This idea began as a curiosity for me as to which micronutrients are most important in the human body and its functions biochemically. Because of my background in statistics, computer sciences, and what is now my degree in data analytics I chose to look for data sets that could offer analytically insight into my questions outside of just literature. It gave me an opportunity to apply skills I’ve learned to a topic I enjoy, in addition to honing my ability to use R. Almost immediately came across the NHANES survey out of the CDC and the data sets they publish bi-yearly from it, I then took up the task of combining them into a single data frame I could analyze.
After my initial exploration of the data I discovered some interesting trends I didn’t expect and wanted to look closer. Namely that while obesity is increasing we appear to consume the same amount of calories. With that said this project transitioned into what factors influence obesity and now more specifically waist sizes increasing over time. It very much is a find the needle idea but it was not only entertaining but informational.
Over this semester I aimed get more data into the file and properly clean it, in addition to completing a PCA of the data as well as other possible regressions. I was helped and advised by Dr. Okamoto during the semester and greatly appreciate his insight.
The project is not complete and I aim to be able to spend more time on it during my masters. I want to look at preforming different clustering methods, dimension reduction techniques, and more ways to deal with multivariate data in hopes it might discover something fun. Currently I worked to just determine what actually plays a role and if there is correlation between obesity or waist size, and recorded metrics in the survey. Some visualizations are present but not all however they will be added in the future.
Currently I have separated files for each of the stage in the project, they individually do something although they could made more efficient. I do aim to improve that. All of them can be found in the repository for more specific detail on what they do or what is included.
The R script below loads any packages I’ve used or tried to use, and then individual functions I wrote for this project.
source('Packages.R') The next, arguably most important file is contains the R coding for the import and reading of all of the individual data sets from the main data folder. The output of this is several large lists with 5 elements. They are labeled by which NHANES survey is contained inside, with each element representing a different survey year/release.
source('DataImport.R')Current Elapsed Time: 5.151193 Seconds
Data Read From Folders
The phrase most of the time preforming an analysis is spent in cleaning unfortunately holds true. I’ve lost track at the total time spent on cleaning, merging, and creating but am happy the difficult parts are done for now.
One of the many annoyances with the NHANES data is that most things are separate files requiring the merging into a single data frame. what made it worse was that I was exploring over the last 10 years, so I had five times as many files as opposed to a single year. The native NHANES files all contain a variable name, typically 6-7 letters or numbers, as well as a written description of each variable as a label attribute. Throughout the process of combining and merging the data I worked to keep both to aid in readability.
By the end I had 5 merged files from the following categories or sections from the CDC: Nutrition Day 1 and 2, Income, Demographics, Body Measurements, and Physical Activity.
source('NutritionCleaning.R')Current Elapsed Time: 0.1759861 Seconds
Finished Step 1: Data Pulled From Files, Years assigned by release yearCurrent Elapsed Time: 53.32734 Seconds
Finished Step 2: Data Frames CreatedCurrent Elapsed Time: 53.33024 Seconds
Finished Step 3: Assigning Labels as Column NamesCurrent Elapsed Time: 54.28219 Seconds
Finished Step 4: Created extra variablesCurrent Elapsed Time: 54.28241 Seconds
Finished Step 5/5: Finished Cleaning
source('DataMerging.R')The two files above handle this task. The nutrition cleaning file was my original file that only combined the nutritional data since that was at the time all I was concerned about. It is extremely inefficient and at some point I hope to remove it entirely.
The Data merging script takes a simpler function (written in the Packages.R script) and combines all of the other files together to create an AllMerge file. The primary merge function in R didn’t work leading to me writing my own. Since the variable names remain the same over the years I was to determine which exist in each data frame and merge them, and then matching respondent id. The file also removes many of the old created files to help with an issue of RStudio crashing.
The remaining file before any creation or removal contains 155 dimensions and a total of 41297 individual records.
source('CreatingVariables.R')
source('CreatingGroupedDFs.R')Using the created data frame I then chose to create additional extra variables from previously existing variables. Previous research and experience loosing weight had led me to look at diet consumption and macronutreient consumption, I created metrics using dplyr that calculated what percent of the diet was fats, proteins, and carbohydrates. I also looked at saturated fats.
