SPSS Statistics

 View Only
  • 1.  #SPSS Statistics Complex Surveys

    Posted Wed May 03, 2023 09:44 AM

    I have been using SPSS's complex surveys module to produce ratio estimates for a survey I'm involved in.

    The design is not traditional as we are sampling one randomly selected day per month, and all trips in that day are surveyed. Basically, it is a cluster design with a sample of 12. Since we are only taking one day in each month, we are not doing a stratified design. The purpose is to produce an annual estimate and not a monthly estimate. We need to allocate the sample over the months with sufficient time in between to avoid correlation between data points. It is essentially a systematic sample of 12 clusters over the year.

    The sample weight is the number of days in the month divided by one since one day in each month is randomly sampled.

    Using SPSS, I have produced almost what I need using the following syntax:

    * Analysis Preparation Wizard.

    CSPLAN ANALYSIS

      /PLAN FILE='G:\7.0 SPSS\OUTPUTS\SURVEY202209_V1.csaplan'

      /PLANVARS ANALYSISWEIGHT=STAGE1_WEIGHT

      /SRSESTIMATOR TYPE=WOR

      /PRINT PLAN

      /DESIGN CLUSTER=Date

      /ESTIMATOR TYPE=WR.

    This approach does not correct for the finite population correction. We should have a fpc since the number of days in a month is small. I couldn't figure out how to include the fpc in SPSS.

    Below is my call function to produce a ratio estimate for a data set.

    * Complex Samples Ratios.

    CSDESCRIPTIVES

      /PLAN FILE='G:\7.0 SPSS\OUTPUTS\SURVEY202209_v1.csaplan'

      /JOINTPROB FILE='FILENAME1'

      /RATIO NUMERATOR=Roundtrip DENOMINATOR=Total

      /STATISTICS SE CV COUNT POPSIZE DEFF DEFFSQRT CIN(95)

      /MISSING SCOPE=ANALYSIS CLASSMISSING=EXCLUDE.

    I am more familiar with SAS's survey means modules and compared the SPSS output to what I would use with SAS. I had a colleague who has a SAS license test the data and approach.

    If I don't include a fpc using SAS's methods, the ratio estimate and confidence intervals output by SPSS and SAS match.

    However, when I include the fpc in SAS, I have as expected a different standard error for the ratio estimate.

    These are the SAS commands:

    TITLE 'FINITE POPULATION CORRECTION';

    PROC SURVEYMEANS DATA=WORK.FILENAME1 TOTAL=MONTHDAYS;

    CLUSTER DATE;

    WEIGHT STAGE1_WEIGHT;

    RATIO ROUNDTRIP / TOTALPASS;

    RUN;

    *TITLE 'NO FINITE POPULATION CORRECTION';

    PROC SURVEYMEANS DATA=WORK.FILENAME1 ;

    CLUSTER DATE;

    WEIGHT STAGE1_WEIGHT;

    RATIO ROUNDTRIP / TOTALPASS;

    RUN;

    Figure 1: Output from SPSS using SURVEY202209_V1.csaplan

    Ratios 1

    Numerator

    Ratio Estimate

    Standard Error

    95% Confidence Interval

    Coefficient of Variation

    Design Effect

    Square Root Design Effect

    Population Size

    Unweighted Count

    Lower

    Upper

    Roundtrip

    Total

    0.548

    0.028

    0.460

    0.636

    0.051

    0.335

    0.579

    876.000

    31



    Figure 2: Output from SAS with no finite population correction.



    Figure 3: Output from SAS with finite population correction.

    My question is what do I need to do when creating my Csplan to include a fpc and match what SAS produces? I have experimented with different csplan configurations but have not been successful. I do find it difficult to follow through the SPSS Complex Survey and can't find an example that fits my sample design. Though that could be my own shortcoming.

    Kind regards,
    May







    ------------------------------
    May Raad
    ------------------------------


  • 2.  RE: #SPSS Statistics Complex Surveys

    Posted Wed May 03, 2023 10:40 AM

    I passed this to one of our statisticians, who replied, "Try changing /ESTIMATOR TYPE=WR to TYPE=WOR."



    ------------------------------
    Rick Marcantonio
    Quality Assurance
    IBM
    ------------------------------



  • 3.  RE: #SPSS Statistics Complex Surveys

    Posted Thu May 04, 2023 10:50 AM

    I don't have a way to contact you apart from this, and henceforth I will be unable to (for reasons beyond my control); however, I sent another note to a statistician and received this reply:

    "I'm actually a bit skeptical about the need for joint inclusion probabilities here, because I don't see evidence that they were used in SAS. The format of the file needed can be seen in recidivism_cs_jointprob.sav, which is from a sampling plan involving stratifying on region and sampling provinces as clusters at the first stage, with five sampled per region. The calculations of joint probabilities are shown in the algorithms for CSSELECT."

    I believe he is referring to the algorithms document, available online, for CSSELECT.




    ------------------------------
    Rick Marcantonio
    Quality Assurance
    IBM
    ------------------------------