heider and other authors call naive psychology.
Harold H. Kelley expresses it in an interesting way in his article: The Processes of Causal Attribution (1973): “I think that social psychology is finally realizing that the appropriate role to take is not to co-ordinate common sense but rather to analyze , refine and expand on it”.
For Kelley, attribution theory “… is about how people make causal explanations, about how people respond to questions starting with” why? “ It focuses on the information that people use to make causal inferences, and what they do with that information to answer the causal questions. “ (Kelley ,1973).
In Grupodot we have integrated social psychology, which tries to understand why people decide how they decide, to complement part of the attribution model (cause / effect). The causality in events has for us a high component in the way people (users) perform an action, for example, the data samples give an account of how people flow within a physical space that belongs to a departments store.
Initial segmentations are made, dividing the groups between those that have a consistent response in all the series and those that don’t. The first group denominate as consistent and we decided that they base their decision mainly on their own traits (internal attribution), and only in these populations do we see the conditions of context, when for some reason, there is inconsistency in their behavior.
This allows global trend analysis to improve, since context and user-specific data of the same weight are not mixed, for the case in which they are consistent or inconsistent users that are likely to be influenced by external variables.
Kelley delved into the topic, focusing on a logical model for judging when a particular action should be attributed to the person’s characteristics or context. He defined the principle of co-variation: “An effect is attributed to one of its possible causes with which, over time, co-varies. This implies that there are times when both, the effect and its cause, are present and others in which both are absent”.
Returning to the example of the department store, almost all the people who visit it, have previously had a similar experience in the same store or in others, for which they have information from multiple observations and in different situations (Christmas, mother’s day, etc).
Therefore, when attribution models are built that seek to analyze the effectiveness of a notice or promotion, located at a specific point in the store, it must initially be taken into account if that user is a frequent and consistent visitor, if so, it can constructed a model that focuses on the co-variation between behavioral history in the store and there looking for the causes, or if on the contrary its an inconsistent user, the analysis should focus on the whole set of context variables inside and outside of the store.
We can say that in department stores there are people for any promotion, at any time of the year or at any point in the store, all are potential business that can not be lost, this answer we will call consistent.
On the other hand, the store has incentive mechanisms: products, discounts, which are use in any time of year or the type of person and the expected actions; we will call this answer consensus.
Finally we have situations in which there is a tremendously good or bad behavior, associated with a product or action that was never deployed with that intention, this will be the distinctive answer.
To answer the question: how does a person know if their perceptions, judgments and evaluations of the world are correct?, Kelley defines a framework of data work in three axes: person / entity / time.
And based on it, formally define the answers like this:
Consistent answer: It means that the answer of this person is maintained over time, ie: successive exposures to stimuli with which it interacts, through different channels and sensory and perceptual modalities. In this way, the data will indicate that similar results or behavior are obtained for the same person regardless of the time and entities with which they interact in different channels.
In this case, what we suggest in Grupodot is to focus on the internal data of the people: socio demography and user / brand interactions), reducing the significant weight of context-based analysis.
Consensus answer: It means that the response of the subject is similar to other people with the same stimulus in the same channel. In this case, the data indicates that it is common that most people, regardless of the moment in time, respond in the same way.
In this way, the data will reflect that it is a successful interaction. What we suggest is to increase the measurement mechanisms on the interaction itself (physical space, article, etc.), to build micro-segmentations about the behavior “inside” of the experience associated with the entity that achieves the consensus.
Distinctive answer: It is the most complex of all, since is only obtained when the person is subjected to a specific stimulus, in a given channel, for a particular time. In this case, actions similar to the consensus response are taken, but there are cases in which, necessarily, the problem must be attacked with causal inference models.
Initial application: Archetypes / Personas
A tool commonly used in marketing and design is the Buyer Persona. A Persona is an archetype, which brings together in a detailed way characteristics that describe and bring the work teams closer to reality, the context, needs and intentions of the target audiences, which will be users and potential customers of the interactions they build.
A first application, from the approach of social psychology, the attribution and the type of answers, is to expand the model and personnel templates used in the conceptualization and design processes, and integrate from that first moment data teams with the design team, to determine how the first groups of users are composed based on the behavior, establishing this way, if the defined persona belongs to an internal or external attribution model and if their type of response in the story is consensus or distinctive
For this, we suggest the following summary scheme of generic classification, based on the initial analysis of the people’s transaction data or interactions:
The suggested steps to apply in the complementary construction of personas are:
- Selection of interaction data: Based on the interaction to be evaluated (printing a video, clicking an ad, filling in a form, approaching a section within a store, commenting on a social network, making a bank transaction, etc.), the description data of the persons must be counted, as well as the transactions executed. Regarding people, isn’t about working with reserved or personal data, the data that can be located socio-demographically are essential, as well as a generic identification (can be Google Analytics userID) that allows differentiating the subject in each one of the transactions that makes in different devices and moments.
- Putting Personas within attribution typologies: The objective is to divide all the Personas into two groups according to the data. Those that have a consistent behavior against the transaction, based on data such as: frequency, amount, type of product, place, time of purchase and those that don’t. The standard on which is the definition of consistent transactions, always issued by the business experts (the client), who know their nature and history.
- Detailed classification within the subjects with a consistent answer: With the second point, two families of Personas will have already been obtained, then the first one can be taken, the one of the consistent ones and look for if within them there are differences in business nature.
- Find the consensus and the distinctive: With the remaining group, that of external attribution, you can find those that have common behaviors as consensus and separate them from the distinctive.
- Extract attributes that describe the people that emerge: With 5 armed groups: two of attribution and three of response, the objective will be to find “emerging” groups with existing data and transactions, using grouping techniques such as cluster analysis.
- Analyze and propose Personas: With the iterative execution of points 1 to 5, it is possible to find and optimize the finding of “real” Personas, people who exist and currently interact with the business in different channels.
This will allow that when advancing in the next step, causal inference, the analysis will be of higher quality since it will be contrasted with the populations that have already been validated.
Once you have a preliminary analysis about the behavior and type of Personas to be included, it’s time to take the next step: build models of causal inference, complemented with preliminary population analysis. In the next installment, we will discuss the basic concepts of these models and then integrate the two approaches into practice.
Finally, as a preview, the following deliveries will focus on the general presentation of the internal components and type results, which can be obtained with the suite of Grupodot’s attribution tools.
