Calculate abstinence using the timeline followback data in substance research.
A Python package to calculate abstinence results using the timeline followback interview data
Use the package in Python
Install the package using the pip tool. If you need instruction on how to install Python, you can find information at the Python website. The pip tool is the most common Python package management tool, and you can find information about its use instruction at the pypa website. The pip tool will be pre-installed if you download Python 3.6+ from python.org directly.
Once your computer has Python and pip installed, you can run the following command in your command line tool, which will install abstcal and its required dependencies, mainly the pandas (for data processing) and the streamlit (for the web app development).
pip install abstcal
If you're not familiar with Python coding, you can run the Jupyter Notebook included on this page on Google Colab, which is an online platform to run your Python code remotely on a server hosted by Google without any cost. With this option, you don't have to worry about installing Python and any Python packages on your local computer.
Web App Interface
If you don't want to use the non-GUI environment, you can use the package's web app, please go to abstcal Hosted by Streamlit. The web app provides the core functionalities of the package. You can find more detailed instructions on the web page.
If you're concerned about data privacy and security associated with using the web app hosted online, you can use the web app hosted locally on your computer. However, it requires the installation of Python and Python packages on your computer. Here's the overall instruction.
- Install Python 3.6+ on your computer (Official Python Downloads).
- Install abstcal on your computer (Using the command line tool, run
pip install abstcal).
- Download the entire package's zip file from the GitHub page.
- Unzip the file to the desired directory on your computer.
- Locate the web app file named calculator_web_app.py and get its full path.
- Launch the web app locally (Using your command line tool, run
streamlit run the_full_path_to_the_web_app.
Overview of the Package
This package is developed to score abstinence using the Timeline Followback (TLFB) and visit data in clinical substance use research. It provides functionalities to preprocess the datasets to remove duplicates and outliers. In addition, it can impute missing data using various criteria.
It supports the calculation of abstinence of varied definitions, including continuous, point-prevalence, and prolonged using either intent-to-treat (ITT) or responders-only assumption. It can optionally integrate biochemical verification data.
The Timeline Followback Data (Required)
The dataset should have three columns: id, date, and amount. The id column stores the subject ids, each of which should uniquely identify a study subject. The date column stores the dates when daily substance uses are collected. The date column can also use day counters using an anchor date for each subject. The amount column stores substance uses for each day.
Using the raw date
Using the day counter
The Biochemical Measures Dataset (Optional)
The dataset should have three columns: id, date, and amount. The id column stores the subject ids, each of which should uniquely identify a study subject. The date column stores the dates when daily substance uses are collected. Similar to the TLFB dataset, the biochemical measures dataset can also use day counters for the date column. The amount column stores the biochemical measures that verify substance use status.
The Visit Data (Required)
It needs to be in one of the following two formats. The long format. The dataset should have three columns: id, visit, and date. The id column stores the subject ids, each of which should uniquely identify a study subject. The visit column stores the visits. The date column stores the dates for the visits.
The wide format. The dataset should have the id column and additional columns with each representing a visit.
For both formats, the date values can be day counters, just as the TLFB dataset.
Supported Abstinence Definitions
The following abstinence definitions have both calculated as the intent-to-treat (ITT) or responders-only options. By default, the ITT option is used.
- Continuous Abstinence: No substance use in the defined time window. Usually, it starts with the target quit date.
- Point-Prevalence Abstinence: No substance use in the defined time window preceding the assessment time point.
- Prolonged Abstinence Without Lapses: No substance use after the grace period (usually 2 weeks) until the assessment time point.
- Prolonged Abstinence With Lapses: Lapses are allowed after the grace period until the assessment time point.
Once you have installed abstcal and prepared your datasets according to the format requirements listed above, you can start to use the tool.
1. Import the Package
from abstcal import TLFBData, VisitData, AbstinenceCalculator, abstcal_utils
abstcal_utils is an optional module, which provides some utility functions as
discussed in the Optional Features section.
