chi-square test for r x n tables with weighted data

[Nazarius Mbona Tumwesigye]

stata can carry out a chi-square test for weighted data in a 2 x 2 table using this command svy: tab var1 var 2 , col pearson.

but it doesnt seem able to do the same for tables with more than 2 levels of outcome or more than two levels of independent variable (RxC).
any idea on how to mitigate this challenge is highly appreciated

[Nick Cox]

What's your null hypothesis? That frequencies of the outcome are just what you would expect from the joint frequencies of the predictors? That's very specific and limited. I guess most researchers would start with Poisson (loglinear) regression instead.

[Nazarius Mbona Tumwesigye]

Thanks Nick. our null hypothesis is that there is no significant difference in family planning use in three rounds of surveys (baseline midline and endline) in a community. Actually modified Poisson regression with round of survey as only independent variable and control for a clustering effect can help but i thought there could be a simple start with chi-sq test for the first table in the write up. just like there is with a two by two table. Naz

[Nazarius Mbona Tumwesigye]

just found that even with linearised poisson the comparison is between two proportions and not more.
the syntax for comparing attitude to FP use in years 2019, 2021 and 2023 is svy: glm attitudegp3 i.year, family(poisson) link(log) eform
and you get such a table like below comparing only 2021 and 2019. 2023 is left out.


Survey: Generalized linear models

Number of strata = 7 Number of obs = 3,006
Number of PSUs = 63 Population size = 3,163.3327
Design df = 56


Linearized
attitudegp3 exp(b) Std. Err. t P>t [95% Conf. Interval]

year

2021 1.302119 .172184 2.00 0.051 .9990979 1.697046

_cons .1667956 .0187672 -15.92 0.000 .1331362 .2089648

[Joseph Coveney]

stata can carry out a chi-square test for weighted data in a 2 x 2 table using this command svy: tab var1 var 2 , col pearson.

but it doesnt seem able to do the same for tables with more than 2 levels of outcome or more than two levels of independent variable (RxC)

If I'm following you correcty, then I cannot replicate your problem. The command works for me with a 4 × 3 tabulation without a hitch. See below.

.

just found that even with linearised poisson the comparison is between two proportions and not more.
the syntax for comparing attitude to FP use in years 2019, 2021 and 2023 is svy: glm attitudegp3 i.year, family(poisson) link(log) eform
. . .2023 is left out.

Ditto, no problem. See below.

.ÿ
.ÿversionÿ18.0

.ÿ
.ÿclearÿ*

.ÿ
.ÿ//ÿseedem
.ÿsetÿseedÿ502374701

.ÿ
.ÿquietlyÿsetÿobsÿ3006

.ÿgenerateÿbyteÿstratumÿ=ÿruniformint(1,ÿ7)

.ÿgenerateÿbyteÿpsuÿ=ÿruniformint(1,ÿ9)

.ÿgenerateÿdoubleÿwgtÿ=ÿruniform(1,ÿ1.1)

.ÿ
.ÿgenerateÿbyteÿyearÿ=ÿmod(_n,ÿ3)

.ÿquietlyÿrecodeÿyearÿ(0=2019)ÿ(1=2021)ÿ(2=2023)

.ÿ
.ÿgenerateÿbyteÿattitudegp3ÿ=ÿruniformint(0,ÿ3)

.ÿ
.ÿ*
.ÿ*ÿBeginÿhere
.ÿ*
.ÿquietlyÿsvysetÿpsu,ÿweight(wgt)ÿstrata(stratum)

.ÿ
.ÿsvy:ÿtabulateÿattitudegp3ÿyear,ÿcolumnÿpearson
(runningÿtabulateÿonÿestimationÿsample)

Numberÿofÿstrataÿ=ÿÿ7ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿNumberÿofÿobsÿÿÿ=ÿÿÿÿÿÿ3,006
NumberÿofÿPSUsÿÿÿ=ÿ63ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿPopulationÿsizeÿ=ÿ3,154.5628
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿDesignÿdfÿÿÿÿÿÿÿ=ÿÿÿÿÿÿÿÿÿ56

--------------------------------------
attitudegÿ|ÿÿÿÿÿÿÿÿÿÿÿÿyearÿÿÿÿÿÿÿÿÿÿÿ
p3ÿÿÿÿÿÿÿÿ|ÿÿ2019ÿÿÿ2021ÿÿÿ2023ÿÿTotal
----------+---------------------------
ÿÿÿÿÿÿÿÿ0ÿ|ÿ.2636ÿÿ.2644ÿÿ.2534ÿÿ.2605
ÿÿÿÿÿÿÿÿ1ÿ|ÿ.2429ÿÿ.2503ÿÿ.2383ÿÿ.2438
ÿÿÿÿÿÿÿÿ2ÿ|ÿ.2206ÿÿ.2151ÿÿ.2566ÿÿ.2308
ÿÿÿÿÿÿÿÿ3ÿ|ÿ.2729ÿÿ.2702ÿÿ.2517ÿÿ.2649
ÿÿÿÿÿÿÿÿÿÿ|ÿ
ÿÿÿÿTotalÿ|ÿÿÿÿÿ1ÿÿÿÿÿÿ1ÿÿÿÿÿÿ1ÿÿÿÿÿÿ1
--------------------------------------
Key:ÿColumnÿproportion

ÿÿPearson:
ÿÿÿÿUncorrectedÿÿÿchi2(6)ÿÿÿÿÿÿÿÿÿ=ÿÿÿÿ6.0128
ÿÿÿÿDesign-basedÿÿF(5.52,ÿ309.37)ÿ=ÿÿÿÿ0.9865ÿÿÿÿÿPÿ=ÿ0.4307

.ÿ
.ÿsvy:ÿglmÿattitudegp3ÿi.year,ÿfamily(poisson)ÿlink(log)ÿeform
(runningÿglmÿonÿestimationÿsample)

Survey:ÿGeneralizedÿlinearÿmodels

Numberÿofÿstrataÿ=ÿÿ7ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿNumberÿofÿobsÿÿÿ=ÿÿÿÿÿÿ3,006
NumberÿofÿPSUsÿÿÿ=ÿ63ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿPopulationÿsizeÿ=ÿ3,154.5628
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿDesignÿdfÿÿÿÿÿÿÿ=ÿÿÿÿÿÿÿÿÿ56

------------------------------------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿÿ|ÿÿÿÿÿÿÿÿÿÿÿÿÿLinearized
ÿattitudegp3ÿ|ÿÿÿÿÿexp(b)ÿÿÿstd.ÿerr.ÿÿÿÿÿÿtÿÿÿÿP>|t|ÿÿÿÿÿ[95%ÿconf.ÿinterval]
-------------+----------------------------------------------------------------
ÿÿÿÿÿÿÿÿyearÿ|
ÿÿÿÿÿÿÿ2021ÿÿ|ÿÿÿ.9922459ÿÿÿ.0337248ÿÿÿÿ-0.23ÿÿÿ0.820ÿÿÿÿÿ.9269356ÿÿÿÿ1.062158
ÿÿÿÿÿÿÿ2023ÿÿ|ÿÿÿ1.002579ÿÿÿ.0354516ÿÿÿÿÿ0.07ÿÿÿ0.942ÿÿÿÿÿ.9340181ÿÿÿÿ1.076173
ÿÿÿÿÿÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿÿÿ_consÿ|ÿÿÿ1.502739ÿÿÿ.0379495ÿÿÿÿ16.13ÿÿÿ0.000ÿÿÿÿÿ1.428608ÿÿÿÿ1.580717
------------------------------------------------------------------------------

.ÿ
.ÿexit

endÿofÿdo-file


.

Entirely possible that I'm not getting something right. But could there be something about your dataset that you're not aware of?

[Nazarius Mbona Tumwesigye]

thanks. let me check with my data set