0
$\begingroup$

When a data set contains a fraction of missing values - Which strategy should be chosen:

  1. first impute data before providing population discriptives, or
  2. give the population insights and cross-tabs without any imputation, and in a subsequent stage use Multiple imputation and Full information Maximum Likelihood, while preparing the model for final analysis.

Thank You.

$\endgroup$
3
  • 2
    $\begingroup$ In my opinion, the controlling most important point is WHY is there missing data? The reason may be so consequential that any subsequent analysis could suggest completely misleading inferences. For example, bias sample selection in the case of heart surgery where the doctor offers the surgery option only to those who will likely survive the operation, based on his experience (many still don't). Subsequent comparisons, on say longevity, to the non-selected candidate population with heart disease is likely distorted. $\endgroup$
    – AJKOER
    Commented Jul 31, 2020 at 20:08
  • $\begingroup$ Is this a question or a statement? $\endgroup$
    – astel
    Commented Aug 1, 2020 at 15:37
  • $\begingroup$ This is a question. i wanted to ask about the things that i can use out of the above two. Thank You $\endgroup$ Commented Aug 2, 2020 at 8:16

1 Answer 1

2
$\begingroup$

UPDATE to the revised question. Your second option, definitely. I'm not commenting on the methods you mention in the second option, as I don't know anything about the data.



Descriptive statistics is reporting the characteristics of a data set, the way the data set is without modification. Imputation comes after that.

You and other researchers need a clear baseline of the data. You may try different imputation methods and other transformations, and could make other choices later on based on new ideas, or new information. That's where the descriptive report comes in. Also, other people can use that report when they are reviewing your transformations and whatever statistical tests you run on your transformed data.

$\endgroup$
4
  • $\begingroup$ So according to you we should present basic frequencies, means and other things before imputing anything and using FIML and then go with these techniques while preparing the model. Then what is recommended when we go for hypothesis testing like comparison of means between groups of Chi-Square test - imputations first to not. Thank You $\endgroup$ Commented Aug 2, 2020 at 8:23
  • $\begingroup$ Not according to me. But yes. Descriptive statistics of a population need to not be distorted by transformations. You can also provide descriptive statistics of characteristics of your transformed data set. $\endgroup$
    – user255758
    Commented Aug 2, 2020 at 14:59
  • $\begingroup$ Don't you think that after imputation cross tabs and means of variables can get change. Thank You for your previous response. $\endgroup$ Commented Aug 3, 2020 at 18:41
  • $\begingroup$ Yes. You can describe that with descriptive statistics. But initially, a descriptive statics report needs to be on the raw data. $\endgroup$
    – user255758
    Commented Aug 6, 2020 at 17:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.