|
|
|
||||
|
|
|||||
Suppose that you are running a regression or a logistic regression, and instead of getting the results that you expect, you find that SUDAAN considers a large portion of your data to be missing. You use proc print or something to see your data set, and it looks fine to you. What has happened? You may have included in your model and on the subgroup statement dummy variables that are coded 0/1.While this is the standard way of coding dummy variables, SUDAAN considers the cases coded as 0 to be missing for variables that are listed on the subgroup statement. In other words, to SUDAAN, non-positive values in variables that are used as categorical independent variables are considered to be missing. Hence, when SUDAAN does a listwise deletion of missing data, a large portion of your cases may be deleted, possibly to the point of making the model unestimatible. (Please see pages 165-166 of the SUDAAN manual for a complete description regarding the use of the subgroup statement, including valid values for subgroups, and below for the example using the subgroup statement.) Consider the example below in which srsex is coded 1/2 and newvar1 is coded 0/1. As you can see, an error is printed in the log and the number of cases used in the analysis is about 4050 fewer than there should be (the 4050 cases that are coded 0 in the data step). You have several ways of dealing with this problem. Perhaps the easiest is to not list the 0/1 variable on the subgroup statement. In many ways the subgroup statement in SUDAAN is like the class statement in SAS. In the same way that you would not list a 0/1 variable on the class statement in SAS, you do not list a 0/1 variable on the subgroup statement in SUDAAN. Another solution is the recode the 0/1 variable to be a 1/2 variable. If you have a variable that is 0/1/2, then you need to recode it. You can do this in a data step before running the procedure.
data temp01; set temp1; newvar1 = 0; if _n_ ge 4050 then newvar1 = 1; run;
proc regress data=temp01 filetype=sas design = jackknife; weight rakedw0; jackwgts rakedw1--rakedw80 / adjjack=1; model ab1 = srsex newvar1 ; subgroup srsex newvar1; levels 2 2; run;
The following message is displayed in the log.
Opened SAS data file TEMP01 for reading. DATA WARNING: The matrix for estimable parameters is singular. The model may be overspecified. You should reduce the number of variables on the right-hand side and refit the model before attempting to draw any conclusions. DATA WARNING : Degrees of freedom for OVERALL contrast are less than maximum number of estimable parameters You may wish to rerun this job with a tolerance (TOL) of 1.000000e-007 and 1.000000e-005
The erroneous output is shown below.
Number of observations read : 55428 Weighted count: 23847415
Observations used in the analysis : 51339 Weighted count: 22067131
Denominator degrees of freedom : 80
Maximum number of estimable parameters for the model is 3
Weighted mean response is 2.504337
Multiple R-Square for the dependent variable AB1: 0.001315
Variance Estimation Method: Replicate Weight Jackknife
Working Correlations: Independent
Link Function: Identity
Response variable AB1: AB1
----------------------------------------------------------------------
Independent P-value
Variables and Beta T-Test
Effects Coeff. SE Beta T-Test B=0 B=0
----------------------------------------------------------------------
Intercept 2.54 0.01 323.02 0.0000
SRSEX
MALE -0.08 0.01 -6.56 0.0000
FEMALE 0.00 0.00 . .
NEWVAR1
1 0.00 0.00 . .
2 0.00 0.00 . .
----------------------------------------------------------------------
-------------------------------------------------------
Contrast Degrees
of P-value
Freedom Wald F Wald F
-------------------------------------------------------
OVERALL MODEL 2 75483.96 0.0000
MODEL MINUS
INTERCEPT 1 43.01 0.0000
INTERCEPT . . .
SRSEX 1 43.01 0.0000
NEWVAR1 . . .
-------------------------------------------------------
This problem is caused by the dummy variable newvar1. If you compare the number of cases used by SUDAAN for the analysis above, 51339, you will see that the 4050 cases coded as 0 in the data step above are missing. Although in the example below we have recoded the problem variable in a data step, you could also use the recode statement in SUDAAN to temporarily recode the variable. If you have many variables that need to be recoded, you may want to use an array in a data step. These options are perhaps most useful when you really want to have the dummy variable listed on the subgroup statement, such as when you are using proc crosstabs. As mentioned above, you could also list only the categorical variables coded with non-zero values on the subgroup statement.
data temp01a; set temp01; if newvar1 = 0 then newvar2 = 1; if newvar1 = 1 then newvar2 = 2; run; proc regress data=temp01a filetype=sas design = jackknife; weight rakedw0; jackwgts rakedw1--rakedw80 / adjjack=1; model ab1 = srsex newvar2 ; subgroup srsex newvar2; levels 2 2; run;
Number of observations read : 55428 Weighted count: 23847415
Observations used in the analysis : 55383 Weighted count: 23829382
Denominator degrees of freedom : 80
Maximum number of estimable parameters for the model is 3
Weighted mean response is 2.502603
Multiple R-Square for the dependent variable AB1: 0.001221
Variance Estimation Method: Replicate Weight Jackknife
Working Correlations: Independent
Link Function: Identity
Response variable AB1: AB1
----------------------------------------------------------------------
Independent P-value
Variables and Beta T-Test
Effects Coeff. SE Beta T-Test B=0 B=0
----------------------------------------------------------------------
Intercept 2.54 0.01 326.26 0.0000
SRSEX
MALE -0.07 0.01 -6.39 0.0000
FEMALE 0.00 0.00 . .
NEWVAR2
1 -0.02 0.02 -0.97 0.3343
2 0.00 0.00 . .
----------------------------------------------------------------------
-------------------------------------------------------
Contrast Degrees
of P-value
Freedom Wald F Wald F
-------------------------------------------------------
OVERALL MODEL 3 60042.62 0.0000
MODEL MINUS
INTERCEPT 2 20.71 0.0000
INTERCEPT . . .
SRSEX 1 40.85 0.0000
NEWVAR2 1 0.94 0.3343
-------------------------------------------------------
In this example we have a 0/1 variable (newvar1) and we are not using it on the subgroup statement. If you want to have the table broken out by the values of newvar1, then you need to recode it to be a 1/2 variable and include it on the subgroup statement and include the number of levels on the levels statement. Note that the recode statement in SUDAAN will create a variable with the first category coded as 0, which means that you can not use this variable on the subgroup statement.
proc descript data=temp01 filetype=sas design = jackknife; weight rakedw0; jackwgts rakedw1--rakedw80 / adjjack=1; var srsex racehpra newvar1; subgroup srsex racehpra; levels 2 2; run;
S U D A A N
Software for the Statistical Analysis of Correlated Data
Copyright Research Triangle Institute January 2003
Release 8.0.2
Number of observations read : 55428 Weighted count : 23847415
Denominator degrees of freedom : 80
Variance Estimation Method: Replicate Weight Jackknife
by: Variable, Self-reported gender.
-----------------------------------------------------------------------------------
| | |
| Variable | | Self-reported gender
| | | Total | MALE | FEMALE |
-----------------------------------------------------------------------------------
| | | | | |
| Self-reported | Sample Size | 55428 | 23002 | 32426 |
| gender | Weighted Size | 23847415.32 | 11631728.37 | 12215686.95 |
| | Total | 36063102.27 | 11631728.37 | 24431373.90 |
| | Mean | 1.51 | 1.00 | 2.00 |
| | SE Mean | 0.00 | 0.00 | 0.00 |
-----------------------------------------------------------------------------------
| | | | | |
| Race - UCLA | Sample Size | 9677 | 4084 | 5593 |
| CHPR Definition | Weighted Size | 5705917.88 | 2866894.01 | 2839023.87 |
| | Total | 5767889.98 | 2897175.85 | 2870714.13 |
| | Mean | 1.01 | 1.01 | 1.01 |
| | SE Mean | 0.00 | 0.00 | 0.00 |
-----------------------------------------------------------------------------------
| | | | | |
| NEWVAR1 | Sample Size | 55428 | 23002 | 32426 |
| | Weighted Size | 23847415.32 | 11631728.37 | 12215686.95 |
| | Total | 22084052.10 | 10772176.06 | 11311876.04 |
| | Mean | 0.93 | 0.93 | 0.93 |
| | SE Mean | 0.00 | 0.00 | 0.00 |
-----------------------------------------------------------------------------------
-----------------------------------------------------------------------------------
| | |
| Variable | | Race - UCLA CHPR Definition
| | | Total | LATINO | PACIFIC |
| | | | | ISLANDER |
-----------------------------------------------------------------------------------
| | | | | |
| Self-reported | Sample Size | 9677 | 9458 | 219 |
| gender | Weighted Size | 5705917.88 | 5643945.79 | 61972.10 |
| | Total | 8544941.75 | 8451279.40 | 93662.35 |
| | Mean | 1.50 | 1.50 | 1.51 |
| | SE Mean | 0.01 | 0.01 | 0.04 |
-----------------------------------------------------------------------------------
| | | | | |
| Race - UCLA | Sample Size | 9677 | 9458 | 219 |
| CHPR Definition | Weighted Size | 5705917.88 | 5643945.79 | 61972.10 |
| | Total | 5767889.98 | 5643945.79 | 123944.19 |
| | Mean | 1.01 | 1.00 | 2.00 |
| | SE Mean | 0.00 | 0.00 | 0.00 |
-----------------------------------------------------------------------------------
| | | | | |
| NEWVAR1 | Sample Size | 9677 | 9458 | 219 |
| | Weighted Size | 5705917.88 | 5643945.79 | 61972.10 |
| | Total | 5275702.50 | 5218781.91 | 56920.60 |
| | Mean | 0.92 | 0.92 | 0.92 |
| | SE Mean | 0.00 | 0.00 | 0.03 |
-----------------------------------------------------------------------------------
An alternative is to use the catlevel statement in proc descript. You will need to list the variable(s) on the var statement as many times as the number of levels for it that you have on the catlevel statement. The output below gives the total, percent and standard error of the percent for each of the levels of the recoded variable ab23.
proc descript data=temp1 filetype=sas design = jackknife; weight rakedw0; jackwgts rakedw1--rakedw80 / adjjack=1; recode ab23 = (50); var ab23 ab23; catlevel 0 1; run;
S U D A A N
Software for the Statistical Analysis of Correlated Data
Copyright Research Triangle Institute January 2003
Release 8.0.2
Number of observations read : 55428 Weighted count : 23847415
Denominator degrees of freedom : 80Variance Estimation Method: Replicate Weight Jackknife
by: Variable, One.
-----------------------------------------------------
| | |
| Variable | | One
| | | 1 |
-----------------------------------------------------
| | | |
| AB23: 0 - HIGH | Sample Size | 3709 |
| | Weighted Size | 1380250.55 |
| | Total | 709750.87 |
| | Percent | 51.42 |
| | SE Percent | 1.03 |
-----------------------------------------------------
| | | |
| AB23: 0 - HIGH | Sample Size | 3709 |
| | Weighted Size | 1380250.55 |
| | Total | 670499.68 |
| | Percent | 48.58 |
| | SE Percent | 1.03 |
-----------------------------------------------------
UCLA Researchers are invited to our Statistical Consulting Services
We recommend others to our list of Other Resources for Statistical Computing Help
These pages are Copyrighted (c) by UCLA Academic Technology Services