GOOD AFTERNOON I WANT TO WELCOME YOU TO THE MIND THE GAP SEMINAR SERIES. THIS EXPLORES RESEARCH AND EVIDENCE AND PRACTDIS AND EVIDENCE WHERE WISDOM MIGHT BE CONTRADICTED BY RECENT EVIDENCE, FROM THE ROLE OF MEDICAL ADVOCATION AND POLICY TO THE IMPORTANCE OF BEHAVIORIAL INTERVENTIONS THE OFFICE OF DISEASE PREVENTION HOPES TO ENGAGE THE PREVENTION RESEARCH COMMUNITY AND THOUGHT PROVOKING DISCUSSIONS TO CHALLENGE WHAT WE THINK WE KNOW AND THINK CRITICALLY ABOUT OUR ROLE IN TODAY'S RESEARCH ENVIRONMENT. BEFORE I BEGIN I HAVE HOUSEKEEPING ITEMS. TO PARTICIPATE BY TWITTER, FOLLOW US @NIHPROVEN AND SUBMIT QUESTIONS. AND YOU CAN ALSO E-MAIL QUESTIONS TO MAIL @NIH.GOV, AT THE CONCLUSION OF TODAY'S TALK, WE WILL OPEN THE FLOOR TO QUESTIONS SUBMITTED VIA E-MAIL AND TWITTER. LASTLY. VISITING THE WEB SITE PAGE AT PREVENTION.NIH.GOV/MIND THE GAP FOLLOWING TODAY'S TALK AND CLICK THE LINK TO THE SEMINAR EVALUATION UNDER THE RESOURCE SECTION TO SUBMIT YOUR FEEDBACK ABOUT THE SEMINAR. AT THIS TIME I WOULD LIKE TO TURN THINGS OVER TO DR. DAVID M. MURRAY ASSOCIATE DIRECTOR OF OFFICE AND DISEASE PREVENTION. >> THANK YOU. IT'S MY PLEASURE TO INTRODUCE OUR SPEAKER TODAY, DR. DAVID MACKINON HAS BEEN EVALUATING HOW INTERVENTIONS WORK FOR 30 YEARS. HE IS A FOUNDATION PROFESSOR AT ARIZONA STATE UNIVERSITY. RECEIVED UNDERGRADUATE DEGREE FROM HARVARD UNIVERSITY AND Ph.D. IN MEASUREMENT AND PSYCHOMETRICS IN UCLA IN 1986 AND ASSISTANT PROFESSOR AT THE INTUITY FOR PREVENTION RESEARCH BEFORE MOVING THE STATE, AND IN EUROPE, AND IN 2011 RECEIVED THE NAN TOKELER REWARD FOR SOCIETY PREVENTION RESEARCH FOR HIS BOOK ON STATISTICAL ANALYSIS WHICH IS THE TOPIC OF THIS CONTROLLER OF BANK, AND HE RECEIVED THE MERIT AND TIME AWARD FROM THE NATIONAL INSTITUTE ON DRUG ABUSE AND MEDIA ANALYSIS AND FEDERAL GRANT REVIEW COMMITTEES AND FORMER CONSULTING EDITOR FOR THE JOURNAL OF PREVENTION SCIENCE AND ON THE EDITORIAL BOARD FOR PSYCHOLOGICAL METHODS AND DR. Mc KINON HAS BEEN THE PRINCIPLE INVESTIGATORS ON SEVERAL GRANTS AND THOMPSON REORGANIZATION OF THE INSTITUTERS HIGHLY CITED RESEARCHER. FELLOW IN THE ASSOCIATION OF PSYCHOLOGICAL SCIENCE AND SOCIETY FOR PREVENTION RESEARCH RESEARCH AND AMERICAN PSYCHOLOGICAL MEASUREMENT AND STATISTICS DIVISION. AT THIS TIME I WOULD LIKE TO WELCOME DR. Mc KINON AND TURN THE SESSION OVER TO HIM AT ARIZONA STATE UNIVERSITY. >> OKAY, GREAT, CAN YOU HEAR ME. >> JUST FINE DAVID. >> THANK YOU DAVID AND RANELLE,--THIS IS AN EXCITING EVENTED EAR AND THANK YOU FOR COMING TO THIS WEBINAR. AND I WOULD LIKE TO THANK NIDA FOR THE CONTINUED SUPPORT TO DEVELOP AND EVALUATE MEDIATION METHOD. SO I WILL DESCRIBE THE MEDIATING EXAMPLES AND APPLICATIONS AND THEN I WILL DISCUSS STATISTICAL MEDIATION ANALYSIS AND THEN MENTION SEVERAL EVENT MODELS THAT ARE RELEVANT FOR PREVENTION RESEARCH AND BRAVELY DISCUSSION THE FUTURE DIRECTIONS. THERE'S A WEB SITE THERE THAT LAST I CHECKED WAS WORKING AND THEN THIS IS A BOOK PUBLISHED IN 2008 AND I'M WORKING ON A SECOND EDITION OF THAT BOOK. SO HERE ARE EXAMPLES WHERE BENEFICIAL EFFECTS ON EXERCISE WHICH LEADS TO REDUCED DEPRESSION, MEDIATOR IS BOLD IN THESE EXAMPLES, SO THAT'S A PREVENTION PROGRAM FOR ANTITOBACCO NORMS AND PRODUCE TOBACCO USE AND SCREENING PROGRAM INCREASES IDENTIFICATION OF EARLY STAGE CANCER WHICH REDUCES CANCER DEATHS, WELLBUTRINREDUCES PARTICIPANTS WILLINGNESS TO QUIT AND SELF--EFFICACY WHICH ARE ASSOCIATE WIDE 1 MONTH ABSTINENCE OF TOBACCO, MAYBE YOU COULD THINK OF A FEW EXAMPLES RIGHT NOW. HERE AIAN DIAGRAM WHERE WE HAVE THE SINGLE MEDIATOR MODEL, INDEPENDENT VARIABLE X AND DEPENDENT VARIABLE Y, AND THEN WE HAVE THE CO EFFICIENT THAT REPRESENTS THE EFFECT OF AXON M AND THE A B CO EFFICIENT THAT REPRESENTS THE XAND Y, SO THAT'S THE INDEPENDENT VARIABLE, X AND Y AND INDIRECTLY AND DIRECTLY THROUGH THE MEDIATOR. HERE ARE MORE MEDIATOR DEFINITIONS, THESE ARE VABL I CAN'T BELIEVE IN A CHAIN WHEREBY AN INDEPENDENT VARIABLE CAUSES THE MEDIATOR WHICH IN TURN CAUSES THE OUTCOME. THE GENRATIVE MECHANISM THROUGH WITH WHICH THE FOCAL INDEPENDENT VARIABLE IS ABLE TO INFLUENCE THE DEPENDENT VARIABLE AND THE LAST 1 IS THE CAUSAL PATHWAY FROM AN INDEPENDENT VARIABLE TO A DEPENDENT VARIABLE. IT CAUSE VARIATION BY THE DEPENDENT VARIABLE AND THIS ILLUSTRATES THE CHALLENGE OF MEDIATION ANALYSIS THAT THIS MEDIATING VARIABLE COMES BETWEEN 2 OTHER VARIABLES AND BOTH DEPENDENT VARIABLE IN 1 EQUATION AND INDEPENDENT VARIABLE IN ANOTHER EQUATION. MOST OF THOSE STATISTICS FOCUS ON THE VARIABLE EFFECTS WHICH IS THE CORRELATION BETWEEN X AND Y AND A CO VARIANT FOR CO EFFICIENT OR ODDS RATIO AND WITH 2 VARIABLES, THERE'S ONLY SO MANY RELATIONSHIPS, RIGHT? WE COULD HAVE X CAUSES WHY, WHY CAUSES X OR THEY'RE RELATED FOR SOME OTHER REASON. THE 3 VARIABLES, THERE ARE MORE DIFFERENT THINGS THAT COULD HAPPEN, WE NOW HAVE X, M, YUSING M TO REPRESENT THE THIRD VARIABLE, WE COULD HAVE YTO Y TO M, OR XX TO Y TO M, SO THERE ARE CAUSES HOW THESE ARE RELATED. WE HAVE SPECIAL NAMES FOR THESE RELATIONSHIPS IN THIRD VARIABLE, 1 IS CONFOUNDER, ANOTHER IS MEDIATOR AND ANOTHER 1 IS MODERATOR INTERACTION SO WE HAVE SPECIAL NAMES FOR THESE VARIABLE EFFECT. WITH 4 VARIABLES, THERE'S LOTS OF POSSIBLE RELATIONS, RIGHT? WE HAVE X, AND Y, WE HAVE Y, TO M, TO Z, TO X, LOTS OF DIFFERENT COMBINATIONS WITH FORGARBLES AND WITH 5 VARIABLES THERE ARE SO MANY COMBIN ANTICIPATIONS, IT WOULD BE EXHAUSTING AND JUST GO HOME AND TAKE A NAP BECAUSE THERE'S SO MANY POSSIBLE RELATIONS BUT WE DON'T TAKE A NAP BUT EVEN THOUGH LOTS OF BEHAVIORS ARE CAUSED BY MANY DIFFERENT VARIABLES WE NEED TO INVESTIGATE THEM. BUT WE SHOULD KEEP IN MIND THERE ARE THESE MANY POSSIBLE RELATIONS AMONG THE VARIABLES IN OUR MODEL. A CONFOUNDER IS A VARIABLE THAT'S RELATED TO 2 VARIABLES OF INTEREST THAT OBSCURES OR EXTENTUATES THE RELATIONSHIP BETWEEN THEM. LIKE THE MEDIATOR THE CAUSAL IS NONAPOPTOTIC IAYQUENCE IS THAT WILL BE A MAIN CHARACTERISTIC OF A MEDIATING VERTICAL IN A CAUSAL SEQUENCE FROM X, TO M, TO Y. IT'S A CONFOUNDER THAT'S RELATED TO X AND Y BUT IT'S NOT IN THAT SEQUENCE, IT CAUSES BOTH OF