Extra Topics¶

Base Python¶

Try / Except - Robustness¶

Errors and warnings are very common while developing code, and an important part of the learning process. In some cases, they can also be useful in designing an algorithm. For example, suppose we have a stream of user entered data that is supposed to contain the user's age in years. You might expect to get a few errors or nonsense entries.

In [1]:
user_ages=["34", "27", "54", "19", "giraffe", "15", "83", "61", "43", "91", "sixteen"]

It would be useful to convert these values to a numeric type to get the average age of our users, but we want to build something that can set non-numeric values aside. We can attempt to convert to numeric and give Python instructions for errors with a try-except statement:

In [2]:
ages = []
problems = []

for age in user_ages:
    try:
        a = int(age)
        ages.append(a)
    except:
        problems.append(age)
        
print(ages)
print(problems)

[34, 27, 54, 19, 15, 83, 61, 43, 91]
['giraffe', 'sixteen']

User-defined Functions¶

While Python (and its available packages) provide a wide variety of functions, sometimes it's useful to create your own. Python's syntax for defining a function is as follows:

def <function_name> ( <arguments> ):
    <code depending on arguments>
    return <value>

The mean function below returns the mean of a list of numbers. (Base Python does not include a function for the mean.)

In [3]:
def mean(number_list):
    s = sum(number_list)
    n = len(number_list)
    m = s/n
    return m

numbers=list(range(1, 51))
print(mean(numbers))

25.5

Pandas¶

Merging and Concatenating Datasets¶

Concatenating Datasets¶

In [4]:
import pandas as pd
import numpy as np

a0 = pd.DataFrame({"StudentID" : [1,2],
                   "GPA_change" : np.random.normal(0,1,2)}) #np.random.normal(0,1,2) pulls 2 random values from a Normal(0,1) distribution
a0
Out[4]:
StudentID GPA_change
0 1 0.466944
1 2 -2.255585
In [5]:
a1 = pd.DataFrame({"StudentID" : [3,4],
                   "GPA_change" : np.random.normal(0,1,2),
                   "Semester" : ["Spring", "Fall"]})
a1
Out[5]:
StudentID GPA_change Semester
0 3 0.585135 Spring
1 4 1.628179 Fall

With the datasets above it seems clear that these DataFramea would be best combined by stacking them on top of eachother, or appending one to the other as additional rows or observations. Panda's pd.concat lets us concatenate a list of DataFramea into a single DataFrame.

In [6]:
pd.concat([a0,a1],axis=0,ignore_index=True)
Out[6]:
StudentID GPA_change Semester
0 1 0.466944 NaN
1 2 -2.255585 NaN
2 3 0.585135 Spring
3 4 1.628179 Fall
  • axis=0 indicates that we want to add the dataframes together row-wise. What happens if we change to axis=1?
  • Pandas thinks about dimensions as rows and columns, in that order. axis=0 refers to rows, whereas axis=1 refers to columns.
  • ignore_index=True resets the DataFrame index to start at 0 and run to 3. Otherwise our row index would be 0 1 0 1, from the indices of the original two DataFrames.

Merging¶

When joining column-wise, we usually can't just concatenate our DataFrames together, instead we use certain key variables to make sure the same observations end up in the same row.

In [7]:
b0 = pd.DataFrame({"name" : ["Marcos", "Crystal"],
                   "year" : [1993,1996]})
b1 = pd.DataFrame({"name" : ["Crystal", "Marcos"],
                   "proj_num" : [6,3]})

pd.concat([b0,b1], axis=1)
Out[7]:
name year name proj_num
0 Marcos 1993 Crystal 6
1 Crystal 1996 Marcos 3

We'll use the pd.merge function to merge datasets on key column variables.

In [8]:
pd.merge(b0,b1)
Out[8]:
name year proj_num
0 Marcos 1993 3
1 Crystal 1996 6

pd.merge automatically uses all column names that appear in both datasets as keys. We can also specify key variables:

In [9]:
#pd.merge(b0,b1, on = "name")
pd.merge(b0,b1, left_on = "name", right_on = "name")
Out[9]:
name year proj_num
0 Marcos 1993 3
1 Crystal 1996 6

Pandas also includes a DataFrame method version of merge:

In [10]:
b0.merge(b1)
Out[10]:
name year proj_num
0 Marcos 1993 3
1 Crystal 1996 6

Note that there is also a join method that focuses on joining using the Pandas indices for the objects in question. It can be useful, but merge is usually more versatile.

In [11]:
b0.join(b1.set_index("name"), on="name")
Out[11]:
name year proj_num
0 Marcos 1993 3
1 Crystal 1996 6

In the examples above, the two DataFrames share the same "name" key values. However, when the values don't completely match, we can use how to choose which values get kept and which values get dropped.