2. Process the TLFB Data
2a. Read the TLFB data
You can either specify the full path of the TLFB data or just the filename if the dataset is in your current work directory. Supported file formats include comma-separated (.csv), tab-delimited (.txt), and Excel spreadsheets (.xls, .xlsx).
Note: If the date column uses day counters, don't forget to set
False to the
# Use the default settings tlfb_data = TLFBData('path_to_tlfb.csv') # Use additional parameters tlfb_data = TLFBData('path_to_tlfb.csv', abst_cutoff=0, included_subjects="all", use_raw_date=True) # abst_cutoff: set the custom abstinence cutoff, default=0 # included_subjects: set the list of subject ids to include in the processed data, default using all subjects # use_raw_date: the TLFB dataset uses the raw dates for the date column when True, if the date column uses day counters, set it to False
2b. Profile the TLFB data
In this step, you will see a report of the data summary, such as the number of records, the number of subjects, and any applicable abnormal data records, including duplicates and outliers. In terms of outliers, you can specify the minimal and maximal values for the substance use amounts. Those values outside the range are considered outliers and are shown in the summary report.
# No outlier identification tlfb_data.profile_data() # Identify outliers that are outside of the range tlfb_data.profile_data(0, 100) # Use the returned values of the function tlfb_summary_overall, tlfb_summary_subject, tlfb_hist_plot = tlfb_data.profile_data() # tlfb_summary_overall: the overall summary of the TLFB data # tlfb_summary_subject: the data summary by subject # tlfb_hist_plot: a histogram of the TLFB amount records # to show the histogram, you can use the utility function, which is just a convenience method to use matplotlib to show the image abstcal_utils.show_figure()
2c. Drop data records with any missing values
Those records with missing id, date, or amount will be removed. The number of removed records will be reported.
2d. Check and remove any duplicate records
Duplicate records are identified based on id and date. There are different ways to remove duplicates: min, max, or mean, which keep the minimal, maximal, or mean of the duplicate records. You can also have the options to remove all duplicates. You can also simply view the duplicates and handle these duplicates manually.
# Check only, no actions for removing duplicates tlfb_data.check_duplicates(None) # Check and remove duplicates by keeping the minimal tlfb_data.check_duplicates("min") # Check and remove duplicates by keeping the maximal tlfb_data.check_duplicates("max") # Check and remove duplicates by keeping the computed mean (all originals will be removed) tlfb_data.check_duplicates("mean") # Check and remove all duplicates tlfb_data.check_duplicates(False)
check_duplicates function will return any duplicate records.
2e. Recode outliers (optional)
Those values outside the specified range are considered outliers. All these outliers will be removed by default. However, if the users set the drop_outliers argument to be False, the values lower than the minimal will be recoded as the minimal, while the values higher than the maximal will be recoded as the maximal.
# Set the minimal and maximal values for outlier detection, by default, the outliers will be dropped tlfb_data.recode_outliers(0, 100) # Alternatively, we can recode outliers by replacing them with bounding values tlfb_data.recode_outliers(0, 100, False)
recode_outliers function returns the summary of the identified outliers.
2f. Impute the missing TLFB data
To calculate the ITT abstinence, the TLFB data will be imputed for the missing records. All contiguous missing intervals will be identified. Each of the intervals will be imputed based on the two values, the one before and the one after the interval.
You can choose to impute the missing values for the interval using the mean of these two values or interpolate the missing values for the interval using the linear values generated from the two values. Alternatively, you can specify a fixed value, which will be used to impute all missing values.
|Imputation Mode||Parameters||Imputed Values|
|Uniform||uniform||Qt = (Q0 + Q1) / 2|
|Linear||linear||Qt = m * (t - t0) + Q0 where m is (Q1 - Q0) / (t1 - t0)|
|Fixed||a numeric value||Use the numeric value to fill all missing gaps|
Note. Q0 and Q1 represent the substance use amount before (t0) and after (t1) the missing TLFB interval. Qt represents the interpolated substance use amount at the time t.
The following figure shows you some examples of these different imputation modes.
