Question on Types of Analysis for a Table 1

Question on Types of Analysis for a Table 1

by Emily Hurstak -

This is probably a dumb statistics question - sorry!  I am doing secondary analysis of a longitudinal cohort.  In my table 1 I have dichotomized three categories of my predictor of interest (head injury) and then want to do descriptive statistics for many variables of interest (other covariates and the ultimate outcome homelessness duration).  

When you do bivariate analyses for a Table 1 do you consider each covariate the outcome (ex: age, gender, education etc) and the column groups (head injury) as the predictor for your tests?  For example: would you do ttest for variable age (continuous) and kwallis for gender (dichotomous)?  

I will ultimately do a logistic regression with my predictor of interest (head injury), covariates, and ultimate outcome (homelessness), but my understanding is this is not what belongs in the table 1.  Does this make sense and am I interpreting this correctly?  

Thank you!

Re: Question on Types of Analysis for a Table 1

by Michael Peters -

Emily,

Most of what you state is correct, but let me clarify some points.  Yes, for a table 1 you would want your two categories (yes head injury and no head injury) as your columns and all your other variables as rows.  Generally you will have both continuous variables and dichotomous variables as rows.  Appropriate analysis for the continuous variables (age) would including either a t-test for parametric data (normal distribution) or a non-parametric mean test such as kwallis or mann-whitney U for data this is not normally distributed.  For your dichotomous variables (gender) your choices include a chi-square or a fisher-exact (fisher being preferred if you have rare outcomes).  Finally, you are correct I would not include logistic regression in your table 1, from the information you described.

 

Make sense?

Michael 

Re: Question on Types of Analysis for a Table 1

by Emily Hurstak -

Thank you!  I think so - but just to clarify:  I should think about the column variable (head injury, dichotomous) as the outcome with each analysis and the covariates as predictors?  I am confused by this as it will ultimately be one of the key predictors of interest in my logistic regression.  

 

If we think about this variable as the outcome for table 1 then I can't use a t-test for a continuous outcome (head injury=dichotomous).  Or alternatively if we think of the row variables as each individual outcome (age, gender etc) then I would use t-test for age, Chi square for gender etc.  I assume for table 1 descriptive statistics you should calculate these one by one ?

 

thanks again!

Re: Question on Types of Analysis for a Table 1

by Michael Peters -

Ok, so Head Injury is your outcome variable.

 

First off, a table 1 should be a description of your subjects.  In your case, it sounds like you have a case control study, i.e. subjects with head injury and subjects without head injury.  It is perfectly acceptable to first just give the description of this without any regression analysis.  Then for you regression you will include (generally speaking) only the variables that are significant after regression.  Now it is important which variable you do include in your model, but that is a topic for biostats 2.  For now, just a table one will be fine.

 

Second, yes you should use your second approach below.  The row variables are individual outcomes, mean differences for continues variables by head injury, and chi square tests for gender.

 

Still ok?

 

MCP

Re: Question on Types of Analysis for a Table 1

by Emily Hurstak -

Sorry for the ridiculous back n forth Michael (btw ignore my question about IF syntax - I figured it out).  I don't just have head injury/no head injury dichotomous.  That is why I am confused I think.  I have 4 categories 1)ever hit head, 2)head injury with LOC, 3)head injury with hospitalization, 4)never hit head and I am trying to do descriptive statistics across many different covariates and see if they are different (comparison) for any of these four groups of "head injury."  Would you use the tabulate, chi square command for this?

 

thanks again 

Re: Question on Types of Analysis for a Table 1

by Michael Peters -

AH!! Four categories!!  Ok, we can still work with this.  For multiple comparisons we need to use an ANOVA test instead of a t-test for continuous variables, and we stick with the chi-2 for categorical variables.  To perform an ANOVA in stata we use the command oneway.  This will give us a p-value if there is a difference between all groups, ie does one group differ from the others.  

for example

oneway age braininjury

The p-value you want to use is the one under Prob > F

In addition, in the future you may want to compare each group, to do this you use a bonferroni correction and the syntax is

oneway age braininjury, bonferroni tab

the p-values are in the table under each comparison.

 

Finally, for categorical variable we can just use the simple tab command followed by chi2

tab gender braininjury, chi2 

This will tell you if you have any differences in any of your groups.  

We good yet? :)

Michael

Re: Question on Types of Analysis for a Table 1

by Emily Hurstak -

You have been super helpful.  I think I may be trying to do something too advanced for this project.  I literally have four different variables (neverhit, everhit, hit with LOC, hit with hosp) all dichotomously coded.  If I did ANOVA with each one it wouldn't compare across all 4 groups - I don't think you can do ANOVA and list four outcomes. Also it wouldn't give the data I want for the actual table (means or proportions).  I guess I was just looking for the statistical test to compare all 4 categories of head injury (4 columns) across these covariates (rows), but that may be too upper level for me at this point.  Perhaps for the purposes of this project I shoudl just dichotomize head injury (0/1) and do it he easy way?

 

Thanks again for your help- I'll bail after this and stop harassing the forum.  I can do the simple way for the purposes of this class.

Re: Question on Types of Analysis for a Table 1

by Michael Peters -

I think I understand, but I might not...

The question I think you are asking is how do I  compare multiple variables across 4 different groups.  The way to do this is just to do the same tests against one group and then to do it again across another group.  You will get means and proportions then for each of the dichotomous outcome variables. 

Now, you may run into issues/questions regarding the age old problem of "correcting for multiple comparisons".  Generally speaking in your context I do not think you need to correct for multiple comparisons.  However, this would require some discussion in regards to your data, and yes for this class I think this is probably not the objective of the assignment.  For this assignment if you just choose one of the dictotomous outcomes and compare it across your other variables you will have completed the assignment.

 

MCP