Feature engineering and variable imputation

This is a follow-up to an exploratory data-analysis post of mine. When I was nearly done with writing that post, some new questions came to my mind.

Here I’ll continue with preprocessing. More basically, can I determine some other variables (out of the 76) that are -apparently- influencing the Blood Pressure in some way?

I am not trying to do groundbreaking medical research here. Hey I’m just a blogger, playing with a dataset for personal entertainment, trying out some of those fancy Machine Learning algorithms that I encountered during MOOC homework assignments. This time I’ll try these techniques out on a larger dataset that I find interesting.

Attribute selection and variable imputation are a topic worth studying, especially on a more complex dataset such as the NHANES data.

The NHANES dataset

NHANES is the US National Health and Nutrition Examination Study. It is a carefully curated, larger medical survey aiming to get a representative sample of the general US population. The survey is carried out periodically.

The NHANES dataset is available as an R package on CRAN. I have used package version 2.1.0, specifically “NHANES 2009-2012 with adjusted weighting”.

Table 1 shows the column names and further below, and in table 5, are some other important metadata of the NHANES dataset.

This dataset contains data from two collection periods, 2009-2010 and 2011-2012. Some attributes weres collected during period 1, but not in period 2, or they were collected in different ways. Therefore the dataset contains many NULL values just because of this. Moreover some attributes are valid only for women (“nBabies”), and other features such as “HeadCirc” (head circumference), were only collected for babies and children.

Feature Engineering

These are the feature selection steps I will apply. Some attributes are redundant, for exaample age-in-years, age-in-months and age-decade. Many blood-pressure values are redundant; keep only the average value. First, remove attributes that were collected for babies, children and generally people from the youngest age groups, less than 20 years of age. This leads to the selection of columns, marked by me as “drop” in Table 1 below.

These are the remaining columns in the NHANES dataset and the number of NA values in each column, respectively:

Variable imputation

Imputation is done with the MICE method: “Multiple Imputation by Chained Equations”.

The MICE method removes all NA values by guessing intelligently. Here I have called the mice() function with default parameters. This is why it created nonsensical values in places, e.g. pregnant fathers., see below.

PregnantNow is the most extreme example- more than 50% NA values were undefined for this column.

This demonstrates that the default choices that the mice() function makes are not always suitable for this dataset. For now, I’ll leave these new values inside the imputed dataset though.

According to the authors of the mice package these decisions need to be made:

1. Assumption: Missing at random satisfied?
2. Target Model / intended Use of the data
3. Set of variables
4. Include dependent variables
5. Order of imputation
6. intiializations and number of iterations
7. number of imputations (how many multiple imputations)

(TBC - imputation done better - the mice package is a lot more flexible and has more to offer to achieve better imputations)

#NHANES_imp_qik <- quickpred(NHANES)
NHANES_imp_qik <- mice (NHANES, pred = quickpred(NHANES, minpuc = 0.25), include = "Age", maxit= 1)
NHANES_imp_qik_complete <- mice::complete(NHANES_imp_qik)
#table(NHANES_imp_qik_complete$PregnantNow, NHANES_imp_qik_complete$Gender)

However, after imputation there are no more NAs in any column of the NHANES Dataset.

Supplementary Materials

These metadata show, for example, the unique distinct values of columns where it makes sense to report them (e.g. male/female for column “gender”).

Top

## $Gender
## [1] "female" "male"  
## 
## $Race1
## [1] "Black"    "Hispanic" "Mexican"  "White"    "Other"   
## 
## $Education
## [1] "8th Grade"      "9 - 11th Grade" "High School"    "Some College"  
## [5] "College Grad"  
## 
## $HomeOwn
## [1] "Own"   "Rent"  "Other"
## 
## $Work
## [1] "Looking"    "NotWorking" "Working"   
## 
## $Diabetes
## [1] "No"  "Yes"
## 
## $HealthGen
## [1] "Excellent" "Vgood"     "Good"      "Fair"      "Poor"     
## 
## $LittleInterest
## [1] "None"    "Several" "Most"   
## 
## $Depressed
## LittleInterest
## 
## $SleepTrouble
## Diabetes
## 
## $PhysActive
## Diabetes
## 
## $Alcohol12PlusYr
## Diabetes
## 
## $Smoke100
## Diabetes
## 
## $Smoke100n
## [1] "Non-Smoker" "Smoker"    
## 
## $PregnantNow
## [1] "Yes"     "No"      "Unknown"

Obligatory Disclaimer

(from the NHANES Package Documentation):

Please note that the data sets provided in this package are derived from the NHANES database and have been adapted for educational purposes. As such, they are NOT suitable for use as a research database.