# Use the mean tlfb_data.impute_data("uniform") # Use the linear interpolation tlfb_data.impute_data("linear") # Use a fixed value, whichever is appropriate to your research question tlfb_data.impute_data(1) tlfb_data.impute_data(5) # A calling that uses all possible features tlfb_data.impute_data("linear", last_record_action="ffill", maximum_allowed_gap_days=30, biochemical_data=bio_data, overridden_amount="infer") # last_record_action: how you interpolate TLFB records using each subject's last record, default="ffill", fill forward # maximum_allowed_gap_days: the maximum allowed days for TLFB data imputation # biochemical_data: the biochemical dataset for abstinence verification (details will be provided later) # overridden_amount: with the presence of biochemical data, how false negative TLFB records will be overridden
3. Process the Visit Data
3a. Read the visit data
Similar to reading the TLFB data, you can read files in .csv, .txt, .xls, or .xlsx format. It's also supported if your visit dataset is in the univariate format, which means that each subject has only one row of data, and the columns are the visits and their dates.
Importantly, it will also detect if any subjects have their visits with the dates that are out of the order. By default, the order is inferred using the numeric or alphabetic order of the visits. These records with incorrect data may result in wrong abstinence calculations.
# Read the visit data in the long format (the default option) visit_data = VisitData("file_path.csv") # Read the visit data in the wide format visit_data = VisitData("file_path.csv", "wide") # Read the visit data and specify the order of the visit visit_data = VisitData("file_path.csv", expected_ordered_visits=[1, 2, 3, 5, 6])
Note: The name of this visit dataset is nominal. It does not only refer to actual in-person and telephone visits, it also refers to other important milestones or timepoints (e.g., Target Quit Day) in clinical cessation trials. Thus, the visit dataset should incluse all these visits that you need to calculate abstinence. Relatedly, this package has a pre-processing tool that allows you to create "virtual" visits based on existing visits. You can find the instruction on this feature at the end of this page.
If you prefer referring to the visit data as time points or milestones, you can do so by creating the visit dataset as following:
# If you prefer using time points timepoint_data = TimePointData("file_path.csv") # If you prefer using milestones milestone_data = MilestoneData("file_path.csv")
Note: If the date column uses the day counters, you'll have to set the
False, just as processing the TLFB data.
# When the dates are day counters visit_data = VisitData("file_path.csv", expected_ordered_visits=[1, 2, 3, 5, 6], use_raw_data=False)
3b. Profile the visit data
You will see a report of the data summary, such as the number of records, the number of subjects, and any applicable abnormal data records, including duplicates and outliers. In terms of outliers, you can specify the minimal and maximal values for the dates. The dates will be inferred from strings. Please use the format mm/dd/yyyy.
# No outlier identification visit_data.profile_data() # Outlier identification visit_data.profile_data("07/01/2000", "12/08/2020") # Use the returned values of the function visit_summary_overall, visit_summary_subject, visit_hist_plot = visit_data.profile_data() # visit_summary_overall: the overall summary of the TLFB data # visit_summary_subject: the data summary by subject # visit_hist_plot: a histogram of the visit records # to show the histogram, you can use the utility function, which is just a convenience method to use matplotlib to show the image abstcal_utils.show_figure()
3c. Drop data records with any missing values
Those records with missing id, visit, or date will be removed. The number of removed records will be reported.
3d. Check and remove any duplicate records
Duplicate records are identified based on id and visit. There are different ways to remove duplicates: min, max, or mean, which keep the minimal, maximal, or mean of the duplicate records. The options are the same as how you deal with duplicates in the TLFB data. Calling this function will return the duplicate records.
