Main Commitment: I will code to do data analysis for at least an hour every day for the next 100 days.
Start Date: 20171211
A companion project to 100 Days of Reading Paper.
Rules
Modification (20180105): I found The 5Day Data Challenge
in Kaggle today . I like Data Challenge
better than Data Analysis with R
. The title of the post was changed from 100 Days Of Code: Data Analysis with R (Round 1)
to 100 Days Of Data Challenge (Round 1)
. Even though no one tweets #100daysOfData
, I want to use it.
 I will code using R or do data analysis with other software (such as Python) for at least an hour every day.
Projects are counted towards the challenge:
 statistical analysis projects in my work
 projects from dissertation
 competitions, such as Kaggle
Activties are counted towards the challenge:
 understand datasets in excel
 import data into R
 clean data
 write codes to do exploratory analysis and apply statistical model
 write statistical reports
Activities are not counted towards the challnege:
 write emails to clients
 meet with clients/supervisors
I will tweet my progress every day, with the hashtag
#100DaysOfCode
,#100daysOfData
and#100DaysOfDataScience
and note which day of the challenge I’m on.I will encourage and support at least two people each day in the
#100DaysOfDataScience
challenge on Twitter. I can read at most 5 tweets about#100DaysOfDataScience
every day. Less is more. Don’t spend more than enough time on the social networking website.3 Options
 Like tweets
 Leave a comment
 (optional) Looking at their projects and giving them feedback (no more than 10 minutes per day)
 I will track my progress here and push to GitHub.
I will only count the days where I spend at least some of my time building projects — not the days where I spend all my coding time working through lessons and tutorials.
I will only skip a day if something important comes up. And when I resume, I won’t count the day I skipped as one of my 100 days.
Some important additional considerations
 Don’t skip two days in a row, and try not to skip more than 1 day in 2 weeks.
Template for Log
1  ### Day : 
LOG
Day 1: 20171217 Sunday
Today’s Progress (achievements and frustrations):
 import data into R
Thoughts and Emotions
The column name is messy after the data was imported into R. Thinking of the idea that I have to rename all columns one by one, I feel scared.
Today I learned parts of tidyverse
package in Lynda.com. I thought it was a good package. When I start using it, I missed the old method. But when I saw base R converted all character variables into factor variable after import, I missed tidyverse
. I had mixed feelings about both base R and tidyverse
.
Day 2: 20180103 Wednesday
Today’s Progress (achievements and frustrations):
One hour work
 Cleaned the dataset column by column
 Wrote a Markdown table to record
Thoughts and Emotions
I did not finish the cleaning. I know cleaning is faster if I use SAS.
I know the one hour work is not productive. But it gave me less pressure.
I set a oneclock timer. I can feel that when it was approaching one hour, the stress level increased.
Every time I work long hour until burning out, it will make it harder to start work next time. My goal is not the speed or productivity but consistency.
Tomorrow’s plan
 Clean data
 Make descriptive data of variables left in the cleaned data
 Write report
 Run simple logistic analysis of batch effect and
Reasons
Day 3: 20180105 Friday
Today’s Progress (achievements and frustrations):
I created a R package project for beginners on RSnippets.com. It is to build a utility package for personal usage.
I started to make one.
Thoughts and Emotions
I did not do this challenge again. The small project can get me started. This is the reason I did a project for beginners, not a big project.
I feel good at the end of onehour session.
Tomorrow’s plan
 Clean data
 Make descriptive data of variables left in the cleaned data
 Write report
 Run simple logistic analysis of batch effect and
Reasons
Day 4: 20180106 Saturday
Today’s Progress (achievements and frustrations):
 Read the post about how changing location will affect the linear regression result.
 Reproduce the example from
Sean
I created 2 toy examples:
 $y = 0.2x$
 $y = 0.2x+x^2$
$x$ is uniform(0.5,0.5). $x$ in the program means uniform(0,1).
If the true model is linear
The above graph shows fitted models look almost the same when the location of $x$ is changed.
The above graph shows the intercept will decrease when the location of $x$ becomes bigger. This conforms to the mathematical proof.
$$y = 0.2x = (0.20.2a) + (x+0.2a)$$
$0.20.2a$ is the intercepts after $x$ moves. It has the linear relationship with $a$.
However, in theory, the coefficient of $x$ remains the same. In fact, it changes a little bit.
If the true model is nonlinear
The above two graphs show that the location has big impact on cofficients for nonlinear models.
 Around 800 and 800,
lm
fails to estimate a coefficient for the quadratic term, & gives a warning about a singular design matrix.  The relationship between intercept and location is not linear anymore.
 Around 800 and 800,
In summary, if the scale of predictors is big, such as over 800, it is a good choice to center them first before performing any data analysis.
R code
1  set.seed(1) 
Thoughts and Emotions
I feel good that I finished the toy examples.
I feel bad that they were supposed to take less time.
Tomorrow’s plan
 Read online posts about normalized data
 Make descriptive data of variables left in the cleaned data
 Write report
 Run simple logistic analysis of batch effect and
Reasons
Day 5: 20180115 Monday
Today’s Progress (achievements and frustrations):
 Data analysis for over 3 hours.
Thoughts and Emotions
The most accomplished thing is to find how to reduce variables required to be created or reduce the number of global variables.
When fitting each model, I have to create intermediate variables, such as a variable fit
to keep the information of fitted models. Every time, I have to think hard to give each variable a different name, or make sure I run all code if the same variable names are reused.
The intermediate variables are all global variables. It is the user’s responsibility to delete unused variables.
It is worse that variables like i
or j
created at the beginning of the loop are not temporary variables!
Now I found a solution: use with
function.
Let me use iris
dataset as an example. By the way, iris
is a global variable name for the dataset. Very bad practice.
1  > head(iris) 
For example, I want to fit two logistic models. They include different independent variables. Model 1 uses length
and width
while Model 2 uses width
only.
1 

Two models used the same variable to fit
to save the information from glm
. fit
is a global variable. In fact, fit
is an intermediate variable and is not useful at the end of data analysis. This is also the reason to reuse it in a different model. It is a poor coding style. I have several choices.
Use a different name for
fit
for different models.Usually, I will fit a lot of models during data analysis. Keeping track of the model fit requires patience.
Set
fit
as a local variable in a functionWrite a function for each model? The same problem exists as above: keeping track of the function requires patience.
Set
fit
as a local variable withinwith
function.My choice.
Let me rewrite the code.
with(iris, {
fit < glm(Species ~ Sepal.Length + Sepal.Width,
family=binomial(link='logit'),
data = iris,
control = list(maxit = 50))
print(summary(fit))
}
)
with(iris, {
fit < glm(Species ~ Sepal.Width,
family=binomial(link='logit'),
data = iris,
control = list(maxit = 50))
print(summary(fit))
})
The fit
variable in each with
function will not affect each other. It will disappear when with
ends. Excellent!
Tomorrow’s plan
 Meeting
 Refine
Future’s plan
 Read source code of Relaxed Lasso’s R package
Day 6: 20180129 Monday
Today’s Progress (achievements and frustrations):
 data analysis
Thoughts and Emotions
I am happy that I created a website for my statistical notes. One drawback is that any knowledge whatever how small it is will take time to write.
When I copy the code from the notebook to do data analysis, I feel the time spent is worthwhile.
It is a delimma. People usually think we should write functions to reduce the reduandancy code. A lot of code in data analysis is used only once or will be different a little bit every time.
Tomorrow’s plan
 meeting
Day 7: 20180201 Thursday
Today’s Progress (achievements and frustrations):
 data analysis
Thoughts and Emotions
In recent years, I paid attention to reproducible research, i.e, reproducible report. Or it means whether the code can produce the same result after several years. The answer usually is “No”. There is only one reason: we cannot know whether the future packages/R version used now will be modified to be so differently from the current version or whether they stop maintainence and cannot be compatible with future version of other packages.
Let me simplify the problem. Assume the failure rate of each package follows binomial distribution with probability of $p$. $p$ is the same for all packages.
If I use n packages in a data analysis project, the probability of code can be reproduced in the future is $(1p)^n$. R itself can be considered as a package, too.
$(1p)^n$ becomes smaller if $n$ increases; more packages in the research decrease the possibility of reproducibility.
Tomorrow’s plan
 meeting
Day 8: 20180202 Friday
Today’s Progress (achievements and frustrations):
 data analysis
Thoughts and Emotions
To reduce the number of packages used, I plan to write some small snippets code in with()
function to create characteristic table.
Tomorrow’s plan
 Take notes
Day 9: 20180204 Sunday
Today’s Progress (achievements and frustrations):
 data analysis
Thoughts and Emotions
It is hard to get started.
I start reading the textbook Applied Linear Statistical Models 5th ed which is the book for my master program. I feel warm from the book and optimistic about the statistics study.
I used the table template from my first published paper and constructed the table on the current project. It gave me a lot of encouragement because it seemed close to publication if the same amount time was spent.
Day 10: 20180206 Monday
Today’s Progress (achievements and frustrations):
 data analysis
Thoughts and Emotions
I was absorbed in doing data analysis + making tables.
Making a satisfying table is timeconsuming. It can take as much time as creating models with R.
I am annoyed by the repeated boring work, such as making tables. The solution is to play music in the background.
Background music is a must for the mind to be absorbed in data analysis because there is a lot of tedious stuff in the process.
Tomorrow’s plan
 Copy output from RStudio to excel to make 3 tables
Day 11: 20180206 Tuesday
Today’s Progress (achievements and frustrations):
 data analysis
 Made 3 tables
 Rereead a paper from others
 Reread a paper from the cooperators
Thoughts and Emotions
I cleaned over 300 emails on medical center’s email inbox. It is boring work. As I said yesterday, the background music is a must for the scenario. Otherwise, “impatience” and intolerable emotion arose.
I was amazed that I finished the things planned yesterday. These two days, I am in the zone though depression sometimes attacked me.
Tomorrow’s plan
 Meet advisor
 Hopefully I can start write the “Method” part of the paper