THEM, SO IF YOU FORGET TO INCLUDE A CONFOUNDER, YOU CAN GET A LONG RESULT. MODERATOR IS A VARIABLE THAT EFFECTS THE STRENGTH OR RELATION BETWEEN 2 VARIABLES X AND Y. AGAIN A MODERATOR IS NOT INTERMEDIATE IN THE CAUSAL SEQUENCE LIKE THE MEDIATOR. A MODERATOR IS A VARIABLE WHERE THE RELATION BETWEEN X AND Y DIFFERS AT DIFFERENT VALUES OF THAT MODERATOR. AS YOU MIGHT EXPECT, WHEN BEING LOAMACYING AT MEDIATION, THERE CAN ALSO BE CONCOMITANT VARIABLES AND THERE CAN ALSO BE MODERATOR VARIABLES AND WE'LL TALK BRIEFLY ABOUT SOME OF THOSE ISSUES BUT THESE ARE OTHER THIRD VARIABLE EFFECTS SO WE HAVE A MEDIATOR EFFECT, CONFOUNDER EFFECT AND A POSSIBLE MODERATOR EFFECT OF HOW A THIRD VARIABLE COULD ANSWER INTO THE RELATIONSHIP BETWEEN X AND Y. SO MEDIATION IS IMPORTANT BECAUSE CENTRAL QUESTIONS IN MANY FIELDS ARE ABOUT MEDIATING PROCESSES OR HOW 1 EFFECTS ANOTHER. IMPORTANT BASIC RESEARCH ON MECHANISMS OF EFFECT. CRITICAL FOR APPLIED RESEARCH ESPECIALLY PREVENTION AND TREATMENT TO IDENTIFY CRITICAL INGREDIENTS LEADING TO MORE EFFICIENT INTERVENTION AND THIS IS HOW I BECAME INTERESTED IN MEDIATION ANALYSIS WHEN I WAS FIRST ASKED TO LOOK AT HOW A TOBACCO, SCHOOL BASED TOBACCO PREVENTION PROGRAM ACHIEVED THIS EFFECT. THEN THERE'S ALL KINDS OF ENTERTAINING STATISTICAL AND MATHEMATICAL ISSUES. I SHOULD ADD ALSO THERE ARE MANY SUBSTANTIVE THEORETICAL ISSUES AND IN FACT MEDIATING VARIABLES TRANSLATE THEORIES IN MANY DIFFERENT FIELDS. THIS FOCUS ON MEDIATING MECHANISMS TO SEE IF THEY HAVE GROWN OVER THE LAST 20 YEARS AND HERE'S A GROWTH FROM THOMAS INSEL, NIMH DIRECTOR THAT FUTURE TRIAL WILL FOLLOW AN EXPERIMENTAL MEDICINE APPROACH IN WHICH INTERVENTIONS SERVE NOT ONLY AS POTENTIAL TREATMENTS BUT AS PROBES TO GENERATE INFORMATION ABOUT THE MECHANISMS UNDERLYING A DISORDER AND OFFERS US A WAY TO UNDERSTAND THE MECHANISMS BY WHICH THESE TREATMENTS ARE LEADING TO A CLINICAL CHANGE AND THIS IS A WAY TO EXPLORE AND INVESTIGATE WHAT THE POSSIBLE MECHANISMS FOR FOR TREATMENT AND ALSO PREVENTION PROGRAMS. NOW 1 OF THE MOST UBIQUITOUS THEORIES IN COLLEGE IS THE SOR THEORY WHERE THE STIMULUS ON A RESPONSE DEPENDS ON THE MECHANISMS IN THE ORGANISM. SO THE MEDIATING MECHANISM TRANSLATE THE STIMUE WELL US TO THE RESPONSE, WE'RE ABLE TO MEASURE THE STIMULUS AND MEASURE THE RESPONSE BUT WE DON'T KNOW WHAT'S GOING ON INSIDE THE ORGANISM, THAT'S WHERE THE MEDIATING PROCESS OCCURS SO ILLUSTRATE THIS FOR EXAMPLE, IF I ASK YOU TO MULTIPLY 24 AND 16, THE STIMULUS IS ASKING YOU TO MULTIPLY THOSE 2 NUMBERS, THE RESPONSE IS YOUR ANSWER WHICH SHOULD BE 384, AND THE MEDIATOR IS YOU, WHAT GOES ON INSIDE YOUR HEAD. IT'S MIDDAY THERE, IT'S PRETTY EARLY HERE IN ARIZONA, SO MAYBE NOT SO MUCH IS GOING ON INSIDE OUR HEADS HERE ON THE WEST COAST. AND THIS IS--FOLLOWING THIS IDEA FOR S. O. R. THEORY WHERE THERE'S A BLACK BOX AND HAVE YOU AN INPUT AND AN OUTPUT TO THE BLACK BOX AND I THOUGHT IT MIGHT BE INTERESTING TO SHOW YOU A BLACK BOX AND SHOW WHAT'S GOING ON IN THERE. SO HERE'S THIS S. O. R. MODEL WHERE WE ARE A STIMULUS AND RESPONSE, BOTH OF THOSE CAN MEASURE WELL MENTAL AND OTHER PROCESSES, WE WOULD HAVE TO FIND SOME WAY TO MEASURE WHAT'S GOING ON THERE AND WE USE QUESTIONNAIRES OR WE CAN USE EEG, USE A VARIETY OF OTHER APPROACHES TO FINISHED WHAT'S GOING ON INSIDE THE ORGANISM. AND THIS MODEL ALSO ILLUSTRATES THE CHALLENGES OF MEDIATING VARIABLES SO WE HAVE TO COME UP WITH A REASONABLE WAY TO MEASURE THE PROCESS AND THE PROCESS COULD BE AT A NEURONAL LEVEL, HIGHER LEVEL, INDIVIDUAL LEVEL AND SO ON AND WOULD HAVE TO SOMEHOW CAPTURE WITH OUR MEASUREMENT WHAT'S GOING ON IN THAT MEDIATING PROCESS. THERE ARE 2 OVERLAPPING APPLICATIONS OF MEDIATION ANALYSIS, 1 OF THEM IS MEDIATION FOR EXPLANATION WHICH IS MUCH OLDER AND THEN THE OTHER 1 IS MEDIATION BY DESIGN. MEDIATION FOR EXPLANATION COMES OUT OF LITERATURE AROUND 1950S OR SO, AND WHAT THE IDEA WAS YOU HAVE AN OBSERVED RELATIONSHIP AND THEN YOU TRY TO EXPLAIN IT BY ADDING A THIRD VARIABLE. AND IN FACT THEY HAVE DIFFERENT NAMES FOR WHAT WE CALL A CO VARIANT. SO IF YOU HAD A THIRD VARIABLE AND YOU CHOOSE BETWEEN X AND Y, THAT WAS A REPLICATION VARIABLE, WHAT WE WOULD WILL CALL A CO VARIANT NOW. IF IT EXPLAINED THE RELATIONSHIP BETWEEN 2 VARIABLES AND WE ADJUSTED FOR A THIRD VARIABLE, THAT WAS CALLED AN EXPLANATION VARIABLE AND NOW WE CALL IT A CONFOUNDER AND IF IT'S INTERVENING IN THE RELATIONSHIP BETWEEN X AND Y, WE CALL IT INTERVENING VARIABLE, WE WOULD CALL NOW A MEDIATOR AND WHAT WE NOW CALL A MODERATOR THEY WOULD WOULD CALL A SPECIFICATION VARIABLE, THE RELATIONSHIP BETWEEN X AND Y IS SPECIFIC TO THE VALUES OF THE THIRD VARIABLE. MEDIATION FOR EXPLANATION, OBSERVE RELATION IS OBLIGATIONS STAINED AND STATISTICAL ANALYSIS UNDERTAKEN TO UNDERSTAND HOW THE THIRD VARIABLE OPERATES, I A MUCH MORE MODERN OR APPLICATION IS MEDIATION BY DESIGN. WHAT MEDIATED VARIABLES THAT ARE CAUSALLY RELATED TO OUTCOME VARIABLE AND THEN WE WILL DESIGN THE INTERVENTION TO CHANGE THE MEDIATOR. IN FACT ALL THE RESEARCHER OFTEN KNOWS THAT THE BEGINNING OF THIS LINE OF RESEARCH ARE RELATED TO THE OUTCOME. AND THEN WE'LL DESIGN INTERVENTION BASED ON THE VARIABLES THAT ARE RELATED TO THAT OUTCOME AND IF THE MEDIATORS ARE CAUSALLY RELATED TO THE UT COME, IF WE CHANGE THAT WITH AN INTERVENTION AND WE WILL CHANGE THE OUTCOME, THIS IS A COMMON MODEL THAT WOULD BE APPLIED RESEARCH AND PREVENTION AND RESEARCH BUT IT SEEMS THEY FELT TO BE EVEN MORE USEFUL AS THE YEARS GO BY AS A WAY TO UNDERSTAND AND ALSO DEVELOP PREVENTION AND TREATMENT PROGRAM. SO HERE'S A DIAGRAM OF AN INTERVENTION MEDIATION MODEL WHERE WE HAVE X AND Y, OR X AND Z INTERVENTION PROGRAM, MEDIATORS AND OUTCOMES AND WE HAVE 2 KINDS OF THEORY IN THIS MEDIATION MODEL. WE HAVE MANIPULATION THEORY AND CONCEPTUAL THEORY. CONCEPTUAL THEORY IS THE USUAL WAY WE THINK ABOUT THEORY, THAT'S WHAT VARIABLES ARE RELATED TO THE OUTCOME, WHAT RISK FACTORS PREDICT AN OUTCOME, SAY SMOKING. MANIPULATION THEORY IS LESS COMMONLY DISCUSSED. THAT'S HOW WE WOULD DESSIGN INTERVENTION TO EFFECT MEDIATORS AND THIS IS PRETTY TRICKY, RIGHT WHAT ARE THE INTERVENTION COMPONENTS THAT WOULD WORK TO MANIPULATE THE SOCIAL NORMS OR BELIEFS ABOUT ALCOHOL USE. OR BELIEFS ABOUT SCREENING. SO THE 2 TYPES OF THEORY, MANIPULATION THEORY AND CONCEPTUAL THEORY. THEN THESE MEDIATORS PLAY A PRIMARY ROLE AND IDENTIFICATION OF THOSE MEDIATOR SYSTEM BASIC FOR A RESEARCH WHERE IF WE CAN DESIGN A STUDY AND FIND EVIDENCE THAT A MEDIATOR DOES SEEM TO BE THE IMPORTANT WOB FOR AN INTERVENTION WE CAN MAKE OUR INTERVENTIONS MORE EFFECTIVE BY TAKING THE COMPONENTS OF THE MEDIATORS THAT ARE MOST IMPORTANT TO TARGET AND THE COST BY REMOVING PARTS OF INTERVENTIONS THAT DO NOT NECESSARILY CHANGE THE MEDIATORS THAT WE'RE INTERESTED IN. OKAY SO TO SUMMARIZE WE HAVE THESE 3 VARIABLES INVOLVED BUT NOW WE'RE GOING TO FOCUS ON MEDIATORS ALTHOUGH THOSE OTHER TERMS WILL COME UP AND THAT THERE ARE 2 KINDS OF THEORIES 1 RELATING X TO M AND ANOTHER RELATING M TO Y, THE CONCEPTUAL THEORY IS USUALLY THE WAY THAT MOST PEOPLE THINK ABOUT THE WAY THIS IS RELATED TO Y BUT THEN THERE'S ANOTHER IMPORTANT THING FOR INTERVENTIONS AND THAT'S HOW WE GO ABOUT CHANGING A MEDIATOR. SO FOR EXAMPLE IMPULSIVITY FOUND TO BE RELATED TO PROBLEM OUTCOME BUT PERHAPS IN OUR 3 SESSION SCHOOL PROGRAM, WE WOULDN'T BE ABLE TO CHANGE IN PULSATIVITY, WE WOULDN'T HAVE ENOUGH SESSIONS TO CHANGE IT SO WE MIGHT TAKE A DIFFERENT MEDIATOR, MORE REALISTIC MEDIATOR AND DECISION WHAT IS CAN BE CHANGED AND WHAT CAN'T BE CHANGED GIVEN THE LIMITATIONS OF INTERVENTION, ENVIRONMENT WOULD BE MANIPULATION THEORY DISCUSSION. OKAY. SO NOW WE'RE GOING TO MOVE TO HOW IT ACTUALLY ESTIMATE SOME OF THESE CO EFFICIENTS. AND THE SINGLE MEDIATOR MODEL USES INFORMATION FROM SOME OR ALL OF 3 REGRESSION EQUATIONS. THESE CO EFFICIENTS MAY BE OBTAINED USING ORDINARY LEAST SQUARES, REGRESSION CO VARIANCE STRUCTURE ANALYSIS OR LOGISTIC REGRESSION, SOME OF THE RULES THAT I'LL TALK ABOUT DON'T DIRECTLY APPLY. WE WILL USE A TEST CALLED A PRODUCT OF CO EFFICIENT TEST THAT WILL BE GENERAL FOR MORE COMPLICATED MODELS OR ANY MODEL WE CAN COME UP WITH, WHY CAN COME UP WITH A PRODUCT OF CO EFESHT TESTS TO EVALUATE THE MEDIATOR. OFTEN THE MOST IMPORTANT ANALYSIS IN THIS EXPERIMENT THAT'S A SIGNIFICANT EXPONENT IN THE OUTCOME. SO THIS CO EFFICIENT WILL BE AN ESTIMATOR OF CAUSAL EFFECT. IT TURNS OUT HOWEVER THOUGH FOR MEDIATION, IT'S NOT NECESSARY THAT THERE BE A SIGNIFICANT RELATION BETWEEN X AND Y. FOR MEDIATION TO EXIST AND I WOULD WILL ARGUE WHEN YOU DON'T GET A SIGNIFICANT EFFECT OF X ON Y, THAT'S WHEN YOU SHOULD MOST DEFINITELY TEST FOR MEDIATION AND I'LL GET TO THAT IN A MINUTE. THE SECOND PART OF THE EQUATION, TEST THE MANIPULATION THEORY AND DETECT OF AXON M REPRESENTED BOY THE A-PATH AND IF X IS RANDOMIZED THAT WILL BALANCE ALL THENS THAT COULD EXPLAIN RELATIONSHIP BETWEEN X AND M. SO THE A-CO EFFICIENT WILL BE AN ESTIMATOR OF A CAUSAL EFFECT. BUT IT'S NOT TRUE FOR THIS THIRD REGRESSION EQUATION WHERE YOU HAVE BOTH X AND M PREDICTING Y. EITHER B OR C PRIME CAN BE CONSIDERED A CAUSAL EFFECT UNLESS CERTAIN ASSUMPTIONS ARE MADE. THE MAIN PROBLEM IS THAT EACH THOUGH IF X WAS RANDOMIZED PEOPLE TO DIFFERENT CONDITIONS WE HAVEN'T RANDOMIZED THEM TO THE VALUE OFLET MEDIATOR. THERE ARE ACTUAL VALUE OF THE MEDIATOR OUR PARTICIPANT PATHS THAT ARE SELECTED THAT ARE DEFINED BY THEIR CHARACTERISTICS. SO MEDIATED EFFECT MEASURES, THERE'S PRODUCT OF CO EFFICIENT MEASURE WHICH IS THE PRODUCT OF THE A-CO EFFICIENT TIMES THE B CO EFFICIENT. THERE'S ANOTHER WAY TO ESTIMATE THE MEDIATED EFFECT USING DIFFERENT IN CO EFFICIENTS AND I THINK YOU COULD ARGUE THAT THE DIFFERENCE MAY BE MORE INTUITIVE WAY TO UNDERSTAND THE MEDIATED EFFECT BECAUSE C IS A RELATION, TOTAL RELATION BETWEEN X AND Y AND C PRIME AND ADJUSTED FOR THE MEDIATOR SO IT MAKES SENSE THAT THE DIFFERENCE BETWEEN THE CO EFFICIENTS WOULD REPRESENT THE MEDIATED EFFECT. THE--ALSO IN THE DIRECT ANALYSIS WE HAVE C PRIME AND A TOTAL EFFECT WHICH WOULD BE THE RELATION OF X AND Y WHICH WILL EQUAL THE MEDIATED EFFECT PLUS THE DIRECT EFFECT: WE CAN FIND A STANDARD ERROR OF THAT MEDIATED EFFECT AB FOR THOSE OF YOU FAMILIAR WITH THE CONCEPTS THAT WOULD JUST BE EQUAL THE SQUARE ROOT OF A, THE SQUARE ROOT OR B SQUARED PLUS B SQUARE PLUS THE STANDARD OF A-SQUARED USING THE DELTA METHOD, YOU IT TEST IT BY TAKING THE EFFECT AND DIVIDING BY THE STANDARD ERROR AND THERE ARE MORE ACCURATE TESTS AND IT TURNS OUT THAT A TIMES B DOESN'T HAVE A NORMAL DISTRIBUTION SO THE Z TEST IS NOT THE BEST TEST, IT HAS A DISTRIBUTION OF THE PLOT, SO FOR USE OF THE DISTRIBUTION OF THE PRODUCT TO DO THESE TESTING, YOU CAN GET MORE ACCURATE INTERVALS MORE ACCURATE TYPE AND MORE POWER. THERE ARE SOME ASSUMPTIONS OF THESE METHODS FOR EACH METHOD OF ESTIMATING MEDIATED EFFECT BASED ON EQUATIONS ON 3 OR 2 AND 3, THAT'S WE ONLY WE NEED EQUATIONS 1 AND 3 OR 2 AND 3 AND AB, WE HAVE RELIABLE AND VALID MEASURES, WE ACCUM THE CO EFFICIENT OF A, B, C-PRIME REFLECT TRUE CAUSAL RELATIONS AND BY CORRECT FUNCTIONAL FORM THAT MEAN THERE IS' A LINEAR RELATIONSHIP BETWEEN VARIABLES. WE ALSO ASSUME THERE AREN'T ANY OMITTED INFLUENCES. IMPORTANT THINGS THAT PREDICT M AND Y SO WE HAVE NOT INCLUDED IN THIS STATISTICAL ANALYSIS. WE ALSO ASSUME MEDIATION CHANGE IS CORRECT THAT THE TEMPORAL ORDERING THAT X COMES BEFORE M AND M BEFORE Y AND WE ALSO ASSUME THAT THE EFFECTS ARE CONSISTENT ACROSS SUBGROUPS THAT THERE AREN'T ANY SUBGROUPS IN OUR DATA WHERE THE RELATION ARE NEXT TO M AND M TO Y DIFFER. ESSENTIALLY WE DON'T HAVE A MODERATOR. I USED TO HAVE 3 BEIGES OF ASSUMPTIONS BUT IT WAS TOO DEPRESSING SO I TICK WITH THAT 1 PAGE THERE. --I STICK WITH THAT 1 PAGE THERE. CONFIRMATIONS CAN GO ON AND ON FOR A WHILE ON THIS TOPIC AND THE BOTTOM LINE IS THAT THE METHODS BASED ON PRODUCT AB APPLIES IN LONGITUDINAL MODEL AND MULTILEVEL MODEL AND ALL KINDS OF DIFFERENT EVENTS, MODELS IT CAN BE MORE CUMBERSOME TO USE THAT C-MINUS, CMARKED FOR IDENTIFICATION PRIME OPROACH IN MORE COMPLICATED MODEL. WANT SO THAT'S TESTS CALLED JOINT SIGNIFICANT TEST WHICH IS TESTING WHETHER THE PATH IS SIGNIFICANT OR WHY THE B PATH IS 8 HOURS ISIVE CANT AND BOTH OF THEM ARE SIGNIFICANT AND THEY INCLUDE SIGNIFICANT MEDIATED EFFECT. MENT THERE AREN'T ANY CONFIDENCES YOU CAN OBTAIN WITH THAT, THE PRODUCT USES THE DISTRIBUTION OF THE PRODUCT TO COME UP WITH CRITICAL VALUES RATHER THAN USING THESE VALUES AND ALSO THE BOOT STRAP WHICH IS A SAMPLING METHOD IN THE DISTRIBUTION OF THE PRODUCT AND THE BOOT STRAP DO THE SAME THING. THEY HANDLE THE PRODUCT THAT THE AB DOES NOT HAVE NORMAL DISTRIBUTION. THIS IS A TABLE WE MADE A WHILE AGO, IN A PAPER IN 2007 THAT LOOKED AT COMPARISON OF THESE DIFFERENT METHODS IN THEIR POWER TO DETECT MEDIATED EFFECT. SO THESE MEMBERS HERE ARE SAMPLE SIZE ESTIMATES, SAMPLE SIZE YOU WOULD NEED TO HAVE .8 POWER TO DETECT MEDIATED EFFECT. AT THE TOP, THE SS REFERS TO SMALL EFFECT FOR THE-PATH, AND SMALL EFFECT OF THE B-PATH AND MEDIUM CORRELATION OF .3 AND LARGE CORRELATION OF .5. SO FOR LL, IT WOULD BE A LARGE EFFECT FOR A AND B AND THOSE NUMBERS ALL LOOK REASONABLE. BUT THE MOST SURPRISING THING IN THIS TABLE IS THERE'S A METHOD THAT WOULD REQUIRE A SERIES OF STEPS, FIRST WOULD REQUIRE THAT X IS SIGNIFICANTLY RELATED TO Y AND X IS RELATED TO M AND M IS SIGNIFICANTLY RELATED TO Y THAT HAS ENORMOUS SAMPLE SIZE REQUIREMENTS 21,000 WHICH WE SAW THIS WHEN WE WERE SHOCKED, BUT THEN WE DID MORE WORK AND THAT NUMBER IS CORRECT, IF YOU THINK ABOUT IT AND THIS IS THE CASE WHERE THERE'S NO DIRECT EFFECT, THE SMALL A PATH AND A SMALL B PATH WHEN MULTIPLIED ARE VERY SMALL EFFECT. AND IF YOU REQUIRE THAT IT'S SIGNIFICANTLY RELATED WHY AND THAT'S SMALL WE NEED A LARGE SAMPLE SIZE TO BE DETECTED. THE BETTER OFF USING THE DISTRIBUTION OF THE PRODUCT, SO, BUT EVEN THE DISTRIBUTION OF PRODUCT THAT THE APATH IS SMALL, THE B-PATH IS SMALL, THE SAMPLE SIZE IS SUBSTANTIAL AND THAT SAMPLE SIZE IS COMPARABLE TO IF WE USE THE BOOT STRAP FOR EXAMPLE. OF COURSE THIS TABLE IS CROSS SECTIONAL, DOESN'T INCLUDE CO VARIANTS OR DOESN'T INCLUDE LONGITUDINAL DATA, OTHER METHODS THAT INCREASE [INDISCERNIBLE] EFFECTS BUT OTHERS REQUIRE A SIGNIFICANT EFFECT OF AXON Y WHEN LOOKING AT MEDIATION WILL BE UNDERPOWERED. AND THIS IS RELATED TO THIS TOPIC, IT'S IMPORTANT TO LOOK AT MEDIATION EVEN WHEN THE EFFECT OF AXON Y IS NOT SIGNIFICANT. --X ON Y IS NOT SIGNIFICANT. IT IS EASY TO SHOW MANY CASES WHERE THE EFFECT OF X ON Y HAS LESS POWER THAN THE TEST OF MEDIATION. THERE'S A FEW PAPERS OUT NOW THAT DEMONSTRATE THAT NOW AND WHEN IT OCCURS. BUT I THINK THE MOST IMPORTANT REASON FOR LOOKING AT MEDIATION WHEN THERE'S NONSIGNIFICANT EFFECT OF X ON Y IS YOU CAN PROVIDE INFORMATION ABOUT MANIPULATION THEORIES, WHETHER YOUR MANIPULATION WAS ABLE TO CHANGE THE MEDIATOR AND INFORMATION ON CONCEPTUAL THEORY WHETHER M WAS RELATED TO Y, THOSE ARE IMPORTANT AND 11 OR BOTH COULD EXPLAIN THE NONSIGNIFICANT EFFECT OF X ON Y AND HELP DECIDE IF YOU WANTED TO MAKE A CHANGE TO YOUR INTERVENTION. OKAY, NOW I'M GOING THROUGH MORE MODELS THAT REALLY JUST KIND OF GO THROUGH THEM QUICKLY. SO WE COULD EXPAND THE SINGLE MEDIATOR MODEL WHOSE HAVE MULTIPLE MEDIATORS, THERE ARE 4 MEDIATORS BUT NOTICE NOW I HAVE TO HAVE SOME NOTATIONS TO KEEP TRACK OF WHICH PATH IS WHICH SO I HAVE AC1, FOR THE EXTRA M, B-1, FOR THE M1 TO Y PATH AND SO ON. SO HERE'S A FLOOR MEDIATOR MODEL AND THERE ARE 4 MEDIATED EFFECTS. THERE'S AC1, B-1, B2, B3, A3, A4, WE WILL GET THE STANDARD ERROR WHICH IS THE SUM OF ALL THE MEDIATED EFFECT AND THE DIRECT EFFECT AND THE TOTAL EFFECT WHICH WOULD BE THE SUM OF ALL THE MEDIATED EFFECT PLUS DIRECT EFFECT. WE CAN TEST SIGNIFICANT FOR MEDIATION AND IF WE USE THE METHOD BASED ON DISTRIBUTION OF PRODUCT WHERE BOOT STRAPPING WILL GET MORE ACCURATE CONS FERENCE ALLELES IN THE STATISTICAL TEST. ANOTHER IMPORTANT ISSUE IS THE KNOCKS OF INCONSISTENT MEDIATION MODEL THAT WE HAD 1 DEVIATED EFFECT IN THE DIFFERENT EFFECT OR OTHER MEDIATED EFFECT. THERE'S MEDIATION. IT'S JUST THAT THE SIGN OF THESE CO EFFICIENTS HAVE THIS MEDIATED EFFECT AND THE DIRECT EFFECT. I THINK THIS IS IMPORTANT BECAUSE INTERVENTIONS STUDIES MAY HAVE A MEDIATOR THAT IS COUNTER PRODUCTIVE AND REALLY THE ONLY WAY TO FIND IT IS WITH THE MEDIATOR MODEL AND THIS IS THE PRINCIPLE INVESTIGATOR, AND THE INTERVENTION STUDIES ARE HIGH SCHOOL FOOTBALL PLAYERS AND PART OF THIS INTERVENTION AND TOP OF THESE STEROIDS RIGHT, YOU WANT TO BE [INDISCERNIBLE] YOU NEED TO SHOW BOTH SIDES AND AS YOU CAN SEE HERE, THE PROGRAM INCREASED THE AUGHTS OF GREATEST USE STEROIDS WHICH THEN ACTUALLY INCREASED ATTENTION TO USE STEROIDS THAT WAS A COUNTER PRODUCTIVE MEDIATION EFFECT, FORTUNATELY ALL THE OTHER COMPONENTS OUTWEIGHED THE--THIS COUNTER PRODUCTIVE COMPONENT AND SO THE EFFECT WAS OVERALL A REDUCTION IN INTENTION TO USE STEROIDS BUT WE HAVE THIS COUNTER PRODUCTIVE MEDIATED EFFECT AND THIS IS BECAUSE OF THE MEDIATED EFFECT IS A PRODUCT OF .753 TIMES .073 AND THE DIRECT EFFECT IS NEGATIVE .81. SO OVER ALL WE WELL A SIGNIFICANT REDUCTION IN ATTENTIONS WE HAD THIS COUNTER PRODUCTIVE MEDIATING PROCESS AND I WOULD EXPECT THERE MAY BE MANY OF THESE INTERVENTIONS AND THE ONLY WAY TO FIND THEM IS TO LOOK AT THESE TYPES OF MEDIATION ANALYSIS. THIS IDEA OF INCONSISTENT MEDIATION IS NOT UNCOMMON FOR EXAMPLE AS PEOPLE GET OLDER, TYPICALLY THERE ARE SITUATIONS WHERE THERE'S NO CHANGE IN THE EFFECT ON OUTCOME. SO HERE FOR EXAMPLE, TYPE AND PROFICIENCY HAS PEOPLE GET OLDER THERE MAY BE CLOSE TO 0 EFFECT OF AGE ON TYPING PROFICIENCY. BECAUSE OF THESE OPPOSING MEDIATED EFFECTS, THERE'S IMMEDIATEIATED EFFECT ON AGE, AS PEOPLE GET OLDER, THE REACTION TIME GETS SLOWER AND THE TYPING PROFICIENCY GOES DOWN AND THEN ALSO AS YOU GET OLDER YOU GET MORE SKILLS AND THAT INCREASES TYPING PROFICIENCY SO PERHAPS OVERALL THERE'S NO TYPE IN PROFICIENCY BECAUSE THERE ARE THESE THEORETICAL AND ACTUAL COUNTER PRODUCTIVE OR OPPOSING MEDIATED EFFECT. THERE'S A THEORY FOR THIS CALLED COMPENSATION THEORY AND AGING RIGHT THERE AS PEOPLE GET OLDER THEY LOSE CAPACITY, THEY FIND OTHER METHODS TO MAKE UP FOR THAT. HERE'S AN EXAMPLE OF MODERATION AND MEDIATION, HERE WE HAVE A MEDIATED EFFECT IN GROUP 1. THE BOX GOES THERE