A python package to read and write sas (sas7bdat, sas7bcat, xport), spps (sav, zsav, por) and stata (dta) data files into/from pandas dataframes.

This module is a wrapper around the excellent Readstat C library by Evan Miller. Readstat is the library used in the back of the R library Haven, meaning pyreadstat is a python equivalent to R Haven.

Detailed documentation on all available methods is in the Module documentation

If you would like to read R RData and Rds files into python in an easy way, take a look to pyreadr, a wrapper around the C library librdata

DISCLAIMER

Pyreadstat is not a validated package. The results may have inaccuracies deriving from the fact most of the data formats are not open. Do not use it for critical tasks such as reporting to the authorities. Pyreadstat is not meant to replace the original applications in this regard.

• Motivation
• Dependencies
• Installation
  • Using pip
  • Using conda
  • From the latest sources
  • Compiling on Windows and Mac
• Usage
  • Basic Usage
    • Reading Files
    • Writing Files
  • More reading options
    • Reading only the headers
    • Reading selected columns
    • Reading files in parallel processes
    • Reading rows in chunks
    • Reading value labels
    • Missing Values
      • SPSS
      • SAS and STATA
    • Reading datetime and date columns
    • Other options
  • More writing options
    • File specific options
    • Writing value labels
    • Writing user defined missing values
    • Setting variable formats
    • Variable type conversion
• Roadmap
• CD/CI and wheels
• Known limitations
• Python 2.7 support.
• Change log
• License
• Contributing
• People

The original motivation came from reading sas7bdat files in python. That is already possible using either the (pure python) package sas7bdat or the (cythonized) method read_sas from pandas. However, those methods are slow (important if you want to read several large files), do not give the possibility to recover value labels (stored in the file itself in the case of spss or stata, or in catalog files in sas), convert both dates and datetime variables to datetime, and you have to specify the encoding otherwise in python 3 instead of strings you get bytes.

This package corrects those problems.

1. Good Performance: Here a comparison of reading a 190 Mb sas7dat file having 202 K rows by 70 columns with numeric, character and date-like columns using different methods. As you can see pyreadstat is the fastest for python and matches the speeds of R Haven.

2. Reading Value Labels Neither sas7bdat and pandas.read_sas gives the possibility to read sas7bcat catalog files. Pyreadstat can do that and also extract value labels from SPSS and STATA files.

3. Reading dates and datetimes sas7bdat and pandas.read_sas convert both date and datetime variables into datetime. That means if you have a date such a '01-01-2018' it will be transformed to '01-01-2018 00:00:00' (it always inserts a time), making it impossible to know looking only at the data if the variable was originally a datetime (if it had a time) or not. Pyreadstat transforms dates to dates and datetimes to datetimes, so that you have a better correspondence with the original data. However, it is possible to keep the original pandas behavior and get always datetimes.

4. Encoding On python 3, pandas.read_sas reads all strings as bytes. If you want strings you have to specify the encoding manually. pyreadstat read strings as str. Thas is possible because readstat extracts the original encoding and translates to utf-8, so that you don't have to care about that anymore. However it is still possible to manually set the encoding.

In addition pyreadstat exposes the variable labels in an easy way (see later). As pandas dataframes cannot handle value labels, you as user will have to take the decision whether to use those values or not. Pandas read_sas reads those labels, but in order to recover them you have to work a bit harder.

Compared to R Haven, pyreadstat offers the possibility to read only the headers: Sometimes you want to take a look to many (sas) files looking for the datasets that contain some specific columns, and you want to do it quick. This package offers the possibility to read only the metadata making it possible a very fast metadata scraping (Pandas read_sas can also do it if you pass the value iterator=True). In addition it offers the capability to read sas7bcat files separately from the sas7bdat files.

More recently there has been a lot of interest from users on using pyreadstat to read SPSS sav files. After improvements in pyreadstat 1.0.3 below some benchmarks are presented. The small file is 200K rows x 100 columns (152 Mb) containing only numeric columns and the big file is 294K rows x 666 columns (1.5 Gb). There are two versions of the big file: one containing numeric columns only and one with a mix of numeric and character. Pyreadstat gives two ways to read files: reading in a single process using read_sav and reading it in multiple processes using read_file_multiprocessing (see later in the readme for more information).

