Maternal_Mortality

Flatiron School // Mod2 Linear Regression Project
By: Samantha Jackson

Purpose

This project aims to test if there is a linear relationship between world development indicators and Maternal Mortality Ratio using an ordinary linear regression model. After a number of iterations, the best results, with an R-squared of .902, came from using a log transformed dependant variable.

Data

Data was pulled from The World Bank Databank, more specifically the world development indicators database. Data was output to a CSV, which made import to a pandas Dataframe easy. The initial dataset included 69 features (development indicators) for 216 countries over 40 years. I also scraped "Region" and "Continent" from Statistics Times.

https://databank.worldbank.org/data/reports.aspx?source=world-development-indicators#

http://statisticstimes.com/geography/countries-by-continents.php

The features included in the model were:

Maternal Mortality Ratio: Maternal deaths per 100K live births
Teen Fertility: Births per 100K women ages 15-19
Immunization Measles: % of children ages 12-23 months immunized for measles
Internet Use: % of population using the internet
GDP_pc: Gross Domestic Product per capita
Continent
IUxEurope: Interaction of Internet Use and Europe

EDA

I narrowed my dataframe down to 6 features based on data availability and an attempt to avoid multicoliniarity.

The only feature that looks somewhat normal is Compulsory Education. Other feature distributions are positively (right) skewed, except for Immunizations. This led me to believe I'd need to try transforming my features for my model.

This pairplot reiterates the need to transform features since few of the relationships between Maternal Mortality Rate and the other features are linear.

Model Training

As a baseline, I threw all untransformed features into a model to see the results, and the R^2 wasn't great. Testing my observations from EDA, I ran the same model with a log-transformed dependant variable and my R^2 increased about 20%.

After a few iterations of the model, I realized I may be able to improve the model by adding Continent as a feature. I took a look at Maternal Mortality Ratios by continent.

I thought the distributions looked different enough to add continent as a feature. However, I wanted to check interactions between continent and the other features.

Continent doesn't have a clear additive relationship for any of the features, which makes sense, so I created interaction terms for each feature.

Final Model

My final model is a product of many iterations of feature testing and cleaning (removing insignificant features). I was able to increase my R^2 from .652 in my base model, to .902. Residuals were mostly normally distributed, with some outliers, and mostly random as shown below.

Coefficient Interpretations

For every 1 unit increase in Teen Pregnancy, Maternal Mortality Ratio increases .4%
1 unit increase Immunization decreases MMR 2%
1 unit increase Internet Use decreases MMR 2%
For 1% increase in GDP per capita, MMR decreases .4%
A country in Asia (vs South America or Africa) decreases MMR 50%
Europe decreases MMR 91%
North America decreases MMR 35%
Oceania decreases MMR 52%

msjacksoniyn / maternal_mortality Goto Github PK