I would like to use Interrupted time series analysis for 2007-2015 data and the policy aimed to reduce the use of certain ‘low-value’ medical procedure after disease diagnosis was implemented in May 2012. I am very new to the interrupted time series analysis and would need the guidance from you all experts.
My outcome, use of low-value medical procedure is a binary variable coded as ‘0’ for patients with no use and ‘1’ for those with use after their disease diagnosis. I created variables time in months (month_dx) (1 through 108 for 2007 to 2015 data), ‘policy’ (0 for pre-policy period, and 1 for post-policy period), and the interaction of time and policy (time_after_policy). I ran the following program (unadjusted) and got the following output where I did not adjust for any covariates.
proc autoreg data = imaging outest=parapst covout;
model imaging = month_dx policy time_after_policy
/method=ml nlag=12 dwprob loglikl covb;
output out=pred p=predict r=resid;
run;
Autoregressive parameters assumed given |
Variable | DF | Estimate | Standard
Error | t Value | Approx
Pr > |t| | Variable Label |
Intercept | 1 | 0.5558 | 0.005242 | 106.04 | <.0001 | |
month_dx | 1 | -0.000457 | 0.000127 | -3.58 | 0.0003 | month of cancer dx |
policy | 1 | -0.003365 | 0.008428 | -0.40 | 0.6897 | policy yes/no |
time_after_policy | 1 | -0.000625 | 0.000297 | -2.10 | 0.0357 | time since policy |
Durbin-Watson value of 2.0016
How do I interpret the values for policy (-0.003365) and time_after_policy (-0.000625)?
In the attached document with figures for diagnostics, I found that the figure for residuals is bimodal as my outcome is binary. Is it normal to have a bimodal figure for residuals with binary outcome? Or do I have to change the outcome data to proportion of patients receiving low-value medical procedure?
After I adjust for the covariates (age, race/ethnicity, socioeconomic characteristics, disease severity, etc.), I get the following output:
proc autoreg data = imaging outest=parapst covout;
model imaging = month_dx policy time_after_policy agegrp region raceeth income educ grade cci gleason psa_level tstage
/method=ml nlag=12 dwprob loglikl covb;
output out=pred p=predict r=resid;
run;
Autoregressive parameters assumed given |
Variable | DF | Estimate | Standard
Error | t Value | Approx
Pr > |t| | Variable Label |
Intercept | 1 | -0.0551 | 0.0176 | -3.13 | 0.0018 | |
month_dx | 1 | 0.001058 | 0.000176 | 6.00 | <.0001 | month of cancer dx |
policy | 1 | -0.0434 | 0.008090 | -5.36 | <.0001 | policy yes/no |
time_after_policy | 1 | 0.001558 | 0.000344 | 4.52 | <.0001 | time since policy |
agegrp | 1 | 0.0176 | 0.001864 | 9.42 | <.0001 | Age groups 4 categories |
region | 1 | -0.0569 | 0.002227 | -25.56 | <.0001 | US region |
raceeth | 1 | 0.003842 | 0.002160 | 1.78 | 0.0753 | 1 wh 2 aa 3 hisp 4 asian 5 oth |
income | 1 | 0.0312 | 0.004939 | 6.31 | <.0001 | income |
educ | 1 | -0.0282 | 0.003466 | -8.14 | <.0001 | education |
grade | 1 | 0.1567 | 0.003447 | 45.45 | <.0001 | grade |
cci | 1 | 0.0528 | 0.002139 | 24.70 | <.0001 | cci index |
gleason | 1 | 0.0561 | 0.003938 | 14.25 | <.0001 | gleason |
psa_level | 1 | 0.0302 | 0.001399 | 21.59 | <.0001 | psa_level |
tstage | 1 | 0.008183 | 0.002040 | 4.01 | <.0001 | T stage |
| | | | | | |
Durbin-Watson value of 2.0008
I see that the estimate for intercept becomes negative in the adjusted regression. Also the estimate of ‘time after policy’ becomes positive (p<0.0001) indicating that policy change increased the use of ‘low-value’ medical procedure.
Please let me know if I followed the correct steps. Also please let me know if we need to include all the covariates in the adjusted regression for interrupted time series. Any guidance will be greatly appreciated.
