We have collected retrospective data on different forms of childhood abuse adult patients, as well as current physical and mental impairment.
The questionnaire on childhood abuse included a time-frame for each question allowing patients to state at which age the abuse took place; the time-frame included 18 boxes for the first 18 years of their life. This way we know for example that a patient got physically abused at the age of 6, 7 and 8 years, and then again at the age of 16 and 18 years.
Our main research question at the moment is, whether the type of abuse or the timing (i.e. at which age the abuse took place) is a better predictor for adult mental and physical health.
We have entered the time frames for each question as 18 dichotomous variables (yes vs. no) for each patients.
The dependent variables (mental and physical health) are coded as nominal (disease present: yes vs. no), as well as metric (e.g. depressive symptoms) variables.
Can anyone recommend a good statistical procedure to approach our question with this kind of data?
Thank you all for your input!!
Maybe this would be of use to you...
Fantuzzo, J., LeBoeuf, W., Brumley, B., & Perlman, S. (2013). A population-based inquiry of homeless episode characteristics and early educational well-being. Children and Youth Services Review, 35(6), 966-972.
Interesting research design, but what about the recall bias? Will the victims exactly remember at what years the abuse occurred in their lives? Is there a way to overcome this in the statistical analysis?
Fascinating study question! What if you simply count numbers of abuse for each of the 18 time boxes to see if there is some clustering and a statistic correlation with mental impairment as adult?
Thank you all for your input and the literature suggestions!
@ Mohamad-Hani Temsah : Very interesting question! Recall bias is a general problem in this field of research, and can only be partially controlled for (my opinion). One helpful approach surely is a larger set of patients, to try to even out recall bias, since we see it as a individual rather than a collective defense mechanism / bias. The problem is, that especially for ACEs prospective studies are incredibly difficult to conduct and so far only a handful have been done. Yet, interestingly enough, the results of the prospective studies do not really differ from the retrospective studies, so the recall bias may be a smaller problem than we initially feared.
@ Martin: thank you :-) I think your suggestion is a good starting point and this is most likely what we will as a first step. The proble is, that there are large dependencies between the ages (i.e. if someone is being abused at the age of 14 there is a large likelihood they will also be abused at the age of 15), which leads to the statistical effect that highly correlated ages will kick each other out of the analysis in the case of stepwise regression. We also have thought about defining evidence based specific age groups (e.g. 5-8 year, 8-10 years, 10-13 years). The problem with this approach obviously would be that there is hardly any good evidence base and the approach would not be exploratory anymore. So I guess i would still have some thining to do ;-)
Again, thank you for your input; any further comments and ideas are of course very welcome!
Best,
David
If you have coded something as 0 and 1 then you have to do logistic regression analysis. Then your research questions goes like this:
What factors (in your case: which age) predict the likelihood that respondents report that they had a problem with being abused
What factors (in your case: which age of abuse) predict the likelihood that respondents report that they had a problem with depression
What factor ( abuse) predict the likelihood that respondents reports that they had a problem with physical health
Has any of you experiences using Random Forest Regression With Conditional Trees? It seems this might be a suitable approach?
- Badges
- Science topic