iid distribution problem
Locked
12
12
 
|
In a survey social scientists used a design where 479 experts where asked to assess 16 fictional biographies (each). That is 7.664 lines /cases.
(How) .. is it possible to use the GLM although the random variables are not iid distributed (each expert assesses 16 times). Thank you.
Dr. Frank Gaeth
|
|
|
Frank,
I'm not sure what you are asking for advice on. So, 479 raters scored 16 objects. Did each rater score (a) the same 16 objects or did the raters score (b) a randomly selected set of 16 object from a pool of N objects. If (a) wouldn't this simply be a one within factor repeated measures model with 16 levels for the within factor? Or, its Mixed equivalent. Gene Maguin -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of drfg2008 Sent: Monday, August 29, 2011 8:49 AM To: [hidden email] Subject: iid distribution problem In a survey social scientists used a design where 479 experts where asked to assess 16 fictional biographies (each). That is 7.664 lines /cases. (How) .. is it possible to use the GLM although the random variables are not iid distributed (each expert assesses 16 times). Thank you. ----- Dr. Frank Gaeth FU-Berlin -- View this message in context: http://spssx-discussion.1045642.n5.nabble.com/iid-distribution-problem-tp474 6052p4746052.html Sent from the SPSSX Discussion mailing list archive at Nabble.com. ===================== 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 |
|
In reply to this post by drfg2008
What do you mean by assess?
What questions are you asking of the data? Art Kendall Social Research Consultants On 8/29/2011 8:48 AM, drfg2008 wrote: ===================== 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 REFCARDIn a survey social scientists used a design where 479 experts where asked to assess 16 fictional biographies (each). That is 7.664 lines /cases. (How) .. is it possible to use the GLM although the random variables are not iid distributed (each expert assesses 16 times). Thank you. ----- Dr. Frank Gaeth FU-Berlin -- View this message in context: http://spssx-discussion.1045642.n5.nabble.com/iid-distribution-problem-tp4746052p4746052.html Sent from the SPSSX Discussion mailing list archive at Nabble.com. ===================== 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
Art Kendall
Social Research Consultants |
|
|
In reply to this post by Maguin, Eugene
My english translation is a bit awkward, sorry for that.
Yes, 479 raters score 16 objects. Each rater score (a) the same 16 objects. I'm not sure, that's why I'm asking: one within factor repeated measures model with 16 levels for the within factor It sounds plausible.
Dr. Frank Gaeth
|
|
|
the case seems to be a bit more complicated (at least for me). Any help and/or syntax appreciated.
Problem: "479 raters score 16 objects. Each rater score (a) the same 16 objects" this is a shortened version of the data with 6 objects and 2 raters... caseNo: number of object DV: dependent variable - has to be explained (object) IV1: 1st independent variable - has to be tested (object) IV2: 2nd independent variable - has to be tested (object) IV3: 3rd independent variable - has to be tested (object) probandNo: the number of the proband (rater) prob_age: age of proband - has to be tested (rater) prob_sex: sex of proband (rater) - has to be tested (rater) probandNo DV IV1 IV2 IV3 caseNo prob_age prob_sex 1000 1 5 4 3 1 42 1 1000 3 4 1 1 2 42 1 1000 4 5 2 4 3 42 1 1000 2 8 4 5 4 42 1 1000 1 7 3 2 5 42 1 1000 1 2 2 2 6 42 1 1001 1 2 2 2 1 32 2 1001 1 7 3 2 2 32 2 1001 2 8 4 5 3 32 2 1001 4 5 2 4 4 32 2 1001 1 5 4 3 5 32 2 1001 3 4 1 1 6 32 2 The underlying question is which of the variables (object: IV1, IV2, IV3 - Rater: prob_age, prob_sex) exerts an influence on DV. Thanks
Dr. Frank Gaeth
|
Administrator
|
Without further explication of the experimental design, measurement properties of your variables etc, IMNSHO, it would be rather imprudent to further advise. Is this your data? Data someone else wants you to advise on? If the latter then you might reconsider your role and refer the query to a qualified statistical resource.
---- What are IV1, IV2, IV3? 1. "properties" of the object? 2. Various conditions in which the objects are rated? 3. What is the nature of this "rating"? If 1. then the values of these variables would be constant within each object (in your data they are not). So, how about providing a few details about the experimental design, the metrical properties of the data etc.
