Practical 2 : targeted marketing

The aim of this practical is to perform a data analysis on the synthetic dataset you generated last week in order to identify a group (or groups) of customers at whom a marketing strategy could be targeted. You will generate summary statistics and data graphics that will characterise what makes your customer groups distinctive from the Leeds population as a whole.

Task 0: Re-visit the datasets

Last week you worked in R and RStudio to create a synthetic population for the c.320,600 households in Leeds using functions contained in the rakeR package. You started with an individual-level survey dataset describing holiday-making behaviours and allocated individuals from this dataset to Output Areas (OAs) in Leeds based on 2011 Census data.

Re-open your .Rproj from last week’s practical. The Environment window in RStudio should list the datasets that you worked upon. Re-familiarise yourslef with these by entering glimpse(<tibble-name>) into the R console.

Create a new RMarkdown file by selecting File > New > New R Markdown..., click OK with the default options, then delete all the placeholder code/text in the new file.

Paste into your RMarkdown file the notes from this document, which contains an outline/skeleton of what you’ll need to complete the practical. Save the file with an appropriate name (practical2).

Task 1: study available variables

This practical session introduces a data analysis to inform a marketing strategyTargeted marketing involves identifying a set of customers that are distinctive in some way and characterising them such that marketing activity could be tailored to them. We are not at all interested in the substantive elements of a marketing strategy in this practical (and module); but rather in the elements of data analysis that could be used to inform and focus strategy development.

. In order to do this, we want to characterise the potential customer population as fully as possible given the data available to us.

Geography is important – where customers of a particular category live and also the types of area in which they live. The simulated dataset allocates individuals to Output Areas (OAs), so we have an indicator of this. These OAs can also be linked to geodemographics.

Demographics clearly play a role – how old and how affluent potential customers are will influence their holiday-making behaviour. In addition to geodemographics, there are known demographic characteristics in the simulated data: sex, age_band, household_income, number_children.

Psychographics is another concept used in targeted marketing, which refers to customer attitudes, aspirations and choices. Psychographics are more difficult to measure because they require some way of ascertaining what people think. In your dataset there are a limited range of these: the destination choice (overseas_airport) suggests something about preference and also how satisfied customers were with their most recent holiday (satisfaction_overall).

Task 1.1: Attach OAC and Airport names

Inspecting the variables in the simulated dataset, you may have noticed that full names are missing for the oac_grp (numeric values rather than supergroup names) and uk_airport and overseas_airport. Lookup tables containing full names are provided as .csv files in the ./data/info directory. We can add these to our simulated dataset, again using dplyr::join.

Task 1.2: Generate frequency plots

An initial means of exploring how the simulated population distributes on these categorical variables is to plot frequencies across those categories. The code snippet below, also in the .Rmd, displays frequencies of the simulated population amongst each age_band, using ggplot2.Plot OAC

ggplot2 specifications have a consistent form that identifies how data are mapped to visuals:

  1. Identify dataset : data=
  2. Specify the mapping : aes(x=..., y=..., fill=..., ...)
  3. Select a geometry (or marks) layer : geom_...

You should relate the form of ggplot2 specifications – the ggplot2 grammar – to the encodings of data and visual channels discussed in GEOG5032’s session on Data Visualization in Tableau.

Given this, now write ggplot specifications of your own for the other variables in the simulated datasetHint: You will need to change the parameterisation to aes().

.

Task 1.3: Generate frequency plots on variables with many categories

Plot destinations

Bar charts are effective at conveying frequencies across categories where the number of levels is reasonably small. You may wish to consider alternatives, such as Cleveland dot plots, for summarising frequencies in data with many levels (destinations in this case).

The code snippet below is a ggplot2 specification for generating a dotplot of frequencies by airport destination. This plot is ordered according to the frequency with which countries are visited, and within countries, the frequency with which airports are visited. In order to effect this ordering, we cast the country variable (dest_airport_country) and airport variable (overseas_airport) as factors and order the factor levels by frequency. There are also some updates to the encodings: since we wish to display frequencies top-to-bottom, airport names (overseas_aiport) are mapped to the y aesthetic (aes()) and frequencies to the x aesthetic, we also use facet() in order to group the plot by destination country.

Task 2: Explore research questions

You now have some knowledge of the data available to you and how the synthetic Leeds population distributes on these key datasets. Next, you should kick-on with your data analysis by identifying research themes and performing exploratory analyses on those themes.

In the example provided here we focus on a single destination, Palma Mallorca (PMI), part of the Balearic Islands, and a destination associated with beach-style holidaysWhen you begin data analysis on your assignment, you should think carefully about the analysis focus. For instance, you may choose to group destinations according to some category – city break, beach holiday, etc. – and characterise the customer populations holidaying to that destination category. You may wish to start with a customer type – for example younger adults without children – and identify destination categories that are accessed by that customer type.

