[FEATURE] Calculate weights model based on a sample - and apply it to all the samples

[talgalili]

Sometimes a user may want to give weights to a large number of respondents (e.g., 10 million), but computation time is too slow.
In such cases, it might be better to create the model based on a smaller sample (say, 100K), and then apply it to all samples (e.g., once this issue is done: #81 )

It would be great to have this capability in some nice flow. E.g., as an extra argument in .adjust() (or something similar)

[wesleytlee]

Alternatively, rather than sampling, we can reduce the sample size by building cells defined by weighting covariates, like in post-stratification, and assigning each cell the appropriate sample weight. For example, with a sample of size 1 million and 600k men and 400k women, convert the sample to two cells, men and women, with weight 0.6 and 0.4.

[talgalili]

For sure, but that only works with low cardinality combination of covariates (e.g. age group and gender, not when combining continues variables - although I guess everything could just be bucketed...)

…

On Mon, Nov 17, 2025, 11:09 PM wesleytlee ***@***.***> wrote: *wesleytlee* left a comment (facebookresearch/balance#155) <#155 (comment)> Alternatively, rather than sampling, we can reduce the sample size by building cells defined by weighting covariates, like in post-stratification, and assigning each cell the appropriate sample weight. For example, with a sample of size 1 million and 600k men and 400k women, convert the sample to two cells, men and women, with weight 0.6 and 0.4. — Reply to this email directly, view it on GitHub <#155 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHOJBSWIZHEK4NE5CVZIPT35I2PHAVCNFSM6AAAAACMMBVH5KVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKNBTHA2TENBZHE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[talgalili]

NOTE: this issue should probably be done AFTER issue #51 is done.