# which statistical test shall I use?

#### Rerun

##### New Member
Hello, I have some doubts about which statistical test shall I use in this case. summarizing:

my main variable of interest is the blood level of a protein, continuous, normally distributed (not completely sure though, should check graphically since in my populations there are only cases, no controls).

I step of analysis

my outcome is the difference between time 1 and time 0 of a variable that expresses whether the person has a symptom or not.
basically it can have 4 values:
- presence symptom at T1 and T0,
- absence symptom at T1 and T0,
- presence symptom at T1 (worsening),
- presence symptom at T0 (improvement).
(I might decide to collapse the first two in one single category since they mean that there has been no change in the situation globally, so it would become a categorical variable with three different values).

I shall run the analysis at unadjusted level and consequently considering few covariates.

I basically would like to investigate whether or not high values of the protein can be predictor of the difference in symptoms at T1 considering T0. My first thought was a multinomial logistic regression (putting as ref level the worsening of situation, since it is a logical evolution of the pathology). But I still have doubts whether or not anova would be a better idea.

II step of the anlysis:

symptoms can be evaluated with a score that goes from 0 to 10 so in the second level of the analysis my outcome is the difference between scores:
outcome: score T1 - score T0 (it goes from 10 to -10). 10 stands for the stronger worsening, - 10 for the stronger improvement. somehow the variable can be considered ordinal but I am not completely sure.
mean and median are both 0 (where median is 0 and mean is 0.2 or something like that) so there is a distribution quite equal at both sides.
My though here was to use tertile of the outcome ( tertile with the greater improvement, middle and greater worsening) and run another multinomial but I have a lot of doubts on this last one. What do you think? I excluded the linear regression (or some form of correlation) as well as the ordinal linear regression because there are quite few values in the outcome, but at same time too many to treat the variable as a categorical at first.

#### Karabiner

##### TS Contributor
whether or not anova would be a better idea.
oneway ANOVA or H-test seems straightforward.
symptoms can be evaluated with a score that goes from 0 to 10 so in the second level of the analysis my outcome is the difference between scores:
Not sure, maybe repeated measures ANOVA with protein as covariate.

With kind regards

K.

#### Rerun

##### New Member
Could your explain the second part better? Level of proteins is still the main variable of interest in my study and the difference between scores the outcome. Why would you treat it as a covariate?

#### Karabiner

##### TS Contributor
Sorry, that was SPSS-speak (where one could put it into the "covariates" box
when setting up an ANOVA), I just ment continuous predictor.

With kind regards

K.

#### Rerun

##### New Member
Regarding the first part of the study I was doubtful about ANOVa because basically the test says whether there are difference or not within groups (am I wrong? sorry statistic is not my background) while I need to know the effect of high or low level of the proteins on the change of clinical condition (so I need to have a coeff or an OR, in order to know the orientation). That`s why I went for the multinomial, but somehow I sensed that could be wrong...

#### Rerun

##### New Member
Sorry, that was SPSS-speak (where one could put it into the "covariates" box
when setting up an ANOVA), I just ment continuous predictor.

With kind regards

K.
Ah ok, now I know what you suggested. The thing is that the sample is slightly more than 100 people, and the score goes from -10 to + 10. In this way I would end up with so many groups but with few records each...

#### Karabiner

##### TS Contributor
Regarding the first part of the study I was doubtful about ANOVa because basically the test says whether there are difference or not within groups
It tells you, basically, whether the mean protein level is different between groups.
If yes, you can perform pairwise comparisons in order to find out which groups
are different from which.
Ah ok, now I know what you suggested. The thing is that the sample is slightly more than 100 people, and the score goes from -10 to + 10. In this way I would end up with so many groups but with few records each...
You should use the score as continuous predictor, not as
grouping factor. It's like you could use it in linear regression,
or in a correlation analysis.

With kind regards

K.