# Check only, no actions for removing duplicates visit_data.check_duplicates(None) # Check and remove duplicates by keeping the minimal visit_data.check_duplicates("min") # Check and remove duplicates by keeping the maximal visit_data.check_duplicates("max") # Check and remove duplicates by keeping the computed mean (all originals will be removed) visit_data.check_duplicates("mean") # Check and remove all duplicates visit_data.check_duplicates(False)
3e. Recode outliers (optional)
Those values outside the specified range are considered outliers. The syntax and usage is the same as what you deal with the TLFB dataset
# Set the minimal and maximal, and outliers will be removed by default visit_data.recode_outliers("07/01/2000", "12/08/2020") # Set the minimal and maximal, but keep the outliers by replacing them with bounding values visit_data.recode_outliers("07/01/2000", "12/08/2020", False)
3f. Impute the missing visit data
To calculate the ITT abstinence, the visit data will be imputed for the missing records. The program will first find the earliest visit date as the anchor visit, which should be non-missing for all subjects. Then it will calculate the difference in days between the later visits and the anchor visit. Based on these difference values, the following two imputation options are available. The "freq" option will use the most frequent difference value, which is the default option. The "mean" option will use the mean difference value.
|Imputation Mode||Parameters||Interpolated Values|
|Frequent||freq||Reference visit’s date + The most frequent interval|
|Mean||mean||Reference visit’s date + The mean interval|
|Dictionary||a dict object||Reference visit’s date + The specified days of interval|
Note. The reference visit is specified by the user, for which all subjects have valid dates. When it is not specified, the calculator will infer the earliest visit as the anchor visit.
The following figure illustrates the different options for imputation. For the sake of a better illustration, the tables use the wide format of the visit data. You don't need to transform you visit data, and everything will be handled under the hood for you.
# Use the most frequent difference value between the missing visit and the anchor visit visit_data.impute_data(impute="freq") # Use the mean difference value between the missing visit and the anchor visit visit_data.impute_data(impute="mean") # Specify which visit should serve as the anchor or reference visit visit_data.impute_data(anchor_visit=1)
4. Calculate Abstinence
4a. Create the abstinence calculator using the TLFB and visit data
To calculate abstinence, you instantiate the calculator by setting the TLFB and visit data. By default, only those who have both TLFB and visit data will be scored.
abst_cal = AbstinenceCalculator(tlfb_data, visit_data)
4b. Check data availability (optional)
You can find out how many subjects have the TLFB data and how many have the visit data.
check_data_availability function returns the data availablility summary.
4c. Calculate abstinence
For all the function calls to calculate abstinence, you can request the calculation to be ITT (intent-to-treat) or RO (responders-only). You can optionally specify the calculated abstinence variable names. By default, the abstinence names will be inferred. Another shared argument is whether you want to include the ending date. Notably, each method will generate the abstinence dataset and a dataset logging first lapses that make a subject nonabstinent for a particular abstinence calculation.
|shared parameter||default value||implication|
|abst_var_names||'infer'||calculated abstinence variables will have names generated automatically based on input|
|including_end||False||the time window used for abstinence calculation will not include the end visit date|
|mode||'itt'||use ITT assumption, if set as 'ro', the responders-only assumption will be used|
To calculate the continuous abstinence, you need to specify the visit when the window starts and the visit when the window ends. To provide greater flexibility, you can specify a series of visits to generate multiple time windows.
# Calculate only one window abst_df, lapse_df = abst_cal.abstinence_cont(2, 5) # Calculate two windows abst_df, lapse_df = abst_cal.abstinence_cont(2, [5, 6]) # Calculate three windows with abstinence names specified abst_df, lapse_df = abst_cal.abstinence_cont(2, [5, 6, 7], ["abst_var1", "abst_var2", "abst_var3"])
To calculate the point-prevalence abstinence, you need to specify the visits. You'll need to specify the number of days preceding the time points. To provide greater flexibility, you can specify multiple visits and multiple numbers of days.
# Calculate only one time point, 7-d point-prevalence abst_df, lapse_df = abst_cal.abstinence_pp(5, 7) # Calculate multiple time points, multiple day conditions abst_df, lapse_df = abst_cal.abstinence_pp([5, 6], [7, 14, 21, 28])
To calculate the prolonged abstinence, you need to specify the quit visit and the number of days for the grace period (the default length is 14 days). You can calculate abstinence for multiple time points. There are several options regarding how a lapse is defined. See below for some examples.