. For a marketing strategy focused on Mallorca, it is clearly useful to profile the known customer population holidaying there (implied by the synthetic population). We will do this in two ways:

  1. Observing frequencies of known PMI holiday-makers by customer group.
  2. Comparing these observed frequencies against the frequencies that would be expected given the Leeds holiday-making population as a whole.

Task 2.1: Calculate totals and proportions

You will notice from the Cleveland dotplot that Palma Mallorca is the destination most frequently visited by the synthetic Leeds population. Using some dplyr functions – filter(), summarise() – I calculated a total of 25,030 households holidaying to Palma Mallorca, c.8% of all Leeds households. You may wish to check that a similar figure is recordedDon’t worry if your calculations do not exactly match mine: there is a stochastic element to the procedure we used for running the spatial microsimulation and your synthetic dataset may have been specified with a different set of constraint variables to mine

 for your synthetic dataset by writing some code to calculate the total number of households holidaying to Palma Mallorca.

Finding that Palma Mallorca, and destinations in Spain, are those most frequently visited by the synthetic population is fine. Far more interesting and useful is information on the characteristics of the c.8% of households that have visited Palma Mallorca and particularly whether those characteristics are discriminating – e.g. the extent to which they are different from the Leeds population as a whole. The code block below makes use of dplyr functions to calculate:

  1. The number of Leeds households holidaying to Palma Mallorca by OAC group expressed as a proportion of all holiday-makers to Palma Mallorca
  2. The number of Leeds households by OAC group expressed as a proportion of all Leeds households

Comparing these numbers allows us to evaluate whether particular OAC groups are over- or under- represented amongst those holidaying to Palma Mallorca.

The code above was deliberately written so as to be re-usable. You may wish, for example, to focus on a destination other than Palma Mallorca and profile customers holidaying there. It is possible to do so simply by changing the value with which dest_focus is initialised. For example, if you wanted to focus on Cancun, Mexico (CUN):

You may later choose to focus your study on a list of destinations. This can be achieved by generating a vector of destination names and supplying this to dest_focus.

Task 2.2: Plot totals and proportions

You will want to profile selected destinations not only on the OAC variables, but by the other demographic and attitudinal variables originally available in the individuals dataset. This could be achieved by re-using the code blocks in the previous section and editing the group_by – e.g. to summarise over the reported satisfaction levels for a given destination group_by(satisfaction_overall).

This would, however, involve much copying-and-pasting of code and here I’ve abstracted the process of calculating proportions grouped on a given variable into my own function. You may have yet to write R functions in LUBS5308, and there are a few syntax things happening here that look odd (!!<parameter-name>)This is to do with the fact that dplyr functions generally use non-standard evaluation, which creates challenges when accessing dplyr programmatically: this post provides an efficient and accessible discussion.

, so do not worry too much about the details here, but a general idea of what the function achieves is useful.

The code snippet below defines the function – on creating this (by running the code), you will notice it listed in the Global Environment window of RStudio:

We then create a new dataset that will hold the summary statistics (control and focus proportions) by variable type.

  1. Create a vector containing the names of grouping variables supplied to calculate_props(). You can change these variables by editing the fields supplied to select(). You can also edit the destination(s) to focus on.
  2. Create a new dataset with the control and focus variables (used by calculate_props()) defined. Note that the control variable is de-facto set to entire population, but you could select a particular subset for comparison here in the same was as the focus variable is defined.
  3. Using a map function, iterate over each grouping variable name and use this to parameterise calculate_props().

Finally plot these data to explore demographic and attitudinal characteristics that are over- or under- represented amongst the destination in focus. I’ve selected Ibiza (IBZ) and the chart does allow us to identify a category of holiday-maker to Ibiza that is discriminating: a bimodal age profile (65+ and 05-49); households with children; Suburbanites OAC groups and not Cosmopolitans.

Plot of control and focus proportions

The ggplot2 specification is below, with a little data munging prior to the ggplot2 call:

  1. gather() the control_prop, focus_prop, diff_drop variables so that they can be passed to facet(). These are then identified with the key stat_type.
  2. In mutate() cast stat_type as a factor so that the levels can be be ordered in the faceted chart and create a new variable stat_sign, which identifies whether stat_value is positive or negative (we do this to differentiate bar colours left and right of the vertical line).

Task 3: Plan your Ass#1 data analysis

This practical provides re-usable code and examples that you could use to perform your Assignment 1 data analysis.

The precise nature and focus of your data analysis is up to you and you should certainly not feel restricted by the techniques made available here. You may wish to explore some machine learning techniques introduced to you as part of your LUBS5308 module: you could for example use decision trees to identify sub-groups of customers associated with a selected destination or set of destinations, CHAID could be a particularly useful technique here; or clustering algorithms to group destination types according to the demographic and attitudinal profile of households holidaying to those locations.

You are not required to use these additional machine learning techniques and only use them if you fully understand them. Your assignment will be marked on the quality and clarity of your data analysis and your ability to make and convey sound analytic thinking.

Some general guidelines for your data analysis:

End of practical

Well done for completing this week’s practical.

Before you leave the practical make sure you save your work: