WP 2014-2

13(20)

Analysis

Using ordinary least squares we estimate the following linear regression

model

(1)

Here

is the number of benefit days for parent

in week

are

intercepts,

is a dummy variable: it is zero but takes value one if the

parent

was assigned to treatment in week

, and

is the error term.

, k = -

are the estimated difference in the paid benefit days

between the treatment and the comparison group. The model is estimated

separately for benefit take-up each week from 52 weeks to 1 week before

the monitoring took place and from 1 week to 52 weeks after the

monitoring took place, in total for 104 weeks. The first 52 coefficients,

< 0

should not differ from zero if the quasi experiment is well designed. The

following 52 coefficients,

> 0, should be negative if individuals are

deterred from using the insurance by the new information of the probability

of being monitored. Fixed effects for the treatment week,

, are added to

correct for the large seasonal and yearly variation in the benefit take-up

rates and variation in the intensity of monitoring. However, dropping fixed

effects for the treatment week does not affect the estimation results.

The question is how we should interpret

, k>0

? There are two concerns

with the given quasi experimental design.

The parent is not contacted if no errors or signs of misuse are detected; it

is possible that many of the monitored parents do not actually know that

they have been monitored.

According to the SSIA’s database errors or

discrepancies are detected in 7 per cent of the monitored applications.

The parents can however also find out about the monitoring from their

employer or from the day care or school, but it is reasonable to assume

that by far all parents know that they have been monitored. As the

behavioral effect only emerges if the parents know about monitoring the

estimated effect will most likely be biased downwards. If we assume that

only 10 per cent of the monitored parents were aware of the monitoring,

the real effect would be ten times higher than the estimated effect.

The second concern is that also the controls may be affected by the

(potential) monitoring of the treated. Monitored parents can tell other

A second, highly related, problem is that we do not know how careful the

caseworker at the SSIA has monitored the application. The caseworker should