--- In [email protected], "authfriend" <[EMAIL PROTECTED]> wrote:

Peter: 
> > The real problem with the study is the design itself.
> > If it had a better design than a simple pre-post
> > (which  makes no sense for research of this sort) non
> > of these question would be discussed.
> 
> What kind of study design would you have used that
> would have avoided these sorts of questions?

As Peter and I have pointed out, from somewhat different angles,
(without putting words in Pete's mouth, here is my view), the study
period needs have some variation in the independnet variables,
particularly the ME parameters. The goal of any regression analysis
(Poisson regression was used in this study) is to explain
(quantitatively) variations of the dependent variable (crime in this
case), by the independent variables(IV). Without some strong variation
in the IVs, then variation in the DV is not explained -- and the
"study" reveals little.  

While a full interuption of the ME is one approach, as Peter has
suggested, it could also simple be varying the number of YF, say 500,
1000, 2000, 500 or so, in two-four week segments. And / or the hours
of YF. 

And the length of the study should be sufficient to capture 
variations in the control variables with the DV (crime). That is,
without the ME variable(s), the control variables should well explain
variations in crime. Some reasonable time period is necessary for
that. Two months is insufficient to capture seasonal effects prevelant
in crime analysis. And a multi-year study would be superior to a
single year study. And/or several such longer-term independent studies
over different time periods in different locations. 

In this contexrt, strong variation (or interuption) of the ME
variable(s) over many months, if not several years, would do several
things. First, it would allow a test of the square root effect. Crime
reduction should quadruple as the group size doubles (after some
threshold.) 

Second, if there is no change in the size or intensity of the ME
variable(s), then there is not much to demonstrate in its ability to
explain variations in the DV (crime).* There will at best be weak
correlation of ME to crime if crime (controlled by other IVS) is
flucuating over time, and ME is not. 

All of the ME studies I have seen are basically one dose/intervention
studies. Which is a bit like testing a drug with one dose on one
patient.  Its exploratory at best. It may provide some plausibility
evidence as justification for "real", further, longer-term  studies
where variation in the intervention IV is structured to vary strongly
over a long-period. 

As an aside, looking at this from another angle, a way of testing a
single or repeated impact of a intervention factor with a set
magnitude/intensity is with a "dummy variable", 1 or 0, depending if
the factor is on or off. Having several intervention points is far
superior to one intervention as are all the ME studies. Better yet, is
to to beyond dummy variables, having actual variations in the ME
variable(s), size and intensity. Over a substantially long period to
capture and contol for seasonal factors. 



---

*the DV in this context is already well controlled for other factors
by the other IVs (such as temperature, income, demographics,
employment levels, lagged abortion rates, etc.) 


Reply via email to