A folder for the Inter-American Development Bank's 2018 Costa Rican Household Poverty Prediction Challenge

Author: Evan Gibson

• Summary
• File Descriptions
• Exploratory Analysis
• Feature Engineering
• Model Results
• Acknowledgments and References

Social programs have a hard time making sure the right people are given enough aid. It makes sense for these programs to establish thresholds for aid "worthiness." Income/expense measurements provide the standard baseline for credit worthiness in most parts of the world. But how can aid programs develop practical criteria when members of their target populations frequently lack "acceptable" income/expense verification? The least affluent members of any society often lack robust income/expense records; therefore, those individuals have a hard time proving that they qualify for certain aid programs.

Microfinance organizations around the world have spent years developing "just right" methods to determine how their funds should be allocated. To make things difficult for those organizations, the poorest segment of any population experiences especially acute outlier circumstances. From income level to the number of mobile phones, the variability of descriptive characteristics among those who occupy the poorest level of society is staggering. In Latin America, one popular method utilizes a unique algorithm to verify income qualification using measurable characteristics outside of income/expense records. It’s called the Proxy Means Test (or PMT). With PMT, agencies use a model that considers a family’s observable household attributes like the material of their walls and ceiling, or the assets found in the home to classify them and predict their level of need.

Even with the nuanced approach of the PMT, accuracy (among other metrics) remains a problem. As the Latin American population grows and poverty declines, reorientation of the PMT is necessary. The Inter-American Development Bank (IDB) has launched a public Kaggle competition to improve the PMT. The IDB is branching away from traditional econometrics in an effort to attain better, more reliable predictions.

Beyond Costa Rica, many countries face this same problem of inaccurately assessing social need.

The present repository will address the following:

• Detail the author's exploration of the given data.
• Attempts to tune an alternative PMT model.
• Investigate the PMT as a concept for the betterment of development practices.

The Target variable is classified (ordinally) as follows:

• 1 = extreme poverty
• 2 = moderate poverty
• 3 = vulnerable households
• 4 = non vulnerable households

• simple_predictions (Folder)
  • All of the enclosed models are the simplest versions of their title purpose. As they contain little parameter tuning, feature engineering, and data considerations, they are most useful for orienting oneself to classifier models that could work for this project and for understanding the focus of the competition.
• images (Folder)
  • This folder contains PNG images that are used in the README
• Exploratory Analysis.ipynb
  • This notebook is used for the exploration of the poverty dataset. It contains descriptive charts and is here to explain some of the nuance within our research context.
• Random Forest and Decision Tree Models.ipynb
  • This file contains tuned Random Forest and Decision Tree models. This file also contains details for visualizing a single tree of a random forest or a decision tree.

Without understanding the data we will use, we cannot hope to derive meaningful, scalable predictions.

The training data contains 143 "unique" variables. For comparison, the test data contains 142 variables. The only difference between the variables in the two files are in the presence of the Target column. The following details some of the most "important" characteristics of both sets:

Income lies at the heart of the PMT. Many organizations use income to determine an individual's worthiness for social programs, interest rates, and credit limits. If our model disregarded income, it would undoubtedly mischaracterize its target. The following is the income distribution for the training set:

It may appear like the high income outliers in this data will skew our model's results. However, if one inspects the data more carefully, it is clear that the high income outliers lie in the "4" bracket of Target. These outliers, if handled non-linearly, present the opportunity for our model to display income thresholds that greatly, and perhaps appropriately, impact Target. The following graphic displays income as it occurs throughout Target groups (in the training data):

On a related note, during exploration of the income, it becomes obvious that no one who owns a fully paid house (tipovivi1) reported income. While this fact does not preclude those who own a house from being considered by our model, it does illuminate some of the data collection context. Why is it that those who own their houses do not report income? It could be the case that the income for the other individuals was self reported as part of an application process. It could be the case that the data for those who own their houses were collected under significantly different circumstances than those who do not own their houses. Ultimately, the answer is unclear. That said, it will be important to understand that there is a possibility that the present data were consolidated from different sources under different methodologies.

Visual exploratory analysis is an important aspect of most statistical methodlogy. By visually representing data, we can identify non-linear patterns that could impact our findings. For example, a simple countplot of the escolari variable shows us that there are two "spikes" in the amount of schooling our population receives (at 6 and at 11 years):

While not enough to paint the whole picture, this graph displays prevalent aspects of the data. If education ends up being at the nexus of the model's most important features, a researcher's understanding of the societal impactors that have driven these spikes might be invaluable to deriving meaningful insight.

In addition to helping researchers understand the characteristics of the data, visual analysis can help researchers determine metrics for model success. The following graph displays the count of individuals belonging to each of the Target groups:

As shown above, there are significantly more individuals belonging to the "4" group than to any other Target group. By understanding this, we can avoid the pitfalls that might come from overreliance on accuracy. Think about this, if the data were 90% comprised of 4 = Target individuals and our model randomly guessed "4" for every prediction, we would have a 90% accuracy. Even though this sounds like a reasonably successful model, it isn't. Using this visualization, we can also see that we might do well to measure our model's weighted F1 score, due to the skewwing towards class "4" (even though the contest makers decided to use macro F1 as the primary performance indicator).

In all, the training and test sets are reasonably similar. That said, if they weren't, would we necesarily have an issue that makes the data sets unusable? The present author's opinion is that differences in the characteristics of training and test data do not necessarily preclude them for usefulness in model-building.

Why?

It is arguably more important for a model's training data to match the collection context and for the test data to match the model's usage context. In most model-building, the training and test data are randomly split at the 80/20 level; therefore, there is generally less concern for collection contexts of training and test data to be different from one another. However, in some cases, broad economic studies use data derived from the general population, rather than from a model's proposed usage context. While this is less of an issue when describing the state of affairs concerning the general population, it can prevent targeted models from attaining the best results in their intended scopes.

Of all the variables included in the present data, there were few that related composite details concerning an individual's household. For instance, the variable qmobilephone indicates the number of mobile phones in a given individual's household. However, feature engineering is required to run a model that considers the proportion of mobile phones per household to a household size (phone_per_person_household). The following illustrates how the author engineered the qmobilephone/hhsize feature for modeling:

df_train['phone_per_person_household'] = df_train['qmobilephone']/df_train['hhsize']

As it turns out, the phone_per_person_household variable ended up being one of the most important variables in the models in which it was employed. By engineering custom features, we can focus modeling efforts to deal with unique variable combinations. In addition to phone_per_person_household, the author engineered 40 other variables.

Most of the engineered variables were built to clarify relationships throughout households. By clarifying the experience an individual has within their immediate familial context, we can better understand our data.

The author also built variables that consolidated some of the underrepresented groups under the assumption that the dimensionality of certain groups do not effectively add insight to the data. For example, the author combined the techozinc, techoentrepiso, techocane, and techootro variables into roof_waste_material because the materials noted in those "techo..." variables are waste materials. It may be the case that the impacts each of the roofing variables has on Target is similiar in characteristic, meaning that the most salient fact about all the people who fall into those groups is that they have roofs made of waste materials.

In the most relevant metrics of evaluation, the tuned random forest model performs the best (against the other models that are present). The following images contain the detailed metrics of the tuned random forest model:

In all of the metrics explored in the visuals above, the random forest model thrives. The following visual shows one of trees in that forest:

Understandably, the tree is rather large. It is included for in-depth exploration purposes.

• Adam P
  • For all things markdown
• Will Koehrsen
  • Will's detailed exploratory analysis and modeling were helpful for building my understanding of the data.