src/onto/longitudinal.lisp

   1
   2 (in-package :data-analysis-ontology)
   3
   4 (session :conference "ENAR" :session "longitudinal data" :year 2009
   5
   6  (talk :name "Jianwen Cai"
   7        :title "Joint Modeling of Longituindal Cat data and Surv data"
   8        :topic "joint survival longitudinal"
   9
  10        (example "liver transplant"
  11                 :goal " baseline-chars disease-hist impact QOL and risk of death postliver transpliatatn"
  12
  13                 (data   :time 4m 1y 2y 3y 4y 5y
  14                         :measure QOL (binary QOL ) (time death)))
  15
  16        (existing methods
  17                  (selection "Tsiatis, T+ Wulfsohn, etc"
  18                             estimate (conditional  (survival-time T)
  19                                                    (longitudinal data X)))
  20                  (pattern mixture models )
  21                  (simultaneous model "Zheng Cai 2005"
  22                                (assume: (longitudinal-outcomes 'normal)))
  23                  ())
  24
  25        (proposal
  26         (model  (GLMM longitudinal outcome, random on subject)
  27                 (PH   T conditionl on random effects above and obsd history before t))
  28
  29         (data
  30          (Ni Yij Xij XTij)
  31          (Zi Di Wi (t) WTi (t)))
  32
  33         (inference  (parameters RE FE baseline CumHaz)
  34                     (MLE 'complete-likelihood  'marginal on RE)
  35
  36                     (compute-with EM 'TLouis-est-var))
  37
  38         (behavior (asympt RE FE CumHaz :consistent (product-norm (theta euclidean-distance)
  39                                                                  (cumhaz sup-norm)))
  40                   (finite-sample :assume  (list  (longitudinal '(regression 'logistic))
  41                                                  (event-time 'PH-model))
  42                                  :results (fine under friendly conditions)) ; FUFC
  43                   )
  44
  45         (example "Liver Transplant"
  46                  :N 582
  47                  :events 76
  48                  :QOL-obsd 1382 )
  49
  50         (summary :weird normal distr assumptions for REs ?)))
  51
  52  (talk :name "Mike Kenward"
  53        :title "Dropout in Long clin trial w/ binary outcome"
  54        :topic "clin trial; longitudinal ; binary response ; dropout/withdrawal"
  55        :goal "compare at endtime, endpoints"
  56
  57        (prior-art "contin outomces with MAR, MVN framework")
  58        (issue "no natural generalization for MV binary resp")
  59        (issue (contrast 'population-average 'subject-specific)) ; marginal conditional
  60
  61        (with-missing-data (assert 'likelihood works) (assert 'estimating-equation issues))
  62
  63        (priot-art
  64         (list  ( Fitzmaurice 'cluster-weighted-gee)
  65                ( Robins/Rotnitsky 'obn-weighted-gee)
  66                ( Paik 'MI-GEE :multiple-imputation followed by 'complete-data-analysis
  67                       (meng 'imputation model conditions))
  68                ( Augmented Inverse Prob Wgt Est :approach " add-term with zero-expectation to incr eff,
  69                            see also scharfstein" :name AIPW)
  70                 ( doubly robust multiple imputation , but restricted
  71         to single var dropout, but
  72                         extended/made practical by Bang and Robins,
  73         :limit (assert monotone) (var-est by bootstrap ) (models for
  74         awkward expectations) )
  75
  76                )
  77         (Proposal
  78          :name "doubly robust MI"
  79          :prop "asympt equiv bang/robins"
  80          :alg (steps  MI over seq of regressions, Rubin's MI var est possible
  81               (not just bootstrap))
  82          :theory ( (obs-weight-gee Aug-obs-weight-gee d ) asympt equiv  for disc data)
  83
  84          )
  85
  86         (conclusion
  87          (MI-GEE more eff than obs-wgt-GEE, up to N=5000)
  88          (aux term incr eff/rob)
  89          (DRMI (sllight less eff than MI-GEE with Y corr spec)
  90                (simple var est)
  91                (practical implem use (simple / clear)))))  )
  92
  93 (talk :name "Xihong Lin"
  94       :title "variable selection using dantig selector"
  95       :topic ""
  96       :goal ""
  97
  98       (prior-art "Candes and Tao 2007" Dantzig Selector
  99                  (resemebles 'estimating-equation based var-select procedure))
 100
 101       (proposal "extend to corr data")
 102       (goal "select coefs which are non-zero")
 103
 104       (prior-art :penalized-likelihood)
 105       (prior-art :regularization)
 106
 107       (control correlation-of-residuals, not size of residuals!)
 108
 109       (subject-to (min (norm 1 beta))
 110                   (norm inf (m* Xt (Y - Xb))))
 111
 112       (note (there-exists (class matrices) st danztig-selector
 113       inconsistent for model selection))
 114
 115       (solution :weight selector through penalization to remove
 116       "droppings")
 117
 118       ;;;
 119       ))
 120
 121 (session "missing data"
 122
 123          (talk :name "Geert M"
 124                :title "MNAR equiv to so MAR"
 125
 126                (example toe nail data, longitudinal measure of serer toenail inferctions)
 127                (example slovenian public opiion survey on independence
 128                         3 q's : seccession-ind attend indep-grp ?)
 129
 130                (frameworks (list selection
 131                                  pattern-mixture
 132                                  shared-parameter)
 133                            :missing-models MCAR->MAR->MNAR)
 134
 135                (assumption "not dependence on future")
 136                (approach "put past responses into future not, AR/markov style")
 137                (everything MNAR model has MAR counterpart)
 138
 139                (explicit guide to construction)
 140
 141                (example :toe-nail infection severity within as a
 142                past/current marker to be managed when missing) )
 143
 144          (talk :name "mike daniels"
 145                :title
 146
 147                (example BCPT
 148                         (endpoints dose-efficacy QOL)
 149                         (rationale clinical-depression assoc with dose)
 150                         :N 11034
 151                         :dropout (> 0.45)
 152
 153                         )
 154                (non-ignoreable missingness in longitudinal trials)
 155
 156                (motivation :priors on missingness structure,
 157                            MAR as a strong point-mass prior)
 158
 159                (assumptions
 160                 (list (assert-hypothesis non-future-dependence 'data)
 161                       (assert-hypothesis selection-model-rpresenttion 'model)))
 162
 163                (model :observed-data 'transistion-model-parameterization
 164                       :missing-data
 165                       )
 166
 167                (issue :priors "spec ORs, but ORs diff to spec, so push to RR")
 168
 169
 170                )
 171
 172          (talk :name "joe hogan"
 173                :title "formulating priors for calibrating priors with incomplete data"
 174
 175
 176
 177                )
 178
 179
 180          )
 181
 182