SPSS Statistics

 View Only
Expand all | Collapse all

Noob: Market Research Statistical Significance And SPSS

  • 1.  Noob: Market Research Statistical Significance And SPSS

    Posted Wed May 05, 2021 10:04 AM

    Hi everyone, 

    I'm a marketer who has limited SPSS and Stats experience. Unfortunately I have the curse of being the one who happens to know the most in my workplace so I am tasked with doing research and stats work. 

    I have attempted to both search the SPSS forum and Google this relatively basic question. I'm sure it's been asked! But I'm having trouble picking out which discussions are relevant as my stats vocabulary is evolved enough to be dangerous. 

    I have what I think are some relatively easy questions:

    First: I ran two quantitative surveys (Wave 1 in March, Wave 2 in April). Wave 1 n=162, Wave 2 n=174. In wave 1, 80/162, or 49.38%, had a Preference for Brand A (Preference being a dichotomous nominal variable: yes/no). In wave 2 (following a TV campaign), 105/174 or 60.34% had a Preference for Brand A.  So it was a 10.96 absolute lift and a 22.2% relative lift from Wave 1 to Wave 2. I can find online statistical significance testers (like this one) that tell me the difference from Wave 1 to Wave 2 is significant at a 95% confidence, two-sided level. These tools give me a p-value of 0.0212 and an observed power of 81.06%. This is great, but I cannot recreate this type of what these online testers are doing in SPSS. I can recreate the N frequency and Mean Scores within Crosstabs pretty easily which suggests my inputs are correct. My hypothesis is that SPSS is applying some level of processing or rounding that is causing the discrepancies. But honestly, my noob limitations are just as likely a culprit here. 

    Second: Assuming there is a way to do it in my first question (and there has to be), let's say that in addition to Preference, my Wave 1 and Wave 2 study measured 30 different attributes (all yes/no). I want to see which of these 30 attributes were different between Wave 1 and Wave 2 at a 95% 2-tailed confidence level. But I kind of want to do it at scale and quickly. Are there any shortcuts that don't involve writing syntax? What workflow would you suggest to make that process go relatively quickly? 

    Third: Anything change if it's a one-tail test? Part of the reason I run these tests is that Wave 1 takes place before TV Advertising. TV Advertising happens, then we run Wave 2. There should be an uplift. That suggests to me I should be running a one-tail test. Is that correct, and if so, what do I do differently with respect to the first two questions? 

    I created a redacted SPSS SAV file and have attached screenshots of my SPSS output as well as the AB link screenshots. 

    Thank you all for your help!

    -Neil



    ------------------------------
    Neil James
    ------------------------------

    #SPSSStatistics

    Attachment(s)

    sav
    Marketing Challenge.sav   3 KB 1 version


  • 2.  RE: Noob: Market Research Statistical Significance And SPSS

    IBM Champion
    Posted Wed May 05, 2021 12:57 PM
    CROSSTABS provides a number of test statistics for significance.  Which are you looking at?
    Also, if you have version 27.0.1, you could do an independent samples proportion test using Analyze > Compare Means > Independent Samples Proportions.  If you have 27.0.0, you can install the Fixpack to get that feature.  If your version is older, you can install the PROPOR extension command via Extensions > Extension Hub for CIs.

    If you have weights, CROSSTABS will by default round the cell counts if you are using the dialog box, but you can (and should) turn that off on the Cells subdialog.

    As for your second question, the dialog box for the proportions test accommodates as many variables as you want, and PROPOR also provides for multiple variables, albeit in a different way.

    On Q3, one would generally do a two-tailed test unless you are quite sure that you want to rule out one tail, but the Independent Samples procedure does provide both.  However, if you are looking at tests for all 30 attributes, you need to take into account problems with multiple hypothesis testing, since the probability of falsely finding a significant difference increases with the number of tests.

    There were no images or data files attached to the email.

    --



    Original Message