Estimating Missing Values and NonParametric Item/Factor Analysis

[Hannah State-Davey]

Dear all,

I have a couple of questions that I am struggling to solve that I hope
someone here can provide some insight. The first reloates to the
imputation of missing values.

I have collected questionnaire data using a 7-point likert scale ranging
from strongly disagree to strongly agree. However, I also included an
additional category of 'Don't Know' to identify items respondents
consistenly could not answer. This has left me with a lot of missing data
that as far as I can tell from SPSS MVA is not MCAR. Now, since it is
Likert data, predictably it has a negative skew.  I have estimated missing
values using EM and Regression but the values imputed fall some way
outside of the range of Likert scale values that I have used (one imputed
value was over 22) and some are also negative values. Can anyone shed any
light on why it has replaced missing values with these types of figures?
Should I have transformed the data before imputation?

My next question relates to running a non-parametric item and factor
analysis. Since my data is ordinal and violates the assumption of
mulitvariate normality I know these non-parametric solutions are the way
to go. I have tried to over-ride Pearson's product-moment correlation in
FA with Spearman's using example syntax found on listserv but I am
struggling to make it work. I also hit the problem of the correlation only
allowing a maximum of 100 variables. I have 129 that standard FA can cope
with so there must someway of increasing this capacity. Does anyone know
if this is possible and how? I also don't know what to over-ride to obtain
a non-parametric item analysis if indeed it is possible.

The syntax I have tried is below. Where am I going wrong with it?

NONPAR CORR
  /VARIABLES= Item1Scor Item2Scor Item3Scor Item4Scor Item5Scor Item6Scor
  Item7Scor Item8Scor Item9Scor Item10Scor Item11Scor Item12Scor Item13Scor
  Item14Scor Item15Scor Item16Scor Item17Scor Item18Scor Item19Scor
Item20Scor
  Item21Scor Item22Scor Item23Scor Item24Scor Item25Scor Item26Scor
Item27Scor
  Item28Scor Item29Scor Item30Scor Item31Scor Item32Scor Item33Scor
Item34Scor
  Item35Scor Item36Scor Item37Scor Item38Scor Item39Scor Item40Scor
Item41Scor
  Item42Scor Item43Scor Item44Scor Item45Scor Item46Scor Item47Scor
Item48Scor
  Item49Scor Item50Scor Item51Scor Item52Scor Item53Scor Item54Scor
Item55Scor
  Item56Scor Item57Scor Item58Scor Item59Scor Item60Scor Item61Scor
Item62Scor
  Item63Scor Item64Scor Item65Scor Item66Scor Item67Scor Item68Scor
Item69Scor
  Item70Scor Item71Scor Item72Scor Item73Scor Item74Scor Item75Scor
Item76Scor
  Item77Scor Item78Scor Item79Scor Item80Scor Item81Scor Item82Scor
Item83Scor
  Item84Scor Item85Scor Item86Scor Item87Scor Item88Scor Item89Scor
Item90Scor
  Item91Scor Item92Scor Item93Scor Item94Scor Item95Scor Item96Scor
Item97Scor
  Item98Scor Item99Scor Item100Scor Item101Scor Item102Scor Item103Scor
  Item104Scor Item105Scor Item106Scor Item107Scor Item108Scor Item109Scor
  Item110scor Item111Scor Item112Scor Item113Scor Item114Scor Item115Scor
  Item116Scor Item117Scor Item118Scor Item119Scor Item120Sco Item121Sco
  Item122Scor Item123Scor Item124Scor Item125Scor Item126Scor Item127Scor
  Item128Scor Item129Scor
  /PRINT=SPEARMAN
  /MATRIX=OUT(*)
 FACTOR
  /MATRIX=IN(COR*)
  /PRINT KMO AIC EXTRACTION ROTATION
  /FORMAT SORT BLANK(0.4)
  /PLOT EIGEN
  /CRITERIA MINEIGEN(1) ITERATE(25)
  /EXTRACTION PC
  /CRITERIA ITERATE(25)
  /ROTATION VARIMAX
  /METHOD=CORRELATION


Many thanks for your help

Hannah

=====================
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

[Clive Downs]

Hi,

May I suggest that you could consider the sample size and perhaps use the
parametric Pearson correlation (even if the distribution is non-normal), if
it is a large sample, as the technique may be quite robust in these
circumstances. You could run both, and compare the results, to see how much
difference it makes, on some of the pairs of scales.

Also, is it perhaps the case that a <don't know> response is not a missing
value? A DK may indicate that the item is difficult for the respondent to
understand. (It may depend what stage of questionnaire development you are
at).

I was interested in your observation about Likert data having negative
skew - do you think it is always the case, or are there particular types of
content that produce such an effect?

Regards,

Clive.

=====================
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