I am in a small disagreement with my statistician and would appreciate some additional perspective. This statistical question is regarding a cohort of participants with cancer. I am interested in investigating whether participants who have changed a behaviour illustrate different sociodemographic/treatment associations, compared to those participants who have not illustrated a behaviour change. To do this, I have created a datapoint from two survey questions I collected.
- Data point 1. Participants indicated 'yes' or 'no' to whether they currently use the internet (cross-sectional timepoint being the completion of the survey).
- Data point 2. Participants were asked retrospectively whether they used the internet for cancer information at a differing timepoint (immediately after their own diagnosis).
- Data point 3. I have created a third data point which is what I wish to analyse against sociodemographic/treatment data - this datapoint asks whether a participant demonstrates a difference in use of internet between Data point 1 and Data point 2?
>Thus, for Data point 3, if a difference does not exist (i.e. 'no'), then a participant has consistently reported 'no' across both Data point 1 and 2, or alternatively, 'yes' across both Data point 1 and 2. If a difference does exist ('yes'), then the participant has at least 1 'yes' and 1 'no', across Data point 1 and 2, regardless of order.
I have created this (data point 3) as I specifically wish to compare these groups (participants who illustrate behaviour change vs. those that do not) against sociodemographic and treatment variables (e.g. Age =65). I believe these populations may differ based on other literature.
My statistician believes that this requires a McNemar test and that this is paired data. I disagree as I feel I am not comparing paired data, i.e. not directly comparing how many men say yes vs no between 2 timepoints.
Instead, my belief is that I have created a new data point (datapoint 3) from the original paired data (data point 1 and 2). Data point 3 is itself a new item which stands alone, and is thus not paired. Ergo, in this instance, when I compare in a 2x2 (e.g. behaviour change status vs age 65) a chi-square test is fine.
Your thoughts? Thank you in advance for reading, and any advice/perspective. Apologies for the length of my query.
Choosing between McNemar's test and a Chi-square test depends on the specific characteristics of your data and research question. Here's a breakdown to help you decide:
McNemar's test:
- Use for:Paired data with two categories (dichotomous). Analyzing changes in proportions within those categories between two measurements (e.g., pre- and post-treatment, two different diagnoses). Specifically tests if the proportion of discordant pairs (changed responses) is significantly different from 50%.
- Advantages:More appropriate for paired data than standard Chi-square. Accounts for chance agreement by focusing on discordant pairs. Simple to calculate and interpret.
- Disadvantages:Can only be used with 2x2 contingency tables. Assumes equal marginal (row and column) totals. May not be accurate for small sample sizes.
Chi-square test:
- Use for:Nominal data with two or more categories (not just dichotomous). Testing for associations or dependence between two categorical variables. Can be used with paired or unpaired data.
- Advantages:More versatile than McNemar's test, applicable to wider data types. Can handle larger contingency tables with more categories. Does not require equal marginal totals.
- Disadvantages:Does not directly address changes within paired data like McNemar's test. Can be sensitive to small expected frequencies in large tables. Requires further analysis of specific cells to understand the nature of the association.
Here are some additional factors to consider:
- Sample size: If your sample size is small (less than 20 pairs), McNemar's test may not be accurate. Consider Fisher's exact test instead.
- Research question: Is your primary focus on changes within paired data (McNemar) or association between two categorical variables (Chi-square)?
- Data type: Is your data truly dichotomous or are there more than two categories?
Once you've considered these factors, you should have a clearer understanding of which test is more appropriate for your analysis.
- Badges
- Science topic
- Similar topics
- Mathematics