Time-varying regression coefficients in Cox PH models (via Grambsch and Therneau residuals) -----------------------------------------------------
^stgtplot^ varlist [^, ci le^vel^(^#^)^ ^me^an graph_options ]
^stgtcalc^ calculates local estimates of the regression coefficients in the Cox model most recently fitted for using @stcox@.
^stgtplot^ plots the estimates and confidence intervals against the failure times. varlist is (a subset of) the independent variables in model fitted by ^stcox^.
^ci^ causes a ^level()^% confidence band to be plotted around the estimates of > the regression coefficients. The pointwise confidence intervals are derived from the running line smoother @running@.
^level(^#^)^ sets the width of a pointwise confidence band for the estimated b(t). Default: current value of system macro ^$S_level^ (usually 95%).
^mean^ uses the ^mean^ option of @running@ to produce a running mean smoother.
graph_options are any of Stata's ^graph, twoway^ options
^stgtcalc^ implements a suggestion of Grambsch and Therneau (1994). For an xvar with estimated regression coefficient b, ^stgtcalc^ calculates
g(T) = b + #deaths * V^^-1 * r(T)
at each failure time T, where V is the estimated dispersion matrix of the regression coefficients, #deaths is the number of uncensored observations and r(T) is the vector of Schoenfeld residuals at T. The g(T)'s are smoothed and the smoothed values are saved. g(T) is a local estimate of beta, so the shape of the smoothed profile may indicate whether beta depends on time. If so, the dataset has non-proportional hazards and the model will need to be refined.
Note that the smoothed g(T) may be unreliable in the initial and final phases of follow-up. This may be partly due to lack of robustness in the smoother.
The values of g(T) for each xvar are saved with names ^GT^xvar1, ^GT^xvar2, etc > .
If more than one xvar is to be plotted by ^stgtplot^, a panel of plots is produ > ced.
. ^stset survtime cens^ . ^stcox x1 x2^ . ^stgtcalc^ . ^stgtplot x1^ . ^stgtplot x1 x2, ci mean^
Grambsch PM, Therneau TM. Proportional hazards tests and diagnostics based on weighted residuals. Biometrika 81: 515-526 (1994).
Patrick Royston Imperial College School of Medicine, London FAX: (011)-44-181-383-8573
Peter Sasieni Imperial Cancer Research Fund, London FAX: (011)-44-171-269-3429
Also see --------
Manual: [R] st stcox On-line: help for @stcox@, @stbtcalc@