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
|
|
In reply to this post by drfg2008
From the Subject: line, you were worried about the 16 objects (or 6)
not being i.i.d. and normal. Or was the concern about some other distributions. - If you are not testing the 16/6, then you don't need to worry about their distributions. What this appears to be, to me, is a Repeated Measures design with varying covariates, for the 3 IVs. You can average across the objects and do a simpler analysis for testing main effects for age and sex. > Date: Wed, 14 Sep 2011 02:38:41 -0700 > From: [hidden email] > Subject: Re: iid distribution problem > To: [hidden email] > > the case seems to be a bit more complicated (at least for me). Any help > and/or syntax appreciated. > > > Problem: > > "479 raters score 16 objects. Each rater score (a) the same 16 objects" > > this is a shortened version of the data with 6 objects and 2 raters... > > *caseNo:* number of object > *DV:* dependent variable - has to be explained (object) > *IV1:* 1st independent variable - has to be tested (object) > *IV2:* 2nd independent variable - has to be tested (object) > *IV3:* 3rd independent variable - has to be tested (object) > *probandNo:* the number of the proband (rater) > *prob_age:* age of proband - has to be tested (rater) > *prob_sex:* sex of proband (rater) - has to be tested (rater) > > > probandNo DV IV1 IV2 IV3 caseNo prob_age prob_sex > 1000 1 5 4 3 1 42 1 > 1000 3 4 1 1 2 42 1 > 1000 4 5 2 4 3 42 1 > 1000 2 8 4 5 4 42 1 > 1000 1 7 3 2 5 42 1 > 1000 1 2 2 2 6 42 1 > > 1001 1 2 2 2 1 32 2 > 1001 1 7 3 2 2 32 2 > 1001 2 8 4 5 3 32 2 > 1001 4 5 2 4 4 32 2 > 1001 1 5 4 3 5 32 2 > 1001 3 4 1 1 6 32 2 > > The underlying question is which of the variables (object: IV1, IV2, IV3 - > Rater: prob_age, prob_sex) exerts an influence on DV. > |
|
|
Thank you for the informations.
Rich Ulrich-2: Actually I did average across the objects and did a simpler analysis for testing main effects for age and sex. However, the problem is to measure all variables in one model at the same time (age, sex, IV1 to IV3). Age and sex occur repeatedly (16 time in the data and 6 times in the example here, because each proband rates 16/6 times). But how to do a repeated measure design in this case. I simply don’t find a solution (if any exists) for that specific problem. Michelle: i.i.d. just means that each proband appears repeatedly in the data (6 times in the example), so the dependent variable (DV) and independent variables (IV1 to IV3) are not independently and identically distributed random variables (this would require that each line in the example data is from a different proband). David: Scale levels according to SPSS standards age: metric age: nominal DV, IV1 to IV3: metric -> age shall be considered N~distributed, also DV, IV1 to IV3. Any idea and/or syntax appreciated. Frank
Dr. Frank Gaeth
|
Automatic reply: iid distribution problem
|
|
I am on leave from the 15th September until 11th October. For a few weeks I will be on a beach in far north Queensland and l will be enjoying the sun. During this time I not be responding to emails. I will
be at a conference from the 3rd October and will have limited access to emails.
|
|
Administrator
|
In reply to this post by drfg2008
LOOK at the MIXED procedure.
I will NOT write your syntax. RTFM and decide which of the various models are applicable to your data. I hope that the presentation order of the 16 objects were randomized, otherwise ...(fill in the blank).
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
|
|
am I right? ...
MIXED DV BY prob_age prob_sex WITH IV1 IV2 IV3 /CRITERIA=CIN(95) MXITER(100) MXSTEP(5) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=| SSTYPE(3) /METHOD=REML /REPEATED=prob_age*prob_sex | SUBJECT(probandNo) COVTYPE(DIAG) /SAVE=FIXPRED /EMMEANS=TABLES(OVERALL).
Dr. Frank Gaeth
|
|
Administrator
|
NOT Highly probable !
> BY prob_age > /REPEATED=prob_age*prob_sex .... You need to hit the books! --- >
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
|
|
never used this procedure before.
Andy Fields "Discovering Statistics using SPSS": MIXED DV BY prob_sex WITH IV1 IV2 IV3 prob_age /CRITERIA=CIN(95) MXITER(100) MXSTEP(5) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=prob_sex IV1 IV2 IV3 prob_age prob_sex*IV1 prob_sex*IV2 prob_sex*IV3 prob_sex*prob_age IV1*IV2 IV1*IV3 IV1*prob_age IV2*IV3 IV2*prob_age IV3*prob_age prob_sex*IV1*IV2 prob_sex*IV1*IV3 prob_sex*IV1*prob_age prob_sex*IV2*IV3 prob_sex*IV2*prob_age prob_sex*IV3*prob_age IV1*IV2*IV3 IV1*IV2*prob_age IV1*IV3*prob_age IV2*IV3*prob_age prob_sex*IV1*IV2*IV3 prob_sex*IV1*IV2*prob_age prob_sex*IV1*IV3*prob_age prob_sex*IV2*IV3*prob_age IV1*IV2*IV3*prob_age prob_sex*IV1*IV2*IV3*prob_age | SSTYPE(3) /METHOD=REML /PRINT=DESCRIPTIVES.
