I'm modelling inequality and my outcome variable is the 'gini' with 503 observations. However, most of the explanatory variables having missing values. Please how can I go about this? I have read through the discussions on 'missing data' and still a long way from knowing how to treat this. I have attached the Stata output for the missing data summary. Any ideas...kindly assist.
before you estimate the Gini you have to impute the missing values (not ignore them).
single imputation: if you are only interested in the point estimates of the Gini
multiple imputation: if you are interested in the variance of the Gini as well
I'm sure that in Stata there are many functions available to impute the data. Otherwise switch to R, and use e.g. the VIM package.
before you estimate the Gini you have to impute the missing values (not ignore them).
single imputation: if you are only interested in the point estimates of the Gini
multiple imputation: if you are interested in the variance of the Gini as well
I'm sure that in Stata there are many functions available to impute the data. Otherwise switch to R, and use e.g. the VIM package.
I have an approach to impute missing values. I wrote some code in R to do this by learning the SVD of a sparse matrix and using that to impute the missing information. It's discussed in my blog where I remove a significant portion of the pixels in a photograph and estimate them using only the information left behind. Let me know if you're interested.
http://beelzebubjones1.blogspot.com.au/2014/07/a-sprinkling-of-pixel-dust.html
I'm overwhelmed with your contributions...thank you al. The only issue is that I'm permitted to use Stata for my analysis so I'm technically stuck with the software. Kindly assist if you have codes on 'multiple imputations' and how to 'count' countries within a 'region' in a panel data structure (for descriptive statistics) using Stata.....thanks a million!
Imputation will enable you to keep all cases, and will also introduce some possible compromises in how natural variation is represented in key variables where values are missing, and also in how those variables relate to others in your analyses.
After examining your data set, and the six or so variables where values are missing in large part, I would take those six variables, and one at a time, model predictors, using regression of the other variables in the data set to see how they predict the cases where the values are not missing, and use those models to impute the missing value for the studied variable. This uses the available variables and non-missing data to empirically construct the values for the missing values.
The simplest alternative is to take the "normative" value based on the distribution of non-missing values for the six or so problematic variables, and substitute the missing values with the value representing the middle of the distribution for the remaining ones.
You could do both and compare results of both imputation methods as part of your analyses. Best of luck.
Edwin and Dasapta.....I am immensely grateful!!! and I'll be doing MI for those missing values...I'm also going through the attached tutorials on MI....thanks again for this assistance.
You can try to calculate the Gini without the variables with missing data and see if it varies a lot (a kind of sensitivity analysis), or you can simply try to calculate a non parametric correlation matrix with Spearman and see if the missing variables are correlated with the ones without missing...
In addition to conducting multiple imputation, I highly suggest that you conduct various bias analyses to quantify systematic errors that can effect your research. I am assuming that the high degree of missingness in some of your variables is likely not at random. This level of missingness in your variables of interest can greatly affect your results even after conduct imputation technique--especially if they are important confounders. See the book below of a good coverage on bias analysis topic:
Applying Quantitative Bias Analysis to Epidemiologic Data
Authors: Lash, Timothy L., Fox, Matthew P., Fink, Aliza K.
I realize this is beyond the question asked, however, I thought it is an important consideration. Hope this helps,
Regards,
Ahmed
Stata is built to handle missing data during analysis but your results will definitely be affected. If there is no way of restoring the missed data, then the options available are li mm item basses on the type of analysis you seek to perform. I would suggest you perform the analysis for the completed data and then analyzed the available variables for the missing data and see if there are significant differences in the output. Alternatively you can use the option for missing data handling in Stata version 11 or 12. All the best.
....thank you guys for your valid contributions........my data was just so 'badly holed' that I honestly realised doing multiple imputations may lead to wrong inferences. So, I am exploring alternative sources of data. My research focus is on income inequality and intentional homicides on a global scale covering 140 countries from 2000-2012, so please if anyone can point me to the right data source(s) for the 'indicators' I will forever be grateful......gracias....:)
I perused the attached Stata output attached, no doubt you have an issue with important variables. Unfortunately I cannot help with data source. Can someone be of assistance please?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics