I am working on three different survey data; each has 20 binary variables for Hierarchical Cluster Analysis (HCA) on SPSS. Out of the 441, 2869 and 2543 total samples for each survey, only 42 (9.5%), 563 (19.6%) and 547 (21.5%) were valid in the HCA. The data consist, individual, behavioural and community characteristics recorded as 1 and 0. For the behavioural variables, 1 stand for a 'presence of risk' and 0 for an 'absence of risk'. Both the individual and community variables have similar codes (as used in Logistic Regression). Is it appropriate to replace missing values as an absence of risk (0), instead of the presence of risk (1) that may exaggerate the risk in the study? I am aware that the former may equally introduce similar error.
How do I manage the missing values and the potential errors?
REMOVAL AS DEFECT...... a researcher should not substitute or create data that is not there. The survey that has no data should be removed as defective due to incompleteness or missing data. This is considered as one of several response defect.
Hi Clement,
Its not recommended at all to replace missing values with Zeros (and particularly for binary data). You can replace them with other value (for example -9) and exclude these value for each analysis you will perform.
Replacing missing data for others induces new biases in estimation procedures. Otherwise there is a problem of nonresponse or absence of answer or observation which induces a bias of estimation too.
The answer depends why the data were missing. For example, if some people were not asked about some risks (because of previous responses, demographics, or just randomly) this is different from if some people skipped certain items by choice.
Thank you all for your swift response to my question. The answers are helpful.
Wright, all the respondents were asked the same questions (others skipped questions about risk). However, while trying to improve the response rate, I excluded two variables with high missing values, the valid samples in the HCA rise to 24.5%, 44.2% and 50.4% against the initial 9.5%, 19.6% and 21.5% . I think the two variables can be sacrificed since their absence may not
affect the study.
It does seem plausible that some respondents may have approached the task as ticking any risk they partake in and not ticking if they do not. Did some of the respondents have only 1s and missing? That would suggest this occurred. This would be something that could be picked up in piloting (e.g., asking why people left items blank). You could even examine this now by getting some people to respond and when some leave items blank ask them why. Of course the other solution is to re-run the study with instructions that make sure they don't leave them blank (or if an online format requiring an answer), but I am assuming that is not a feasible option.
This is a very important question many people ask. My answer is: Yes, you can. Then, you obtain a new variable which gives you the following information: "In most of the cases I don't know" (this is the interpretation of your new zero value). What can you do with such a "source of information"? So, surely you can, but you shouldn't do that
No, you can not put zero for missing data in binary analysis. Then it will be very complicated for your research. This is one kind of biases also.
Shafiq Ahmad depending on the person who creates the dataset, but, zeros does not mean a missing value.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS