Chi-squared test for trend using nominal, weighted data

[thadb]

I find several examples of how to code for chi-squared test for trend using nominal/ordinal data for data that is not weighted. However, the data I use are weighted and using the syntax "trend" after the tables statement produces an error with only the following options. Anyone aware of how to perform the same test using "proc surveyfreq"?

711 proc surveyfreq data=y15;

712 table bmi*q69 bmi*intentions intentions*exercise/chisq row cl trend;

-----

22

202

ERROR 22-322: Syntax error, expecting one of the following: ;, (, AGREE, ALPHA, CELLCHI2, CHISQ,

CHISQ1, CL, CLWT, COL, COLUMN, COV, COVP, CV, CVWT, DDF, DEFF, DEFFCOL, DEFFROW,

DEVIATION, DF, EXPECTED, KAPPA, LRCHISQ, LRCHISQ1, NOCELLPERCENT, NOFREQ,

NOPERCENT, NOPRINT, NOSPARSE, NOSTD, NOTOTAL, NOWT, ODDSRATIO, OR, PEARSONRES,

PEARSONRESID, PLOTS, RELRISK, RISK, RISK1, RISK2, RISKDIFF, RISKDIFF1, RISKDIFF2,

ROW, RSCHISQ, RSCHISQ1, TABLE, TESTP, VAR, VARWT, WALD, WCHISQ, WLLCHISQ, WT,

WTFREQ, WTKAP, WTKAPPA.

ERROR 202-322: The option or parameter is not recognized and will be ignored.

713 weight weight;

714 cluster psu;

715 strata stratum;

716 run;

[ballardw]

I don't believe that TREND is a valid Tables option in Surveyfreq.

[thadb]

Very true, that's the problem.  I am using one nominal variable and one ordinal variable with three levels, not sure else how to test it.

[ballardw]

You might describe your data a bit such as which variables are nominal and which ordinal, how many levels of each and what kind of "trend" you are looking to identify.

You may want to think more of "association" if your number of levels for each is small.

[thadb]

One variable is "intention" 1. to gain weight, 2. to lose weight, and 3. do nothing.  The second variable is how many days of the week they engage in physical activity 0, 1, 2, 3, 4, 5, 6, 7.  Basically looking to see if there is an increase days of activity when moving from one intention to another.

[ballardw]

Consider surveymeans with your categorical variable sich as intention as a Domain and your ordinal as a VAR variable, means and confidence limits should get you started.

[thadb]

Thank you.  I ran proc surveymeans and the mean value for each level was the same as the row percent for each level when I ran proc surveyfreq, which I guess is no surprise.  My only question is how to interpret the mean values as a test for trend?

Thanks for your input

[ballardw]

---

@thadb wrote:

Thank you.  I ran proc surveymeans and the mean value for each level was the same as the row percent for each level when I ran proc surveyfreq, which I guess is no surprise.  My only question is how to interpret the mean values as a test for trend?

Thanks for your input

---

Not a trend. But if the mean for physical activity days in one domain such as "gain weight" is smaller than for "lose weight" you have an indication. In lieu of a formal means test if the confidence limits don't overlap you have a very conservative test that the difference is statistically significant.

I am also not sure that 1. to gain weight, 2. to lose weight, and 3. do nothing is actually even ordinal so "trend" across these wouldn't make sense (at least to me as the implied coding goes pehaps if the coding were 3. to gain weight, 1 to lose weight, and 2. do nothing).

One minor problem I see with your specific domain "intention" is that both gain and lose weight could have the same  change in days of activity but the type of activity could vary based on the intent. Gain weight may be lifting weights with heavy weights and lower reps while lose weight may be more aerobic exercise.

[thadb]

Agreed, thank you again for your input.