AND A MEDIATED EFFECT IN GROUP 2, INTERVENTION FOR MALES AND FEMALES AND WE CAN DO COMPARISONS OF THE BOX FOR THE CONCEPTUAL THEORY BY COMPARING B PATH AND WE CAN EVEN TEST FOR MEDIATED EFFECT ACROSS GROUPS. SO THIS WOULD BE A WAY TO INCLUDE INDIVIDUAL DIFFERENCES IN A MEDIATION ANALYSIS. THERE ARE ALSO LONG NUDEINAL MEDIATION MODELS, MEDIATION IS A LONGITUDINAL MODEL. WE ASSUME WE'VE GOT THE CORRECT ORDERING THAT X IS BEFORE M AND M IS BEFORE WHY, EVEN WHEN WE HAVE CROSS SECTIONAL DATA AND WITH CROSS SECTIONAL DATA WE HAVE TO MAKE AN ASSUMPTION THAT THE OBSERVED RELATIONS ARE REAL AND NOT SLOWLY WHEN THEY'RE MEASURED, FOR EXAMPLE WE MEASURE THINGS LATER WE WOULD GET A DIFFERENT MODEL. PART OF LONGITUDINAL MEDIATION ANALYSIS REQUIRES THE CORRECT TIMING AND SPACING OF MEASURES. IT REQUIRES ANSWERING A QUESTION WHEN DOES X EFFECT M AND M EFFECT WHY. THESE ARE THEORETICAL QUESTIONS THAT SHOULD BE ADDRESSED TO STUDY, THERE'S LOTS OF DIFFERENT POTENTIAL PROCESSES FOR HOW IT MIGHT OCCUR THAT COULD BE TRIGGERING OR CASCADING OF THE X TO M ORE M TO Y RELATION. THE PROCESS BY WHICH X EFFECTS M AND M EFFECTS Y COULD DIFFER. SO WE COULD HAVE A TRIGGERING RELATION BETWEEN X AND M AND SOMEWHERE IN THE RELATIONSHIP IN CHANGE OF MON Y. SO TIMING IS EFFECTING AND WHEN WE COLLECT THE LONG NUDEINAL DATA AND ALIGN A LONGITUDINAL STUDY TO CAPTURE WHEN THESE THINGS TEND TO OCCUR. WITH 2 MEASURES OF X, M, AND Y THERE ARE A FEW OPTIONS, YOU COMPETE A DIFFERENT SCORE IN COVA, AND YOU CAN RESIDUALIZE CHANGE IF YOU 3 OR MORE TIME POINTS THERE'S A LOT OF DIFFERENT MODELS THAT ARE AGGRESSIVE FOR CHANGE FOR SURVIVAL MODELS FOR INTERVENTION, THIS IS TYPICALLY X IS MEASURED ONCE BUT THEM SOMETIMES X IS MEASURED OTHER TIMES SO THERE ARE--ADDITIONAL INTERVENTIONS THAT ARE GIVEN BUT TYPICALLY X IS MEASURED AFTER BASELINE OR X REPRESENTS INTERVENTION AND IT'S MEASURED ONCE OR DELIVERED ONCE, ALTHOUGH THERE ARE SOME INTERVENTION STUDIES THAT HAVE BOOSTERS SO THEN X WOULD BE REPEATED AND TYPICALLY WITH INTERVENTION STUDIES X IS MEASURED ONCE. HERE'S AN EXAMPLE OF AN OTHER AGGRESSIVE MODEL OF TIME ORDERED MEDIATION. WHERE NOW WE HAVE X 1, MEDIATOR IS NOW M1, M2 AND M3. WE HAVE M1, Y2, WHY 3, THIS IS AN EXAMPLE WHERE WE HAVE 1 INTERVENTION X AND THEN WE HAVE AN AC1 PATH TO RELATES X 1 TO X 2 AND X WAS BEFORE M2 AND M2 WAS RELATED TO Y 3. AND SO M2 COMES BEFORE WHY 3. THIS WOULD BE CALLED A LONGITUDINAL MEDIATED EFFECT BECAUSE THE TIMES IS CORRECT. IT'S TIME 1, TIME 2 MEDIATOR TO TIME 3 OUTCOME VARIABLE. MEDIATED EFFECT IS TIME 1 TO B 1 AND USE THOSE STATISTICAL METHODS FOR THAT AC1 B-1. AS YOU CAN IMAGINE, YOU COULD ADD MORE WAYS OF DATA, THIS WAS REPEATED IT WOULD MEDIATED EFFECTS THAT OCCUR ON OTHER TIMES AND SO ON. THE X IN THIS MODEL BY THE WAY REFERS TO THE STABILITY OF M1 AND Y AND Y OVER TIME. ONE OF THE MOST ACTIVE AREAS IN MEDIATION RIGHT NOW IS CAUSAL INFERENCE. IN A 1979 PRESIDENTIAL ADDRESS BLALOCK'S SUGGEST THAD FOR SOCIOLOGICAL LOGICAL PHENOMENON 50 VARIABLES ARE INVOLVED. THAT'S A LOT OF VARIABLES. THERE ARE ALSO THESE COMPREHENSIVE PSYCHIATRIST MODELS THAT WOULD CONSIDER ALL THE VARIABLES THAT ARE RELATED TO HEALTH BEHAVIOR, FOR EXAMPLE, ALL THE VARIABLES THAT WOULD BE RELATED TO FLOSSING FOR EXAMPLE, YOU MIGHT THINK HOW MANY ARE IMPORTANT FOR YOUR RESEARCH PROJECT THAT WOULD HAVE A NON0 RELATION OF THE OUTCOME VARIABLE? THAT'S A LOT OF VARIABLES. TOO MANY VARIABLES. REMEMBER EARLIER WE TALKED ABOUT ALL THE POSSIBLE RELATIONS, EVEN JUST 5 VARIABLES THAT THINK ABOUT THE POSSIBLE 1S. THE OTHER ISSUE WITH MEDIATION ANALYSIS THAT I MENTIONED EARLIER IS THAT BECAUSE EVEN THOUGH WE RANDOMIZED TRIAL AMLY ASSIGNED PEOPLE TO LEVELS OF X WE DON'T RANDOMLY ASSIGN THEM TO A VALUE OF M, BECAUSE M IS TYPICALLY SELF-SELECTED IN THAT CONTEXT. MOST OF THESE MODERN MODELS ARE KIND OF FACTUAL OR POTENTIAL OUTCOME MODELS, MODELS CONSIDER FOR EXAMPLE THE TREATMENT PARTICIPANT IF INSTEAD THEY WERE IN IT THE CONTROL GROUP, OR A CONTROL PARTICIPANT IF INSTEAD THEY WERE IN THE TREATMENT GROUP. SO IF EACH PERSON, YOU THINK ABOUT THE POSSIBLE OR POTENTIAL CONDITIONS THAT THEY COULD BE IN. AND IN THAT SITUATION, ALL THESE POSSIBLE COUNTER FACT ULTIMATELY AND OTHER CONDITIONS OF AN EXPERIMENT ARE CONSIDERED AND THE STATISTICAL MODEL IS BASED ON ALL OF THESE POTENTIAL OUTCOMES. IN A DESIGN FOR THESE CAUSAL APPROACHES THEY WOULD LIKE TO HAVE THE SAME PERSON IN BOTH GROUPS AT THE SAN FRANCISCO TIME WHICH CAN'T HAPPEN AND THAT'S CALLED THE FUNDAMENTAL PROBLEM OF CAUSAL INFERENCE BECAUSE YOU WOULD LIKE TO HAVE THE SAME PERSON SIMULTANEOUSLY SERVE IN THE TREATMENT AND CONTROL CONDITION. HOWEVER IF WE RANDOMIZE A LARGE NUMBER OF PERSONS WE CAN TAKE THE AVERAGE IN EACH GROUP, AS LONG AS WE RANDOMIZATION WORKED, SUPPOSEDLY WORKED, THEN THE DIFFERENCE IN THE MEANS BETWEEN THOSE 2 GROUPS, IS A CAUSAL EFFECT. AND THIS IS WHY THE A-PATH AND THE C-PATH EARLY EROZAN CAUSAL EVENTS WITH RANDOMIZATION BECAUSE BY RANDOMIZING WE BALANCED ALL THE POSSIBLE CONFOUNDERS SO WE CAN INTERPRET THOSE CO EFFICIENTS AS A CAUSAL EFFECT. BUT THESE PRIMES DON'T HAVE A CAUSAL RELATION BECAUSE THIS IS NOT UNDER EXPERIMENTAL CONTROL. B AND C DO NOT NECESSARILY REPRESENT CAUSAL EFFECTS BUT WE REALLY NEED IS WE NEED THE RELATION BETWEEN M& Y FOR PARTICIPANTS IN THE TREATMENT GROUP, CONTROL GROUP, RELATIONSHIP BETWEEN M AND Y FOR CONTROL PARTICIPANTS IF THEY INSTEAD WERE IN THE TREATMENT GROUP. THE PRIME WERE NOT CAUSAL EVENT BECAUSE M IS NOT RANDOMIZED: SO WHAT CAN YOU DO ABOUT THAT? THIS IS A DIAGRAM THAT REPRESENTS THIS ISSUE WHERE WE HAVE THE VISUAL MODEL WHERE WE TALKED ABOUT BUT NOW WE HAVE THESE RED BOXES, AND I THINK ABOUT THE TV SHOWS THAT HAVE THE çNGEL AND THE DEVIL AND 1 ON EACH SHOULDER, THIS IS REALLY 2 DEVILS ON EACH SHOULDER. WE HAVE POSSIBLE CONFOUNDERS OF THE X TO M RELATIONSHIP AND POSSIBLE CONFOUNDERS OF THE M TO Y RELATIONSHIP. IF WE RANDOMIZE X, THEN THE CONFOUNDER OF THE X M RELATION WILL GO AWAY BECAUSE WE BALANCED ALL THOSE CONFOUNDERS BETWEEN THESE, SO THIS FOUNDER OF THE M TO Y RELATIONSHIP COULD STILL BE