# Lapse isn't allowed abst_df, lapse_df = abst_cal.abstinence_prolonged(3, [5, 6], False) # Lapse is defined as exceeding a defined amount of substance use abst_df, lapse_df = abst_cal.abstinence_prolonged(3, [5, 6], '5 cigs') # Lapse is defined as exceeding a defined number of substance use days abst_df, lapse_df = abst_cal.abstinence_prolonged(3, [5, 6], '3 days') # Lapse is defined as exceeding a defined amount of substance use over a time window abst_df, lapse_df = abst_cal.abstinence_prolonged(3, [5, 6], '5 cigs/7 days') # Lapse is defined as exceeding a defined number of substance use days over a time window abst_df, lapse_df = abst_cal.abstinence_prolonged(3, [5, 6], '3 days/7 days') # Combination of these criteria abst_df, lapse_df = abst_cal.abstinence_prolonged(3, [5, 6], ('5 cigs', '3 days/7 days'))
4d. Responders-only abstinence calculation
By default, the calculation of the above-mentioned abstinence is based on the ITI assumption. To calculate responders-only abstinence, you need to set the mode parameter to "ro" when you call these calculation-related functions.
abst_cal.abstinence_pp(5, 7, mode="ro")
The above function call will calculate visit=5's 7-day point-prevalance abstinence with the assumption of responders-only. Under the hood, the calculator will consider abstinent only if 1) the subject had 7 TLFB data records before v5 2) the subject did not smoke at all in these 7 days. If a subject had less than 7 TLFB data records before v5, he or she is considered a non-responder, and the abstinence outcome will be N/A. If a subject had 7 TLFB data records and smoked any day, he or she is considered non-abstinent.
5. Output Datasets
5a. The abstinence datasets
To output the abstinence datasets that you have created from calling the abstinence calculation methods, you can use the following method to create a combined dataset, something like below.
# The output data will merge these individual DataFrame objects, and save it to the file that you specify. abst_cal.merge_abst_data([abst_df0, abst_df1, abst_df2], "merged_abstinence_data.csv") # Merge DataFrame objects only, no data will be saved to your computer abst_cal.merge_abst_data([abst_df0, abst_df1, abst_df2])
5b. The lapse datasets
To output the lapse datasets that you have created from calling the abstinence calculation methods, you can use the following method to create a combined dataset, something like below.
# The output data will merge these individual DataFrame objects, and save it to the file that you specify. abst_cal.merge_lapse_data([lapse_df0, lapse_df1, lapse_df2], "merged_lapse_data.csv") # Merge DataFrame objects only, no data will be saved to your computer abst_cal.merge_abst_data([abst_df0, abst_df1, abst_df2])
I. Integration of Biochemical Verification Data
If your study has collected biochemical verification data, such as carbon monoxide for smoking or breath alcohol concentration for alcohol intervention, these biochemical data can be integrated into the TLFB data. In this way, non-honest reporting can be identified (e.g., self-reported of no use, but biochemically un-verified), the self-reported value will be overridden, and the updated record will be used in later abstinence calculation.
The following code shows you a possible work flow. Please note that the biochemical measures dataset should have the
same data structure as you TLFB dataset. In other words, it should have three columns: id, date, and
amount. The biochemical data model shares the same data model with the TLFB data, both of which uses the
Note: If day counters are used for the date column, please set
True when you create the
biochemical_data variable below.
Ia. Prepare the Biochemical Dataset
A key operation to prepare the biochemical dataset is to interpolate extra meaningful records based on the exiting
records using the
interpolate_biochemical_data function, as shown below.
