140 likes | 333 Views
D/RS 1013. Data Screening/Cleaning/ Preparation for Analyses. Data entered in computer. assuming reasonable care was taken scanner probably most "error free" checking physical forms against file verifying any recoding or score calculations "list cases"(mac) or "case summaries” (windows).
E N D
D/RS 1013 Data Screening/Cleaning/ Preparation for Analyses
Data entered in computer • assuming reasonable care was taken • scanner probably most "error free" • checking physical forms against file • verifying any recoding or score calculations • "list cases"(mac) or "case summaries” (windows)
Data screening • descriptives: look for out of range values • check values against original forms • correct data in file
Missing data • respondents will not answer all questions on a survey • what to do about items where data is missing? • several options to consider/ways to address
Missing data (cont.) • single variable - is systematic bias present in the kinds of people who fail to answer an item? • if the amount of missing data is small don't really need to worry • use pairwise deletion • pairwise can cause problems
Missing data (cont.) • drop subject's data completely • if missing data on unimportant variable don't analyze • if a reasonable guess can be made based on other available variables, do it • numerical variable - use average
Missing data (cont.) • correlation between answered and unanswered questions • regression equation to predict values on one variable based on others for which we have data • new variable that flags whether they answered question or not • analyze for possible differences on some other variable.
Outliers • exert influence on the mean • inflate variance of the sample • identify - look at a graph or run explore requesting outliers • rule out some kind of data problem • can dump and not use • compromise is to move outlier • residual analysis and detecting multivariate outliers when we move on to multiple regression (e.g. Mahalanobis Dist.)
Normality • assessing univariate normality • look at graph • skew and kurtosis values • can test significance • divide by standard error • result is a z score
Normality (cont.) • tells us whether skew/kurtosis is significantly different than "0” • does not necessarily mean it is a problem • Kline's (1998) recommendations skewness values > 3 and kurtosis > 10 • If seriously violated transforming is an option
Linearity of relationship • relationship between variables reasonably summarized by straight line • check scatterplot • may be curvilinear
Homoscedasticity • assumption that variation in one variable is constant across range of another variable • check scatterplot