I am doing a study where I will get a measure for 20 individual patients pre and post an intervention. The measure is a figure (could be between 0.6 and 24 let's say) and so i assume this is continuous data?! What statistical tests should I use to analysis the measure pre and post intervention; I initially thought a paired sample students t test but now wondering about random co efficient regression models. There is no control arm but obviously the 20 will have varying baseline characteristics. THANK YOU for any help!
Outside United States program evaluation
There are several components to your question. First, with regard to the nature of the data: You don’t say how the scores are calculated or what they represent, so from the information provided I cannot say whether the data are continuous. However, since they take on decimal values (with a range from 0.6 to 24), it appears that there is probably a large number of potential score values. If this is indeed true, e.g., more than about 7 or 8 discrete potential values, then the scores can be treated as continuous.
However, a second critical question concerns the shape of the distribution. Does it approximate normal, or is it irregular or highly skewed? The answer to that question will determine whether you can use a parametric statistical test such as a paired-sample t test. If the scores do not approximate a normal distribution and can’t be treated as such, then you can use a nonparametric test for paired samples, such as the Wilcoxon signed-rank test. The information you provide does not address how the scores are distributed, but you should examine your scores and see how closely they approximate normality. Calculate the skew and kurtosis, and make sure that those values are both close to zero.
Finally, you ask whether a t test or a regression model would be most appropriate. It seems that your primary evaluation question concerns whether the patient scores have increased from pre to post. In that case, a paired pre-post test will be most appropriate, either a t-test (if the data are treated as normally distributed) or Wilcoxon (if a nonparametric test is needed). I don’t think a regression model makes sense, because you are not trying to predict scores on one variable from a collection of other variables.
From a conceptual standpoint, regression analysis could be used in a secondary way, to determine what scores on other characteristics predict the scores on your outcome of interest. That kind of analysis could help you determine which subgroups might be most responsive to your intervention. However, in your particular case, you have a very small number of cases (N=20) so any regression analyses would have very low power and is probably not advisable.