As you see performance degrades in pyreadstat when reading a table with both numeric and character types. This is because numpy and pandas do not have a native type for strings but they use a generic object type which brings a big hit in performance. The situation can be improved tough by reading files in multiple processes.

The module depends on pandas, which you normally have installed if you got Anaconda (highly recommended.)

In order to compile from source you will need a C compiler (see installation). Only if you want to do changes to the cython source code, you will need cython (normally not necessary). If you want to compile for python 2.7 or windows, you will need cython (see python 2.7 support later).

Readstat depends on the C library iconv to handle character encodings. On mac, the library is found on the system, but users have sometimes reported problems. In those cases it may help to install libiconv with conda (see later, compilation on mac). Readstat also depends on zlib; it was reported not to be installed by default on Lubuntu. If you face this problem installing the library solves it.

Probably the easiest way: from your conda, virtualenv or just base installation do:

pip install pyreadstat

If you are running on a machine without admin rights, and you want to install against your base installation you can do:

pip install pyreadstat --user

At the moment we offer pre-compiled wheels for windows, mac and linux. Look at the pypi webpage to find out which python versions are currently supported. If there is no pre-compiled wheel available, pip will attempt to compile the source code.

The package is also available in conda-forge for windows, mac and linux 64 bit. Visit the Conda forge webpage to find out which python versions are currently supported.

In order to install:

conda install -c conda-forge pyreadstat

Download or clone the repo, open a command window and type:

python3 setup.py install

If you don't have admin privileges to the machine (for example on Bee) do:

python3 setup.py install --user

You can also install from the github repo directly (without cloning). Use the flag --user if necessary.

pip install git+https://github.com/Roche/pyreadstat.git

You need a working C compiler and cython >=3.0.0.

Compiling on linux is very easy, but on windows you need some extra preparation. Some instructions are found here

Compiling on mac is usually easy. Readstat depends however on the C library iconv to handle character encodings; while on linux is part of gclib, on mac it is a separated shared library found on the system (h file is in /usr/include and shared library on /usr/lib). While compiling against this usually works fine, some users have reported problems (for example missing symbol _iconv, or libiconv version too old). In those cases it helped to install libiconv with conda:

conda install libiconv

and then recompile again (be sure to delete any cache, if using pip do pip --no-cache-dir, if using setup.py remove the folder build, otherwise you may be installing the old compilation again).

Pass the path to a file to any of the functions provided by pyreadstat. It will return a pandas data frame and a metadata object.
The dataframe uses the column names. The metadata object contains the column names, column labels, number_rows, number_columns, file label (if any), file encoding (if applicable), notes and objects about value labels (if present). Be aware that file_label and file_encoding may be None, not all columns may have labels, notes may not be present and there may be no value labels.

For example, in order to read a sas7bdat file:

import pyreadstat

df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat')

# done! let's see what we got
print(df.head())
print(meta.column_names)
print(meta.column_labels)
print(meta.column_names_to_labels)
print(meta.number_rows)
print(meta.number_columns)
print(meta.file_label)
print(meta.file_encoding)
# there are other metadata pieces extracted. See the documentation for more details.

You can replace the column names by column labels very easily (but check first that all columns have distinct labels!):

# replace column names with column labels
df.columns = meta.column_labels
# to go back to column names
df.columns = meta.column_names

Pyreadstat can write STATA (dta), SPSS (sav and zsav, por currently nor supported) and SAS (Xport, sas7bdat and sas7bcat currently not supported) files from pandas data frames.

write functions take as first argument a pandas data frame (other data structures are not supported), as a second argument the path to the destination file. Optionally you can also pass a file label and a list with column labels.

import pandas as pd
import pyreadstat

df = pd.DataFrame([[1,2.0,"a"],[3,4.0,"b"]], columns=["v1", "v2", "v3"])
# column_labels can also be a dictionary with variable name as key and label as value
column_labels = ["Variable 1", "Variable 2", "Variable 3"]
pyreadstat.write_sav(df, "path/to/destination.sav", file_label="test", column_labels=column_labels)

Some special arguments are available depending on the function. write_sav can take also notes as string, wheter to compress or not as zsav or apply row compression, variable display widths and variable measures. write_dta can take a stata version. write_xport a name for the dataset. User defined missing values and value labels are also supported. See the Module documentation for more details.

