Read SPSS .sav files in Python: GSS 2024

Social Surveys

Python

Data I/O

How to open an SPSS .sav file in Python with svy, using the 2024 General Social Survey: read it straight from the zip, keep its variable and value labels, see what happens to declared missing values, pick the right weight and get design-based estimates.

Author

Mamadou S. Diallo, Ph.D.

Published

October 15, 2026

Modified

October 15, 2026

Keywords

read sav file Python, SPSS sav Python, GSS Python, General Social Survey 2024, SPSS missing values Python, SPSS value labels Python, complex survey analysis Python

An SPSS .sav file carries variable labels, value labels and declared missing values. With svy you read all three in one call and go straight to survey estimation.

The example is the 2024 cross-section of the General Social Survey (GSS): 3,986 respondents, 989 variables.

Get the file

NORC publishes each year as a zip on the GSS SPSS page:

from urllib.request import urlretrieve

urlretrieve(
    "https://gss.norc.org/content/dam/gss/get-the-data/documents/spss/2024_spss.zip",
    "data/2024_spss.zip",
)

Read it

pip install svy

svy.read_spss returns a polars DataFrame. It reads the .sav inside the zip, which also holds the codebook PDFs, so there is nothing to unpack:

import svy

gss = svy.read_spss("data/2024_spss.zip")
print(gss.shape)
gss.select("id", "age", "sex", "degree", "happy", "vstrat", "vpsu", "wtssnrps_as").head(3)

(3986, 989)

shape: (3, 8)

id	age	sex	degree	happy	vstrat	vpsu	wtssnrps_as
f64	f64	f64	f64	f64	f64	f64	f64
1.0	33.0	1.0	3.0	3.0	3401.0	1.0	1.663284
2.0	64.0	1.0	4.0	2.0	3401.0	1.0	1.506088
3.0	69.0	2.0	2.0	2.0	3401.0	1.0	1.363951

Read only what you need with columns=:

svy.read_spss("data/2024_spss.zip", columns=["id", "happy"]).shape

(3986, 2)

svy.read_sav is the same function under another name.

Keep the labels

svy.create_from_spss reads the data and the labels into a Sample:

sample = svy.create_from_spss("data/2024_spss.zip")
print(sample.meta.var_labels["happy"])
{code: text for code, text in sample.meta.value_labels["happy"].items() if not code.startswith("-")}

General happiness

{'1': 'very happy', '2': 'pretty happy', '3': 'not too happy'}

Missing values are already null

GSS codes “don’t know”, “no answer”, “skipped on web” and similar as negative numbers, and the file declares them as SPSS missing values. svy reads declared missing values as nulls, so there is nothing to recode:

sample.data["happy"].value_counts().sort("happy")

shape: (4, 2)

happy	count
f64	u32
null	32
1.0	797
2.0	2303
3.0	854

The 32 nulls are those answers. Stata and SAS files have nothing like this, so codes like these arrive as ordinary numbers that you have to clear yourself (see the Stata note). Pass drop_nulls=True when estimating to leave them out.

Set the design and the weight

The GSS is a stratified, clustered sample with variance strata and PSUs in the file:

sample = sample.set_design(
    svy.Design(stratum="vstrat", psu="vpsu", wgt="wtssnrps_as")
)
print(sample.n_strata, "strata,", sample.n_psus, "PSUs")

68 strata, 137 PSUs

The weight needs care. The 2024 file adds 677 respondents from an oversample of Hispanic, Black and Asian American, Native Hawaiian and Pacific Islander adults. The usual weight, wtssnrps, is null for them. NORC built wtssnrps_as so that all 3,986 respondents can be analysed together:

print("wtssnrps    nulls:", gss["wtssnrps"].null_count())
print("wtssnrps_as nulls:", gss["wtssnrps_as"].null_count())

wtssnrps    nulls: 677
wtssnrps_as nulls: 0

GSS weights add up to the sample size, not the population. Use them for means and proportions, not totals.

Estimates now carry design-based standard errors and the value labels:

print(sample.estimation.prop("happy", drop_nulls=True))

╭──────────────────── Estimate: PROP (TAYLOR) ─────────────────────╮
│                                                                  │
│  General happiness      est       se      lci      uci   cv (%)  │
│  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━  │
│  not too happy       0.2094   0.0094   0.1912   0.2288     4.50  │
│  pretty happy        0.5606   0.0106   0.5393   0.5817     1.90  │
│  very happy          0.2300   0.0113   0.2082   0.2533     4.91  │
│                                                                  │
╰──────────────────────────────────────────────────────────────────╯

print(sample.estimation.prop("happy", by="sex", drop_nulls=True))

╭───────────────────────────── Estimate: PROP (TAYLOR) ──────────────────────────────╮
│                                                                                    │
│  Respondents sex   General happiness      est       se      lci      uci   cv (%)  │
│  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━  │
│  female            not too happy       0.2018   0.0117   0.1795   0.2261     5.80  │
│  female            pretty happy        0.5671   0.0139   0.5392   0.5945     2.45  │
│  female            very happy          0.2311   0.0152   0.2022   0.2629     6.58  │
│  male              not too happy       0.2174   0.0152   0.1887   0.2492     6.98  │
│  male              pretty happy        0.5544   0.0164   0.5216   0.5868     2.95  │
│  male              very happy          0.2281   0.0129   0.2034   0.2549     5.67  │
│                                                                                    │
╰────────────────────────────────────────────────────────────────────────────────────╯

Other formats

The same steps work for the other two formats. Only the reader changes, along with what each format keeps:

SAS: Read NHANES DEMO_L.xpt in Python. Transport files have variable labels but no value labels.
Stata: Read Stata .dta files in Python: MEPS 2023. Value labels come through, but reserve codes are ordinary numbers.

Data: General Social Survey 2024, Release 3a, NORC at the University of Chicago.