Dr. Frank Gaeth
|
|
Administrator
|
Isn't this a repeated measures data set?
Don't see how that is incorporated. Are you sure all these factors should be fixed (Again we have no idea what IV1, IV2, IV3 are)? Are you sure you want to waste DF on 3,4 and 5 way interactions? Do you have any theory guiding this analysis or are you on a GIGO adventure. Seriously Frank, is some silly NDA preventing you from spilling the beans on WTF the data are all about?
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
|
|
Isn't this a repeated measures data set?
-> Isn't this rather a hierarchical model? Ok, 3,4 and 5 way interactions may be unnecessary. Frank
Dr. Frank Gaeth
|
|
Administrator
|
Hmmm Frank,
Hard to discern this as you haven't revealed diddly squat about the design/data collection methodology etc. What specific Hierarchical model are you attempting to fit (please be specific and no hand waving BS). I don't see any hierarchical model in your current flustercluck.
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
|
|
According to the researcher this design is a Factorial Survey (Hox/Kreft/Hermkens: The Analysis of Factorial Surveys, 1991) [1] also known as Vignette Analysis. The statistical methodology suggested is a multilevel / hierarchical regression model (p.495, p.499).
Frank [1] http://smr.sagepub.com/content/19/4/493 DOI: 10.1177/0049124191019004003 Sociological Methods & Research 1991 19: 493 JOOP J. HOX, ITA G. G. KREFT and PIET L. J. HERMKENS
Dr. Frank Gaeth
|
|
Administrator
|
Quote from your link:
"Factorial surveys constitute a specific technique for introducing experimental designs in sample surveys. Respondents are presented with descriptions (vignettes) of a constructed world in which important factors are built in experimentally. Using balanced designs well known from the multivariate experimental tradition, it is possible to build in a relatively large number of factors and levels. Within this context, the normal hypothesis is that responses are consistent on the individual level, but not totally idiosyncratic. In the analysis, it is important to determine the influence of both the vignette and the respondent variables. Analysis models for this type of data should reflect the fact that factorial surveys produce data pertaining to two distinct levels: the individual level and the vignette level. Such models are available and are generally known as multilevel analysis models. This article discusses the properties of factorial survey designs and some analysis models that address the multilevel aspects of the data. An example is presented using data on judgments on the fairness of incomes." So how does this relate to your question? What is manipulated in the 16 vignettes? What multilevel model are you proposing to fit to the data? Have you bothered to research HLM?
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
Administrator
|
Off-Topic
This comment is not about the statistics. Rather, it is concerned with what conclusions one can draw from studies that have people read vignettes and then indicate how they would behave or respond in that situation. Reading the quote below reminded me of B.F. Skinner's 1985 article "Cognitive Science and Behaviourism" (British J of Psychology, Vol 76, No. 3, pp. 291-301). See point 3 in the abstract: --- Abstract from Skinner (1985) --- In this paper it is argued that cognitive scientists, claiming the support of brain science and computer simulation, have revived a traditional view that behaviour is initiated by an internal, autonomous mind. In doing so, they have (1) misused the metaphor of storage and retrieval, (2) given neurology a misleading assignment, (3) frequently replaced controlled experimental conditions with mere descriptions of conditions and the assessment of behaviour with statements of expectations and intentions, (4) given feelings and states of mind the status of causes of behaviour rather than the products of the causes, and (5) failed to define many key terms in dimensions acceptable to science. --- End of abstract --- It's been a while since I read the article, but as I recall, Skinner's point was that people's expectations and intentions about how they would act in a given situation often do not match their actual behaviour. So if you want to know how people actually behave in some situation (rather than their expectations or intentions), you have to put them in the situation and observe. Suppose Milgram had asked people how much shock they would administer to someone rather than putting them in the situation and observing, for example. No doubt the results would have been dramatically different. http://psychology.about.com/od/historyofpsychology/a/milgram.htm
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
Following up on Bruce's point below on Milgram's study: Milgram
*did* ask subjects what they would do. In the original report (ref below), on page 375 in the first paragraph ("Preliminary Notions") of the Results section Milgram reports asking 14 Yale seniors (all psychology majors) what would they do in the situation that matched that of the actual experiment. Milgram wrote: | There was considerable agreement among the respondents | on the expected behavior of hypothetical subjects. All | respondent predicted that only an insignificant minority would | go through to the end of the shock series. (The estimates | ranged from 0% to 3%; i.e., the most "pessimistic" member | of the class predicted that of 100 persons, 3 would continues | to the most potent shock available on the shock generator -- 450 | volts.) The class mean was 1.2%. The question was also | posed informally to colleagues of the author, and the most | general