MeasuringChangeinTwo-WaveStudies.pptVIP

* * * * * * * * * * * * * * * * * * * * Measuring Change in Two-Wave Studies David A. Kenny December 15, 2013 * The Focus Here in Measuring Change To understand the causes of change, e.g., intervention analysis. Not: To understand the variable of interest and how it changes. To measure who changes more or less in the sample. * A Bit of Notation Two times Time 1 measure: Y1 Time 2 measure: Y2 X a potential cause of change in Y easier to think of as a dichotomy (0 and 1) but not required X may be randomized or not * Change Score Analysis (CSA) Time 2 minus Time 1 Seen by some as a na?ve model. Can compute the difference score more indirectly (keep Time 1 in the model). * Control for Baseline (CfB) Adjust out for the effect of Time 1, i.e. empirically weight Time 1 and not fix its coefficient to one. Generally preferred method. * Equations Change Score Analysis (CSA) Y2 – Y1 = a + bX + e Control for Baseline (CfB) Y2 = a + bY1 + bX + e (will refer to b later) * The Coefficient for Y1 Note that we can rewrite CSA as Y2 = a + 1Y1 + bX + e Thus, CSA fixes the coefficient for Y1 to one, whereas CfB empirically estimates it. For CfB usually, though not always, b is less than 1. In fact, if the variances of Y1 and Y2 are equal, b cannot be bigger than 1 (given that b = 0). * Lord’s Paradox CSA and CfB yield different results when b is different from 1 and there is correlation between X and Y1. Why Different? * How Different? * CSA: The Gap Persists * CfB: The Gap Narrows (b 1) * CSA vs. CfB * CfB * Randomized Studies Lord’s paradox happens only when there is some association between X and Y1, i.e. a gap. If there is randomization, there is no gap in the population. By chance, there is some gap in the sample. * Estimate of the Treatment Effect (b) MD1: The mean difference between treatment groups at time 1. MD2: The mean difference between treatment groups at time 2. CSA: b = MD2 – MD1 CfB: b = MD2 - bMD1 If b 1, then CfB adjusts less than CSA.