John Paravantis recently posed an interesting data-parsing challenge. It came to me as list-owner and not to the list because it had as attachment a very big dataset. I'll quote (most of) John's query below and append my analysis. John says: I'm not quite sure what DAYS measures (it ranges from -5 to 118) but it clearly is _not_ an index of the half-hour periods, of which there are several thousand. As elaborated below, DATE is a string variable identifying the calendar day. To answer the last question first: No, it is not at all OK to proceed as if a panel dataset were a continuous time series; any results would be meaningless statistical garbage, as I'm sure John knows. Consider the symptom John mentions: gretl says that the index variables he selected "do not represent a panel structure". That's quite right. The "unit" variable must constitute a unique ID for (in the case) the household, and he's OK in that respect, he has HOUSENUM. But in addition the "time" variable must constitute a unique identifier for the time period, and he's not OK in that respect: there's no variable in the dataset that fulfils that role. There's a bunch of time-related variables -- DAYS, DATE, MONTH, DAY, TIME, HOUR, etc., too many in fact! -- but no unique period ID for the half-hourly frequency. The variables that look most promising for forming such a series are DATE and TIME. These are both string variables, with formats DATE: "DD-MM-YYYY" TIME: "hh:mm" I'm adding below a script that creates a period ID using these variables. Perhaps it ought to be possible to do this more easily -- and maybe in fact it is (Jack?) but anyway it's something of a showcase for gretl's current string-handling capabilities. The strategy is to write out these two variables as .txt files, read them back in as strings, and parse them to create a series that (a) has distinct values for every half hour and (b) is actually increasing in the true time dimension. <hansl> open <John's data file> # we need to exceed the default max number of # "while" iterations below set loop_maxiter 0 # write DATE (string: DD-MM-YYY) out to text # with no added "obs" column or header store dates.txt DATE --omit-obs --no-header # and read it back as a big string string s = readfile("dates.txt") string line scalar y, m, d series ymd # Loop across the dates.txt lines, parsing out day, # month and year and reassembling as a proper sequence. # We scan line + 1 to skip the leading quote scalar i = 1 loop while getline(s, line) --quiet sscanf(line + 1, "%d-%d-%d", d, m, y) ymd[i] = y*10000 + m*100 + d i++ endloop # now write out TIME (string: hh:mm) as text store time.txt TIME --omit-obs --no-header # and read back in as a string s = readfile("time.txt") scalar h series hm # loop across the time.txt lines, parsing out hour, # minute and reassembling as a numerical sequence i = 1 loop while getline(s, line) --quiet sscanf(line + 1, "%d:%d", h, m) hm[i] = h*100 + m i++ endloop # now we're ready to construct a proper "time" index # for these half-hourly data series realtime = 10000*ymd + hm # take a look and check print DATE ymd TIME hm realtime --byobs # use "realtime" as the panel time variable: OK! setobs HOUSENUM realtime --panel-vars </hansl> Allin Cottrell