Proper analysis
Locked
 
|
In a repeated measures design, they presented participants with 10 pictures, half the pictures were of a male and half female. Further, the pictures each varied in the ethnicity of the model (Caucasian, Hispanic, etc., 5 ethnicities in all). So, there was a male-caucasian pic, a female-caucasian pic, a male-hispanic pic, a female-hispanic pic, and so on. They asked participants to assign to each picture a college major from a list of 15 majors. The study was examining gender- and racial-stereotypes of people’s choice of college majors. The data set includes 119 participants, each assigning a major (nominal variable) to each of the 10 pictures. They want to know if there are differences depending on the gender and race of the pictures. Other than a pure qualitative analysis reporting the percentage people who assigned a given major to each pic, is there a statistical procedure to test for any patterns or differences from chance? If so, can someone provide a quick example of how to set up such an analysis in SPSS? Thanks in advance. I hope this made sense! Fred Fredric E. Rose, Ph.D. Associate Professor of Psychology Palomar College 760-744-1150 x2344 frose@... ====================To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
This is a really interesting question, and I certainly don't know of an
ideal solution. I'm not speaking from experience on working with this type of data, but this has captured my imagination and here are my thoughts: If it were not for the repeated measures aspect, a chi-squared goodness of fit would allow you to test deviance from a 'chance' allocation of majors (i.e. whether cell frequencies differ from 119/15=7.93. This could be done in SPSS easily. However the observations aren't independent, so this would not be valid. You could conduct multiple GoF tests for the 10 independent pictures, but this doesn't really allow you to make conclusions about effects of gender or ethnicity, because you cannot compare between the tests (also corrections for multiple comparisons would have to be made). Another way to get around the repeated measures problem is to code response patterns rather than individual repeated responses. I.e. collapse across multiple repetitions so people who chose A,A = pattern 1, A,B = pattern 2, B,A = pattern 3 and B,B = pattern 4. You could then assess goodness of fit on the patterns. However this is impossible with 10 repetitions and 15 majors as the number of possible response patterns is colossal and therefore unsuited to chi-square analysis. Even reducing the question being asked - e.g. Is there a difference between the proportions of different majors assigned to male and female caucasians? coding response pattern here would produce 15*15=225 possible response patterns: expected frequencies for a GoF test would be well below the 5 recommended for chi-square! The only way I can think of to extract usable quantitative data is to change the question asked of the dependent variable. Could the hypothesis be focused? I.e. could the majors be recoded into categories: stereotypically male vs stereotypcially female vs gender neutral for example? With 2 or 3 response options (rather than 15) the response pattern coding for repeated measures becomes more viable. Even better, could the majors be ranked in terms of 'femaleness', 'maleness', 'perceived difficulty' or 'perceived potential future earnings' or similar potential explanations of differences. Ranked data could be treated as continuous or used in non-parametric ranked methodology either of which would allow ANOVA to be conducted. I'd be interested to hear how you get around this problem, Andrew Lawrence Rose, Fred wrote: > I’ve been asked by some undergraduate students for advice on how to > analyze the following data and frankly, I’m a bit stumped. The design: > > In a repeated measures design, they presented participants with 10 > pictures, half the pictures were of a male and half female. Further, > the pictures each varied in the ethnicity of the model (Caucasian, > Hispanic, etc., 5 ethnicities in all). So, there was a male-caucasian > pic, a female-caucasian pic, a male-hispanic pic, a female-hispanic > pic, and so on. > > They asked participants to assign to each picture a college major from > a list of 15 majors. The study was examining gender- and > racial-stereotypes of people’s choice of college majors. > > The data set includes 119 participants, each assigning a major > (nominal variable) to each of the 10 pictures. They want to know if > there are differences depending on the gender and race of the pictures. > > Other than a pure qualitative analysis reporting the percentage people > who assigned a given major to each pic, is there a statistical > procedure to test for any patterns or differences from chance? If so, > can someone provide a quick example of how to set up such an analysis > in SPSS? Thanks in advance. I hope this made sense! > > Fred > > Fredric E. Rose, Ph.D. > Associate Professor of Psychology > Palomar College > 760-744-1150 x2344 > [hidden email] > =================== To manage your subscription to SPSSX-L, send a > message to [hidden email] (not to SPSSX-L), with no body > text except the command. To leave the list, send the command SIGNOFF > SPSSX-L For a list of commands to manage subscriptions, send the > command INFO REFCARD -- Andrew J. Lawrence Research Psychologist Centre for Clinical Neuroscience St George's University of London Cranmer Terrace London SW17 0RE [hidden email] ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
Yes, it is interesting.
Each participant (one of 119) will have chosen 10 majors from 15 each linked to a gender from 2 and an ethnic origin from 5. Once the major is chosen, so is the gender and ethnicity. There will be 10 possible outcomes: 1M, 1F, 2M, 2F...........5M, 5F, where 1 to 5 codes for ethnicity . Would Friedman or Quade test be appropriate, structuring the test as 2-way non-parametric, with related (rather than independent) samples ? An extra thought: are you controlling for the gender and ethnicity of the participants ? Best Wishes, Martin Holt "The study was examining gender- and> racial-stereotypes of people’s choice of college majors." ----- Original Message ----- From: "Andrew Lawrence" <[hidden email]> To: <[hidden email]> Sent: Friday, November 20, 2009 12:08 PM Subject: Re: Proper analysis > This is a really interesting question, and I certainly don't know of an > ideal solution. I'm not speaking from experience on working with this > type of data, but this has captured my imagination and here are my > thoughts: > > If it were not for the repeated measures aspect, a chi-squared goodness > of fit would allow you to test deviance from a 'chance' allocation of > majors (i.e. whether cell frequencies differ from 119/15=7.93. This > could be done in SPSS easily. However the observations aren't > independent, so this would not be valid. > > You could conduct multiple GoF tests for the 10 independent pictures, > but this doesn't really allow you to make conclusions about effects of > gender or ethnicity, because you cannot compare between the tests (also > corrections for multiple comparisons would have to be made). > > Another way to get around the repeated measures problem is to code > response patterns rather than individual repeated responses. I.e. > collapse across multiple repetitions so people who chose A,A = pattern > 1, A,B = pattern 2, B,A = pattern 3 and B,B = pattern 4. You could then > assess goodness of fit on the patterns. However this is impossible with > 10 repetitions and 15 majors as the number of possible response patterns > is colossal and therefore unsuited to chi-square analysis. Even reducing > the question being asked - e.g. Is there a difference between the > proportions of different majors assigned to male and female caucasians? > coding response pattern here would produce 15*15=225 possible response > patterns: expected frequencies for a GoF test would be well below the 5 > recommended for chi-square! > > The only way I can think of to extract usable quantitative data is to > change the question asked of the dependent variable. > > Could the hypothesis be focused? I.e. could the majors be recoded into > categories: stereotypically male vs stereotypcially female vs gender > neutral for example? With 2 or 3 response options (rather than 15) the > response pattern coding for repeated measures becomes more viable. Even > better, could the majors be ranked in terms of 'femaleness', 'maleness', > 'perceived difficulty' or 'perceived potential future earnings' or > similar potential explanations of differences. Ranked data could be > treated as continuous or used in non-parametric ranked methodology > either of which would allow ANOVA to be conducted. > > I'd be interested to hear how you get around this problem, > > Andrew Lawrence > > > > > > Rose, Fred wrote: >> I’ve been asked by some undergraduate students for advice on how to >> analyze the following data and frankly, I’m a bit stumped. The design: >> >> In a repeated measures design, they presented participants with 10 >> pictures, half the pictures were of a male and half female. Further, >> the pictures each varied in the ethnicity of the model (Caucasian, >> Hispanic, etc., 5 ethnicities in all). So, there was a male-caucasian >> pic, a female-caucasian pic, a male-hispanic pic, a female-hispanic >> pic, and so on. >> >> They asked participants to assign to each picture a college major from >> a list of 15 majors. The study was examining gender- and >> racial-stereotypes of people’s choice of college majors. >> >> The data set includes 119 participants, each assigning a major >> (nominal variable) to each of the 10 pictures. They want to know if >> there are differences depending on the gender and race of the pictures. >> >> Other than a pure qualitative analysis reporting the percentage people >> who assigned a given major to each pic, is there a statistical >> procedure to test for any patterns or differences from chance? If so, >> can someone provide a quick example of how to set up such an analysis >> in SPSS? Thanks in advance. I hope this made sense! >> >> Fred >> >> Fredric E. Rose, Ph.D. >> Associate Professor of Psychology >> Palomar College >> 760-744-1150 x2344 >> [hidden email] >> =================== To manage your subscription to SPSSX-L, send a >> message to [hidden email] (not to SPSSX-L), with no body >> text except the command. To leave the list, send the command SIGNOFF >> SPSSX-L For a list of commands to manage subscriptions, send the >> command INFO REFCARD > > -- > Andrew J. Lawrence > Research Psychologist > Centre for Clinical Neuroscience > St George's University of London > Cranmer Terrace > London > SW17 0RE > > [hidden email] > > ===================== > To manage your subscription to SPSSX-L, send a message to > [hidden email] (not to SPSSX-L), with no body text except the > command. To leave the list, send the command > SIGNOFF SPSSX-L > For a list of commands to manage subscriptions, send the command > INFO REFCARD > ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |