Of individuals who received a letter, 6.00 percent applied for benefits in the nine months
after the letters were sent out, compared with 0.96 percent of individuals who did not
receive a letter. After adjusting for age, socioeconomic status, and other factors that might
influence whether an individual applies for benefits, we estimated that letters resulted in
an increase of 5.09 percentage points (p
=.000, 95% CI [5.02, 5.17]).
Note two things about how the treatment effect is described in this example. First, the text makes
clear that the estimated treatment effect (an increase of 5.09 percentage points) is not just the
difference between the observed means in the two experimental groups but rather is adjusted for
covariates included in the statistical model. Second, the text is clear that this treatment effect on a
binary outcome is measured in percentage points (as opposed to “an increase of 5.09 percent”).
Effects on binary outcomes should be described in percentage points. If it is also useful to describe
the effect in relative terms as a percentage of the baseline outcome (“a 530% increase”), then this
can be done too, but should be done in addition to, not instead of, an estimate in percentage
points.
Multiple treatment conditions and a control condition
Many evaluations include two or more treatment conditions. When this is the case, the preferred
elements to include in text are:
● The observed mean (proportion for binary responses) for the control group
● The point estimate (regression coefficient) for the treatment effect for each treatment
group, relative to the control/reference group. Depending on the research design and
statistical model, these may be covariate-adjusted treatment effects.
● The 95% confidence intervals for these point estimates
● The p
-values for these treatment effects, rounded to 2 or 3 decimal places as appropriate.
Report the actual p
-values, not just whether they fall below a threshold.
● If adjusting p
-values for multiple comparisons, report the un
adjusted values in text and
note the adjusted values in a footnote (read more about OES guidance on multiple
comparisons).
Depending on the research questions and comparisons specified in the analysis plan, differences
among the treatment arms may be relevant. In each case, the (potentially covariate-adjusted)
difference should be reported as described above (point estimate of the difference, 95% CI, and
actual p
-value).
Predicted means for the treatment arms may also be explicitly reported if relevant, but this is
optional (these are likely depicted graphically; see guidance on figures below).
Here’s an example from project 1738, which evaluated the effect of postcard reminders to seniors
on vaccine uptake. The trial involved a stepped-wedge design where timing of receiving postcards
Update September 2020