I am designing a pain study. The outcome is change in pain from a Numeric Rating Scale (NRS), 0-11, pre- and post- treatment. Our original thought was to hypothesize a change of at least 2.0, with a standard deviation of 4.0 yielding an effect size of 0.5, requiring 34 subjects. However, we thought about changing this to a change in NRS of 30%, so 1-3 would be at least one point, 4-6 at least 2 points, 7-10 at least 3 points. This makes the outcome binary. There is no control group. I was unsure of how to translate this into an odds ratio for the hypothesis and power calculation. If there's a direct translation or an approximate "medium" odds ratio that could be justified in the same way an effect size of 0.5 is "medium", and how to set up the other parameters: P(Y=1|X=1) Also what the R^2 Other X is, and the X parm pi (proportion). I'm using G*Power 3.1.9.4 and logistic regression, trying to get the entire confidence interval over 1.
Study Design & Power for Pain Research
- Thread starter DavidB
- Start date
So pre-data collection of a pain score then pain score collection in the SAME people again or a completely new group. If the same people, which I am assuming, is moving from a 3 down to 1 the same thing as moving from a 10 to an 8? Also, is time between measurements uniform or do you need to control for the differences between measurement. Most retest experts recommend modeling: post ~ period + pre-value, and you may need to add a term for time between measurements.
I wouldn't dichotomize scores, since you will lose information. Also, I did not completely follow your above description, but you can't have an outcome with odds ratios and R^2. You can get R^2 out of a logistic regression, but they are considered useless.
The easiest was to do this would be simulations of pre data with a change added to those values, then conduct your test statistic and repeat this process and repeat this process for multiple sample sizes.
I wouldn't dichotomize scores, since you will lose information. Also, I did not completely follow your above description, but you can't have an outcome with odds ratios and R^2. You can get R^2 out of a logistic regression, but they are considered useless.
The easiest was to do this would be simulations of pre data with a change added to those values, then conduct your test statistic and repeat this process and repeat this process for multiple sample sizes.
--> Thanks for the reply. Yes, pain measures are matched. They're measured before and after being given a drug. We're considering a 30% reduction in pain after being given the drug to be "success", so going from 3 to 2 would be the same as going from 9 to 6, both 33% reduction in pain. So this includes the baseline score in the measure. You make a good point about losing information from dichotomizing the variable. The issue is that some patients are anticipated to respond to the drug while others are not - likely bimodal (or somewhat bimodal) data. We are hoping to collect data to determine who will be responders, but we don't know yet if the continuous confounders will be significantly associated. Since the data are actually integer (11 points, 0-10), not continuous, and since only a drop of 30% or more is considered clinically meaningful, I thought making it dichotomous might solve some problems. A nonparametric paired test (i.e. Wilcoxon Signed Rank Test) would work, but wouldn't allow us to incorporate other confounding variables (eventually). So I'm ambivalent. I could make a better decision once we have collected the data, but we have to do the power analysis and study design proposal now.
--> So if I do logistic regression, I'm a little confused about how to set up the power analysis. (The R^2 from G*Power referred to the x variable parameter, not the outcome.) I don't know what my main predictor is, because the data are paired. The predictor isn't receiving the drug, because everyone receives it. Should I be using a different test all together? And if so, can it eventually incorporate other confounders? The one-tailed one-sample binomial test would work initially, but doesn't allow for incorporating confounders.
--> So if I do logistic regression, I'm a little confused about how to set up the power analysis. (The R^2 from G*Power referred to the x variable parameter, not the outcome.) I don't know what my main predictor is, because the data are paired. The predictor isn't receiving the drug, because everyone receives it. Should I be using a different test all together? And if so, can it eventually incorporate other confounders? The one-tailed one-sample binomial test would work initially, but doesn't allow for incorporating confounders.