To chi square or not to chi square?

lisasimps

I would love some expert opinion.
I'm working with spss to analyse results of a survey.
I've two data sets with multiple categories, always far more than 2.

I was going to run chi square but there are a lot of cells with expected count less than 5.

Next choice would be fishers exact but spss won't run that test unless 2×2?

Any advice?

lisasimps

lisasimps wrote: »

I would love some expert opinion.
I'm working with spss to analyse results of a survey.
I've two data sets with multiple categories, always far more than 2.

I was going to run chi square but there are a lot of cells with expected count less than 5.

Next choice would be fishers exact but spss won't run that test unless 2×2?

Any advice?

Also, I'm hoping to get the results published so need to ensure I choose the right tests

Black Swan

lisasimps wrote: »

I would love some expert opinion.
I'm working with spss to analyse results of a survey.

Need more information regarding your data sets and the types of questions you asked on your survey questionnaire.

What kinds of questions did you ask on your survey questionnaire? Nominal yes or no, male or female, etc., type questions? Ordinal data level greater than/less than Likert scales or other types of scales? Questions with ranged quantities? Combinations of different data level questions (nominal, ordinal, interval, and ratio)? Or what?

lisasimps wrote: »

I was going to run chi square but there are a lot of cells with expected count less than 5.

Lots of cells with less than 5 may result in no significant outcomes, so I would not run chi square. But why are you running a test like chi square, or other such test to see if there are differences between groups? Are you testing hypotheses, or answering research questions, or testing a conceptual framework, all the above, or what?

How did you select your subjects to receive your survey? Randomly or by some non-random method? How many surveys did you send? How many surveys were returned completed? Was the survey handed-out, mailed, or online, how?

Coincidentally, I am working with a lad that needs help with his survey design and analytic methods this weekend, so I can visit your thread this weekend with the same frame of mind. So give me more details, and I'll see if I can help you.

lisasimps

Its nominal data with independent data sets. Only ordinal is age, ive grouped into pools of 15 years so I can use that as categorical or ordinal, can't I?

Ive spent all day looking in to it and its looking like I go with Chi squared likeliness ratio when there are more than 20% of cells less than 5?

Its looking for relationships between independent data sets - does area of body bitten affect type of medical treatment, does hiring of trainer relate to likelihood of dog biting again etc.

It's questions with lots of variables, on dog bites, so it could be "where on body bitten" versus "medical treatment"with multiple variables.

The sample size is 315 - survey was online, posters circulated in hospital and gp waiting rooms, shared through social media and promoted on radio interviews. Participants had to have been victims of dog bites in republic of ireland over any time period.


Not that i'd be any good, but if I can be any help to you (seriously unlikely!) let me know

Black Swan

Sorry I could not get back to you sooner. The lad I was working with wanted some hands-on collaboration in his data analysis, which took more time than expected.

Without having unrestricted access to the data sets, as well as all the details of prior research design, survey administration, etc., the below is my best guess. Of course, I would not recommend sharing confidential medical data on the net, so if I stumble a bit, or miss a vital aspect, apologies.

lisasimps wrote: »

Its looking for relationships...

Nominal data, and non-random convenience sample (n = 315), therefore nonparametric statistical analysis (e.g., Phi coefficients). Results of study can only address subjects participating in the survey, and cannot be generalised to the dog bitten population of the Republic of Ireland.

(HR1) Research Hypothesis 1: There is a relationship between area of body bitten and medical treatment.

(HN1) Null Hypothesis 1: There is no relationship between area of body bitten and medical treatment.

Note: By convention, the null hypothesis is tested, not the research hypothesis. If the relationship between independent and dependent variables is significant at the p<.05 level of confidence, then the null is rejected and the research hypothesis is supported.

Compute SPSS Phi Coefficient:

Independent Variable --> | Dependent Variable
Area of body bitten | Medical treatment

(HR2) Research Hypothesis 2: There is a relationship between hiring of a trainer and likelihood of dog biting again.

(HN2) Null Hypothesis 2: There is no relationship between hiring of a trainer and likelihood of dog biting again.

Compute SPSS Phi Coefficient:

Independent Variable --> | Dependent Variable
Hiring of trainer | Likelihood of dog biting again


Additional HRs and HNs until you exhaust the relationships you wish to examine.

Display data in a Phi coefficient matrix table of independent and dependent variables. Discuss results in terms of p<.05 significance levels for each HN and its HR, as well as the direction and strength of the correlations.

The Phi coefficient does not suffer from the recommended cell minimum of 5 found in chi square. Phi also addresses significant relationships rather than differences, so the approach to hypotheses testing is a bit different, but may be workable for your analysis.

Helpful online SPSS procedure, example links, and citation: Phi Coefficient, Research Methods by Graziano & Raulin, 8th edition.

Hopefully the above is helpful and makes some sense. Given that I'm 8 time zones west of Eire and it's 5AM Pacific, if this all comes out as silly babble, sorry. If you wish to chat more about your research problems, theory/conceptual framework for explanations, analysis alternatives, etc., I'll visit your thread again Sunday after getting a bit of needed sleep. Cheers, Black Swan.

Black Swan

lisasimps wrote: »

Also, I'm hoping to get the results published so need to ensure I choose the right tests

Oh, another question about your data analysis. Have you run Cronbach's alpha to measure the internal consistency of your survey questionnaire items? It's the most widely used objective measure of reliability. If you are going for publication, you need to include all the measures that may increase the confidence in your results.

Useful reference: Making sense of Cronbach's alpha. Mohsen Tavakol and Reg Dennick, International Journal of Medical Education, 2011; 2: 53–55.