Here there is a relation of all functions available. You can also check the Module documentation.

All functions accept a keyword argument "metadataonly" which by default is False. If True, then no data will be read, but still both the metadata and the dataframe will be returned. The metadata will contain all fields as usual, but the dataframe will be emtpy, although with the correct columns names. Sometimes number_rows may be None if it was not possible to determine the number of rows without reading the data.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat', metadataonly=True)

All functions accept a keyword "usecols" which should be a list of column names. Only the columns which names match those in the list will be imported (case sensitive). This decreases memory consumption and speeds up the process. Usecols must always be a list, even if there is only one member.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat', usecols=["variable1", "variable2"])

A challenge when reading large files is the time consumed in the operation. In order to alleviate this pyreadstat provides a function "read_file_multiprocessing" to read a file in parallel processes using the python multiprocessing library. As it reads the whole file in one go you need to have enough RAM for the operation. If that is not the case look at Reading rows in chunks (next section)

Speed ups in the process will depend on a number of factors such as number of processes available, RAM, content of the file etc.

import pyreadstat

fpath = "path/to/file.sav" 
df, meta = pyreadstat.read_file_multiprocessing(pyreadstat.read_sav, fpath, num_processes=4) 

num_processes is the number of workers and it defaults to 4 (or the number of cores if less than 4). You can play with it to see where you get the best performance. You can also get the number of all available workers like this:

import multiprocessing
num_processes = multiprocessing.cpu_count()

Notes for Xport, Por and some defective SAV files not having the number of rows in the metadata

1. In all Xport, Por and some defective SAV files, the number of rows cannot be determined from the metadata. In such cases, you can use the parameter num_rows to be equal or larger to the number of rows in the dataset. This number can be obtained reading the file without multiprocessing, reading in another application, etc.

Notes for windows

1. For this to work you must include a name == "main" section in your script. See this issue for more details.

import pyreadstat

if __name__ == "__main__":
     df, meta = pyreadstat.read_file_multiprocessing(pyreadstat.read_sav, 'sample.sav')

2. If you include too many workers or you run out of RAM you main get a message about not enough page file size. See this issue

Reading large files with hundred of thouseds of rows can be challenging due to memory restrictions. In such cases, it may be helpful to read the files in chunks.

Every reading function has two arguments row_limit and row_offset that help achieving this. row_offset makes to skip a number of rows before start reading. row_limit makes to stop after a number of rows are read. Combining both you can read the file in chunks inside or outside a loop.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat("/path/to/file.sas7bdat", row_offset=1, row_limit=1)
# df will contain only the second row of the file

Pyreadstat also has a convienence function read_file_in_chunks, which returns a generator that helps you to iterate through the file in chunks. This function takes as first argument a pyreadstat reading function and a second argument a path to a file. Optionally you can change the size of the chunks with chunksize (default to 100000), and also add an offset and limit. You can use any keyword argument you wish to pass to the pyreadstat reading function.

import pyreadstat
fpath = "path/to/file.sas7bdat"
reader = pyreadstat.read_file_in_chunks(pyreadstat.read_sas7bdat, fpath, chunksize= 10, offset=2, limit=100, disable_datetime_conversion=True)

for df, meta in reader:
    print(df) # df will contain 10 rows except for the last one
    # do some cool calculations here for the chunk

For very large files it may be convienient to speed up the process by reading each chunks in parallel. For this purpose you can pass the argument multiprocess=True. This is a combination of read_file_in_chunks and read_file_multiprocessing. Here you can use the arguments row_offset and row_limit to start reading the file from an offest and stop after a row_offset+row_limit.

import pyreadstat
fpath = "path/to/file.sav"
reader = pyreadstat.read_file_in_chunks(pyreadstat.read_sav, fpath, chunksize= 10000, multiprocess=True, num_processes=4)

for df, meta in reader:
    print(df) # df will contain 10000 rows except for the last one
    # do some cool calculations here for the chunk

If using multiprocessing, please read the notes in the previous section regarding Xport, Por and some defective SAV files not having the number of rows in the metadata

For Windows, please check the notes on the previous section reading files in parallel processes

For sas7bdat files, value labels are stored in separated sas7bcat files. You can use them in combination with the sas7bdat or read them separately.

If you want to read them in combination with the sas7bdat files, pass the path to the sas7bcat files to the read_sas7bdat function. The original values will be replaced by the values in the catalog.

import pyreadstat

# formats_as_category is by default True, and it means the replaced values will be transformed to a pandas category column. There is also formats_as_ordered_category to get an ordered category, this by default is False.
df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat', catalog_file='/path/to/a/file.sas7bcat', formats_as_category=True, formats_as_ordered_category=False)

If you prefer to read the sas7bcat file separately, you can apply the formats later with the function set_catalog_to_sas. In this way you can have two copies of the dataframe, one with catalog and one without.

import pyreadstat

# this df will have the original values
df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat')
# read_sas7bdat returns an emtpy data frame and the catalog
df_empty, catalog = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bcat')
# enrich the dataframe with the catalog
# formats_as_category is by default True, and it means the replaced values will be transformed to a pandas category column. formats_as_ordered_category is by default False meaning by default categories are not ordered.
df_enriched, meta_enriched = pyreadstat.set_catalog_to_sas(df, meta, catalog, 
                             formats_as_category=True, formats_as_ordered_category=False)

For SPSS and STATA files, the value labels are included in the files. You can choose to replace the values by the labels when reading the file using the option apply_value_formats, ...

import pyreadstat

# apply_value_formats is by default False, so you have to set it to True manually if you want the labels
# formats_as_category is by default True, and it means the replaced values will be transformed to a pandas category column. formats_as_ordered_category is by default False meaning by default categories are not ordered.
df, meta = pyreadstat.read_sav("/path/to/sav/file.sav", apply_value_formats=True, 
                                formats_as_category=True, formats_as_ordered_category=False)

... or to do it later with the function set_value_labels:

import pyreadstat

# This time no value labels.
df, meta = pyreadstat.read_sav("/path/to/sav/file.sav", apply_value_formats=False)
# now let's add them to a second copy
df_enriched = pyreadstat.set_value_labels(df, meta, formats_as_category=True, formats_as_ordered_category=False)

Internally each variable is associated with a label set. This information is stored in meta.variable_to_label. Each label set contains a map of the actual value in the variable to the label, this informtion is stored in meta.variable_value_labels. By combining both you can get a dictionary of variable names to a dictionary of actual values to labels.

For SPSS and STATA:

import pyreadstat

df, meta = pyreadstat.read_sav("test_data/basic/sample.sav")
# the variables mylabl and myord are associated to the label sets labels0 and labels1 respectively
print(meta.variable_to_label)
#{'mylabl': 'labels0', 'myord': 'labels1'}

# labels0 and labels1 contain a dictionary of actual value to label
print(meta.value_labels)
#{'labels0': {1.0: 'Male', 2.0: 'Female'}, 'labels1': {1.0: 'low', 2.0: 'medium', 3.0: 'high'}}

# both things have been joined by pyreadstat for convienent use
print(meta.variable_value_labels)
#{'mylabl': {1.0: 'Male', 2.0: 'Female'}, 'myord': {1.0: 'low', 2.0: 'medium', 3.0: 'high'}}

SAS is very similar except that meta.variable_to_label comes from the sas7bdat file and meta.value_labels comes from the sas7bcat file. That means if you read a sas7bdat file and a sas7bcat file togheter meta.variable_value_labels will be filled in. If you read only the sas7bdat file only meta.variable_to_label will be available and if you read the sas7bcat file only meta.value_labels will be available. If you read a sas7bdat file and there are no associated label sets, SAS will assign by default the variable format as label sets.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat("test_data/sas_catalog/test_data_linux.sas7bdat")
meta.variable_to_label
{'SEXA': '$A', 'SEXB': '$B'}

df2, meta2 = pyreadstat.read_sas7bcat("test_data/sas_catalog/test_formats_linux.sas7bcat")
meta2.value_labels
{'$A': {'1': 'Male', '2': 'Female'}, '$B': {'2': 'Female', '1': 'Male'}}

There are two types of missing values: system and user defined. System are assigned by the program by default. User defined are valid values that the user decided to give the meaning of missing in order to differentiate between several situations.For example if one has a categorical variable representing if the person passed a test, you could have 0 for did not pass, 1 for pass, and as user defined missing variables 2 for did not show up for the test, 3 for unable to process the results, etc.

By default both cases are represented by NaN when read with pyreadstat. Notice that the only possible missing value in pandas is NaN (Not a Number) for both string and numeric variables, date, datetime and time variables have NaT (Not a Time).

In the case of SPSS sav files, the user can assign to a numeric variable either up to three discrete missing values or one range plus one discrete missing value. As mentioned by default all of these possiblities are translated into NaN, but one can get those original values by passing the argument user_missing=True to the read_sav function:

# user set with default missing values
import pyreadstat
df, meta = pyreadstat.read_sav("/path/to/file.sav")
print(df)
>> test_passed
   1
   0
   NaN
   NaN

Now, reading the user defined missing values:

# user set with user defined missing values
import pyreadstat
df, meta = pyreadstat.read_sav("/path/to/file.sav", user_missing=True)
print(df)
>> test_passed
   1
   0
   2
   3

As you see now instead o NaN the values 2 and 3 appear. In case the dataset had value labels, we could bring those in

# user set with user defined missing values and labels
import pyreadstat
df, meta = pyreadstat.read_sav("/path/to/file.sav", user_missing=True, apply_value_formats=True)
print(df)
>> test_passed
   "passed"
   "not passed"
   "not shown"
   "not processed"

Finally, the information about what values are user missing is stored in the meta object, in the variable missing_ranges. This is a dicitonary with the key being the name of the variable, and as value a list of dictionaries, each dictionary contains the elements "hi" and "lo" to represent the lower and upper bound of the range, however for discrete values as in the example, both boundaries are also present although the value is the same in both cases.

# user set with default missing values
import pyreadstat
df, meta = pyreadstat.read_sav("/path/to/file.sav", user_missing=True, apply_value_formats=True)
print(meta.missing_ranges)
>>> {'test_passed':[{'hi':2, 'lo':2}, {'hi':3, 'lo':3}]}

SPSS sav files also support up to 3 discrete user defined missing values for non numeric (character) variables. Pyreadstat is able to read those and the behavior is the same as for discrete numerical user defined missing values. That means those values will be translated as NaN by default and to the correspoding string value if user_missing is set to True. meta.missing_ranges will show the string value as well.

If the value in a character variable is an empty string (''), it will not be translated to NaN, but will stay as an empty string. This is because the empty string is a valid character value in SPSS and pyreadstat preserves that property. You can convert empty strings to nan very easily with pandas if you think it is appropiate for your dataset.

In SAS the user can assign values from .A to .Z and ._ as user defined missing values. In Stata values from .a to .z. As in SPSS, those are normally translated to NaN. However, using user_missing=True with read_sas7bdat or read_dta will produce values from A to Z and _ for SAS and a to z for dta. In addition a variable missing_user_values will appear in the metadata object, being a list with those values that are user defined missing values.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat("/path/to/file.sas7bdat", user_missing=True)

df, meta = pyreadstat.read_dta("/path/to/file.dta", user_missing=True)

print(meta.missing_user_values)

The user may also assign a label to user defined missing values. In such case passing the corresponding sas7bcat file to read_sas7bdat or using the option apply_value_formats to read_dta will show those labels instead of the user defined missing value.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat("/path/to/file.sas7bdat", catalog_file="/path/to/file.sas7bcat", user_missing=True)

df, meta = pyreadstat.read_dta("/path/to/file.dta", user_missing=True, apply_value_formats=True)

Empty strings are still transtaled as empty strings and not as NaN.

The information about what values are user missing is stored in the meta object, in the variable missing_user_values. This is a list listing all user defined missing values.

User defined missing values are currently not supported for file types other than sas7bdat, sas7bcat and dta.

SAS, SPSS and STATA represent datetime, date and other similar concepts as a numeric column and then applies a display format on top. Roughly speaking, internally there are two possible representations: one for concepts with a day or lower granularity (date, week, quarter, year, etc.) and those with a higher granularity than a day (datetime, time, hour, etc). The first group is suceptible to be converted to a python date object and the second to a python datetime object.