feeling was that few if many would go beyond | designation Very Strong Shock. Milgram found, however, that in his experimental situation, 26 of 40 subjects went to the top shock (i.e., 450 volts). The other subjects went up to "intense shock" (300 volts) to "Danger: Severe Shock" (375 volts). As Bruce points out, one of the main lessons from Milgram's study is that you can ask what people would do in a particular situation but you don't really know what they would do unless they are put into that situation. Milgram's study was done at a time in the 1960s and 1970s when social psychologists like to produced counterintuitive results such as Milgram's (i.e., if an authority figure takes responsibility for a person's actions, that person will do some really extreme things), "bystander apathy" (i.e., the more people in a situation where someone needs help, the less likely that person will be helped), and the Stanford Prison Experiment (i.e., people will assume a role and take it to extremes). -Mike Palij New York University [hidden email] Reference Milgram, S. (1963). Behavioral study of obedience. Journal of Abnormal and Social Psychology, 67(4), 371-378. ----- Original Message ----- From: "Bruce Weaver" <[hidden email]> To: <[hidden email]> Sent: Saturday, September 17, 2011 8:02 PM Subject: Re: iid distribution problem > *Off-Topic* > > This comment is not about the statistics. Rather, it is concerned with what > conclusions one can draw from studies that have people read vignettes and > then indicate how they would behave or respond in that situation. Reading > the quote below reminded me of B.F. Skinner's 1985 article "Cognitive > Science and Behaviourism" (British J of Psychology, Vol 76, No. 3, pp. > 291-301). See point 3 in the abstract: > > --- Abstract from Skinner (1985) --- > In this paper it is argued that cognitive scientists, claiming the support > of brain science and computer simulation, have revived a traditional view > that behaviour is initiated by an internal, autonomous mind. In doing so, > they have (1) misused the metaphor of storage and retrieval, (2) given > neurology a misleading assignment, (3) frequently replaced controlled > experimental conditions with mere descriptions of conditions and the > assessment of behaviour with statements of expectations and intentions, (4) > given feelings and states of mind the status of causes of behaviour rather > than the products of the causes, and (5) failed to define many key terms in > dimensions acceptable to science. > --- End of abstract --- > > It's been a while since I read the article, but as I recall, Skinner's point > was that people's expectations and intentions about how they would act in a > given situation often do not match their actual behaviour. So if you want > to know how people /actually/ behave in some situation (rather than their > expectations or intentions), you have to put them in the situation and > observe. Suppose Milgram had asked people how much shock they would > administer to someone rather than putting them in the situation and > observing, for example. No doubt the results would have been dramatically > different. > > http://psychology.about.com/od/historyofpsychology/a/milgram.htm > > > > David Marso wrote: >> >> Quote from your link: >> "Factorial surveys constitute a specific technique for introducing >> experimental designs in sample surveys. Respondents are presented with >> descriptions (vignettes) of a constructed world in which important factors >> are built in experimentally. Using balanced designs well known from the >> multivariate experimental tradition, it is possible to build in a >> relatively large number of factors and levels. Within this context, the >> normal hypothesis is that responses are consistent on the individual >> level, but not totally idiosyncratic. In the analysis, it is important to >> determine the influence of both the vignette and the respondent variables. >> Analysis models for this type of data should reflect the fact that >> factorial surveys produce data pertaining to two distinct levels: the >> individual level and the vignette level. Such models are available and are >> generally known as multilevel analysis models. This article discusses the >> properties of factorial survey designs and some analysis models that >> address the multilevel aspects of the data. An example is presented using >> data on judgments on the fairness of incomes." >> >> So how does this relate to your question? >> What is manipulated in the 16 vignettes? >> What multilevel model are you proposing to fit to the data? >> Have you bothered to research HLM? >> >> >> >> drfg2008 wrote: >>> >>> According to the researcher this design is a Factorial Survey >>> (Hox/Kreft/Hermkens: The Analysis of Factorial Surveys, 1991) [1] also >>> known as Vignette Analysis. The statistical methodology suggested is a >>> multilevel / hierarchical regression model (p.495, p.499). >>> >>> Frank >>> >>> [1] >>> >>> http://smr.sagepub.com/content/19/4/493 >>> DOI: 10.1177/0049124191019004003 >>> Sociological Methods & Research 1991 19: 493 >>> JOOP J. HOX, ITA G. G. KREFT and PIET L. J. HERMKENS >>> >> > > > ----- > -- > Bruce Weaver > [hidden email] > http://sites.google.com/a/lakeheadu.ca/bweaver/ > > "When all else fails, RTFM." > > NOTE: My Hotmail account is not monitored regularly. > To send me an e-mail, please use the address shown above. > > -- > View this message in context: http://spssx-discussion.1045642.n5.nabble.com/iid-distribution-problem-tp4746052p4814999.html > Sent from the SPSSX Discussion mailing list archive at Nabble.com. > > ===================== > 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 |