Following exploration of those variables and additional files from NHANES, I created variables for the following:
-
RFM Metric - Newer alternative calculation for predicting body fat percentage. Works better then BMI
-
RFM Obese - Binary variable that depicts if someone is obese or not based on the RFM metric and the cut off point for their specific sex
-
Weight to Waist Size - Ratio for Weight to Waist, testing if this would be efficient when compared to height to waist
-
BMR - Basal Metabolic rate based on listed height and weigh of subject
-
BMR - Kcal - measurement of BMR subtracting the total number of calories consumed.
-
Height to Waist Ratio - Simple ratio of height to waist size. From my findings general consensus lists a ratio of .5 or under as healthy.
-
BMI Obese metric - Just like with RFM, a variable showing obesity however it is broken down 0:4 to show what stage they are. Someone who is underweight is a 0, normal is 1, overweight is 2, obese is 3, and clinically obese would be 4.
-
Under 18 - Binary to show if someone is over or under 18, used to separate children data from adults since it was skewing and affecting the averages.
-
Under 15 - Binary for over or under 15 years of age. This was the year where the rate of growth showed slowing in both height and weight, lessening the impact in averages with adults.
-
Age Group Numeric - Separated age into 18 groupings. the groups span about 5 years of age but gave a glimpse into dietary changes in different age groups. It helped explain some things, for example caffeine looked to increase obesity. However when you looked at the data by age it was that caffeine consumption increased with age and there happened to be a lot of older people who were obese driving up the average.
Other important cleaning done here was the removal of 43 rows from the original AllMerge data frame, leaving a total of 129 dimensions. For the most part it came down to irrelevancy or them just not being important for what I was trying to see.
I also implemented quick NA handling where if a specific row is missing more then 75% of the possible data points, the row is removed. At the point of implementation it was working however looking at it now, I’m not sure if it is changing anything or removing rows.
So far in my explorations I’ve been able to come to the single conclusion that data just doesn’t make sense. At the beginning I was assuming that I would see a trend of more calories being eaten overtime and there was not particularly a significant change which was odd considering that we were still increasing in obesity rates. A partial explanation here would be we are on average eating more then we need so we will end up storing more, but it led me to questions as what else is missing or included in our diet or lifestyle that is leading to an increase in adoposity. Genetics and so many other factors ultimately play into this.
In a general sense, we seem to be lacking in vitamin consumption across the board. Biochemically this could be acting as a catalyst for the obesity we’re seeing. There was specific dietary trends which also could point to why things are getting worse. We know sugar consumption can lead to storage of fat, we also know that metabolism tends to slow as we age. In the data it was showing that children consume significantly more sugar, in excess of 10 grams more then an adult. This could contribute to their obesity, but also if they are setting a foundation where they hold more weight earlier in their lives it makes it harder to loose later in life.
Waist size has increased about four centimeters over the 10 year span, with the height to waist ratio raising from .577 to .602. We also on average are four kilo’s heavier. Finding this more confirmed that we in fact still seeing this increase in obesity in the US, interestingly those who were considered to have a healthy obesity level appeared to consume more calories. Well then I thought, maybe they’re exercising more and they didn’t appear to be by that much on average. It added to the list of things not particularly adding up.
In line with my goals I preformed a PCA using the prcomp function in R.
AllMerge[is.na(AllMerge)] <- 0
AllMerge <- AllMerge[,-1]
NonNumeric <- which(!sapply(AllMerge, is.numeric))
PCA <- prcomp(AllMerge[,-NonNumeric])
PCASum <- summary(PCA)
tidy(t(PCASum$importance[,1:12]))# A tibble: 12 × 1
x[,"Standard deviation"] [,"Proportion of Variance"] [,"Cumulative Proporti…¹
<dbl> <dbl> <dbl>
1 7889. 0.605 0.605
2 4688. 0.214 0.819
3 2702. 0.0710 0.890
4 2310. 0.0519 0.941
5 1399. 0.0190 0.960
6 822. 0.00657 0.967
7 772. 0.0058 0.973
8 706. 0.00484 0.978
9 687. 0.00459 0.982
10 637. 0.00394 0.986
11 565. 0.0031 0.989
12 440. 0.00188 0.991
# ℹ abbreviated name: ¹[,"Cumulative Proportion"]
fviz_pca_var(PCA,
col.var = "contrib", select.var = list(contrib = 7),geom = c("point", "text"),
gradient.cols = c("#00AFBB", "#E7B800", "#FC4E07"),repel = TRUE)
fviz_pca_var(PCA, axes = c(2,3),
col.var = "contrib", select.var = list(contrib = 15),geom = c("point", "text"),
gradient.cols = c("#00AFBB", "#E7B800", "#FC4E07"),repel = TRUE)
fviz_pca_var(PCA, axes = c(3,4),
col.var = "contrib", select.var = list(contrib = 10),geom = c("point", "text"),
gradient.cols = c("#00AFBB", "#E7B800", "#FC4E07"),repel = TRUE)
The function pops out 125 PC’s for each dimension however the top tweleve will explain 99% of the variability, with the first 5 containing over 96%.