Pyreadstat attempts to read columns with datetime, date and time formats that are convertible to python datetime, date and time objects automatically. However there are other formats that are not fully convertible to any of these formats, for example SAS "YEAR" (displaying only the year), "MMYY" (displaying only month and year), etc. Because there are too many of these formats and these keep changing, it is not possible to implement a rule for each of those, therefore these columns are not transformed and the user will obtain a numeric column.

In order to cope with this issue, there are two options for each reader function: extra_datetime_formats and extra_date_formats that allow the user to pass these datetime or date formats, to transform the numeric values into datetime or date python objects. Then, the user can format those columns appropiately; for example extracting the year only to an integer column in the case of 'YEAR' or formatting it to a string 'YYYY-MM' in the case of 'MMYY'. The choice between datetime or date format depends on the granularity of the data as explained above.

This arguments are also useful in the case you have a valid datetime, date or time format that is currently not recognized in pyreadstat. In those cases, feel free to file an issue to ask those to be added to the list, in the meantime you can use these arguments to do the conversion.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat', extra_date_formats=["YEAR", "MMYY"])

You can set the encoding of the original file manually. The encoding must be a iconv-compatible encoding. This is absolutely necessary if you are handling old xport files with non-ascii characters. Those files do not have stamped the encoding in the file itself, therefore the encoding must be set manually.

import pyreadstat

df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat', encoding="LATIN1")

You can preserve the original pandas behavior regarding dates (meaning dates are converted to pandas datetime) with the dates_as_pandas_datetime option

import pyreadstat

df, meta = pyreadstat.read_sas7bdat('/path/to/a/file.sas7bdat', dates_as_pandas_datetime=True)

You can get a dictionary of numpy arrays instead of a pandas dataframe when reading any file format. In order to do that, set the parameter output_format='dict' (default is 'pandas'). This is useful if you want to transform the data to some other format different to pandas, as transforming the data to pandas is a costly process both in terms of speed and memory. Here for example an efficient way to transform the data to a polars dataframe:

import pyreadstat
import polars

dicdata, meta = pyreadstat.read_sav('/path/to/a/file.sav', output_format='dict')
df = polars.DataFrame(dicdata)

For more information, please check the Module documentation.

Some special arguments are available depending on the function. write_sav can take also notes as string, wheter to compress or not as zsav or apply row compression, variable display widths and variable measures. write_dta can take a stata version. write_xport a name for the dataset. See the Module documentation for more details.

The argument variable_value_labels can be passed to write_sav and write_dta to write value labels. This argument must be a dictionary where keys are variable names (names must match column names in the pandas data frame). Values are another dictionary where keys are the value present in the dataframe and values are the labels (strings).

import pandas as pd
import pyreadstat
df = pd.DataFrame([[1,1],[2,2],[1,3]], columns=['mylabl', 'myord'])
variable_value_labels = {'mylabl': {1: 'Male', 2: 'Female'}, 'myord': {1: 'low', 2: 'medium', 3: 'high'}}
path = "/path/to/somefile.sav"
pyreadstat.write_sav(df, path, variable_value_labels=variable_value_labels)

The argument missing_ranges can be passed to write_sav to write user defined missing values. This argument be a dictionary with keys as variable names matching variable names in the dataframe. The values must be a list. Each element in that list can either be either a discrete numeric or string value (max 3 per variable) or a dictionary with keys 'hi' and 'lo' to indicate the upper and lower range for numeric values (max 1 range value + 1 discrete value per variable). hi and lo may also be the same value in which case it will be interpreted as a discrete missing value. For this to be effective, values in the dataframe must be the same as reported here and not NaN.

import pandas as pd
import pyreadstat
df = pd.DataFrame([["a",1],["c",2],["c",3]], columns=['mychar', 'myord'])
missing_ranges = {'mychar':['a'], 'myord': [{'hi':2, 'lo':1}]}
path = "/path/to/somefile.sav"
pyreadstat.write_sav(df, path, missing_ranges=missing_ranges)

The argument missing_user_values can be passed to write_dta to write user defined missing values only for numeric variables. This argument be a dictionary with keys as variable names matching variable names in the dataframe. The values must be a list of missing values, valid values are single character strings between a and z. Optionally a value label can also be attached to those missing values using variable_value_labels.

