Designing observational studies to approximate carefully implemented (e.g., low dropout rates) randomized experiments must be the goal of objective causal inference based on non-randomized data, and the propensity score has important roles to play in achieving this goal. First, the propensity score should be used to help create treatment and control groups that have "balance" -- similar distributions of background covariates within strata, as similar as, or more similar than, in a equivalent randomized experiment. This design effort must be executed without access to any outcome data, just as in a randomized experiment. Second, once the first use of propensity scores has achieved this balance, the propensity score should be used to help specify the statistical analysis plan for estimating causal effects on outcome data, also without access to ultimate outcome data, just as in a randomized experiment. Both of these steps could be termed "design" activities because both take place before seeing any ultimate outcome data, but for clarity, we refer to the second as the "specification" step for the ultimate analysis plan. Both the design and specification steps are critical to activities if the resulting observational study can be argued to be as objective as a carefully designed, specified, and conducted (e.g., low dropout), randomized experiment.