PRESENT. LET AND IN FACT, IF THE MODEL LOOKED LIKE THIS WE NEED TO HAVE THESE DECO EFFICIENTS AS THERE AS TRYING TO CORRECTLY INVESTIGATE THE MEDIATED EFFECT 6789 THERE'S A COUPLE OPTIONS THAT CAN BE DONE IN THIS SITUATION, RIGHT? BECAUSE OFTEN WE DON'T KNOW WE HAVEN'T MEASURED THESE CONFOUNDERS, WE CAN DO A KIND OF SENSITIVITY ANALYSIS, THE IDEA IS TO SEE HOW MUCH THE MEDIATION EFFECT WOULD CHANGE OR A POSSIBLE CONFOUNDER OF A CERTAIN SIZE. ONE OF THESE METHODS WE'VE ADAPTED WHICH IS EASIER TO DESCRIBE WHERE YOU COME COME UP WITH WHAT THE MEDIATED EFFECT WOULD BE IF YOU HAD A CONFOUNDER THAT HAD A CERTAIN CORRELATION WITH Y AND A CONFOUNDER HAD A CERTAIN CORRELATION EVENT AND AGAIN THIS IS A CONFOUNDER YOU HAD IT MEASURED AND THEN YOU SEE AS YOU VARY THOSE CORRELATIONS FOR THE UNOBSERVED CONFOUNDER HOW THE MEDIATED EFFECT CHANGES. THIS IS ANOTHER METHOD THAT CAN BE APPLIED FOR ANY STATISTICAL METHOD, THAT LOOKS AT SIMILARLY THE RELATIONSHIP OF CONFOUNDER, AND THE OUTCOME WITHIN THE PROPORTION OF PEOPLE WHO HAVE THAT CONFOUNDER VARIABLE. AND AGAIN THESE ARE--YOU WOULD WOULD HAVE TO GUESS THOSE POSSIBLE VALUES IN THE UNOBSERVED CONFOUNDER AND THE IDEA WOULD BE TO SEE HOW EFFECTS CHANGE AS YOU VARY THIS AND THEN WITH THE METHOD BASED ON CORRELATION BETWEEN ERROR TERMS THAT WOULD REFLECT CONFOUNDING. ANOTHER METHOD TO DEAL WITH THIS TYPE OF CONFOUNDING WOULD BE STATISTICAL METHODS, AND EACH 1 OF THESE HAS A LONG HISTORY AND AGAIN A WAY TO DEAL WITH THE METHOD IS A INSTRUMENTAL VARIABLE METHOD WHICH WOULD BE RELEVANT IN INTERVENTION TRIALS IF YOU COULD FIND AN INTERVENTION OR MEDIATOR THAT WAS A COMPLETE MEDIATOR THAT THERE WAS NO DIRECT EFFECT OF X ON Y ONCE YOU ADJUST FOR THE MEDIATOR. METHODS BASED ON STRATIFYING OR THEORETICAL STRATIFICATIONS OF HOW PEOPLE WOULD BE LIKELY TO BE EFFECTED BY THE INTERVENTION, FOR THE PRINCIPLE STRATIFICATION METHOD, A METHOD CALLED INVERSE PROBABILITY RATING WHICH I'LL DESCRIBE ON THE NEXT PAGE. AND THEN FINALLY A METHOD CALLED G-ESTIMATION, WORK ON THESE METHODS IS ACTIVE AND THE DEVELOPMENTS EVERY MONTH OR SO OF DIFFERENT APPROACHES TO IMPROVE CONFERENCE BASED ON THESE---CAUSAL INFERENCE ON THESE METHODS. I'D LIKE TO TALK BRIEFLY ON INVERSE PROBABILITY WEIGHTING, THE METHOD TO ADJUST RESULTS TO CONFOUNDERS IF YOU CONFUSE ALL YOUR MEASURES AT BASELINE FOR EXAMPLE, AS A WAY TO DEAL WITH PROBABLE CONFOUNDERS. IT CONDUCTS A WEIGHTING TYPE METHOD THAT COULD ALSO BE USED WITH MISSING DATA AND SO ON, BUT THE IDEA OF THESE WEIGHTS THEY USE TO ADJUST EACH INDIVIDUAL'S CONTRIBUTION TO THE ANALYSIS, DEPENDING ON HOW MUCH THERE'S CONFOUNDING OF THE MTO Y RELATIONSHIP. AGAIN THAT'S A USEFUL RELATIVELY STRAIGHT FORWARD WAY TO ADJUST FOR CONFOUNDING IN YOUR ANALYSIS AND OFTEN PEOPLE HAVE LOTS OF BASELINE MEASURES SO THEY CHRONIC LIVER DISEASE BE INCLUDED IN THE STATISTICAL ANALYSIS. THERE ARE ALSO DESIGN APPROACHES IN IMPROVING CAUSAL INFERENCE. SO 1 WAY TO DO MEDIATION IN ALL OF THIS IS THAT LET'S SAY YOU HAVE A MEASURE OF X AND YOU HAVE A MEASURE OF Y, THEN YOU ALSO HAVE A MEASURE OF THE MEDIATOR, WHAT'S THE BEST WAY TO USE THAT MEASURE OF THE MEDIATOR. I THINK MOST PEOPLE WOULD AGREE, IF YOU HAVE A MEASURE OF X, M, Y AND MORE INFORMATION AND YOU HAVE X AND Y, SO THE MEDIATION ANALYSIS REALLY IS THE OPTIMAL WAY TO USE THAT MEASURE OF THE MEDIATING VARIABLE TO TRY TO UNDERSTAND THE MEDIATING PROCESS. BUT ANOTHER QUESTION WOULD BE, AFTER YOU CONSIDERED 1 STUDY, WHAT'S THE NEXT BEST STUDY THAT YOU COULD DO TO VALIDATE OR GAIN MORE INFORMATION ABOUT A PARTICULAR MEDIATING VARIABLE. AND THERE ARE 2 GENERAL TYPES OF DESIGNS YOU COULD USE. ONE OF THEM WOULD BE CONSISTENCY WHERE YOU DEMONSTRATE THAT THE MEDIATED EFFECT IS CONSISTENT WITH OTHER GROUPS WITH SIMILAR VARIABLES AND SIMILAR MEASURES OF THE MEDIATOR FOR EXAMPLE. ANOTHER APPROACH WOULD BE SPECIFICITY TYPE STUDIES WHERE YOU DEMON TRAIT THAT THE MEDIATED EFFECT IS SPECIFIC TO 1 MEDIATOR AND NOT SPECIFIC TO ANOTHER MEDIATOR. SO THERE'S 2--DESIGN APPROACHES TO IMPROVE INFERENCE ABOUT A MEDIATING PROCESS. AND THEN HERE ARE FUTURE DIRECTIONS, THINGS THAT ARE GOING NOW, MEDIATION, META-ANALYSIS, APPROACH WHERE IS YOU COULD USE INFORMATION FROM PRIOR STUDIES THAT LOOKED AT MEDIATION, INCLUDING STUDIES HOOKED AT THE APOT AND LOOKED AT THE B-PATH AS A WAY TO COMBINE THAT INFORMATION IN THE MOST REASONABLE WAY POSSIBLE. SO OUR PERSONAL ORIENTED METHODS THAT WILL INVESTIGATE THINGS LIKE ARE THERE PEOPLE WHO SEEM TO FOLLOW THIS MEDIATION PROCESS AND ARE THERE PEOPLE WHO DON'T FOLLOW THIS MEDIATING PROCESS. QUALITATIVE MEDIATION METHODS. OTHER WAYS TO PROVIDE USEFUL INFORMATION ABOUT HOW MEDIATING VARIABLE PROCESS OPERATED IN A RESEARCH STUDY FOR EXAMPLE. BY INTERVIEWING PARTICIPANTS IN THE STUDY AND SO ON. AND A NARRATIVE OF HOW THE STORIES AND SUGGEST HOW THIS MEDIATING PROCESS WORKS. A LOT MORE WORK NOW, MEDIATION FOR NONLINEAR MODELS, LOGISTIC REGRESSION AND SURVIVAL ANAL IS AND POTENTIAL OUTCOMES MODEL HAS PROVIDED GROUND BREAKING WAYS TO LOOK AT THESE TYPES--LOOK AT MEDIATION IN THESE TYPES OF NONLINEAR MODEL, AND JUST THE MEDIATION ANALYSIS, WOULD BE A WAY TO INCLUDE PRIOR REMEDIATION AND REMEDIATION ANALYSIS. ONE WAY--1 INTERESTING THING WOULD BE TO LINK THE APPROACH WITH THE METAANAL AND SYNTHESIS WHERE YOU WOULD USE INFORMATION ON ALL PRIOR STUDIES IN YOUR ANALYSIS OF THE CURRENT STUDY. >> OKAY, SO TO WRAP IT UP MEDIATION IS IMPORTANT BECAUSE IT PROVIDES INFORMATION ON HOW VARIABLES ARE RELATED AND HOW INTERVENTION ACHIEVES ITS EFFECTS AND EFFECTS WILL UNFOLD OVER TIME AND IN TACK MEDIATION ANALYSIS CAN PROVIDE INFORMATION ON MANIPULATION THEORY THAT IS HOW INTERVENTION CHANGES M AND ALSO PERCEPTUAL THEOR O WHICH VARIABLES ARE CAUSALLY RELATED TO THE OUTCOME AND KIND OF INTERESTING THAT MOST RESEARCHERS DESIGN INTERVENTIONS WHERE THE ONLY THEY THINK THEY REALLY KNOW IS THE B-PATH AND THEN THE B-PATH IS THE 1 THAT'S EFFECTED OR COULD BE EFFECTED BY CONFOUNDERS. THAT'S MEDIATION BASED ON THE PRODUCT AB, WITH THIS DISTRIBUTION OF PRODUCT BOOT STRAP ARE MOST ACCURATE, YOU CAN HAVE MULTIPLE MEDIATOR MODELS, YOU CAN HAVE MODERATION AND MEDIATION, LONGITUDINAL MODEL TALKED ABOUT HERE BUT THERE'S MULTILEVEL MEDIATION MODEL AND SO ON, THAT AGAIN APPLY THESE--BUT THIS PRODUCT METHOD TO INVESTIGATE MEDIATOR EFFECTS AND THEN CAUSAL INFERENCE AND ACTIVATING RESEARCH GENERATE FOR CONFOUNDER BIAS AND THERE ARE ALSO EXPERIMENTAL DESIGNS AVAILABLE. SO HERE'S HYPOTH SCIZZ THE EFFECTS OF THIS PRESENTATION. WE HAVE THE PRESENTATION, I HOPE THEY WOULD WILL INCREASE YOUR [INDISCERNIBLE] OF INCRETESSING VARIABLES AND INCREASING MODERN METHODS AND INCREASE KNOWLEDGE OF LONGITUDINAL DESIGNS AND YOU THINK THE VARIABLES ARE FUN WITH THE IDEA THAT MAYBE THEY'LL IDENTIFY MORE OF THESE MEDIATING MECHANISMS IN FUTURE RESEARCH. THANK YOU. >> THANK YOU DAVID VERY MUCH EXCELLENT PRESENTATION. LOTS OF MATERIAL IN 40 MINUTES. WE HAVE QUESTIONS. I HAVE SOME, WE'RE PULLING IN QUESTIONS VIA TWITTER AND E-MAIL. SO LET'S GET START WIDE THOSE IN THE TIME THAT WE HAVE LEFT. ONE OBSERVE OF MINE, MOST OF THE CIRCUMSTANCES THAT YOU DESCRIBED IN IN THE MATERIAL AS YOU WERE PRESENTING IT ASSUME THAT YOU HAD A RANDOMIZED EXPERIMENT AND INDIVIDUALS ARE RANDOMIZED TO STUDY AND THEN LOOK AT MEDIATION IN THOSE INVESTMENT EXPERIMENTS WHAT ABOUT QUASI EXPERIMENTS AND PEOPLE ARE ASSIGNED THROUGH A RANDOMIZED PROCESS TO RECEIVE THE INTERVENTION OR NOT, DOES MEDIATION TAKE THE APPROACH, DOES IT WORK THE SAME WAY, CAN YOU TELL US ABOUT THAT? >> IT'S A GREAT, GREAT, QUESTION AND I JUST FOCUSED ON INTERVENTIONS BECAUSE THAT SEEMED TO BE WHAT WOULD BE THE MOST RELEVANT HERE, RIGHT. ALL THE ANALYSIS METHODS STILL APPLY BUT THEN THE CHALLENGE FOR THE X TO M RELATIONSHIP IS BECAUSE OF THAT QUASI EXPERIMENTAL ASSIGNMENT WHERE THERE COULD BE CONFOUNDERS OF THE X TO M RELATIONSHIP. BUT ALL THESE WOULD APPLY THERE BUT THE CHALLENGE WOULD BE GREATER IN THAT SITUATION. >> YOU MENTIONED THAT MEDIATION IS IMPORTANT EVEN IF THE INTERVENTION EFFECT, THE MAIN EFFECT THE INVESTIGATOR WILL BE FOCUSED ON IS NOT SIGNIFICANT. SO I'M IMAGINE NOTHING A SITUATION WHERE YOU'VE DONE THE ANALYSIS AND THE INTERVENTION DOES NOT HAVE A SIGNIFICANT EFFECT ON THE OUTCOME BUT YOU HAVE A REMEDIATION EFFECT. LET WHAT IS THE IN NEXT STEP, WHAT DOES IT MEAN IN TERMS OF THE NEXT THING TO DO? >> WELL FIRST OF ALL WHAT CAN OCCUR IS IN SOME SITUATIONS YOU HAVE POWER TO DETECT THE MEDIATED EFFECT THAN THE OUTCOME EFFECT, 1 THING WITH REPUDIATION WILL HAVE UNEXPLAINED VARIABLE IN MAND Y, THAT WOULD BE WHY MAYBE THE OVERALL EFFECT OF X ON Y. YEAH, I THINK IT'S NOT THE BEST OUTCOME IN THAT SITUATION. BUT IT DOES SUGGEST THAT THE MEDIATING PROCESS IS CORRECT. AND MAYBE THERE COULD BE ESPECIALLY IN THAT SITUATION, THESE COUNTEDDER PRODUCTIVE MEDIATED EFFECTS THAT MAY HAVE WORKED AGAINST THE OVERALL EFFECT OF X ON Y AND THE WAY TO TRY TO UNDERSTAND THOSE IS TO LOOK AT THE MULTIPLE MEDIATOR MODEL. IF YOU DO THAT AND FIND COUNTER PRODUCTIVE EFFECT YOU COULD MODIFY PREVENTION BY ADDRESSING BOTH, STRENGTHENING 1 AND WEAKENING OTHERS AND TRY IT AGAIN. >> EXACTLY AND IN THAT EXAMPLE I TALKED ABOUT THE FOR THE STEROID PREVENTION PROGRAM FOR HIGH SCHOOL FOOTBALL PLAYERS, WE HAD A LOST DISCUSSION AS YOU MIGHT EXPECT--A LOT OF DISCUSSION AS YOU MIGHT EXPECT AND THE SITUATION WAS MADE THAT WE NEED TO ADDRESS THE USE OF AN BOLETIC STEROID USE TO INCREASE ACCOUNTABLE FOR THESE YOUNG ADOLESCENT KIDS TO MAKE IT MORE BELIEVABLE, MORE CREDIBLE. SO I THINK THAT'S--I THINK THERE ARE PROBABLY MORE OF THESE EFFECTS ON THE INTERVENTIONS THAN WE REALIZE AND IF WE COULD STUDY THEM MORE AND EXPECT THAT INFORMATION ABOUT THEM WE WOULD WOULD HAVE BETTER INTERVENTIONS. ! >> ANOTHER KIND OF DESIGN THAT YOU MENTIONED A COUPLE OF TIMES BUT DIDN'T SPEAK TO DIRECTLY ARE RANDOMIZED CLUSTERED TRIALS AND ARE EXAMPLES FOR THAT MULTILEVEL DESIGN? >> YEAH, THERE ARE GREAT 1S. AND AS YOU KNOW MORE THAN ANYONE, PRETTY MUCH EVERY PREVENTION STUDY HAS A MULTILEVEL THAT OFTEN UNITS ARE ASSIGNED BY LEVELS OR SCHOOLS OR CLINIC AND THERE'S REALLY FASCINATING ASPECTS OF MEDIATION IN THE MULTILEVEL MODEL. OF COURSE THERE'S 2 WAYS THE MULTILEVEL PART FIRST [INDISCERNIBLE], AS A STATISTICAL THING THAT YOU HAVE TO DEAL WITH AND THE OTHER AS A THEORETICALLY IMPORTANT EFFECT SO YOU COULD LOOK AT MEDIATED EFFECT LET'S SAY AT THE SCHOOL LEVEL AND THEN ALSO AT AN INDIVIDUAL LEVEL, THERE'S A LOT MORE INFORMATION THAT CAN BE GAINED FROM THE STUDY IF YOU HAVE THAT MULTILEVEL STRUCTURALLY UPPER TO IT SO THERE ARE CHALLENGES IN THE ANALYSIS. >> THE ANALYSIS IS MORE COMPLICATED, IT SOUNDS LIKE THE MEDE ANALYSIS IS MORE COMPLICATED BUT EQUALLY APPLICABLE AND IMPORTANT. >> RIGHT. , 1 OF THE MOST INTERESTING MODELS NOW ARE THESE MODELS THAT WOULD HAVE REPEATED MEASURES OF LIKE, FOR ECOLOGICALLY MOMENTARY ASSESSMENT DATA THAT WOULD WOULD HAVE MEASURES FOR X, M, Y FOR EACH PERSON AND COULD COULD LOOKA MEDIATED EFFECTS IN A PERSON AND THE OVERALL AVERAGE MEDIATOR EFFECT ACROSS PEOPLE, THAT'S INTERESTING EXAMPLES OF THAT NOW IN THE LAST PROBABLY 5 YEARS OR SO. >> OUR NEXT QUESTION IS WHAT ABOUT YOUR THOUGHTS OF HAZE METHODS PROCESS ANALYSIS AND THAT FULL PARTIAL MEDIATION IS NO LONGER MEANINGFUL OR NEEDED. >> WERE YOU ABLE TO HEAR THAT? >> YEAH, I HEARD IT. WELL I SHOWED YOU 1 SLIDE OF A CONDITIONAL MEDIATION MODEL RIGHT THAT HAD THE 2 GROUPS FOR THE MEDIATED EFFECT COULD BE AT 1--1 GROUP BUT NOT ANOTHER OR THE A-PATH COULD DIFFER. MENT YEAH, I'M NOT SURE I UNDERSTOOD THE QUESTION ABOUT PARTIAL MEDIATION BUT IT WOULD BE IDEAL IF IT WAS COMPLOATLY THE CASE BUT I DON'T THINK THAT EVER--IN OUR--WITH REAL RESEARCH STUDIES OFTEN THERE WILL BE A DIRECT EFFECT, IT'S VERY DIFFICULT TO EXPLAIN ALL OF THE EFFECT OF X ON Y. >> OKAY, IS THERE ANY SOFTWARE FOR THE CAUSAL INFERENCE YOU MENTIONED? >> RIGHT. THERE'S THE BSS PROGRAMS WRITTEN BY LINDA AND TYLER. THE M+ PROGRAM BY LINDA [INDISCERNIBLE] AND ALL WILL CAUSAL ESTIMATORS HAS ONLY BEEN DEVELOPED FOR THE SINGLE MEDIATOR MODEL. THERE'S RECENT WORK ON DIFFERENT APPROACHES WITH MULTIPLE MEDIATOR. AND THAT, I HOPE THAT THERE WILL BE MORE SOFTWARE AVAILABLE FOR THAT OVER THE NEXT 2 YEARS. I MEAN THE PROBLEM WITH MULTIPLE MEDIATORS IS REMEMBER WE TALKED ABOUT ALL THE POTENTIAL OUTCOMES SO IF YOU HAD ANOTHER MEDIATOR THERE'S ALL THOSE POTENTIAL OUTCOMES OF A PERSON THAT COULD HAVE A VALUE OF MEDIATOR 1 AND A DIFFERENT VALUE OF MEDIATOR 2 SO IT GETS VERY LARGE ALL THE POSSIBLE, POTENTIAL OUTCOMES GET, VERY, VERY LARGE. >> YOU DESCRIBED A NUMBER OF ASSUMPTIONS THAT UNDERLIE THESE METHODS THAT YOU HAD PRESENTED TODAY. IN FACT, YOU SAID YOU USED TO HAVE A MUCH LONGER LIST OF ASSUMPTIONS AND IT WAS FRIGHTENING, SO YOU--YOU COULD ONLY PRESENT SOME OF THEM NOW. A LOT OF PEOPLE EVER UNCOMFORTABLE MAKING TOO MANY ASSUMPTIONS AND WOULD LIKE TO AT LEAST TEST THEM OR AT LEAST THE MOST IMPORTANT 1S. ARE THERE METHODS AVAILABLE TO EVALUATE THOSE ASSUMPTIONS TO ASSURE YOURSELF THAT YOU'RE NOT GOING TOO FAR ASTRAY. >> RIGHT. GOOD POINT. MENT YEAH, THERE ARE WAYS TO DEAL WITH EACH OF THOSE ASSUMPTIONS AND THERE'S--THERE'S IDENTIFICATION OF ASSUMPTIONS WHICH IS A SLIDE I TOOK OUT THAT MERELY RELATE TO THE CONFOUNDERS, OTHER CONFOUNDERS OF X ON M, OTHER CONFOUNDERS OF X ON Y, OTHER CONFOUNDERS OF M TO Y AND ARE THERE ANY EFFECTS OF AN INTERVENTION THAT THEN CONFOUND THE M TO Y RELATION. SO THERE ARE ACTUALLY ONLY 4 IDENTIFICATIONS OF ASSUMPTIONS THAT ARE LISTED THERE, BUT EXAMPLE MORE RELIABLE RELATIONS OR MEASUREMENT OF WORK TO GET MORE MEASURES, WE CAN TEST FOR MODERATION AND WE CAN LOOK AT OMIT A VARIABLE BY LOOKING AT THE SENSITIVITY FOR CONFOUNDERS AND SO ON, AND I THINK THE--WHAT EVERYONE SAYS I GUESS IS THAT LOOKING AT MEDIATING VARIABLES IS A CUMULATIVE PROCESS THAT REQUIRES LOTS OF EXPERIMENTS. OF COURSE THAT'S DIFFICULT IN PREVENTION OFTEN BECAUSE WE MAGIC LOAMACYY HAVE THESE LARGE STUDIES THAT ARE DIFFICULT TO REPEAT. SO WE REALLY NEED TO GET THE MAXIMUM AMOUNT OF INFORMATION FROM EACH 1 OF OUR STUDIES. AND--BUT THEN HOPEFULLY ASPECTS OF THAT COULD BE REPLICATED. >> DO YOU RECOMMEND DOING SENSITIVITY ANALYSIS AROUND MEDIATION MODELS AS A WAY OF GETTING A SENSE OF HOW FAR OFF YOU WOULD WOULD HAVE TO BE WITH A SENSE OF ASSUMPTIONS TO HAVE A SERIOUS CHANGE IN THE INTERPRETATION. >> EXACTLY. >> YEAH, SENSITIVITY ANALYSIS FOR THESE ASSUMPTIONS IS PROBABLY THE BEST WAY TO GO. I MEAN IT'S NOT LIKE, MEDIATED EFFECTS ARE REAL EASY TO FIND IN INTERVENTIONS. IN MY EXPERIENCE, THERE'S BEEN CONSISTENT. FOR EXAMPLE, THEY START PREVENTION NORMS, ESPECIALLY NORMS ON THE CRITICAL INGREDIENT OR PARENTING LIKE CONSISTENT DISCIPLINE AND PARENTAL WARMTH OR CRITICAL, THINGS THAT INTERMEDIATE COME UP AND LOTS OF DIFFERENT STUDIES AND THEN FOR LIKE, POETIC TACK LOW AND ALCOHOL TREATMENT CRAVING AS AN IMPORTANT MEDIATOR TO TARGET. SO YOU HOPE THAT MAYBE EVEN EACH STUDY HAS PLOTS THAT IN THE NEXT STUDY IF YOU DESIGN THE INTERVENTION, TO TARGET THE CRAVING OR YOU COULD GET THE BENEFICIALS EFFECTS THROUGH THAT. >> HOW CAN PEOPLE LEARN MORE ABOUT MEDIATION. THEY'VE GOTTEN THIS GOOD INTRODUCTION FROM YOU TODAY. MENT THEY'RE INTERESTED IN THIS LEARNING MORE. OBVIOUSLY THEY CAN LOOK A LOOK AT THE BOOK YOU'VE WRITTEN AND IT SOUNDS LIKE YOU MAY BE PREPARING A NEW ADDITION. THERE ARE COURSES AVAILABLE? WORKSHOPS OFFERED AROUND THE COUNTRY? HOW CAN PEOPLE LEARN MORE ABOUT THIS? , YEAH, THERE ARE WORKSHOPS OUT THERE, THERE'S A GREAT BOOK BY [INDISCERNIBLE] ON THE CAUSAL INFERENCE AND MEDIATION. LET'S SEE, THERE'S--WORKSHOPS, I'VE GIVEN SOME, AND I THINK THERE'S MORE AND MORE ANDREW HAYES GIVES WORKSHOPS ON MODERATION AND MEDIATION, I GUESS THAT WOULD BE THE WAY TO DO AND THEN THERE ARE SOME WORKSHOPS GIVEN FOR EXAMPLE, I THINK EACH YEAR I'VE GIVEN ABOUT 2 OF THEM AND THEN OTHER PEOPLE GIVE THEM AS WELL. THE OTHER QUESTION WAS--ALL THIS STUFF WE'RE MAKING A LOT OF THIS STUFF AVAILABLE ONLINE ALSO. I THINK THE AMOUNT OF INFORMATION IS REALLY GROWING, MORE WEB SITES ARE INCLUDING PROGRAM SCRIPTS FOR EXAMPLE IN M+ AND STUFF. >> IF PEOPLE CAME TO WAR WEB SITE, THEY COULD FIND INFORMATION ABOUT WHERE WORKSHOPS ARE BEING OFFERED? >> YES. >> I'LL PUTTED THAT ON THERE. >> I HAVE A QUESTION ABOUT MEDIATION AS IT RELATES TO VARIABLES THAT ARE OBTAIN INDEED MACHINE LEARNING ALGORITHMS, CAN YOU SPEAK TO THAT A LITTLE BIT? MORE INFORMATION? >> I NEED MORE INFORMATION BUT THE--NO MATTER WHERE THE DATA ARE OBTAINED FROM, THESE WOULD BE THE MODELS YOU WOULD TRY TO ADDRESS AND IT COULD BE LIKE AN ENORMOUS AMOUNT OF DATA WHICH I WOULD THINK SHOULD IMPROVE THE MEASUREMENT OF THE VARIABLES AND LEAD TO MORE ACCURATE RESULTS. SOME OF THE CAUSAL INFERENCE METHODS ARE CAUSED MACHINE LEARNING METHODS. >> ANY CLOSING REMARKS YOU WOULD LIKE TO OFFER, ANY ADVICE TO THE AUDIENCE? >> WELL, THANKS FOR LISTENING AND I THANK YOU FOR THE INVITE. I GUESS WHEN I STARTED THIS 30 YEARS AGO, IT WAS VERY DIFFICULT TO GET PEOPLE INTERESTED IN MEDIATION AND NOW IT'S GREAT TO SEE THAT PEOPLE KIND OF REALIZE HOW MUCH THIS INFORMATION THAT'S IN OUR DATA THAT WE CAN EXTRACT AND PERHAPS, YOU KNOW OUR METHODS AREN'T PERFECT BUT WE'LL MAKE THE MOST OUT OF THIS INFORMATION AND IN FACT, THE INFORMATION CAN BE VERY, VERY USEFUL FOR BOTH DEVELOPING PROGRAMS AND THEN OF COURSE EVALUATE. >> WELL, I REMEMBER A CONVERSATION YOU AND I HAD MANY YEARS AGO WHEN YOU WERE THINKING ABOUT WRITING THE BOOK ABOUT THIS, AND I'M GLAD YOU DID, AND THE FIELD IS GLAD YOU DID AND WE WANT TO THANK YOU VERY MUCH FOR YOUR PRESENTATION TODAY. >> THANK YOU. >> THANK YOU! >> IF YOU HAVE ANYMORE QUESTIONS FORAUR AUDIENCE, PLEASE FEEL FREE TO E-MAIL DR. Mc KINNON, IF YOU HAVE SPECIFIC QUESTIONS YOU WANT ANSWERED THAT WE DEPENDENT GET TO TODAY. >> SO THANK YOU FOR YOUR USEFUL INFORMATION AND EVERYONE WHO PARTICIPATED TODAY. OT MEDICINE MIND THE GAP WEB SITE WHICH IS PREVENTION.NIH.GOV/MINDTHEGAP. YOU WILL FIND SLIDE, REFERENCES, AND LINK TO COMPLETE EVALUATION, FEEDBACK IS IMPORTANT TO US AS WE PLAN THE REMAINING SESSIONS FOR 2016. OUR IN ACCIDENT MIND THE GAP THAT WAS SCHEDULED FOR APRIL 26 HAS BEEN POSTPONED. THERE WILL BE MORE INFORMATION POSTED AT A LATER DATE. THANK YOU AGAIN FOR YOUR TIME. >> THANK YOU.