import pandas as pd
import pyreadstat
df = pd.DataFrame([["a", 1],[2.2, 2],[3.3, "b"]], columns=['Var1', 'Var2'])
variable_value_labels = {'Var1':{'a':'a missing value'}
missing_ranges = {'Var1':['a'], 'Var2': ['b']}
path = "/path/to/somefile.sav"
pyreadstat.write_sav(df, path, missing_ranges=missing_ranges, variable_value_labels=variable_value_labels)

Numeric types in SPSS, SAS and STATA can have formats that affect how those values are displayed to the user in the application. Pyreadstat automatically sets the formatting in some cases, as for example when translating dates or datetimes (which in SPSS/SAS/STATA are just numbers with a special format). The user can however specify custom formats for their columns with the argument "variable_format", which is a dictionary with the column name as key and a string with the format as values:

import pandas as pd
import pyreadstat

path = "path/where/to/write/file.sav"
df = pd.DataFrame({'restricted':[1023, 10], 'integer':[1,2]})
formats = {'restricted':'N4', 'integer':'F1.0'}
pyreadstat.write_sav(df, path, variable_format=formats)

The appropiate formats to use are beyond the scope of this documentation. Probably you want to read a file produced in the original application and use meta.original_value_formats to get the formats. Otherwise look for the documentation of the original application.

In the case of SPSS we have some presets for some formats:

• restricted_integer: with leading zeros, equivalent to N + variable width (e.g N4)
• integer: Numeric with no decimal places, equivalent to F + variable width + ".0" (0 decimal positions). A pandas column of type integer will also be translated into this format automatically.

import pandas as pd
import pyreadstat

path = "path/where/to/write/file.sav"
df = pd.DataFrame({'restricted':[1023, 10], 'integer':[1,2]})
formats = {'restricted':'restricted_integer', 'integer':'integer'}
pyreadstat.write_sav(df, path, variable_format=formats)

There is some information about the possible formats here.

The following rules are used in order to convert from pandas/numpy/python types to the target file types:

Columns with mixed types are translated to character. This does not apply to column cotaining np.nan, where the missing values are correctly translated. It also does not apply to columns with user defined missing values in stata/sas where characters (a to z, A to Z, _) will be recorded as numeric.

• Include latest releases from Readstat as they come out.

A CD/CI pipeline producing the wheels is available here. Contributions are welcome.

pyreadstat builds on top of Readstat and therefore inherits its limitations. Currently those include:

• Cannot write SAS sas7bdat. Those files can be written but not read in SAS and therefore are not supported in pyreadstat. (see here)

Converting data types from foreign applications into python some times also bring some limitations:

• Pyreadstat transforms date, datetime and time like variables which are internally represented in the original application as numbers to python datetime objects. Python datetime objects are however limited in the range of dates they can represent (for example the max year is 10,000), while in other applications it is possible (although probably an error in the data) to have very high or very low dates. In this cases pyreadstat would raise an error:

OverflowError: date value out of range

The workaround is to deal with this include using the keyword argument disable_datetime_conversion so that you will get numbers instead of datetime objects or skipping reading such columns with the argument usecols.

As version 1.2.3 Python 2.7 is not supported. In previous versions it was possible to compile it for mac and linux but not for windows, but no wheels were provided. In linux and mac it will fail if the path file contains non-ascii characters.

A log with the changes for each version can be found here

pyreadstat is distributed under Apache 2.0 license. Readstat is distributed under MIT license. See the License file for more information.

Contributions are welcome! Those include corrections to the documentation, bugs reporting, testing, and of course code pull requests. For code pull requests please consider opening an issue explaining what you plan to do, so that we can get aligned before you start investing time on it (this also avoids duplicated efforts).

The ReadStat code in this repo (under the subfolder src) is coming from the main Readstat trunk and should not be modified in order to keep full compatibility with the original. In that way improvements in ReadStat can be taken here with almost no effort. If you would like to propose new features involving changes in the ReadStat code, please submit a pull request to ReadStat first.

Otto Fajardo - author, maintainer

Matthew Brett - contributor python wheels

Jonathon Love - contributor: open files with international characters. Function to open files for writing.

Clemens Brunner - integration with pandas.read_spss

Thomas Grainger - corrections and suggestions to source code

benjello, maxwell8888, drcjar, labenech: improvements to documentation

alchemyst: improvements to docstrings

bmwiedemann, toddrme2178 , Martin Thorsen Ranang: improvements to source code