Compared with traditional interviews or drop-off/mail-back surveys, web surveys are advantageous for extracting particular subsamples or panel surveys with short intervals. These advantages are expected to bring about methodological advances in voting behavior research. However, web surveys with screening or intensive panel design inevitably have to be based on purposive sampling, and this purposive sampling brings about a serious deviation from probabilistic survey data. Using two datasets sharing certain variables and which were collected in the same period, we investigated the effectiveness of propensity score adjustment for web surveys. One set of data was from a web panel survey based on purposive sampling with short intervals, and the other set was from personal interview surveys based on probabilistic random sampling. The web panel survey ran for three days, starting two days before the voting day of the national election of the House of Councilors (upper house) in 2007; i.e., July 27th, 28th, and 29th (voting day). The respondents were purposively screened from a vast pool of registrants on the condition that they were usually exposed to information about political and social issues on the Internet. The personal interview survey data was collected right after the election of the House of Councilors in 2007 by probabilistic random sampling using the electoral rolls. Setting party identification and the parties to which respondents actually voted as dependent variables, the covariates for calculating the propensity scores were selected on the basis of the “strongly ignorable treatment assignment” condition (Rosenbaum & Rubin, 1983). Using three sets of covariates, three propensity scores were calculated and their effectivenesses in adjusting dependent variables were compared. The results of propensity score adjustment indicated that the distribution of parties to which respondents actually voted was effectively adjusted. However, propensity scores failed to adjust the distribution of party identification. Conditions on which propensity scores can effectively adjust web survey data are discussed. In particular, the need for enough covariates and further research into stable covariates are emphasized.