In [12]:
c0 = pd.DataFrame({"name" : ["Marcos","Crystal","Devin","Lilly"],
                   "year" : [1993,1996,1985,2001]})
c1 = pd.DataFrame({"name" : ["Marcos","Crystal","Devin","Tamera"],
                   "project_num" : [6,3,9,8]})
In [13]:
pd.merge(c0,c1, how="inner") #default merge type - keeps only observations that appear in both DataFrames
Out[13]:
name year project_num
0 Marcos 1993 6
1 Crystal 1996 3
2 Devin 1985 9
In [14]:
pd.merge(c0,c1, how="left") #The "left" join keeps only the observations in the first named DataFrame
Out[14]:
name year project_num
0 Marcos 1993 6.0
1 Crystal 1996 3.0
2 Devin 1985 9.0
3 Lilly 2001 NaN
In [15]:
pd.merge(c0,c1, how="right") #The "right" join keeps only the observations in the second named DataFrame
Out[15]:
name year project_num
0 Marcos 1993.0 6
1 Crystal 1996.0 3
2 Devin 1985.0 9
3 Tamera NaN 8
In [16]:
pd.merge(c0,c1, how="outer") #The "full" or "outer" join keeps all observations
Out[16]:
name year project_num
0 Marcos 1993.0 6.0
1 Crystal 1996.0 3.0
2 Devin 1985.0 9.0
3 Lilly 2001.0 NaN
4 Tamera NaN 8.0

The "cross" option joins every key value to every other key value (a Cartesian product): every possible pair of names appears.

In [17]:
pd.merge(c0,c1, how="cross").head(10)
Out[17]:
name_x year name_y project_num
0 Marcos 1993 Marcos 6
1 Marcos 1993 Crystal 3
2 Marcos 1993 Devin 9
3 Marcos 1993 Tamera 8
4 Crystal 1996 Marcos 6
5 Crystal 1996 Crystal 3
6 Crystal 1996 Devin 9
7 Crystal 1996 Tamera 8
8 Devin 1985 Marcos 6
9 Devin 1985 Crystal 3

Reshaping Data¶

Reshaping a DataFrame can have lots of benefits across data cleanup, analysis, and communication. Here are three different ways to structure the same data.

In [18]:
raw = pd.DataFrame({"City" : ["Raleigh","Durham","Chapel Hill"],
                    "2000" : np.random.normal(0,1,3).round(2),
                    "2001" : np.random.normal(0,1,3).round(2),
                    "2002" : np.random.normal(0,1,3).round(2),
                    "2003" : np.random.normal(0,1,3).round(2)})
raw
Out[18]:
City 2000 2001 2002 2003
0 Raleigh -0.29 0.04 0.57 -0.22
1 Durham 0.85 2.31 0.88 -0.01
2 Chapel Hill 0.70 0.02 -0.58 0.03
In [19]:
longer = raw.melt(id_vars = ["City"], value_vars = ["2000","2001","2002","2003"],
        var_name = "Year")
longer
Out[19]:
City Year value
0 Raleigh 2000 -0.29
1 Durham 2000 0.85
2 Chapel Hill 2000 0.70
3 Raleigh 2001 0.04
4 Durham 2001 2.31
5 Chapel Hill 2001 0.02
6 Raleigh 2002 0.57
7 Durham 2002 0.88
8 Chapel Hill 2002 -0.58
9 Raleigh 2003 -0.22
10 Durham 2003 -0.01
11 Chapel Hill 2003 0.03
In [20]:
wide_again = longer.pivot(index = "Year", columns = "City", values = "value")  #reset_index moves Year out of the index and back into the DataFrame proper
wide_again = wide_again.reset_index(col_level=1).rename_axis(columns={"City":None})  #cleanup to remove axis names - not always necessary
In [21]:
wide_again
Out[21]:
Year Chapel Hill Durham Raleigh
0 2000 0.70 0.85 -0.29
1 2001 0.02 2.31 0.04
2 2002 -0.58 0.88 0.57
3 2003 0.03 -0.01 -0.22

"Big Data" and iteration in pandas¶

Reminder: Download CountyHealthData_2014-2015.csv

Pandas can read csvs in smaller chunks to help deal with files that are too large to be read into RAM.

In the code below, setting chunksize and iterator=True generates a flow of 1000 row chunks out of the main dataset. This isn't really necessary in our 6109 row dataset, but might be critical to working with a 61 million row dataset.

In [22]:
#Create an empty list for storing chunks.
chunk_list = []

#Read in 1000 rows at a time and store only NC rows as separate chunks in chunk_list.
for chunk in pd.read_csv("CountyHealthData_2014-2015.csv", chunksize=1000, iterator=True):
    nc_rows = chunk[chunk["State"]=="NC"]
    chunk_list.append(nc_rows)

#Combine NC chunks into single data frame and view top rows.
nc_df = pd.concat(chunk_list, ignore_index=True)
nc_df.head(3)
Out[22]:
State Region Division County FIPS GEOID SMS Region Year Premature death Poor or fair health ... Drug poisoning deaths Uninsured adults Uninsured children Health care costs Could not see doctor due to cost Other primary care providers Median household income Children eligible for free lunch Homicide rate Inadequate social support
0 NC South South Atlantic Alamance County 37001 37001 Region 20 1/1/2014 7123.0 0.192 ... 10.48 0.259 0.073 8640.0 0.167 46.0 41394 0.444 4.94 0.202
1 NC South South Atlantic Alamance County 37001 37001 Region 20 1/1/2015 7291.0 0.192 ... 12.38 0.249 0.088 9050.0 0.167 56.0 43001 0.455 4.60 NaN
2 NC South South Atlantic Alexander County 37003 37003 Region 20 1/1/2014 7974.0 0.178 ... 22.74 0.240 0.077 9316.0 0.205 30.0 39655 0.417 6.27 0.273

3 rows × 64 columns

Alternative: The csv package¶

Python also comes packaged with package for reading Comma Separated Values (csv) files. This can sometimes be easier to work with if you don't need the extra functionality of pandas or would prefer base Python objects to work wtih!

In [23]:
import csv

This package provides two major ways to read csv files:

  • csv.reader: reads the csv into a list of lists where each row is represented by a list within a master list object.
  • csv.DictReader: reads the csv into a list of dicts where each row is a dictionary with keys derived from the first row of the dataset.

The syntax for each command is similar:

In [24]:
list_of_lists = []
with open("CountyHealthData_2014-2015.csv","r") as csvfile:
    reader=csv.reader(csvfile)
    for row in reader:
        list_of_lists.append(row)
In [25]:
list_of_dicts = []
with open("CountyHealthData_2014-2015.csv","r") as csvfile:
    reader=csv.DictReader(csvfile)
    for row in reader:
        list_of_dicts.append(row)

Notice that each process reads the csv in row by row - this can be easily adapted with an if condition to filter out specific rows from a dataset that might be too large to open all at once.

Let's take a look at the differences between the output from each of these processes.

In [26]:
print(list_of_lists[1])

['AK', 'West', 'Pacific', 'Aleutians West Census Area', '2016', '02016', 'Insuff Data', '1/1/2014', '', '0.122', '2.1', '2.1', '', '0.267', '0.3', '7.002', '0.234', '0.896', '0.266', '', '290.7', '21.1', '0.355', '91', '50', '99', '', '', '', '', '0.466', '0.091', '0.087', '', '0.289', '', '322.06', '', '', '0.03', '0.221', '0.272', '0', '5547', '0.078', '1', '0.067', '', '181', '', '', '0.17', '0.075', '', '', '0.374', '0.25', '3791', '0.185', '216', '69192', '0.127', '', '0.287']
In [27]:
print(list_of_dicts[1])

{'State': 'AK', 'Region': 'West', 'Division': 'Pacific', 'County': 'Aleutians West Census Area', 'FIPS': '2016', 'GEOID': '02016', 'SMS Region': 'Insuff Data', 'Year': '1/1/2015', 'Premature death': '', 'Poor or fair health': '0.122', 'Poor physical health days': '2.1', 'Poor mental health days': '2.1', 'Low birthweight': '0.04', 'Adult smoking': '0.267', 'Adult obesity': '0.329', 'Food environment index': '6.6', 'Physical inactivity': '0.22', 'Access to exercise opportunities': '0.896', 'Excessive drinking': '0.266', 'Alcohol-impaired driving deaths': '', 'Sexually transmitted infections': '288.4', 'Teen births': '21.6', 'Uninsured': '0.293', 'Primary care physicians': '36', 'Dentists': '73', 'Mental health providers': '163', 'Preventable hospital stays': '', 'Diabetic screening': '', 'Mammography screening': '', 'High school graduation': '', 'Some college': '0.474', 'Unemployment': '0.088', 'Children in poverty': '0.076', 'Income inequality': '3.907', 'Children in single-parent households': '0.289', 'Social associations': '9.014', 'Violent crime': '317.35', 'Injury deaths': '47.2', 'Air pollution - particulate matter': '', 'Drinking water violations': '0.026', 'Severe housing problems': '0.207', 'Driving alone to work': '0.347', 'Long commute - driving alone': '0', '2011 population estimate': '5511', 'Population that is not proficient in English': '0.08', 'Population living in a rural area': '1', 'Diabetes': '0.065', 'HIV prevalence rate': '', 'Premature age-adjusted mortality': '173.7', 'Infant mortality': '', 'Child mortality': '', 'Food insecurity': '0.173', 'Limited access to healthy foods': '0.075', 'Motor vehicle crash deaths': '', 'Drug poisoning deaths': '', 'Uninsured adults': '0.314', 'Uninsured children': '0.176', 'Health care costs': '4837', 'Could not see doctor due to cost': '0.185', 'Other primary care providers': '254', 'Median household income': '74088', 'Children eligible for free lunch': '0.133', 'Homicide rate': '', 'Inadequate social support': ''}

Read more about the csv package here: https://docs.python.org/3/library/csv.html

Learn more¶

  • pandas provides a quick introduction here
  • Python Data Science Handbook provides more detail and integration with other software.
  • A full list of attributes and methods available for DataFrames is available here.