As shown in the charts (which I’m not sure I fully understand) Lycopene is way out to the left. I’m planning at going back and scaling the variables to see how things are changed, however for now because it gives an error message I will not.
I didn’t find this are beneficial as I maybe though it would be.
I next ran a stepwise regression and simple linear model on the data. Height to Waist ratio (HTW) was set as the response variable; additional columns were removed due to their direct explanation or contribution to the ratio (Ex: Height or the weight of someone)
intercept <- lm(HWR~1, data = AllMerge[,-c(1:9, 118:122,124:128, 83)])
glm <- lm(HWR~., data = AllMerge[,-c(1:9, 118:122,124:128, 83)])
Stepwise <- step(intercept, trace=0, steps = 10, direction = "both", scope = formula(glm))
glance(glm)# A tibble: 1 × 12
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC
<dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
1 0.133 0.131 0.160 60.9 0 104 17146. -34081. -33166.
# ℹ 3 more variables: deviance <dbl>, df.residual <int>, nobs <int>
glance(Stepwise)# A tibble: 1 × 12
r.squared adj.r.squared sigma statistic p.value df logLik AIC BIC
<dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
1 0.121 0.120 0.161 566. 0 10 16846. -33668. -33564.
# ℹ 3 more variables: deviance <dbl>, df.residual <int>, nobs <int>
tidy(Stepwise)# A tibble: 11 × 5
term estimate std.error statistic p.value
<chr> <dbl> <dbl> <dbl> <dbl>
1 (Intercept) 0.453 0.00304 149. 0
2 PAQ650 0.0329 0.00168 19.6 6.07e-85
3 FatPercent 0.00128 0.0000665 19.2 5.07e-82
4 DRQSDT1 0.0679 0.00330 20.6 1.14e-93
5 RIDRETH1 -0.0135 0.000643 -21.0 1.00e-97
6 DMDEDUC2 0.00738 0.000563 13.1 3.37e-39
7 RIDAGEYR 0.000665 0.0000493 13.5 1.89e-41
8 PAQ620 -0.0160 0.00145 -11.0 2.93e-28
9 DR1TMOIS 0.0000107 0.000000827 13.0 1.60e-38
10 DR1TMAGN -0.0000872 0.00000755 -11.5 9.22e-31
11 CarbPercent 0.000442 0.0000475 9.30 1.54e-20
Overall the base linear model only explained 13.3% and the stepwise model which only contained 10 variables explained 12.1%. I will need to go and access model fit and assumptions which don’t look optimal. I also will need to separate children from adults in the data frames when running the models due to their points being noticeably different and impactful.
par(mfrow = c(2, 2))
plot(Stepwise)
The variables below were included in the model based on the output from the stepwise regression. Because of the 40,000 different data points it makes plotting them individually difficult and I still haven’t determined the best way to visualize all these variables without taking up a ton of space.
-
Vigorous Rec activities
-
Fat Percent
-
On weight loss or low cal diet
-
Race/Hispanic Origin
-
Education level (over 20)
-
Age in Years
-
Moderate work activity
-
Moisture
-
Magnesium
-
Carbohydrate Percent
Most of this makes sense in terms of how they may affect ones HTW. I found it particularly interesting that magnesium was the specific micronutrient included since it was not significant when running the model all at once.
As for the other variables I want to look closer at ethnicity but also Vigorous Rec activities. Ethnicity makes sense to me due to dietary differences and even possible genetic differences within groups of people. As for Vigorous rec, in the model it actually has a positive coefficient so I will want to look and see if that is completely correct or not.
I’ve spoken a bit on some of what I would like to explore above. Other aspects I would like to try are different clustering methods, splines, and possibly exploring more computationally intensive methods in python as I better learn analytics packages.
I also want to start focusing on more literature surrounding the topic, I will likely have more access to free material at the university of Minnesota.