Long time reader, first time poster-
I received a journal review and am trying to address their comments but running into some issues with which types of tests are the most appropriate, thus, this is more best practices in statistics rather than an issue with STATA but I have exhausted my social networks, read through too many postings, and books on the subject(s) and am in need of some advice/motivation
![]( "Smile")
.
Quick summary: I have 42 years of data where I want to test whether a trend is significantly increasing/decreasing. I have created separate groups from the categorical data (e.g. White Male) and examined the proportion over time to control for changes in rates/counts. I am examining crime over time within the same race (proportion White on White vs not; Black on Black vs not) and need to determine the best type of testing for a bunch of panel models (for each race, gender, and crime type so about 24 different groups' proportions with 42 years/data points each). I am not sure if simple nonparametric tests are best (ologit, etc.) or if I need to assess using time series analyses (unit root then ARIMA).
More detail:
Data are structured for proportion over time for each group (e.g. White male crime perpetrated by another White male). Crime is largely intraracial so White males, for example, experience a crime committed by another White male about 75% of the time. The range of the data is a low of 66 to a high of 80% (or proportion of .66 and .80 respectively). The trend bounces around over the 42 years of data but the question is, has it changed significantly over time? How best do I address this?
Some of the trends seem linear (increasing, decreasing, or flat), others curvilinear (~) when potted over time (visual inspection). Including linear trend lines show most groups have little change over time but some slopes are greater than others. Initial dfuller tests suggest some of the group trends are stationary and some are not so I would use the first difference term 0 1 0.
It does not seem the reviewers want a simple nonparametric test but time series analysis (specifically stated I should assess if the trend is "stable" or "stationary" in review). While I outlined that time series analyses were not appropriate for so few data points, there is a precedent for begrudgingly performing these tests with limited data points (a similar published article's footnote states they did a unit root test and a "positive time trends were found to be significant with and without the autoregressive error process included". To me, this means they did an Augmented Dickey Fuller test and a ARIMA model with and without the p term (ARIMA 0 0 0 and 1 0 0)).
This is the first time I have worked with time series data and have few colleagues who have either so I am running out of options. I am starting to wonder what is the point of any an all of these tests and am feeling discouraged about the time spent and lack of progress. I have googled the crap out of the issue, searched this forum (e.g.
here), have a very long syntax/do file with all types of test but no idea which is best/better (e.g. nonparametric: ologit, nptrend, spearman, ktau, jonter; stationarity: dfuller with Varsoc, dfgls, lomackinlay, kpss, pperron ; ARIMA: twoway, ac, pac to assess model fit), read and replicated Box-Steffensmiere book (
link), and just need someone to help if possible. I am not even sure why I need to do a full ARIMA model or if I just need to assess stationarity or just do a more simple nonparametric test.
Any and all advice or references would be greatly appreciated because my brain hurts and I am feeling discouraged.
.