# First read the biochemical verification data biochemical_data = TLFBData("test_co.csv", included_subjects=included_subjects, abst_cutoff=4) biochemical_data.profile_data() # Interpolate biochemical records based on the half-life biochemical_data.interpolate_biochemical_data(half_life_in_days=0.5, maximum_days_to_interpolate=1) # half_life_in_days: the half life of the biochemical measure in days # maximum_days_to_interpolate: the maximum number of days to interpolate before the measurement day # Other data cleaning steps biochemical_data.drop_na_records() biochemical_data.check_duplicates()
Ib. Integrate the Biochemical Dataset with the TLFB data
The following code shows you how the integration can be performed. Everything else stays the same, except that in the
impute_data method, you need to specify the
tlfb_data = TLFBData("test_tlfb.csv", included_subjects=included_subjects) tlfb_sample_summary, tlfb_subject_summary, tlfb_hist_plot = tlfb_data.profile_data() tlfb_data.drop_na_records() tlfb_data.check_duplicates() tlfb_data.recode_data() tlfb_data.impute_data(biochemical_data=biochemical_data)
II. Calculate Retention Rates
You can also calculate the retention rate with the visit data with a simple function call, as shown below. If a filepath is specified, it will write to a file.
# Just show the retention rates results visit_data.get_retention_rates() # Write the retention rates to an external file visit_data.get_retention_rates('retention_rates.csv')
III. Calculate Abstinence Rates
You can calculate the computed abstinence by providing the list of pandas DataFrame objects.
# Calculate abstinence by various definitions abst_pp, lapses_pp = abst_cal.abstinence_pp([9, 10], 7, including_end=True) abst_pros, lapses_pros = abst_cal.abstinence_prolonged(4, [9, 10], '5 cigs') abst_prol, lapses_prol = abst_cal.abstinence_prolonged(4, [9, 10], False) # Calculate abstinence rates for each abst_cal.calculate_abstinence_rates([abst_pp, abst_pros, abst_prol]) abst_cal.calculate_abstinence_rates([abst_pp, abst_pros, abst_prol], 'abstinence_results.csv')
It will create the following DataFrame as the output. If a filepath is specified, it will write to a file.
|Abstinence Name||Abstinence Rate|
Data Converision Tool (wide to long format)
The package is best to work with datasets in the long format. If your datasets are in the wide format (one subject per row with columns storing data), you can use the following function.
# import the module if you've not done this yet from abstcal import abstcal_utils long_df = abstcal_utils.from_wide_to_long("filepath_to_wide.csv", data_source_type="tlfb", subject_col_name="id") # data_source_type: specify the data source is tlfb or visit, using which the function will use the desired column names after the transformation # subject_col_name: the original name for the subject column
from_wide_to_long function will return the DataFrame in the long format with correctly named columns.
Date Masking Tool
For privacy concerns, you may want to mask the dates in the datasets. To provide consistent mapping between all related datasets, you need to map TLFB, Visit, and Biochemical (optional) datasets altogether.
# Use a particular visit as reference (each subject's date for the visit will be used) abstcal_utils.mask_dates("path_to_tlfb.csv", "path_to_bio.csv", "path_to_visit.csv", 0) # Use a date (mm/dd/yyyy) as reference for all subjects abstcal_utils.mask_dates("path_to_tlfb.csv", "path_to_bio.csv", "path_to_visit.csv", "12/29/2020") # If you don't have biochemical data, please specify the second parameter as None abstcal_utils.mask_dates("path_to_tlfb.csv", None, "path_to_visit.csv", 0)
mask_dates function returns the masked datasets.
Visit Date Creation Tool
Sometimes, we need to create extra "virtual visit" dates that use existing visits plus a specific number of days' difference. This is possible with that
abstcal_utils.add_additional_visit_dates("path_to_visit.csv", [('TQD', 'v0', 7), ('v7', 'v8', -5)], use_raw_date=True)
The above example will read the long-format visit data from the specified path and add two new visit variables. The first one will be named TQD, which is equal to each subject's v0 date plus 7 days, and the other will be named v7, which is each subject's v8 date plus -5 days. The
use_raw_date parameter just specifies whether the visit data uses raw dates or day counters.
Output DataFrame to Files
Many of these data processing functions produce DataFrame objects as the return value. If you want to save these DataFrame objects to external files on your computer, use the
abstcal_utils.write_data_to_path(df, "filepath_to_output.csv", index=False) # index: when True, the output speadsheet will keep the index column, while False, it won't
Questions or Comments
If you have any questions about this package or would like to contribute to this project, please feel free to leave comments here or send me an email to firstname.lastname@example.org.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.