3

Title explains it all. I want to know which test is the most appropriate to track weight loss / weight gain on a weekly basis in order to create a confidence interval, to see if my expected rate of loss/gain is in that confidence interval. It makes it easier to see if I'm stalling and have to adjust calories.

Some extra info:

  • Measurements will be taken each morning, so no bias (although water weight can be a bitch)
    • Although there are 7 days in a week, I can't take a measurement every morning. Some weeks will have 7 out of 7, other 6 out of 7, ...

I'm not sure if the T-Test is the most reliable for this because the measurements are not independent. I'm also not sure about normality since it's only a sample size of maximum 7. Any idea which test is the most appropriate?

Thanks for reading

Anon
  • 31
  • 2

2 Answers2

1

Neither a t-test nor ANOVA is appropriate here.

  1. Your measurements at different times are not independent
  2. You have a continuous covariate--time.

One approach would be OLS regression. It is possible that a mixed-model would be even more appropriate.

Suppose you are measuring the value of variable $W$ representing the weight. In addition, you have an ID for each subject $ID$. You likely also have some additional covariates such as sex $S$, diet $D$, breed $B$, etc. Finally, you have the time-stamp of each measurement of $T$.

Forget which week and how many times during the week. I'm assuming that it is the raw effect of time you are interested in. So just put a time stamp on each measurement. It won't matter if some measurements are skipped. The time stamp will still keep them in the correct order.

You could then perform the regression

W ~ T + S + D + B + ID

and any interactions you think might be involved.

This tutorial leads you through the steps of a repeated-measures design analysis, along with post-hoc analysis (don't forget the post-hoc!).

https://rcompanion.org/handbook/I_09.html

Many might argue that a mixed-model approach would be superior, with the subject ID as the random factor. In a simple before/after study you assume the effect of the treatment (if any exists) has occurred by the time you take the "after" measurement. However, the rate at which different subjects respond might be different although in the same trend. Treating this trend as a random effect increases your poser.

You can read about these types of models here:

https://www.jmp.com/en_dk/learning-library/mixed-models.html

https://www.stat.ncsu.edu/people/davidian/courses/st790/notes/chap3.pdf

Additional Note:

This question just popped up on the main feed:

Difference between Repeated measures ANOVA, ANCOVA and Linear mixed effects model

abalter
  • 770
  • 6
  • 18
0

You likely need a paired t-test here, which is appropriate when you have paired (dependent) data (e.g., before/after some new drug was administered, before/after some new diet regimen). t-tests are appropriate for small samples, so I believe this will suit you well.

compbiostats
  • 1,017
  • 9
  • 18
  • Thanks for the answer! It's just that I'm using this personally and it's sometimes not possible to have 7 out of 7 measurements every week (work, ...). How would you compare for example week one (with 6 samples) with week two (with 5 samples)? – Anon Sep 19 '17 at 19:50
  • @Anon The paired t-test requires the same number of samples for each group. For situations like 5 samples in one week and 6 samples in the other, perhaps Friedman's (non-parametric) test for repeated measures (ANOVA) would be most appropriate... but ANOVA is for more than 2 groups. – compbiostats Sep 19 '17 at 19:53
  • I guess I will have to start experimenting with that. Thanks! – Anon Sep 19 '17 at 20:01
  • The paired t-test is not appropriate here. For one thing, you have more than two groups because of all of the different times. Or, if you did a paired t-test between groups at each time, then you would be in danger of creating multiple comparisons hell, not to mention ignoring the random effect of measurement time on the results. Finally, you would have to exclude times where you did not perform all measurements. – abalter Jul 23 '20 at 03:55